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         tb = get_tb();
 220         if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
 221                 per_cpu(wd_timer_tb, cpu) = tb;
 222
 223                 wd_smp_lock(&flags);
 224                 if (cpumask_test_cpu(cpu, &wd_smp_cpus_stuck)) {
 225                         wd_smp_unlock(&flags);
 226                         goto out;
 227                 }
 228                 set_cpu_stuck(cpu, tb);
 229
 230                 pr_emerg("Watchdog CPU:%d Hard LOCKUP\n", cpu);
 231                 print_modules();
 232                 print_irqtrace_events(current);
 233                 if (regs)
 234                         show_regs(regs);
 235                 else
 236                         dump_stack();
 237
 238                 wd_smp_unlock(&flags);
 239
 240                 if (sysctl_hardlockup_all_cpu_backtrace)
 241                         trigger_allbutself_cpu_backtrace();
 242
 243                 if (hardlockup_panic)
 244                         nmi_panic(regs, "Hard LOCKUP");
 245         }
 246         if (wd_panic_timeout_tb < 0x7fffffff)
 247                 mtspr(SPRN_DEC, wd_panic_timeout_tb);
 248
 249 out:
 250         nmi_exit();
 251 }
 252
 253 static void wd_timer_reset(unsigned int cpu, struct timer_list *t)
 254 {
 255         t->expires = jiffies + msecs_to_jiffies(wd_timer_period_ms);
 256         if (wd_timer_period_ms > 1000)
 257                 t->expires = __round_jiffies_up(t->expires, cpu);
 258         add_timer_on(t, cpu);
 259 }
 260
 261 static void wd_timer_fn(unsigned long data)
 262 {
 263         struct timer_list *t = this_cpu_ptr(&wd_timer);
 264         int cpu = smp_processor_id();
 265
 266         watchdog_timer_interrupt(cpu);
 267
 268         wd_timer_reset(cpu, t);
 269 }
 270
 271 void arch_touch_nmi_watchdog(void)
 272 {
 273         unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
 274         int cpu = smp_processor_id();
 275
 276         if (get_tb() - per_cpu(wd_timer_tb, cpu) >= ticks)
 277                 watchdog_timer_interrupt(cpu);
 278 }
 279 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
 280
 281 static void start_watchdog_timer_on(unsigned int cpu)
 282 {
 283         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 284
 285         per_cpu(wd_timer_tb, cpu) = get_tb();
 286
 287         setup_pinned_timer(t, wd_timer_fn, 0);
 288         wd_timer_reset(cpu, t);
 289 }
 290
 291 static void stop_watchdog_timer_on(unsigned int cpu)
 292 {
 293         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 294
 295         del_timer_sync(t);
 296 }
 297
 298 static int start_wd_on_cpu(unsigned int cpu)
 299 {
 300         unsigned long flags;
 301
 302         if (cpumask_test_cpu(cpu, &wd_cpus_enabled)) {
 303                 WARN_ON(1);
 304                 return 0;
 305         }
 306
 307         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 308                 return 0;
 309
 310         if (watchdog_suspended)
 311                 return 0;
 312
 313         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 314                 return 0;
 315
 316         wd_smp_lock(&flags);
 317         cpumask_set_cpu(cpu, &wd_cpus_enabled);
 318         if (cpumask_weight(&wd_cpus_enabled) == 1) {
 319                 cpumask_set_cpu(cpu, &wd_smp_cpus_pending);
 320                 wd_smp_last_reset_tb = get_tb();
 321         }
 322         wd_smp_unlock(&flags);
 323
 324         start_watchdog_timer_on(cpu);
 325
 326         return 0;
 327 }
 328
 329 static int stop_wd_on_cpu(unsigned int cpu)
 330 {
 331         unsigned long flags;
 332
 333         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 334                 return 0; /* Can happen in CPU unplug case */
 335
 336         stop_watchdog_timer_on(cpu);
 337
 338         wd_smp_lock(&flags);
 339         cpumask_clear_cpu(cpu, &wd_cpus_enabled);
 340         wd_smp_unlock(&flags);
 341
 342         wd_smp_clear_cpu_pending(cpu, get_tb());
 343
 344         return 0;
 345 }
 346
 347 static void watchdog_calc_timeouts(void)
 348 {
 349         wd_panic_timeout_tb = watchdog_thresh * ppc_tb_freq;
 350
 351         /* Have the SMP detector trigger a bit later */
 352         wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
 353
 354         /* 2/5 is the factor that the perf based detector uses */
 355         wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
 356 }
 357
 358 void watchdog_nmi_reconfigure(void)
 359 {
 360         int cpu;
 361
 362         watchdog_calc_timeouts();
 363
 364         for_each_cpu(cpu, &wd_cpus_enabled)
 365                 stop_wd_on_cpu(cpu);
 366
 367         for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
 368                 start_wd_on_cpu(cpu);
 369 }
 370
 371 /*
 372  * This runs after lockup_detector_init() which sets up watchdog_cpumask.
 373  */
 374 static int __init powerpc_watchdog_init(void)
 375 {
 376         int err;
 377
 378         watchdog_calc_timeouts();
 379
 380         err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "powerpc/watchdog:online",
 381                                 start_wd_on_cpu, stop_wd_on_cpu);
 382         if (err < 0)
 383                 pr_warn("Watchdog could not be initialized");
 384
 385         return 0;
 386 }
 387 arch_initcall(powerpc_watchdog_init);
 388
 389 static void handle_backtrace_ipi(struct pt_regs *regs)
 390 {
 391         nmi_cpu_backtrace(regs);
 392 }
 393
 394 static void raise_backtrace_ipi(cpumask_t *mask)
 395 {
 396         unsigned int cpu;
 397
 398         for_each_cpu(cpu, mask) {
 399                 if (cpu == smp_processor_id())
 400                         handle_backtrace_ipi(NULL);
 401                 else
 402                         smp_send_nmi_ipi(cpu, handle_backtrace_ipi, 1000000);
 403         }
 404 }
 405
 406 void arch_trigger_cpumask_backtrace(const cpumask_t *mask, bool exclude_self)
 407 {
 408         nmi_trigger_cpumask_backtrace(mask, exclude_self, raise_backtrace_ipi);
 409 }