1 // SPDX-License-Identifier: GPL-2.0
3 * Watchdog support on powerpc systems.
5 * Copyright 2017, IBM Corporation.
7 * This uses code from arch/sparc/kernel/nmi.c and kernel/watchdog.c
10 #define pr_fmt(fmt) "watchdog: " fmt
12 #include <linux/kernel.h>
13 #include <linux/param.h>
14 #include <linux/init.h>
15 #include <linux/percpu.h>
16 #include <linux/cpu.h>
17 #include <linux/nmi.h>
18 #include <linux/module.h>
19 #include <linux/export.h>
20 #include <linux/kprobes.h>
21 #include <linux/hardirq.h>
22 #include <linux/reboot.h>
23 #include <linux/slab.h>
24 #include <linux/kdebug.h>
25 #include <linux/sched/debug.h>
26 #include <linux/delay.h>
27 #include <linux/smp.h>
29 #include <asm/interrupt.h>
34 * The powerpc watchdog ensures that each CPU is able to service timers.
35 * The watchdog sets up a simple timer on each CPU to run once per timer
36 * period, and updates a per-cpu timestamp and a "pending" cpumask. This is
39 * Then there are two systems to check that the heartbeat is still running.
40 * The local soft-NMI, and the SMP checker.
42 * The soft-NMI checker can detect lockups on the local CPU. When interrupts
43 * are disabled with local_irq_disable(), platforms that use soft-masking
44 * can leave hardware interrupts enabled and handle them with a masked
45 * interrupt handler. The masked handler can send the timer interrupt to the
46 * watchdog's soft_nmi_interrupt(), which appears to Linux as an NMI
47 * interrupt, and can be used to detect CPUs stuck with IRQs disabled.
49 * The soft-NMI checker will compare the heartbeat timestamp for this CPU
50 * with the current time, and take action if the difference exceeds the
53 * The limitation of the soft-NMI watchdog is that it does not work when
54 * interrupts are hard disabled or otherwise not being serviced. This is
55 * solved by also having a SMP watchdog where all CPUs check all other
58 * The SMP checker can detect lockups on other CPUs. A gobal "pending"
59 * cpumask is kept, containing all CPUs which enable the watchdog. Each
60 * CPU clears their pending bit in their heartbeat timer. When the bitmask
61 * becomes empty, the last CPU to clear its pending bit updates a global
62 * timestamp and refills the pending bitmask.
64 * In the heartbeat timer, if any CPU notices that the global timestamp has
65 * not been updated for a period exceeding the watchdog threshold, then it
66 * means the CPU(s) with their bit still set in the pending mask have had
67 * their heartbeat stop, and action is taken.
69 * Some platforms implement true NMI IPIs, which can be used by the SMP
70 * watchdog to detect an unresponsive CPU and pull it out of its stuck
71 * state with the NMI IPI, to get crash/debug data from it. This way the
72 * SMP watchdog can detect hardware interrupts off lockups.
75 static cpumask_t wd_cpus_enabled __read_mostly;
77 static u64 wd_panic_timeout_tb __read_mostly; /* timebase ticks until panic */
78 static u64 wd_smp_panic_timeout_tb __read_mostly; /* panic other CPUs */
80 static u64 wd_timer_period_ms __read_mostly; /* interval between heartbeat */
82 static DEFINE_PER_CPU(struct hrtimer, wd_hrtimer);
83 static DEFINE_PER_CPU(u64, wd_timer_tb);
85 /* SMP checker bits */
86 static unsigned long __wd_smp_lock;
87 static cpumask_t wd_smp_cpus_pending;
88 static cpumask_t wd_smp_cpus_stuck;
89 static u64 wd_smp_last_reset_tb;
91 static inline void wd_smp_lock(unsigned long *flags)
94 * Avoid locking layers if possible.
95 * This may be called from low level interrupt handlers at some
98 raw_local_irq_save(*flags);
99 hard_irq_disable(); /* Make it soft-NMI safe */
100 while (unlikely(test_and_set_bit_lock(0, &__wd_smp_lock))) {
101 raw_local_irq_restore(*flags);
102 spin_until_cond(!test_bit(0, &__wd_smp_lock));
103 raw_local_irq_save(*flags);
108 static inline void wd_smp_unlock(unsigned long *flags)
110 clear_bit_unlock(0, &__wd_smp_lock);
111 raw_local_irq_restore(*flags);
114 static void wd_lockup_ipi(struct pt_regs *regs)
116 int cpu = raw_smp_processor_id();
119 pr_emerg("CPU %d Hard LOCKUP\n", cpu);
120 pr_emerg("CPU %d TB:%lld, last heartbeat TB:%lld (%lldms ago)\n",
121 cpu, tb, per_cpu(wd_timer_tb, cpu),
122 tb_to_ns(tb - per_cpu(wd_timer_tb, cpu)) / 1000000);
124 print_irqtrace_events(current);
130 /* Do not panic from here because that can recurse into NMI IPI layer */
133 static void set_cpumask_stuck(const struct cpumask *cpumask, u64 tb)
135 cpumask_or(&wd_smp_cpus_stuck, &wd_smp_cpus_stuck, cpumask);
136 cpumask_andnot(&wd_smp_cpus_pending, &wd_smp_cpus_pending, cpumask);
137 if (cpumask_empty(&wd_smp_cpus_pending)) {
138 wd_smp_last_reset_tb = tb;
139 cpumask_andnot(&wd_smp_cpus_pending,
144 static void set_cpu_stuck(int cpu, u64 tb)
146 set_cpumask_stuck(cpumask_of(cpu), tb);
149 static void watchdog_smp_panic(int cpu, u64 tb)
155 /* Double check some things under lock */
156 if ((s64)(tb - wd_smp_last_reset_tb) < (s64)wd_smp_panic_timeout_tb)
158 if (cpumask_test_cpu(cpu, &wd_smp_cpus_pending))
160 if (cpumask_weight(&wd_smp_cpus_pending) == 0)
163 pr_emerg("CPU %d detected hard LOCKUP on other CPUs %*pbl\n",
164 cpu, cpumask_pr_args(&wd_smp_cpus_pending));
165 pr_emerg("CPU %d TB:%lld, last SMP heartbeat TB:%lld (%lldms ago)\n",
166 cpu, tb, wd_smp_last_reset_tb,
167 tb_to_ns(tb - wd_smp_last_reset_tb) / 1000000);
169 if (!sysctl_hardlockup_all_cpu_backtrace) {
171 * Try to trigger the stuck CPUs, unless we are going to
172 * get a backtrace on all of them anyway.
174 for_each_cpu(c, &wd_smp_cpus_pending) {
177 smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
181 /* Take the stuck CPUs out of the watch group */
182 set_cpumask_stuck(&wd_smp_cpus_pending, tb);
184 wd_smp_unlock(&flags);
188 * printk_safe_flush() seems to require another print
189 * before anything actually goes out to console.
191 if (sysctl_hardlockup_all_cpu_backtrace)
192 trigger_allbutself_cpu_backtrace();
194 if (hardlockup_panic)
195 nmi_panic(NULL, "Hard LOCKUP");
200 wd_smp_unlock(&flags);
203 static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
205 if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
206 if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
207 struct pt_regs *regs = get_irq_regs();
212 pr_emerg("CPU %d became unstuck TB:%lld\n",
214 print_irqtrace_events(current);
220 cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
221 wd_smp_unlock(&flags);
225 cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
226 if (cpumask_empty(&wd_smp_cpus_pending)) {
230 if (cpumask_empty(&wd_smp_cpus_pending)) {
231 wd_smp_last_reset_tb = tb;
232 cpumask_andnot(&wd_smp_cpus_pending,
236 wd_smp_unlock(&flags);
240 static void watchdog_timer_interrupt(int cpu)
244 per_cpu(wd_timer_tb, cpu) = tb;
246 wd_smp_clear_cpu_pending(cpu, tb);
248 if ((s64)(tb - wd_smp_last_reset_tb) >= (s64)wd_smp_panic_timeout_tb)
249 watchdog_smp_panic(cpu, tb);
252 DEFINE_INTERRUPT_HANDLER_NMI(soft_nmi_interrupt)
255 int cpu = raw_smp_processor_id();
258 /* should only arrive from kernel, with irqs disabled */
259 WARN_ON_ONCE(!arch_irq_disabled_regs(regs));
261 if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
264 __this_cpu_inc(irq_stat.soft_nmi_irqs);
267 if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
269 if (cpumask_test_cpu(cpu, &wd_smp_cpus_stuck)) {
270 wd_smp_unlock(&flags);
273 set_cpu_stuck(cpu, tb);
275 pr_emerg("CPU %d self-detected hard LOCKUP @ %pS\n",
276 cpu, (void *)regs->nip);
277 pr_emerg("CPU %d TB:%lld, last heartbeat TB:%lld (%lldms ago)\n",
278 cpu, tb, per_cpu(wd_timer_tb, cpu),
279 tb_to_ns(tb - per_cpu(wd_timer_tb, cpu)) / 1000000);
281 print_irqtrace_events(current);
284 wd_smp_unlock(&flags);
286 if (sysctl_hardlockup_all_cpu_backtrace)
287 trigger_allbutself_cpu_backtrace();
289 if (hardlockup_panic)
290 nmi_panic(regs, "Hard LOCKUP");
292 if (wd_panic_timeout_tb < 0x7fffffff)
293 mtspr(SPRN_DEC, wd_panic_timeout_tb);
298 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
300 int cpu = smp_processor_id();
302 if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
303 return HRTIMER_NORESTART;
305 if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
306 return HRTIMER_NORESTART;
308 watchdog_timer_interrupt(cpu);
310 hrtimer_forward_now(hrtimer, ms_to_ktime(wd_timer_period_ms));
312 return HRTIMER_RESTART;
315 void arch_touch_nmi_watchdog(void)
317 unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
318 int cpu = smp_processor_id();
321 if (tb - per_cpu(wd_timer_tb, cpu) >= ticks) {
322 per_cpu(wd_timer_tb, cpu) = tb;
323 wd_smp_clear_cpu_pending(cpu, tb);
326 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
328 static void start_watchdog(void *arg)
330 struct hrtimer *hrtimer = this_cpu_ptr(&wd_hrtimer);
331 int cpu = smp_processor_id();
334 if (cpumask_test_cpu(cpu, &wd_cpus_enabled)) {
339 if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
342 if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
346 cpumask_set_cpu(cpu, &wd_cpus_enabled);
347 if (cpumask_weight(&wd_cpus_enabled) == 1) {
348 cpumask_set_cpu(cpu, &wd_smp_cpus_pending);
349 wd_smp_last_reset_tb = get_tb();
351 wd_smp_unlock(&flags);
353 *this_cpu_ptr(&wd_timer_tb) = get_tb();
355 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
356 hrtimer->function = watchdog_timer_fn;
357 hrtimer_start(hrtimer, ms_to_ktime(wd_timer_period_ms),
358 HRTIMER_MODE_REL_PINNED);
361 static int start_watchdog_on_cpu(unsigned int cpu)
363 return smp_call_function_single(cpu, start_watchdog, NULL, true);
366 static void stop_watchdog(void *arg)
368 struct hrtimer *hrtimer = this_cpu_ptr(&wd_hrtimer);
369 int cpu = smp_processor_id();
372 if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
373 return; /* Can happen in CPU unplug case */
375 hrtimer_cancel(hrtimer);
378 cpumask_clear_cpu(cpu, &wd_cpus_enabled);
379 wd_smp_unlock(&flags);
381 wd_smp_clear_cpu_pending(cpu, get_tb());
384 static int stop_watchdog_on_cpu(unsigned int cpu)
386 return smp_call_function_single(cpu, stop_watchdog, NULL, true);
389 static void watchdog_calc_timeouts(void)
391 wd_panic_timeout_tb = watchdog_thresh * ppc_tb_freq;
393 /* Have the SMP detector trigger a bit later */
394 wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
396 /* 2/5 is the factor that the perf based detector uses */
397 wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
400 void watchdog_nmi_stop(void)
404 for_each_cpu(cpu, &wd_cpus_enabled)
405 stop_watchdog_on_cpu(cpu);
408 void watchdog_nmi_start(void)
412 watchdog_calc_timeouts();
413 for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
414 start_watchdog_on_cpu(cpu);
418 * Invoked from core watchdog init.
420 int __init watchdog_nmi_probe(void)
424 err = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
425 "powerpc/watchdog:online",
426 start_watchdog_on_cpu,
427 stop_watchdog_on_cpu);
429 pr_warn("could not be initialized");