aboutsummaryrefslogtreecommitdiffstats
path: root/kernel/watchdog_hld.c
diff options
context:
space:
mode:
Diffstat (limited to 'kernel/watchdog_hld.c')
-rw-r--r--kernel/watchdog_hld.c227
1 files changed, 227 insertions, 0 deletions
diff --git a/kernel/watchdog_hld.c b/kernel/watchdog_hld.c
new file mode 100644
index 000000000000..84016c8aee6b
--- /dev/null
+++ b/kernel/watchdog_hld.c
@@ -0,0 +1,227 @@
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 <asm/irq_regs.h>
17#include <linux/perf_event.h>
18
19static DEFINE_PER_CPU(bool, hard_watchdog_warn);
20static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
21static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
22
23/* boot commands */
24/*
25 * Should we panic when a soft-lockup or hard-lockup occurs:
26 */
27unsigned int __read_mostly hardlockup_panic =
28 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
29static unsigned long hardlockup_allcpu_dumped;
30/*
31 * We may not want to enable hard lockup detection by default in all cases,
32 * for example when running the kernel as a guest on a hypervisor. In these
33 * cases this function can be called to disable hard lockup detection. This
34 * function should only be executed once by the boot processor before the
35 * kernel command line parameters are parsed, because otherwise it is not
36 * possible to override this in hardlockup_panic_setup().
37 */
38void hardlockup_detector_disable(void)
39{
40 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
41}
42
43static int __init hardlockup_panic_setup(char *str)
44{
45 if (!strncmp(str, "panic", 5))
46 hardlockup_panic = 1;
47 else if (!strncmp(str, "nopanic", 7))
48 hardlockup_panic = 0;
49 else if (!strncmp(str, "0", 1))
50 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
51 else if (!strncmp(str, "1", 1))
52 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
53 return 1;
54}
55__setup("nmi_watchdog=", hardlockup_panic_setup);
56
57void touch_nmi_watchdog(void)
58{
59 /*
60 * Using __raw here because some code paths have
61 * preemption enabled. If preemption is enabled
62 * then interrupts should be enabled too, in which
63 * case we shouldn't have to worry about the watchdog
64 * going off.
65 */
66 raw_cpu_write(watchdog_nmi_touch, true);
67 touch_softlockup_watchdog();
68}
69EXPORT_SYMBOL(touch_nmi_watchdog);
70
71static struct perf_event_attr wd_hw_attr = {
72 .type = PERF_TYPE_HARDWARE,
73 .config = PERF_COUNT_HW_CPU_CYCLES,
74 .size = sizeof(struct perf_event_attr),
75 .pinned = 1,
76 .disabled = 1,
77};
78
79/* Callback function for perf event subsystem */
80static void watchdog_overflow_callback(struct perf_event *event,
81 struct perf_sample_data *data,
82 struct pt_regs *regs)
83{
84 /* Ensure the watchdog never gets throttled */
85 event->hw.interrupts = 0;
86
87 if (__this_cpu_read(watchdog_nmi_touch) == true) {
88 __this_cpu_write(watchdog_nmi_touch, false);
89 return;
90 }
91
92 /* check for a hardlockup
93 * This is done by making sure our timer interrupt
94 * is incrementing. The timer interrupt should have
95 * fired multiple times before we overflow'd. If it hasn't
96 * then this is a good indication the cpu is stuck
97 */
98 if (is_hardlockup()) {
99 int this_cpu = smp_processor_id();
100
101 /* only print hardlockups once */
102 if (__this_cpu_read(hard_watchdog_warn) == true)
103 return;
104
105 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
106 print_modules();
107 print_irqtrace_events(current);
108 if (regs)
109 show_regs(regs);
110 else
111 dump_stack();
112
113 /*
114 * Perform all-CPU dump only once to avoid multiple hardlockups
115 * generating interleaving traces
116 */
117 if (sysctl_hardlockup_all_cpu_backtrace &&
118 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
119 trigger_allbutself_cpu_backtrace();
120
121 if (hardlockup_panic)
122 nmi_panic(regs, "Hard LOCKUP");
123
124 __this_cpu_write(hard_watchdog_warn, true);
125 return;
126 }
127
128 __this_cpu_write(hard_watchdog_warn, false);
129 return;
130}
131
132/*
133 * People like the simple clean cpu node info on boot.
134 * Reduce the watchdog noise by only printing messages
135 * that are different from what cpu0 displayed.
136 */
137static unsigned long cpu0_err;
138
139int watchdog_nmi_enable(unsigned int cpu)
140{
141 struct perf_event_attr *wd_attr;
142 struct perf_event *event = per_cpu(watchdog_ev, cpu);
143
144 /* nothing to do if the hard lockup detector is disabled */
145 if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
146 goto out;
147
148 /* is it already setup and enabled? */
149 if (event && event->state > PERF_EVENT_STATE_OFF)
150 goto out;
151
152 /* it is setup but not enabled */
153 if (event != NULL)
154 goto out_enable;
155
156 wd_attr = &wd_hw_attr;
157 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
158
159 /* Try to register using hardware perf events */
160 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
161
162 /* save cpu0 error for future comparision */
163 if (cpu == 0 && IS_ERR(event))
164 cpu0_err = PTR_ERR(event);
165
166 if (!IS_ERR(event)) {
167 /* only print for cpu0 or different than cpu0 */
168 if (cpu == 0 || cpu0_err)
169 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
170 goto out_save;
171 }
172
173 /*
174 * Disable the hard lockup detector if _any_ CPU fails to set up
175 * set up the hardware perf event. The watchdog() function checks
176 * the NMI_WATCHDOG_ENABLED bit periodically.
177 *
178 * The barriers are for syncing up watchdog_enabled across all the
179 * cpus, as clear_bit() does not use barriers.
180 */
181 smp_mb__before_atomic();
182 clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
183 smp_mb__after_atomic();
184
185 /* skip displaying the same error again */
186 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
187 return PTR_ERR(event);
188
189 /* vary the KERN level based on the returned errno */
190 if (PTR_ERR(event) == -EOPNOTSUPP)
191 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
192 else if (PTR_ERR(event) == -ENOENT)
193 pr_warn("disabled (cpu%i): hardware events not enabled\n",
194 cpu);
195 else
196 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
197 cpu, PTR_ERR(event));
198
199 pr_info("Shutting down hard lockup detector on all cpus\n");
200
201 return PTR_ERR(event);
202
203 /* success path */
204out_save:
205 per_cpu(watchdog_ev, cpu) = event;
206out_enable:
207 perf_event_enable(per_cpu(watchdog_ev, cpu));
208out:
209 return 0;
210}
211
212void watchdog_nmi_disable(unsigned int cpu)
213{
214 struct perf_event *event = per_cpu(watchdog_ev, cpu);
215
216 if (event) {
217 perf_event_disable(event);
218 per_cpu(watchdog_ev, cpu) = NULL;
219
220 /* should be in cleanup, but blocks oprofile */
221 perf_event_release_kernel(event);
222 }
223 if (cpu == 0) {
224 /* watchdog_nmi_enable() expects this to be zero initially. */
225 cpu0_err = 0;
226 }
227}
generated by cgit v1.2.2 (git 2.25.0) at 2026-01-05 18:40:56 -0500