1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Watchdog support on powerpc systems.
4 *
5 * Copyright 2017, IBM Corporation.
6 *
7 * This uses code from arch/sparc/kernel/nmi.c and kernel/watchdog.c
8 */
9
10#define pr_fmt(fmt) "watchdog: " fmt
11
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/processor.h>
28#include <linux/smp.h>
29
30#include <asm/interrupt.h>
31#include <asm/paca.h>
32#include <asm/nmi.h>
33
34/*
35 * The powerpc watchdog ensures that each CPU is able to service timers.
36 * The watchdog sets up a simple timer on each CPU to run once per timer
37 * period, and updates a per-cpu timestamp and a "pending" cpumask. This is
38 * the heartbeat.
39 *
40 * Then there are two systems to check that the heartbeat is still running.
41 * The local soft-NMI, and the SMP checker.
42 *
43 * The soft-NMI checker can detect lockups on the local CPU. When interrupts
44 * are disabled with local_irq_disable(), platforms that use soft-masking
45 * can leave hardware interrupts enabled and handle them with a masked
46 * interrupt handler. The masked handler can send the timer interrupt to the
47 * watchdog's soft_nmi_interrupt(), which appears to Linux as an NMI
48 * interrupt, and can be used to detect CPUs stuck with IRQs disabled.
49 *
50 * The soft-NMI checker will compare the heartbeat timestamp for this CPU
51 * with the current time, and take action if the difference exceeds the
52 * watchdog threshold.
53 *
54 * The limitation of the soft-NMI watchdog is that it does not work when
55 * interrupts are hard disabled or otherwise not being serviced. This is
56 * solved by also having a SMP watchdog where all CPUs check all other
57 * CPUs heartbeat.
58 *
59 * The SMP checker can detect lockups on other CPUs. A global "pending"
60 * cpumask is kept, containing all CPUs which enable the watchdog. Each
61 * CPU clears their pending bit in their heartbeat timer. When the bitmask
62 * becomes empty, the last CPU to clear its pending bit updates a global
63 * timestamp and refills the pending bitmask.
64 *
65 * In the heartbeat timer, if any CPU notices that the global timestamp has
66 * not been updated for a period exceeding the watchdog threshold, then it
67 * means the CPU(s) with their bit still set in the pending mask have had
68 * their heartbeat stop, and action is taken.
69 *
70 * Some platforms implement true NMI IPIs, which can be used by the SMP
71 * watchdog to detect an unresponsive CPU and pull it out of its stuck
72 * state with the NMI IPI, to get crash/debug data from it. This way the
73 * SMP watchdog can detect hardware interrupts off lockups.
74 */
75
76static cpumask_t wd_cpus_enabled __read_mostly;
77
78static u64 wd_panic_timeout_tb __read_mostly; /* timebase ticks until panic */
79static u64 wd_smp_panic_timeout_tb __read_mostly; /* panic other CPUs */
80
81static u64 wd_timer_period_ms __read_mostly; /* interval between heartbeat */
82
83static DEFINE_PER_CPU(struct hrtimer, wd_hrtimer);
84static DEFINE_PER_CPU(u64, wd_timer_tb);
85
86/* SMP checker bits */
87static unsigned long __wd_smp_lock;
88static unsigned long __wd_reporting;
89static unsigned long __wd_nmi_output;
90static cpumask_t wd_smp_cpus_pending;
91static cpumask_t wd_smp_cpus_stuck;
92static u64 wd_smp_last_reset_tb;
93
94#ifdef CONFIG_PPC_PSERIES
95static u64 wd_timeout_pct;
96#endif
97
98/*
99 * Try to take the exclusive watchdog action / NMI IPI / printing lock.
100 * wd_smp_lock must be held. If this fails, we should return and wait
101 * for the watchdog to kick in again (or another CPU to trigger it).
102 *
103 * Importantly, if hardlockup_panic is set, wd_try_report failure should
104 * not delay the panic, because whichever other CPU is reporting will
105 * call panic.
106 */
107static bool wd_try_report(void)
108{
109 if (__wd_reporting)
110 return false;
111 __wd_reporting = 1;
112 return true;
113}
114
115/* End printing after successful wd_try_report. wd_smp_lock not required. */
116static void wd_end_reporting(void)
117{
118 smp_mb(); /* End printing "critical section" */
119 WARN_ON_ONCE(__wd_reporting == 0);
120 WRITE_ONCE(__wd_reporting, 0);
121}
122
123static inline void wd_smp_lock(unsigned long *flags)
124{
125 /*
126 * Avoid locking layers if possible.
127 * This may be called from low level interrupt handlers at some
128 * point in future.
129 */
130 raw_local_irq_save(*flags);
131 hard_irq_disable(); /* Make it soft-NMI safe */
132 while (unlikely(test_and_set_bit_lock(0, &__wd_smp_lock))) {
133 raw_local_irq_restore(*flags);
134 spin_until_cond(!test_bit(0, &__wd_smp_lock));
135 raw_local_irq_save(*flags);
136 hard_irq_disable();
137 }
138}
139
140static inline void wd_smp_unlock(unsigned long *flags)
141{
142 clear_bit_unlock(nr: 0, addr: &__wd_smp_lock);
143 raw_local_irq_restore(*flags);
144}
145
146static void wd_lockup_ipi(struct pt_regs *regs)
147{
148 int cpu = raw_smp_processor_id();
149 u64 tb = get_tb();
150
151 pr_emerg("CPU %d Hard LOCKUP\n", cpu);
152 pr_emerg("CPU %d TB:%lld, last heartbeat TB:%lld (%lldms ago)\n",
153 cpu, tb, per_cpu(wd_timer_tb, cpu),
154 tb_to_ns(tb - per_cpu(wd_timer_tb, cpu)) / 1000000);
155 print_modules();
156 print_irqtrace_events(current);
157 if (regs)
158 show_regs(regs);
159 else
160 dump_stack();
161
162 /*
163 * __wd_nmi_output must be set after we printk from NMI context.
164 *
165 * printk from NMI context defers printing to the console to irq_work.
166 * If that NMI was taken in some code that is hard-locked, then irqs
167 * are disabled so irq_work will never fire. That can result in the
168 * hard lockup messages being delayed (indefinitely, until something
169 * else kicks the console drivers).
170 *
171 * Setting __wd_nmi_output will cause another CPU to notice and kick
172 * the console drivers for us.
173 *
174 * xchg is not needed here (it could be a smp_mb and store), but xchg
175 * gives the memory ordering and atomicity required.
176 */
177 xchg(&__wd_nmi_output, 1);
178
179 /* Do not panic from here because that can recurse into NMI IPI layer */
180}
181
182static bool set_cpu_stuck(int cpu)
183{
184 cpumask_set_cpu(cpu, dstp: &wd_smp_cpus_stuck);
185 cpumask_clear_cpu(cpu, dstp: &wd_smp_cpus_pending);
186 /*
187 * See wd_smp_clear_cpu_pending()
188 */
189 smp_mb();
190 if (cpumask_empty(srcp: &wd_smp_cpus_pending)) {
191 wd_smp_last_reset_tb = get_tb();
192 cpumask_andnot(dstp: &wd_smp_cpus_pending,
193 src1p: &wd_cpus_enabled,
194 src2p: &wd_smp_cpus_stuck);
195 return true;
196 }
197 return false;
198}
199
200static void watchdog_smp_panic(int cpu)
201{
202 static cpumask_t wd_smp_cpus_ipi; // protected by reporting
203 unsigned long flags;
204 u64 tb, last_reset;
205 int c;
206
207 wd_smp_lock(flags: &flags);
208 /* Double check some things under lock */
209 tb = get_tb();
210 last_reset = wd_smp_last_reset_tb;
211 if ((s64)(tb - last_reset) < (s64)wd_smp_panic_timeout_tb)
212 goto out;
213 if (cpumask_test_cpu(cpu, cpumask: &wd_smp_cpus_pending))
214 goto out;
215 if (!wd_try_report())
216 goto out;
217 for_each_online_cpu(c) {
218 if (!cpumask_test_cpu(cpu: c, cpumask: &wd_smp_cpus_pending))
219 continue;
220 if (c == cpu)
221 continue; // should not happen
222
223 __cpumask_set_cpu(cpu: c, dstp: &wd_smp_cpus_ipi);
224 if (set_cpu_stuck(c))
225 break;
226 }
227 if (cpumask_empty(srcp: &wd_smp_cpus_ipi)) {
228 wd_end_reporting();
229 goto out;
230 }
231 wd_smp_unlock(flags: &flags);
232
233 pr_emerg("CPU %d detected hard LOCKUP on other CPUs %*pbl\n",
234 cpu, cpumask_pr_args(&wd_smp_cpus_ipi));
235 pr_emerg("CPU %d TB:%lld, last SMP heartbeat TB:%lld (%lldms ago)\n",
236 cpu, tb, last_reset, tb_to_ns(tb - last_reset) / 1000000);
237
238 if (!sysctl_hardlockup_all_cpu_backtrace) {
239 /*
240 * Try to trigger the stuck CPUs, unless we are going to
241 * get a backtrace on all of them anyway.
242 */
243 for_each_cpu(c, &wd_smp_cpus_ipi) {
244 smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
245 __cpumask_clear_cpu(cpu: c, dstp: &wd_smp_cpus_ipi);
246 }
247 } else {
248 trigger_allbutcpu_cpu_backtrace(exclude_cpu: cpu);
249 cpumask_clear(dstp: &wd_smp_cpus_ipi);
250 }
251
252 if (hardlockup_panic)
253 nmi_panic(NULL, msg: "Hard LOCKUP");
254
255 wd_end_reporting();
256
257 return;
258
259out:
260 wd_smp_unlock(flags: &flags);
261}
262
263static void wd_smp_clear_cpu_pending(int cpu)
264{
265 if (!cpumask_test_cpu(cpu, cpumask: &wd_smp_cpus_pending)) {
266 if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
267 struct pt_regs *regs = get_irq_regs();
268 unsigned long flags;
269
270 pr_emerg("CPU %d became unstuck TB:%lld\n",
271 cpu, get_tb());
272 print_irqtrace_events(current);
273 if (regs)
274 show_regs(regs);
275 else
276 dump_stack();
277
278 wd_smp_lock(flags: &flags);
279 cpumask_clear_cpu(cpu, dstp: &wd_smp_cpus_stuck);
280 wd_smp_unlock(flags: &flags);
281 } else {
282 /*
283 * The last CPU to clear pending should have reset the
284 * watchdog so we generally should not find it empty
285 * here if our CPU was clear. However it could happen
286 * due to a rare race with another CPU taking the
287 * last CPU out of the mask concurrently.
288 *
289 * We can't add a warning for it. But just in case
290 * there is a problem with the watchdog that is causing
291 * the mask to not be reset, try to kick it along here.
292 */
293 if (unlikely(cpumask_empty(&wd_smp_cpus_pending)))
294 goto none_pending;
295 }
296 return;
297 }
298
299 /*
300 * All other updates to wd_smp_cpus_pending are performed under
301 * wd_smp_lock. All of them are atomic except the case where the
302 * mask becomes empty and is reset. This will not happen here because
303 * cpu was tested to be in the bitmap (above), and a CPU only clears
304 * its own bit. _Except_ in the case where another CPU has detected a
305 * hard lockup on our CPU and takes us out of the pending mask. So in
306 * normal operation there will be no race here, no problem.
307 *
308 * In the lockup case, this atomic clear-bit vs a store that refills
309 * other bits in the accessed word wll not be a problem. The bit clear
310 * is atomic so it will not cause the store to get lost, and the store
311 * will never set this bit so it will not overwrite the bit clear. The
312 * only way for a stuck CPU to return to the pending bitmap is to
313 * become unstuck itself.
314 */
315 cpumask_clear_cpu(cpu, dstp: &wd_smp_cpus_pending);
316
317 /*
318 * Order the store to clear pending with the load(s) to check all
319 * words in the pending mask to check they are all empty. This orders
320 * with the same barrier on another CPU. This prevents two CPUs
321 * clearing the last 2 pending bits, but neither seeing the other's
322 * store when checking if the mask is empty, and missing an empty
323 * mask, which ends with a false positive.
324 */
325 smp_mb();
326 if (cpumask_empty(srcp: &wd_smp_cpus_pending)) {
327 unsigned long flags;
328
329none_pending:
330 /*
331 * Double check under lock because more than one CPU could see
332 * a clear mask with the lockless check after clearing their
333 * pending bits.
334 */
335 wd_smp_lock(flags: &flags);
336 if (cpumask_empty(srcp: &wd_smp_cpus_pending)) {
337 wd_smp_last_reset_tb = get_tb();
338 cpumask_andnot(dstp: &wd_smp_cpus_pending,
339 src1p: &wd_cpus_enabled,
340 src2p: &wd_smp_cpus_stuck);
341 }
342 wd_smp_unlock(flags: &flags);
343 }
344}
345
346static void watchdog_timer_interrupt(int cpu)
347{
348 u64 tb = get_tb();
349
350 per_cpu(wd_timer_tb, cpu) = tb;
351
352 wd_smp_clear_cpu_pending(cpu);
353
354 if ((s64)(tb - wd_smp_last_reset_tb) >= (s64)wd_smp_panic_timeout_tb)
355 watchdog_smp_panic(cpu);
356
357 if (__wd_nmi_output && xchg(&__wd_nmi_output, 0)) {
358 /*
359 * Something has called printk from NMI context. It might be
360 * stuck, so this triggers a flush that will get that
361 * printk output to the console.
362 *
363 * See wd_lockup_ipi.
364 */
365 printk_trigger_flush();
366 }
367}
368
369DEFINE_INTERRUPT_HANDLER_NMI(soft_nmi_interrupt)
370{
371 unsigned long flags;
372 int cpu = raw_smp_processor_id();
373 u64 tb;
374
375 /* should only arrive from kernel, with irqs disabled */
376 WARN_ON_ONCE(!arch_irq_disabled_regs(regs));
377
378 if (!cpumask_test_cpu(cpu, cpumask: &wd_cpus_enabled))
379 return 0;
380
381 __this_cpu_inc(irq_stat.soft_nmi_irqs);
382
383 tb = get_tb();
384 if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
385 /*
386 * Taking wd_smp_lock here means it is a soft-NMI lock, which
387 * means we can't take any regular or irqsafe spin locks while
388 * holding this lock. This is why timers can't printk while
389 * holding the lock.
390 */
391 wd_smp_lock(flags: &flags);
392 if (cpumask_test_cpu(cpu, cpumask: &wd_smp_cpus_stuck)) {
393 wd_smp_unlock(flags: &flags);
394 return 0;
395 }
396 if (!wd_try_report()) {
397 wd_smp_unlock(flags: &flags);
398 /* Couldn't report, try again in 100ms */
399 mtspr(SPRN_DEC, 100 * tb_ticks_per_usec * 1000);
400 return 0;
401 }
402
403 set_cpu_stuck(cpu);
404
405 wd_smp_unlock(flags: &flags);
406
407 pr_emerg("CPU %d self-detected hard LOCKUP @ %pS\n",
408 cpu, (void *)regs->nip);
409 pr_emerg("CPU %d TB:%lld, last heartbeat TB:%lld (%lldms ago)\n",
410 cpu, tb, per_cpu(wd_timer_tb, cpu),
411 tb_to_ns(tb - per_cpu(wd_timer_tb, cpu)) / 1000000);
412 print_modules();
413 print_irqtrace_events(current);
414 show_regs(regs);
415
416 xchg(&__wd_nmi_output, 1); // see wd_lockup_ipi
417
418 if (sysctl_hardlockup_all_cpu_backtrace)
419 trigger_allbutcpu_cpu_backtrace(exclude_cpu: cpu);
420
421 if (hardlockup_panic)
422 nmi_panic(regs, "Hard LOCKUP");
423
424 wd_end_reporting();
425 }
426 /*
427 * We are okay to change DEC in soft_nmi_interrupt because the masked
428 * handler has marked a DEC as pending, so the timer interrupt will be
429 * replayed as soon as local irqs are enabled again.
430 */
431 if (wd_panic_timeout_tb < 0x7fffffff)
432 mtspr(SPRN_DEC, wd_panic_timeout_tb);
433
434 return 0;
435}
436
437static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
438{
439 int cpu = smp_processor_id();
440
441 if (!(watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED))
442 return HRTIMER_NORESTART;
443
444 if (!cpumask_test_cpu(cpu, cpumask: &watchdog_cpumask))
445 return HRTIMER_NORESTART;
446
447 watchdog_timer_interrupt(cpu);
448
449 hrtimer_forward_now(timer: hrtimer, interval: ms_to_ktime(ms: wd_timer_period_ms));
450
451 return HRTIMER_RESTART;
452}
453
454void arch_touch_nmi_watchdog(void)
455{
456 unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
457 int cpu = smp_processor_id();
458 u64 tb;
459
460 if (!cpumask_test_cpu(cpu, cpumask: &watchdog_cpumask))
461 return;
462
463 tb = get_tb();
464 if (tb - per_cpu(wd_timer_tb, cpu) >= ticks) {
465 per_cpu(wd_timer_tb, cpu) = tb;
466 wd_smp_clear_cpu_pending(cpu);
467 }
468}
469EXPORT_SYMBOL(arch_touch_nmi_watchdog);
470
471static void start_watchdog(void *arg)
472{
473 struct hrtimer *hrtimer = this_cpu_ptr(&wd_hrtimer);
474 int cpu = smp_processor_id();
475 unsigned long flags;
476
477 if (cpumask_test_cpu(cpu, cpumask: &wd_cpus_enabled)) {
478 WARN_ON(1);
479 return;
480 }
481
482 if (!(watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED))
483 return;
484
485 if (!cpumask_test_cpu(cpu, cpumask: &watchdog_cpumask))
486 return;
487
488 wd_smp_lock(flags: &flags);
489 cpumask_set_cpu(cpu, dstp: &wd_cpus_enabled);
490 if (cpumask_weight(srcp: &wd_cpus_enabled) == 1) {
491 cpumask_set_cpu(cpu, dstp: &wd_smp_cpus_pending);
492 wd_smp_last_reset_tb = get_tb();
493 }
494 wd_smp_unlock(flags: &flags);
495
496 *this_cpu_ptr(&wd_timer_tb) = get_tb();
497
498 hrtimer_init(timer: hrtimer, CLOCK_MONOTONIC, mode: HRTIMER_MODE_REL);
499 hrtimer->function = watchdog_timer_fn;
500 hrtimer_start(timer: hrtimer, tim: ms_to_ktime(ms: wd_timer_period_ms),
501 mode: HRTIMER_MODE_REL_PINNED);
502}
503
504static int start_watchdog_on_cpu(unsigned int cpu)
505{
506 return smp_call_function_single(cpuid: cpu, func: start_watchdog, NULL, wait: true);
507}
508
509static void stop_watchdog(void *arg)
510{
511 struct hrtimer *hrtimer = this_cpu_ptr(&wd_hrtimer);
512 int cpu = smp_processor_id();
513 unsigned long flags;
514
515 if (!cpumask_test_cpu(cpu, cpumask: &wd_cpus_enabled))
516 return; /* Can happen in CPU unplug case */
517
518 hrtimer_cancel(timer: hrtimer);
519
520 wd_smp_lock(flags: &flags);
521 cpumask_clear_cpu(cpu, dstp: &wd_cpus_enabled);
522 wd_smp_unlock(flags: &flags);
523
524 wd_smp_clear_cpu_pending(cpu);
525}
526
527static int stop_watchdog_on_cpu(unsigned int cpu)
528{
529 return smp_call_function_single(cpuid: cpu, func: stop_watchdog, NULL, wait: true);
530}
531
532static void watchdog_calc_timeouts(void)
533{
534 u64 threshold = watchdog_thresh;
535
536#ifdef CONFIG_PPC_PSERIES
537 threshold += (READ_ONCE(wd_timeout_pct) * threshold) / 100;
538#endif
539
540 wd_panic_timeout_tb = threshold * ppc_tb_freq;
541
542 /* Have the SMP detector trigger a bit later */
543 wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
544
545 /* 2/5 is the factor that the perf based detector uses */
546 wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
547}
548
549void watchdog_hardlockup_stop(void)
550{
551 int cpu;
552
553 for_each_cpu(cpu, &wd_cpus_enabled)
554 stop_watchdog_on_cpu(cpu);
555}
556
557void watchdog_hardlockup_start(void)
558{
559 int cpu;
560
561 watchdog_calc_timeouts();
562 for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
563 start_watchdog_on_cpu(cpu);
564}
565
566/*
567 * Invoked from core watchdog init.
568 */
569int __init watchdog_hardlockup_probe(void)
570{
571 int err;
572
573 err = cpuhp_setup_state_nocalls(state: CPUHP_AP_ONLINE_DYN,
574 name: "powerpc/watchdog:online",
575 startup: start_watchdog_on_cpu,
576 teardown: stop_watchdog_on_cpu);
577 if (err < 0) {
578 pr_warn("could not be initialized");
579 return err;
580 }
581 return 0;
582}
583
584#ifdef CONFIG_PPC_PSERIES
585void watchdog_hardlockup_set_timeout_pct(u64 pct)
586{
587 pr_info("Set the NMI watchdog timeout factor to %llu%%\n", pct);
588 WRITE_ONCE(wd_timeout_pct, pct);
589 lockup_detector_reconfigure();
590}
591#endif
592

source code of linux/arch/powerpc/kernel/watchdog.c