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_cpu_stuck(int cpu, u64 tb)
 105 {
 106         cpumask_set_cpu(cpu, &wd_smp_cpus_stuck);
 107         cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
 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
 116 static void watchdog_smp_panic(int cpu, u64 tb)
 117 {
 118         unsigned long flags;
 119         int c;
 120
 121         wd_smp_lock(&flags);
 122         /* Double check some things under lock */
 123         if ((s64)(tb - wd_smp_last_reset_tb) < (s64)wd_smp_panic_timeout_tb)
 124                 goto out;
 125         if (cpumask_test_cpu(cpu, &wd_smp_cpus_pending))
 126                 goto out;
 127         if (cpumask_weight(&wd_smp_cpus_pending) == 0)
 128                 goto out;
 129
 130         pr_emerg("Watchdog CPU:%d detected Hard LOCKUP other CPUS:%*pbl\n",
 131                         cpu, cpumask_pr_args(&wd_smp_cpus_pending));
 132
 133         /*
 134          * Try to trigger the stuck CPUs.
 135          */
 136         for_each_cpu(c, &wd_smp_cpus_pending) {
 137                 if (c == cpu)
 138                         continue;
 139                 smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
 140         }
 141         smp_flush_nmi_ipi(1000000);
 142
 143         /* Take the stuck CPU out of the watch group */
 144         for_each_cpu(c, &wd_smp_cpus_pending)
 145                 set_cpu_stuck(c, tb);
 146
 147         wd_smp_unlock(&flags);
 148
 149         printk_safe_flush();
 150         /*
 151          * printk_safe_flush() seems to require another print
 152          * before anything actually goes out to console.
 153          */
 154         if (sysctl_hardlockup_all_cpu_backtrace)
 155                 trigger_allbutself_cpu_backtrace();
 156
 157         if (hardlockup_panic)
 158                 nmi_panic(NULL, "Hard LOCKUP");
 159
 160         return;
 161
 162 out:
 163         wd_smp_unlock(&flags);
 164 }
 165
 166 static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 167 {
 168         if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
 169                 if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
 170                         unsigned long flags;
 171
 172                         pr_emerg("Watchdog CPU:%d became unstuck\n", cpu);
 173                         wd_smp_lock(&flags);
 174                         cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
 175                         wd_smp_unlock(&flags);
 176                 }
 177                 return;
 178         }
 179         cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
 180         if (cpumask_empty(&wd_smp_cpus_pending)) {
 181                 unsigned long flags;
 182
 183                 wd_smp_lock(&flags);
 184                 if (cpumask_empty(&wd_smp_cpus_pending)) {
 185                         wd_smp_last_reset_tb = tb;
 186                         cpumask_andnot(&wd_smp_cpus_pending,
 187                                         &wd_cpus_enabled,
 188                                         &wd_smp_cpus_stuck);
 189                 }
 190                 wd_smp_unlock(&flags);
 191         }
 192 }
 193
 194 static void watchdog_timer_interrupt(int cpu)
 195 {
 196         u64 tb = get_tb();
 197
 198         per_cpu(wd_timer_tb, cpu) = tb;
 199
 200         wd_smp_clear_cpu_pending(cpu, tb);
 201
 202         if ((s64)(tb - wd_smp_last_reset_tb) >= (s64)wd_smp_panic_timeout_tb)
 203                 watchdog_smp_panic(cpu, tb);
 204 }
 205
 206 void soft_nmi_interrupt(struct pt_regs *regs)
 207 {
 208         unsigned long flags;
 209         int cpu = raw_smp_processor_id();
 210         u64 tb;
 211
 212         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 213                 return;
 214
 215         nmi_enter();
 216         tb = get_tb();
 217         if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
 218                 per_cpu(wd_timer_tb, cpu) = tb;
 219
 220                 wd_smp_lock(&flags);
 221                 if (cpumask_test_cpu(cpu, &wd_smp_cpus_stuck)) {
 222                         wd_smp_unlock(&flags);
 223                         goto out;
 224                 }
 225                 set_cpu_stuck(cpu, tb);
 226
 227                 pr_emerg("Watchdog CPU:%d Hard LOCKUP\n", cpu);
 228                 print_modules();
 229                 print_irqtrace_events(current);
 230                 if (regs)
 231                         show_regs(regs);
 232                 else
 233                         dump_stack();
 234
 235                 wd_smp_unlock(&flags);
 236
 237                 if (sysctl_hardlockup_all_cpu_backtrace)
 238                         trigger_allbutself_cpu_backtrace();
 239
 240                 if (hardlockup_panic)
 241                         nmi_panic(regs, "Hard LOCKUP");
 242         }
 243         if (wd_panic_timeout_tb < 0x7fffffff)
 244                 mtspr(SPRN_DEC, wd_panic_timeout_tb);
 245
 246 out:
 247         nmi_exit();
 248 }
 249
 250 static void wd_timer_reset(unsigned int cpu, struct timer_list *t)
 251 {
 252         t->expires = jiffies + msecs_to_jiffies(wd_timer_period_ms);
 253         if (wd_timer_period_ms > 1000)
 254                 t->expires = __round_jiffies_up(t->expires, cpu);
 255         add_timer_on(t, cpu);
 256 }
 257
 258 static void wd_timer_fn(unsigned long data)
 259 {
 260         struct timer_list *t = this_cpu_ptr(&wd_timer);
 261         int cpu = smp_processor_id();
 262
 263         watchdog_timer_interrupt(cpu);
 264
 265         wd_timer_reset(cpu, t);
 266 }
 267
 268 void arch_touch_nmi_watchdog(void)
 269 {
 270         unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
 271         int cpu = smp_processor_id();
 272
 273         if (get_tb() - per_cpu(wd_timer_tb, cpu) >= ticks)
 274                 watchdog_timer_interrupt(cpu);
 275 }
 276 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
 277
 278 static void start_watchdog_timer_on(unsigned int cpu)
 279 {
 280         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 281
 282         per_cpu(wd_timer_tb, cpu) = get_tb();
 283
 284         setup_pinned_timer(t, wd_timer_fn, 0);
 285         wd_timer_reset(cpu, t);
 286 }
 287
 288 static void stop_watchdog_timer_on(unsigned int cpu)
 289 {
 290         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 291
 292         del_timer_sync(t);
 293 }
 294
 295 static int start_wd_on_cpu(unsigned int cpu)
 296 {
 297         if (cpumask_test_cpu(cpu, &wd_cpus_enabled)) {
 298                 WARN_ON(1);
 299                 return 0;
 300         }
 301
 302         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 303                 return 0;
 304
 305         if (watchdog_suspended)
 306                 return 0;
 307
 308         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 309                 return 0;
 310
 311         cpumask_set_cpu(cpu, &wd_cpus_enabled);
 312         if (cpumask_weight(&wd_cpus_enabled) == 1) {
 313                 cpumask_set_cpu(cpu, &wd_smp_cpus_pending);
 314                 wd_smp_last_reset_tb = get_tb();
 315         }
 316         smp_wmb();
 317         start_watchdog_timer_on(cpu);
 318
 319         return 0;
 320 }
 321
 322 static int stop_wd_on_cpu(unsigned int cpu)
 323 {
 324         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 325                 return 0; /* Can happen in CPU unplug case */
 326
 327         stop_watchdog_timer_on(cpu);
 328
 329         cpumask_clear_cpu(cpu, &wd_cpus_enabled);
 330         wd_smp_clear_cpu_pending(cpu, get_tb());
 331
 332         return 0;
 333 }
 334
 335 static void watchdog_calc_timeouts(void)
 336 {
 337         wd_panic_timeout_tb = watchdog_thresh * ppc_tb_freq;
 338
 339         /* Have the SMP detector trigger a bit later */
 340         wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
 341
 342         /* 2/5 is the factor that the perf based detector uses */
 343         wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
 344 }
 345
 346 void watchdog_nmi_reconfigure(void)
 347 {
 348         int cpu;
 349
 350         watchdog_calc_timeouts();
 351
 352         for_each_cpu(cpu, &wd_cpus_enabled)
 353                 stop_wd_on_cpu(cpu);
 354
 355         for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
 356                 start_wd_on_cpu(cpu);
 357 }
 358
 359 /*
 360  * This runs after lockup_detector_init() which sets up watchdog_cpumask.
 361  */
 362 static int __init powerpc_watchdog_init(void)
 363 {
 364         int err;
 365
 366         watchdog_calc_timeouts();
 367
 368         err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "powerpc/watchdog:online",
 369                                 start_wd_on_cpu, stop_wd_on_cpu);
 370         if (err < 0)
 371                 pr_warn("Watchdog could not be initialized");
 372
 373         return 0;
 374 }
 375 arch_initcall(powerpc_watchdog_init);
 376
 377 static void handle_backtrace_ipi(struct pt_regs *regs)
 378 {
 379         nmi_cpu_backtrace(regs);
 380 }
 381
 382 static void raise_backtrace_ipi(cpumask_t *mask)
 383 {
 384         unsigned int cpu;
 385
 386         for_each_cpu(cpu, mask) {
 387                 if (cpu == smp_processor_id())
 388                         handle_backtrace_ipi(NULL);
 389                 else
 390                         smp_send_nmi_ipi(cpu, handle_backtrace_ipi, 1000000);
 391         }
 392 }
 393
 394 void arch_trigger_cpumask_backtrace(const cpumask_t *mask, bool exclude_self)
 395 {
 396         nmi_trigger_cpumask_backtrace(mask, exclude_self, raise_backtrace_ipi);
 397 }