arch/powerpc/kernel/watchdog.c

   1 /*
   2  * Watchdog support on powerpc systems.
   3  *
   4  * Copyright 2017, IBM Corporation.
   5  *
   6  * This uses code from arch/sparc/kernel/nmi.c and kernel/watchdog.c
   7  */
   8 #include <linux/kernel.h>
   9 #include <linux/param.h>
  10 #include <linux/init.h>
  11 #include <linux/percpu.h>
  12 #include <linux/cpu.h>
  13 #include <linux/nmi.h>
  14 #include <linux/module.h>
  15 #include <linux/export.h>
  16 #include <linux/kprobes.h>
  17 #include <linux/hardirq.h>
  18 #include <linux/reboot.h>
  19 #include <linux/slab.h>
  20 #include <linux/kdebug.h>
  21 #include <linux/sched/debug.h>
  22 #include <linux/delay.h>
  23 #include <linux/smp.h>
  24
  25 #include <asm/paca.h>
  26
  27 /*
  28  * The watchdog has a simple timer that runs on each CPU, once per timer
  29  * period. This is the heartbeat.
  30  *
  31  * Then there are checks to see if the heartbeat has not triggered on a CPU
  32  * for the panic timeout period. Currently the watchdog only supports an
  33  * SMP check, so the heartbeat only turns on when we have 2 or more CPUs.
  34  *
  35  * This is not an NMI watchdog, but Linux uses that name for a generic
  36  * watchdog in some cases, so NMI gets used in some places.
  37  */
  38
  39 static cpumask_t wd_cpus_enabled __read_mostly;
  40
  41 static u64 wd_panic_timeout_tb __read_mostly; /* timebase ticks until panic */
  42 static u64 wd_smp_panic_timeout_tb __read_mostly; /* panic other CPUs */
  43
  44 static u64 wd_timer_period_ms __read_mostly;  /* interval between heartbeat */
  45
  46 static DEFINE_PER_CPU(struct timer_list, wd_timer);
  47 static DEFINE_PER_CPU(u64, wd_timer_tb);
  48
  49 /*
  50  * These are for the SMP checker. CPUs clear their pending bit in their
  51  * heartbeat. If the bitmask becomes empty, the time is noted and the
  52  * bitmask is refilled.
  53  *
  54  * All CPUs clear their bit in the pending mask every timer period.
  55  * Once all have cleared, the time is noted and the bits are reset.
  56  * If the time since all clear was greater than the panic timeout,
  57  * we can panic with the list of stuck CPUs.
  58  *
  59  * This will work best with NMI IPIs for crash code so the stuck CPUs
  60  * can be pulled out to get their backtraces.
  61  */
  62 static unsigned long __wd_smp_lock;
  63 static cpumask_t wd_smp_cpus_pending;
  64 static cpumask_t wd_smp_cpus_stuck;
  65 static u64 wd_smp_last_reset_tb;
  66
  67 static inline void wd_smp_lock(unsigned long *flags)
  68 {
  69         /*
  70          * Avoid locking layers if possible.
  71          * This may be called from low level interrupt handlers at some
  72          * point in future.
  73          */
  74         raw_local_irq_save(*flags);
  75         hard_irq_disable(); /* Make it soft-NMI safe */
  76         while (unlikely(test_and_set_bit_lock(0, &__wd_smp_lock))) {
  77                 raw_local_irq_restore(*flags);
  78                 spin_until_cond(!test_bit(0, &__wd_smp_lock));
  79                 raw_local_irq_save(*flags);
  80                 hard_irq_disable();
  81         }
  82 }
  83
  84 static inline void wd_smp_unlock(unsigned long *flags)
  85 {
  86         clear_bit_unlock(0, &__wd_smp_lock);
  87         raw_local_irq_restore(*flags);
  88 }
  89
  90 static void wd_lockup_ipi(struct pt_regs *regs)
  91 {
  92         pr_emerg("Watchdog CPU:%d Hard LOCKUP\n", raw_smp_processor_id());
  93         print_modules();
  94         print_irqtrace_events(current);
  95         if (regs)
  96                 show_regs(regs);
  97         else
  98                 dump_stack();
  99
 100         if (hardlockup_panic)
 101                 nmi_panic(regs, "Hard LOCKUP");
 102 }
 103
 104 static void set_cpumask_stuck(const struct cpumask *cpumask, u64 tb)
 105 {
 106         cpumask_or(&wd_smp_cpus_stuck, &wd_smp_cpus_stuck, cpumask);
 107         cpumask_andnot(&wd_smp_cpus_pending, &wd_smp_cpus_pending, cpumask);
 108         if (cpumask_empty(&wd_smp_cpus_pending)) {
 109                 wd_smp_last_reset_tb = tb;
 110                 cpumask_andnot(&wd_smp_cpus_pending,
 111                                 &wd_cpus_enabled,
 112                                 &wd_smp_cpus_stuck);
 113         }
 114 }
 115 static void set_cpu_stuck(int cpu, u64 tb)
 116 {
 117         set_cpumask_stuck(cpumask_of(cpu), tb);
 118 }
 119
 120 static void watchdog_smp_panic(int cpu, u64 tb)
 121 {
 122         unsigned long flags;
 123         int c;
 124
 125         wd_smp_lock(&flags);
 126         /* Double check some things under lock */
 127         if ((s64)(tb - wd_smp_last_reset_tb) < (s64)wd_smp_panic_timeout_tb)
 128                 goto out;
 129         if (cpumask_test_cpu(cpu, &wd_smp_cpus_pending))
 130                 goto out;
 131         if (cpumask_weight(&wd_smp_cpus_pending) == 0)
 132                 goto out;
 133
 134         pr_emerg("Watchdog CPU:%d detected Hard LOCKUP other CPUS:%*pbl\n",
 135                         cpu, cpumask_pr_args(&wd_smp_cpus_pending));
 136
 137         /*
 138          * Try to trigger the stuck CPUs.
 139          */
 140         for_each_cpu(c, &wd_smp_cpus_pending) {
 141                 if (c == cpu)
 142                         continue;
 143                 smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
 144         }
 145         smp_flush_nmi_ipi(1000000);
 146
 147         /* Take the stuck CPUs out of the watch group */
 148         set_cpumask_stuck(&wd_smp_cpus_pending, tb);
 149
 150         wd_smp_unlock(&flags);
 151
 152         printk_safe_flush();
 153         /*
 154          * printk_safe_flush() seems to require another print
 155          * before anything actually goes out to console.
 156          */
 157         if (sysctl_hardlockup_all_cpu_backtrace)
 158                 trigger_allbutself_cpu_backtrace();
 159
 160         if (hardlockup_panic)
 161                 nmi_panic(NULL, "Hard LOCKUP");
 162
 163         return;
 164
 165 out:
 166         wd_smp_unlock(&flags);
 167 }
 168
 169 static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 170 {
 171         if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
 172                 if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
 173                         unsigned long flags;
 174
 175                         pr_emerg("Watchdog CPU:%d became unstuck\n", cpu);
 176                         wd_smp_lock(&flags);
 177                         cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
 178                         wd_smp_unlock(&flags);
 179                 }
 180                 return;
 181         }
 182         cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
 183         if (cpumask_empty(&wd_smp_cpus_pending)) {
 184                 unsigned long flags;
 185
 186                 wd_smp_lock(&flags);
 187                 if (cpumask_empty(&wd_smp_cpus_pending)) {
 188                         wd_smp_last_reset_tb = tb;
 189                         cpumask_andnot(&wd_smp_cpus_pending,
 190                                         &wd_cpus_enabled,
 191                                         &wd_smp_cpus_stuck);
 192                 }
 193                 wd_smp_unlock(&flags);
 194         }
 195 }
 196
 197 static void watchdog_timer_interrupt(int cpu)
 198 {
 199         u64 tb = get_tb();
 200
 201         per_cpu(wd_timer_tb, cpu) = tb;
 202
 203         wd_smp_clear_cpu_pending(cpu, tb);
 204
 205         if ((s64)(tb - wd_smp_last_reset_tb) >= (s64)wd_smp_panic_timeout_tb)
 206                 watchdog_smp_panic(cpu, tb);
 207 }
 208
 209 void soft_nmi_interrupt(struct pt_regs *regs)
 210 {
 211         unsigned long flags;
 212         int cpu = raw_smp_processor_id();
 213         u64 tb;
 214
 215         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 216                 return;
 217
 218         nmi_enter();
 219
 220         __this_cpu_inc(irq_stat.soft_nmi_irqs);
 221
 222         tb = get_tb();
 223         if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
 224                 per_cpu(wd_timer_tb, cpu) = tb;
 225
 226                 wd_smp_lock(&flags);
 227                 if (cpumask_test_cpu(cpu, &wd_smp_cpus_stuck)) {
 228                         wd_smp_unlock(&flags);
 229                         goto out;
 230                 }
 231                 set_cpu_stuck(cpu, tb);
 232
 233                 pr_emerg("Watchdog CPU:%d Hard LOCKUP\n", cpu);
 234                 print_modules();
 235                 print_irqtrace_events(current);
 236                 if (regs)
 237                         show_regs(regs);
 238                 else
 239                         dump_stack();
 240
 241                 wd_smp_unlock(&flags);
 242
 243                 if (sysctl_hardlockup_all_cpu_backtrace)
 244                         trigger_allbutself_cpu_backtrace();
 245
 246                 if (hardlockup_panic)
 247                         nmi_panic(regs, "Hard LOCKUP");
 248         }
 249         if (wd_panic_timeout_tb < 0x7fffffff)
 250                 mtspr(SPRN_DEC, wd_panic_timeout_tb);
 251
 252 out:
 253         nmi_exit();
 254 }
 255
 256 static void wd_timer_reset(unsigned int cpu, struct timer_list *t)
 257 {
 258         t->expires = jiffies + msecs_to_jiffies(wd_timer_period_ms);
 259         if (wd_timer_period_ms > 1000)
 260                 t->expires = __round_jiffies_up(t->expires, cpu);
 261         add_timer_on(t, cpu);
 262 }
 263
 264 static void wd_timer_fn(unsigned long data)
 265 {
 266         struct timer_list *t = this_cpu_ptr(&wd_timer);
 267         int cpu = smp_processor_id();
 268
 269         watchdog_timer_interrupt(cpu);
 270
 271         wd_timer_reset(cpu, t);
 272 }
 273
 274 void arch_touch_nmi_watchdog(void)
 275 {
 276         unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
 277         int cpu = smp_processor_id();
 278
 279         if (get_tb() - per_cpu(wd_timer_tb, cpu) >= ticks)
 280                 watchdog_timer_interrupt(cpu);
 281 }
 282 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
 283
 284 static void start_watchdog_timer_on(unsigned int cpu)
 285 {
 286         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 287
 288         per_cpu(wd_timer_tb, cpu) = get_tb();
 289
 290         setup_pinned_timer(t, wd_timer_fn, 0);
 291         wd_timer_reset(cpu, t);
 292 }
 293
 294 static void stop_watchdog_timer_on(unsigned int cpu)
 295 {
 296         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 297
 298         del_timer_sync(t);
 299 }
 300
 301 static int start_wd_on_cpu(unsigned int cpu)
 302 {
 303         unsigned long flags;
 304
 305         if (cpumask_test_cpu(cpu, &wd_cpus_enabled)) {
 306                 WARN_ON(1);
 307                 return 0;
 308         }
 309
 310         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 311                 return 0;
 312
 313         if (watchdog_suspended)
 314                 return 0;
 315
 316         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 317                 return 0;
 318
 319         wd_smp_lock(&flags);
 320         cpumask_set_cpu(cpu, &wd_cpus_enabled);
 321         if (cpumask_weight(&wd_cpus_enabled) == 1) {
 322                 cpumask_set_cpu(cpu, &wd_smp_cpus_pending);
 323                 wd_smp_last_reset_tb = get_tb();
 324         }
 325         wd_smp_unlock(&flags);
 326
 327         start_watchdog_timer_on(cpu);
 328
 329         return 0;
 330 }
 331
 332 static int stop_wd_on_cpu(unsigned int cpu)
 333 {
 334         unsigned long flags;
 335
 336         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 337                 return 0; /* Can happen in CPU unplug case */
 338
 339         stop_watchdog_timer_on(cpu);
 340
 341         wd_smp_lock(&flags);
 342         cpumask_clear_cpu(cpu, &wd_cpus_enabled);
 343         wd_smp_unlock(&flags);
 344
 345         wd_smp_clear_cpu_pending(cpu, get_tb());
 346
 347         return 0;
 348 }
 349
 350 static void watchdog_calc_timeouts(void)
 351 {
 352         wd_panic_timeout_tb = watchdog_thresh * ppc_tb_freq;
 353
 354         /* Have the SMP detector trigger a bit later */
 355         wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
 356
 357         /* 2/5 is the factor that the perf based detector uses */
 358         wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
 359 }
 360
 361 void watchdog_nmi_reconfigure(void)
 362 {
 363         int cpu;
 364
 365         watchdog_calc_timeouts();
 366
 367         for_each_cpu(cpu, &wd_cpus_enabled)
 368                 stop_wd_on_cpu(cpu);
 369
 370         for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
 371                 start_wd_on_cpu(cpu);
 372 }
 373
 374 /*
 375  * This runs after lockup_detector_init() which sets up watchdog_cpumask.
 376  */
 377 static int __init powerpc_watchdog_init(void)
 378 {
 379         int err;
 380
 381         watchdog_calc_timeouts();
 382
 383         err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "powerpc/watchdog:online",
 384                                 start_wd_on_cpu, stop_wd_on_cpu);
 385         if (err < 0)
 386                 pr_warn("Watchdog could not be initialized");
 387
 388         return 0;
 389 }
 390 arch_initcall(powerpc_watchdog_init);
 391
 392 static void handle_backtrace_ipi(struct pt_regs *regs)
 393 {
 394         nmi_cpu_backtrace(regs);
 395 }
 396
 397 static void raise_backtrace_ipi(cpumask_t *mask)
 398 {
 399         unsigned int cpu;
 400
 401         for_each_cpu(cpu, mask) {
 402                 if (cpu == smp_processor_id())
 403                         handle_backtrace_ipi(NULL);
 404                 else
 405                         smp_send_nmi_ipi(cpu, handle_backtrace_ipi, 1000000);
 406         }
 407 }
 408
 409 void arch_trigger_cpumask_backtrace(const cpumask_t *mask, bool exclude_self)
 410 {
 411         nmi_trigger_cpumask_backtrace(mask, exclude_self, raise_backtrace_ipi);
 412 }