]>
Commit | Line | Data |
---|---|---|
73ce0511 BM |
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> | |
b17b0153 IM |
16 | #include <linux/sched/debug.h> |
17 | ||
73ce0511 BM |
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 | ||
73ce0511 | 25 | static unsigned long hardlockup_allcpu_dumped; |
73ce0511 | 26 | |
f2e0cff8 | 27 | void arch_touch_nmi_watchdog(void) |
73ce0511 BM |
28 | { |
29 | /* | |
30 | * Using __raw here because some code paths have | |
31 | * preemption enabled. If preemption is enabled | |
32 | * then interrupts should be enabled too, in which | |
33 | * case we shouldn't have to worry about the watchdog | |
34 | * going off. | |
35 | */ | |
36 | raw_cpu_write(watchdog_nmi_touch, true); | |
73ce0511 | 37 | } |
f2e0cff8 | 38 | EXPORT_SYMBOL(arch_touch_nmi_watchdog); |
73ce0511 BM |
39 | |
40 | static struct perf_event_attr wd_hw_attr = { | |
41 | .type = PERF_TYPE_HARDWARE, | |
42 | .config = PERF_COUNT_HW_CPU_CYCLES, | |
43 | .size = sizeof(struct perf_event_attr), | |
44 | .pinned = 1, | |
45 | .disabled = 1, | |
46 | }; | |
47 | ||
48 | /* Callback function for perf event subsystem */ | |
49 | static void watchdog_overflow_callback(struct perf_event *event, | |
50 | struct perf_sample_data *data, | |
51 | struct pt_regs *regs) | |
52 | { | |
53 | /* Ensure the watchdog never gets throttled */ | |
54 | event->hw.interrupts = 0; | |
55 | ||
b94f5118 DZ |
56 | if (atomic_read(&watchdog_park_in_progress) != 0) |
57 | return; | |
58 | ||
73ce0511 BM |
59 | if (__this_cpu_read(watchdog_nmi_touch) == true) { |
60 | __this_cpu_write(watchdog_nmi_touch, false); | |
61 | return; | |
62 | } | |
63 | ||
64 | /* check for a hardlockup | |
65 | * This is done by making sure our timer interrupt | |
66 | * is incrementing. The timer interrupt should have | |
67 | * fired multiple times before we overflow'd. If it hasn't | |
68 | * then this is a good indication the cpu is stuck | |
69 | */ | |
70 | if (is_hardlockup()) { | |
71 | int this_cpu = smp_processor_id(); | |
72 | ||
73 | /* only print hardlockups once */ | |
74 | if (__this_cpu_read(hard_watchdog_warn) == true) | |
75 | return; | |
76 | ||
77 | pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu); | |
78 | print_modules(); | |
79 | print_irqtrace_events(current); | |
80 | if (regs) | |
81 | show_regs(regs); | |
82 | else | |
83 | dump_stack(); | |
84 | ||
85 | /* | |
86 | * Perform all-CPU dump only once to avoid multiple hardlockups | |
87 | * generating interleaving traces | |
88 | */ | |
89 | if (sysctl_hardlockup_all_cpu_backtrace && | |
90 | !test_and_set_bit(0, &hardlockup_allcpu_dumped)) | |
91 | trigger_allbutself_cpu_backtrace(); | |
92 | ||
93 | if (hardlockup_panic) | |
94 | nmi_panic(regs, "Hard LOCKUP"); | |
95 | ||
96 | __this_cpu_write(hard_watchdog_warn, true); | |
97 | return; | |
98 | } | |
99 | ||
100 | __this_cpu_write(hard_watchdog_warn, false); | |
101 | return; | |
102 | } | |
103 | ||
104 | /* | |
105 | * People like the simple clean cpu node info on boot. | |
106 | * Reduce the watchdog noise by only printing messages | |
107 | * that are different from what cpu0 displayed. | |
108 | */ | |
8dcde9de PB |
109 | static unsigned long firstcpu_err; |
110 | static atomic_t watchdog_cpus; | |
73ce0511 BM |
111 | |
112 | int watchdog_nmi_enable(unsigned int cpu) | |
113 | { | |
114 | struct perf_event_attr *wd_attr; | |
115 | struct perf_event *event = per_cpu(watchdog_ev, cpu); | |
8dcde9de | 116 | int firstcpu = 0; |
73ce0511 BM |
117 | |
118 | /* nothing to do if the hard lockup detector is disabled */ | |
119 | if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED)) | |
120 | goto out; | |
121 | ||
122 | /* is it already setup and enabled? */ | |
123 | if (event && event->state > PERF_EVENT_STATE_OFF) | |
124 | goto out; | |
125 | ||
126 | /* it is setup but not enabled */ | |
127 | if (event != NULL) | |
128 | goto out_enable; | |
129 | ||
8dcde9de PB |
130 | if (atomic_inc_return(&watchdog_cpus) == 1) |
131 | firstcpu = 1; | |
132 | ||
73ce0511 BM |
133 | wd_attr = &wd_hw_attr; |
134 | wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh); | |
135 | ||
136 | /* Try to register using hardware perf events */ | |
137 | event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL); | |
138 | ||
8dcde9de PB |
139 | /* save the first cpu's error for future comparision */ |
140 | if (firstcpu && IS_ERR(event)) | |
141 | firstcpu_err = PTR_ERR(event); | |
73ce0511 BM |
142 | |
143 | if (!IS_ERR(event)) { | |
8dcde9de PB |
144 | /* only print for the first cpu initialized */ |
145 | if (firstcpu || firstcpu_err) | |
73ce0511 BM |
146 | pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n"); |
147 | goto out_save; | |
148 | } | |
149 | ||
150 | /* | |
151 | * Disable the hard lockup detector if _any_ CPU fails to set up | |
152 | * set up the hardware perf event. The watchdog() function checks | |
153 | * the NMI_WATCHDOG_ENABLED bit periodically. | |
154 | * | |
155 | * The barriers are for syncing up watchdog_enabled across all the | |
156 | * cpus, as clear_bit() does not use barriers. | |
157 | */ | |
158 | smp_mb__before_atomic(); | |
159 | clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled); | |
160 | smp_mb__after_atomic(); | |
161 | ||
162 | /* skip displaying the same error again */ | |
8dcde9de | 163 | if (!firstcpu && (PTR_ERR(event) == firstcpu_err)) |
73ce0511 BM |
164 | return PTR_ERR(event); |
165 | ||
166 | /* vary the KERN level based on the returned errno */ | |
167 | if (PTR_ERR(event) == -EOPNOTSUPP) | |
168 | pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu); | |
169 | else if (PTR_ERR(event) == -ENOENT) | |
170 | pr_warn("disabled (cpu%i): hardware events not enabled\n", | |
171 | cpu); | |
172 | else | |
173 | pr_err("disabled (cpu%i): unable to create perf event: %ld\n", | |
174 | cpu, PTR_ERR(event)); | |
175 | ||
176 | pr_info("Shutting down hard lockup detector on all cpus\n"); | |
177 | ||
178 | return PTR_ERR(event); | |
179 | ||
180 | /* success path */ | |
181 | out_save: | |
182 | per_cpu(watchdog_ev, cpu) = event; | |
183 | out_enable: | |
184 | perf_event_enable(per_cpu(watchdog_ev, cpu)); | |
185 | out: | |
186 | return 0; | |
187 | } | |
188 | ||
189 | void watchdog_nmi_disable(unsigned int cpu) | |
190 | { | |
191 | struct perf_event *event = per_cpu(watchdog_ev, cpu); | |
192 | ||
193 | if (event) { | |
194 | perf_event_disable(event); | |
195 | per_cpu(watchdog_ev, cpu) = NULL; | |
196 | ||
197 | /* should be in cleanup, but blocks oprofile */ | |
198 | perf_event_release_kernel(event); | |
8dcde9de | 199 | |
73ce0511 | 200 | /* watchdog_nmi_enable() expects this to be zero initially. */ |
8dcde9de PB |
201 | if (atomic_dec_and_test(&watchdog_cpus)) |
202 | firstcpu_err = 0; | |
73ce0511 BM |
203 | } |
204 | } |