kernel/watchdog_hld.c

   1 /*
   2  * Detect hard lockups on a system
   3  *
   4  * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
   5  *
   6  * Note: Most of this code is borrowed heavily from the original softlockup
   7  * detector, so thanks to Ingo for the initial implementation.
   8  * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
   9  * to those contributors as well.
  10  */
  11
  12 #define pr_fmt(fmt) "NMI watchdog: " fmt
  13
  14 #include <linux/nmi.h>
  15 #include <linux/module.h>
  16 #include <linux/sched/debug.h>
  17
  18 #include <asm/irq_regs.h>
  19 #include <linux/perf_event.h>
  20
  21 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
  22 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
  23 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
  24
  25 /* boot commands */
  26 /*
  27  * Should we panic when a soft-lockup or hard-lockup occurs:
  28  */
  29 unsigned int __read_mostly hardlockup_panic =
  30                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  31 static unsigned long hardlockup_allcpu_dumped;
  32 /*
  33  * We may not want to enable hard lockup detection by default in all cases,
  34  * for example when running the kernel as a guest on a hypervisor. In these
  35  * cases this function can be called to disable hard lockup detection. This
  36  * function should only be executed once by the boot processor before the
  37  * kernel command line parameters are parsed, because otherwise it is not
  38  * possible to override this in hardlockup_panic_setup().
  39  */
  40 void hardlockup_detector_disable(void)
  41 {
  42         watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
  43 }
  44
  45 static int __init hardlockup_panic_setup(char *str)
  46 {
  47         if (!strncmp(str, "panic", 5))
  48                 hardlockup_panic = 1;
  49         else if (!strncmp(str, "nopanic", 7))
  50                 hardlockup_panic = 0;
  51         else if (!strncmp(str, "0", 1))
  52                 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
  53         else if (!strncmp(str, "1", 1))
  54                 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
  55         return 1;
  56 }
  57 __setup("nmi_watchdog=", hardlockup_panic_setup);
  58
  59 void arch_touch_nmi_watchdog(void)
  60 {
  61         /*
  62          * Using __raw here because some code paths have
  63          * preemption enabled.  If preemption is enabled
  64          * then interrupts should be enabled too, in which
  65          * case we shouldn't have to worry about the watchdog
  66          * going off.
  67          */
  68         raw_cpu_write(watchdog_nmi_touch, true);
  69 }
  70 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
  71
  72 static struct perf_event_attr wd_hw_attr = {
  73         .type           = PERF_TYPE_HARDWARE,
  74         .config         = PERF_COUNT_HW_CPU_CYCLES,
  75         .size           = sizeof(struct perf_event_attr),
  76         .pinned         = 1,
  77         .disabled       = 1,
  78 };
  79
  80 /* Callback function for perf event subsystem */
  81 static void watchdog_overflow_callback(struct perf_event *event,
  82                  struct perf_sample_data *data,
  83                  struct pt_regs *regs)
  84 {
  85         /* Ensure the watchdog never gets throttled */
  86         event->hw.interrupts = 0;
  87
  88         if (atomic_read(&watchdog_park_in_progress) != 0)
  89                 return;
  90
  91         if (__this_cpu_read(watchdog_nmi_touch) == true) {
  92                 __this_cpu_write(watchdog_nmi_touch, false);
  93                 return;
  94         }
  95
  96         /* check for a hardlockup
  97          * This is done by making sure our timer interrupt
  98          * is incrementing.  The timer interrupt should have
  99          * fired multiple times before we overflow'd.  If it hasn't
 100          * then this is a good indication the cpu is stuck
 101          */
 102         if (is_hardlockup()) {
 103                 int this_cpu = smp_processor_id();
 104
 105                 /* only print hardlockups once */
 106                 if (__this_cpu_read(hard_watchdog_warn) == true)
 107                         return;
 108
 109                 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 110                 print_modules();
 111                 print_irqtrace_events(current);
 112                 if (regs)
 113                         show_regs(regs);
 114                 else
 115                         dump_stack();
 116
 117                 /*
 118                  * Perform all-CPU dump only once to avoid multiple hardlockups
 119                  * generating interleaving traces
 120                  */
 121                 if (sysctl_hardlockup_all_cpu_backtrace &&
 122                                 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
 123                         trigger_allbutself_cpu_backtrace();
 124
 125                 if (hardlockup_panic)
 126                         nmi_panic(regs, "Hard LOCKUP");
 127
 128                 __this_cpu_write(hard_watchdog_warn, true);
 129                 return;
 130         }
 131
 132         __this_cpu_write(hard_watchdog_warn, false);
 133         return;
 134 }
 135
 136 /*
 137  * People like the simple clean cpu node info on boot.
 138  * Reduce the watchdog noise by only printing messages
 139  * that are different from what cpu0 displayed.
 140  */
 141 static unsigned long firstcpu_err;
 142 static atomic_t watchdog_cpus;
 143
 144 int watchdog_nmi_enable(unsigned int cpu)
 145 {
 146         struct perf_event_attr *wd_attr;
 147         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 148         int firstcpu = 0;
 149
 150         /* nothing to do if the hard lockup detector is disabled */
 151         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 152                 goto out;
 153
 154         /* is it already setup and enabled? */
 155         if (event && event->state > PERF_EVENT_STATE_OFF)
 156                 goto out;
 157
 158         /* it is setup but not enabled */
 159         if (event != NULL)
 160                 goto out_enable;
 161
 162         if (atomic_inc_return(&watchdog_cpus) == 1)
 163                 firstcpu = 1;
 164
 165         wd_attr = &wd_hw_attr;
 166         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 167
 168         /* Try to register using hardware perf events */
 169         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 170
 171         /* save the first cpu's error for future comparision */
 172         if (firstcpu && IS_ERR(event))
 173                 firstcpu_err = PTR_ERR(event);
 174
 175         if (!IS_ERR(event)) {
 176                 /* only print for the first cpu initialized */
 177                 if (firstcpu || firstcpu_err)
 178                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 179                 goto out_save;
 180         }
 181
 182         /*
 183          * Disable the hard lockup detector if _any_ CPU fails to set up
 184          * set up the hardware perf event. The watchdog() function checks
 185          * the NMI_WATCHDOG_ENABLED bit periodically.
 186          *
 187          * The barriers are for syncing up watchdog_enabled across all the
 188          * cpus, as clear_bit() does not use barriers.
 189          */
 190         smp_mb__before_atomic();
 191         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
 192         smp_mb__after_atomic();
 193
 194         /* skip displaying the same error again */
 195         if (!firstcpu && (PTR_ERR(event) == firstcpu_err))
 196                 return PTR_ERR(event);
 197
 198         /* vary the KERN level based on the returned errno */
 199         if (PTR_ERR(event) == -EOPNOTSUPP)
 200                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 201         else if (PTR_ERR(event) == -ENOENT)
 202                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 203                          cpu);
 204         else
 205                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 206                         cpu, PTR_ERR(event));
 207
 208         pr_info("Shutting down hard lockup detector on all cpus\n");
 209
 210         return PTR_ERR(event);
 211
 212         /* success path */
 213 out_save:
 214         per_cpu(watchdog_ev, cpu) = event;
 215 out_enable:
 216         perf_event_enable(per_cpu(watchdog_ev, cpu));
 217 out:
 218         return 0;
 219 }
 220
 221 void watchdog_nmi_disable(unsigned int cpu)
 222 {
 223         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 224
 225         if (event) {
 226                 perf_event_disable(event);
 227                 per_cpu(watchdog_ev, cpu) = NULL;
 228
 229                 /* should be in cleanup, but blocks oprofile */
 230                 perf_event_release_kernel(event);
 231
 232                 /* watchdog_nmi_enable() expects this to be zero initially. */
 233                 if (atomic_dec_and_test(&watchdog_cpus))
 234                         firstcpu_err = 0;
 235         }
 236 }