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         local_irq_save(*flags);
  75         while (unlikely(test_and_set_bit_lock(0, &__wd_smp_lock)))
  76                 cpu_relax();
  77 }
  78
  79 static inline void wd_smp_unlock(unsigned long *flags)
  80 {
  81         clear_bit_unlock(0, &__wd_smp_lock);
  82         local_irq_restore(*flags);
  83 }
  84
  85 static void wd_lockup_ipi(struct pt_regs *regs)
  86 {
  87         pr_emerg("Watchdog CPU:%d Hard LOCKUP\n", raw_smp_processor_id());
  88         print_modules();
  89         print_irqtrace_events(current);
  90         if (regs)
  91                 show_regs(regs);
  92         else
  93                 dump_stack();
  94
  95         if (hardlockup_panic)
  96                 nmi_panic(regs, "Hard LOCKUP");
  97 }
  98
  99 static void set_cpu_stuck(int cpu, u64 tb)
 100 {
 101         cpumask_set_cpu(cpu, &wd_smp_cpus_stuck);
 102         cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
 103         if (cpumask_empty(&wd_smp_cpus_pending)) {
 104                 wd_smp_last_reset_tb = tb;
 105                 cpumask_andnot(&wd_smp_cpus_pending,
 106                                 &wd_cpus_enabled,
 107                                 &wd_smp_cpus_stuck);
 108         }
 109 }
 110
 111 static void watchdog_smp_panic(int cpu, u64 tb)
 112 {
 113         unsigned long flags;
 114         int c;
 115
 116         wd_smp_lock(&flags);
 117         /* Double check some things under lock */
 118         if ((s64)(tb - wd_smp_last_reset_tb) < (s64)wd_smp_panic_timeout_tb)
 119                 goto out;
 120         if (cpumask_test_cpu(cpu, &wd_smp_cpus_pending))
 121                 goto out;
 122         if (cpumask_weight(&wd_smp_cpus_pending) == 0)
 123                 goto out;
 124
 125         pr_emerg("Watchdog CPU:%d detected Hard LOCKUP other CPUS:%*pbl\n",
 126                         cpu, cpumask_pr_args(&wd_smp_cpus_pending));
 127
 128         /*
 129          * Try to trigger the stuck CPUs.
 130          */
 131         for_each_cpu(c, &wd_smp_cpus_pending) {
 132                 if (c == cpu)
 133                         continue;
 134                 smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
 135         }
 136         smp_flush_nmi_ipi(1000000);
 137
 138         /* Take the stuck CPU out of the watch group */
 139         for_each_cpu(c, &wd_smp_cpus_pending)
 140                 set_cpu_stuck(c, tb);
 141
 142 out:
 143         wd_smp_unlock(&flags);
 144
 145         printk_safe_flush();
 146         /*
 147          * printk_safe_flush() seems to require another print
 148          * before anything actually goes out to console.
 149          */
 150         if (sysctl_hardlockup_all_cpu_backtrace)
 151                 trigger_allbutself_cpu_backtrace();
 152
 153         if (hardlockup_panic)
 154                 nmi_panic(NULL, "Hard LOCKUP");
 155 }
 156
 157 static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 158 {
 159         if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
 160                 if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
 161                         unsigned long flags;
 162
 163                         pr_emerg("Watchdog CPU:%d became unstuck\n", cpu);
 164                         wd_smp_lock(&flags);
 165                         cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
 166                         wd_smp_unlock(&flags);
 167                 }
 168                 return;
 169         }
 170         cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
 171         if (cpumask_empty(&wd_smp_cpus_pending)) {
 172                 unsigned long flags;
 173
 174                 wd_smp_lock(&flags);
 175                 if (cpumask_empty(&wd_smp_cpus_pending)) {
 176                         wd_smp_last_reset_tb = tb;
 177                         cpumask_andnot(&wd_smp_cpus_pending,
 178                                         &wd_cpus_enabled,
 179                                         &wd_smp_cpus_stuck);
 180                 }
 181                 wd_smp_unlock(&flags);
 182         }
 183 }
 184
 185 static void watchdog_timer_interrupt(int cpu)
 186 {
 187         u64 tb = get_tb();
 188
 189         per_cpu(wd_timer_tb, cpu) = tb;
 190
 191         wd_smp_clear_cpu_pending(cpu, tb);
 192
 193         if ((s64)(tb - wd_smp_last_reset_tb) >= (s64)wd_smp_panic_timeout_tb)
 194                 watchdog_smp_panic(cpu, tb);
 195 }
 196
 197 void soft_nmi_interrupt(struct pt_regs *regs)
 198 {
 199         unsigned long flags;
 200         int cpu = raw_smp_processor_id();
 201         u64 tb;
 202
 203         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 204                 return;
 205
 206         nmi_enter();
 207         tb = get_tb();
 208         if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
 209                 per_cpu(wd_timer_tb, cpu) = tb;
 210
 211                 wd_smp_lock(&flags);
 212                 if (cpumask_test_cpu(cpu, &wd_smp_cpus_stuck)) {
 213                         wd_smp_unlock(&flags);
 214                         goto out;
 215                 }
 216                 set_cpu_stuck(cpu, tb);
 217
 218                 pr_emerg("Watchdog CPU:%d Hard LOCKUP\n", cpu);
 219                 print_modules();
 220                 print_irqtrace_events(current);
 221                 if (regs)
 222                         show_regs(regs);
 223                 else
 224                         dump_stack();
 225
 226                 wd_smp_unlock(&flags);
 227
 228                 if (sysctl_hardlockup_all_cpu_backtrace)
 229                         trigger_allbutself_cpu_backtrace();
 230
 231                 if (hardlockup_panic)
 232                         nmi_panic(regs, "Hard LOCKUP");
 233         }
 234         if (wd_panic_timeout_tb < 0x7fffffff)
 235                 mtspr(SPRN_DEC, wd_panic_timeout_tb);
 236
 237 out:
 238         nmi_exit();
 239 }
 240
 241 static void wd_timer_reset(unsigned int cpu, struct timer_list *t)
 242 {
 243         t->expires = jiffies + msecs_to_jiffies(wd_timer_period_ms);
 244         if (wd_timer_period_ms > 1000)
 245                 t->expires = __round_jiffies_up(t->expires, cpu);
 246         add_timer_on(t, cpu);
 247 }
 248
 249 static void wd_timer_fn(unsigned long data)
 250 {
 251         struct timer_list *t = this_cpu_ptr(&wd_timer);
 252         int cpu = smp_processor_id();
 253
 254         watchdog_timer_interrupt(cpu);
 255
 256         wd_timer_reset(cpu, t);
 257 }
 258
 259 void arch_touch_nmi_watchdog(void)
 260 {
 261         int cpu = smp_processor_id();
 262
 263         watchdog_timer_interrupt(cpu);
 264 }
 265 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
 266
 267 static void start_watchdog_timer_on(unsigned int cpu)
 268 {
 269         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 270
 271         per_cpu(wd_timer_tb, cpu) = get_tb();
 272
 273         setup_pinned_timer(t, wd_timer_fn, 0);
 274         wd_timer_reset(cpu, t);
 275 }
 276
 277 static void stop_watchdog_timer_on(unsigned int cpu)
 278 {
 279         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 280
 281         del_timer_sync(t);
 282 }
 283
 284 static int start_wd_on_cpu(unsigned int cpu)
 285 {
 286         if (cpumask_test_cpu(cpu, &wd_cpus_enabled)) {
 287                 WARN_ON(1);
 288                 return 0;
 289         }
 290
 291         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 292                 return 0;
 293
 294         if (watchdog_suspended)
 295                 return 0;
 296
 297         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 298                 return 0;
 299
 300         cpumask_set_cpu(cpu, &wd_cpus_enabled);
 301         if (cpumask_weight(&wd_cpus_enabled) == 1) {
 302                 cpumask_set_cpu(cpu, &wd_smp_cpus_pending);
 303                 wd_smp_last_reset_tb = get_tb();
 304         }
 305         smp_wmb();
 306         start_watchdog_timer_on(cpu);
 307
 308         return 0;
 309 }
 310
 311 static int stop_wd_on_cpu(unsigned int cpu)
 312 {
 313         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 314                 return 0; /* Can happen in CPU unplug case */
 315
 316         stop_watchdog_timer_on(cpu);
 317
 318         cpumask_clear_cpu(cpu, &wd_cpus_enabled);
 319         wd_smp_clear_cpu_pending(cpu, get_tb());
 320
 321         return 0;
 322 }
 323
 324 static void watchdog_calc_timeouts(void)
 325 {
 326         wd_panic_timeout_tb = watchdog_thresh * ppc_tb_freq;
 327
 328         /* Have the SMP detector trigger a bit later */
 329         wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
 330
 331         /* 2/5 is the factor that the perf based detector uses */
 332         wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
 333 }
 334
 335 void watchdog_nmi_reconfigure(void)
 336 {
 337         int cpu;
 338
 339         watchdog_calc_timeouts();
 340
 341         for_each_cpu(cpu, &wd_cpus_enabled)
 342                 stop_wd_on_cpu(cpu);
 343
 344         for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
 345                 start_wd_on_cpu(cpu);
 346 }
 347
 348 /*
 349  * This runs after lockup_detector_init() which sets up watchdog_cpumask.
 350  */
 351 static int __init powerpc_watchdog_init(void)
 352 {
 353         int err;
 354
 355         watchdog_calc_timeouts();
 356
 357         err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "powerpc/watchdog:online",
 358                                 start_wd_on_cpu, stop_wd_on_cpu);
 359         if (err < 0)
 360                 pr_warn("Watchdog could not be initialized");
 361
 362         return 0;
 363 }
 364 arch_initcall(powerpc_watchdog_init);
 365
 366 static void handle_backtrace_ipi(struct pt_regs *regs)
 367 {
 368         nmi_cpu_backtrace(regs);
 369 }
 370
 371 static void raise_backtrace_ipi(cpumask_t *mask)
 372 {
 373         unsigned int cpu;
 374
 375         for_each_cpu(cpu, mask) {
 376                 if (cpu == smp_processor_id())
 377                         handle_backtrace_ipi(NULL);
 378                 else
 379                         smp_send_nmi_ipi(cpu, handle_backtrace_ipi, 1000000);
 380         }
 381 }
 382
 383 void arch_trigger_cpumask_backtrace(const cpumask_t *mask, bool exclude_self)
 384 {
 385         nmi_trigger_cpumask_backtrace(mask, exclude_self, raise_backtrace_ipi);
 386 }