diff options
author | Cyrill Gorcunov <gorcunov@gmail.com> | 2008-05-24 11:36:41 -0400 |
---|---|---|
committer | Thomas Gleixner <tglx@linutronix.de> | 2008-05-25 16:32:51 -0400 |
commit | 1798bc22b2790bf2a956588e6b17c36ef79ceff7 (patch) | |
tree | 68dd3275aa52221fa4edeca4982f05be8e21a437 /arch | |
parent | fd5cea02de100197a4c26d9e103508cf09b50a82 (diff) |
x86: nmi_32/64.c - merge down nmi_32.c and nmi_64.c to nmi.c
Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>
Cc: hpa@zytor.com
Cc: mingo@redhat.com
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Diffstat (limited to 'arch')
-rw-r--r-- | arch/x86/kernel/Makefile | 2 | ||||
-rw-r--r-- | arch/x86/kernel/nmi.c (renamed from arch/x86/kernel/nmi_32.c) | 101 | ||||
-rw-r--r-- | arch/x86/kernel/nmi_64.c | 477 |
3 files changed, 65 insertions, 515 deletions
diff --git a/arch/x86/kernel/Makefile b/arch/x86/kernel/Makefile index 5e618c3b4720..d5dd6668225f 100644 --- a/arch/x86/kernel/Makefile +++ b/arch/x86/kernel/Makefile | |||
@@ -53,7 +53,7 @@ obj-$(CONFIG_X86_32_SMP) += smpcommon.o | |||
53 | obj-$(CONFIG_X86_64_SMP) += tsc_sync.o smpcommon.o | 53 | obj-$(CONFIG_X86_64_SMP) += tsc_sync.o smpcommon.o |
54 | obj-$(CONFIG_X86_TRAMPOLINE) += trampoline_$(BITS).o | 54 | obj-$(CONFIG_X86_TRAMPOLINE) += trampoline_$(BITS).o |
55 | obj-$(CONFIG_X86_MPPARSE) += mpparse.o | 55 | obj-$(CONFIG_X86_MPPARSE) += mpparse.o |
56 | obj-$(CONFIG_X86_LOCAL_APIC) += apic_$(BITS).o nmi_$(BITS).o | 56 | obj-$(CONFIG_X86_LOCAL_APIC) += apic_$(BITS).o nmi.o |
57 | obj-$(CONFIG_X86_IO_APIC) += io_apic_$(BITS).o | 57 | obj-$(CONFIG_X86_IO_APIC) += io_apic_$(BITS).o |
58 | obj-$(CONFIG_X86_REBOOTFIXUPS) += reboot_fixups_32.o | 58 | obj-$(CONFIG_X86_REBOOTFIXUPS) += reboot_fixups_32.o |
59 | obj-$(CONFIG_KEXEC) += machine_kexec_$(BITS).o | 59 | obj-$(CONFIG_KEXEC) += machine_kexec_$(BITS).o |
diff --git a/arch/x86/kernel/nmi_32.c b/arch/x86/kernel/nmi.c index a94dbedf2b74..69a839fc1eb0 100644 --- a/arch/x86/kernel/nmi_32.c +++ b/arch/x86/kernel/nmi.c | |||
@@ -11,10 +11,11 @@ | |||
11 | * Mikael Pettersson : PM converted to driver model. Disable/enable API. | 11 | * Mikael Pettersson : PM converted to driver model. Disable/enable API. |
12 | */ | 12 | */ |
13 | 13 | ||
14 | #include <linux/nmi.h> | ||
15 | #include <linux/mm.h> | ||
14 | #include <linux/delay.h> | 16 | #include <linux/delay.h> |
15 | #include <linux/interrupt.h> | 17 | #include <linux/interrupt.h> |
16 | #include <linux/module.h> | 18 | #include <linux/module.h> |
17 | #include <linux/nmi.h> | ||
18 | #include <linux/sysdev.h> | 19 | #include <linux/sysdev.h> |
19 | #include <linux/sysctl.h> | 20 | #include <linux/sysctl.h> |
20 | #include <linux/percpu.h> | 21 | #include <linux/percpu.h> |
@@ -22,13 +23,15 @@ | |||
22 | #include <linux/cpumask.h> | 23 | #include <linux/cpumask.h> |
23 | #include <linux/kernel_stat.h> | 24 | #include <linux/kernel_stat.h> |
24 | #include <linux/kdebug.h> | 25 | #include <linux/kdebug.h> |
25 | #include <linux/slab.h> | ||
26 | 26 | ||
27 | #include <asm/smp.h> | 27 | #include <asm/smp.h> |
28 | #include <asm/nmi.h> | 28 | #include <asm/nmi.h> |
29 | #include <asm/proto.h> | ||
29 | #include <asm/timer.h> | 30 | #include <asm/timer.h> |
30 | 31 | ||
31 | #include "mach_traps.h" | 32 | #include <asm/mce.h> |
33 | |||
34 | #include <mach_traps.h> | ||
32 | 35 | ||
33 | int unknown_nmi_panic; | 36 | int unknown_nmi_panic; |
34 | int nmi_watchdog_enabled; | 37 | int nmi_watchdog_enabled; |
@@ -45,19 +48,25 @@ atomic_t nmi_active = ATOMIC_INIT(0); /* oprofile uses this */ | |||
45 | static int panic_on_timeout; | 48 | static int panic_on_timeout; |
46 | 49 | ||
47 | unsigned int nmi_watchdog = NMI_DEFAULT; | 50 | unsigned int nmi_watchdog = NMI_DEFAULT; |
48 | static unsigned int nmi_hz = HZ; | ||
49 | 51 | ||
52 | static unsigned int nmi_hz = HZ; | ||
50 | static DEFINE_PER_CPU(short, wd_enabled); | 53 | static DEFINE_PER_CPU(short, wd_enabled); |
51 | |||
52 | static int endflag __initdata = 0; | 54 | static int endflag __initdata = 0; |
53 | 55 | ||
54 | static inline unsigned int get_nmi_count(int cpu) | 56 | static inline unsigned int get_nmi_count(int cpu) |
55 | { | 57 | { |
58 | #ifdef CONFIG_X86_64 | ||
59 | return cpu_pda(cpu)->__nmi_count; | ||
60 | #else | ||
56 | return nmi_count(cpu); | 61 | return nmi_count(cpu); |
62 | #endif | ||
57 | } | 63 | } |
58 | 64 | ||
59 | static inline int mce_in_progress(void) | 65 | static inline int mce_in_progress(void) |
60 | { | 66 | { |
67 | #if defined(CONFIX_X86_64) && defined(CONFIG_X86_MCE) | ||
68 | return atomic_read(&mce_entry) > 0; | ||
69 | #endif | ||
61 | return 0; | 70 | return 0; |
62 | } | 71 | } |
63 | 72 | ||
@@ -67,8 +76,12 @@ static inline int mce_in_progress(void) | |||
67 | */ | 76 | */ |
68 | static inline unsigned int get_timer_irqs(int cpu) | 77 | static inline unsigned int get_timer_irqs(int cpu) |
69 | { | 78 | { |
79 | #ifdef CONFIG_X86_64 | ||
80 | return read_pda(apic_timer_irqs) + read_pda(irq0_irqs); | ||
81 | #else | ||
70 | return per_cpu(irq_stat, cpu).apic_timer_irqs + | 82 | return per_cpu(irq_stat, cpu).apic_timer_irqs + |
71 | per_cpu(irq_stat, cpu).irq0_irqs; | 83 | per_cpu(irq_stat, cpu).irq0_irqs; |
84 | #endif | ||
72 | } | 85 | } |
73 | 86 | ||
74 | /* Run after command line and cpu_init init, but before all other checks */ | 87 | /* Run after command line and cpu_init init, but before all other checks */ |
@@ -76,26 +89,33 @@ void nmi_watchdog_default(void) | |||
76 | { | 89 | { |
77 | if (nmi_watchdog != NMI_DEFAULT) | 90 | if (nmi_watchdog != NMI_DEFAULT) |
78 | return; | 91 | return; |
92 | #ifdef CONFIG_X86_64 | ||
93 | nmi_watchdog = NMI_NONE; | ||
94 | #else | ||
79 | if (lapic_watchdog_ok()) | 95 | if (lapic_watchdog_ok()) |
80 | nmi_watchdog = NMI_LOCAL_APIC; | 96 | nmi_watchdog = NMI_LOCAL_APIC; |
81 | else | 97 | else |
82 | nmi_watchdog = NMI_IO_APIC; | 98 | nmi_watchdog = NMI_IO_APIC; |
99 | #endif | ||
83 | } | 100 | } |
84 | 101 | ||
85 | #ifdef CONFIG_SMP | 102 | #ifdef CONFIG_SMP |
86 | /* The performance counters used by NMI_LOCAL_APIC don't trigger when | 103 | /* |
104 | * The performance counters used by NMI_LOCAL_APIC don't trigger when | ||
87 | * the CPU is idle. To make sure the NMI watchdog really ticks on all | 105 | * the CPU is idle. To make sure the NMI watchdog really ticks on all |
88 | * CPUs during the test make them busy. | 106 | * CPUs during the test make them busy. |
89 | */ | 107 | */ |
90 | static __init void nmi_cpu_busy(void *data) | 108 | static __init void nmi_cpu_busy(void *data) |
91 | { | 109 | { |
92 | local_irq_enable_in_hardirq(); | 110 | local_irq_enable_in_hardirq(); |
93 | /* Intentionally don't use cpu_relax here. This is | 111 | /* |
94 | to make sure that the performance counter really ticks, | 112 | * Intentionally don't use cpu_relax here. This is |
95 | even if there is a simulator or similar that catches the | 113 | * to make sure that the performance counter really ticks, |
96 | pause instruction. On a real HT machine this is fine because | 114 | * even if there is a simulator or similar that catches the |
97 | all other CPUs are busy with "useless" delay loops and don't | 115 | * pause instruction. On a real HT machine this is fine because |
98 | care if they get somewhat less cycles. */ | 116 | * all other CPUs are busy with "useless" delay loops and don't |
117 | * care if they get somewhat less cycles. | ||
118 | */ | ||
99 | while (endflag == 0) | 119 | while (endflag == 0) |
100 | mb(); | 120 | mb(); |
101 | } | 121 | } |
@@ -106,7 +126,7 @@ int __init check_nmi_watchdog(void) | |||
106 | unsigned int *prev_nmi_count; | 126 | unsigned int *prev_nmi_count; |
107 | int cpu; | 127 | int cpu; |
108 | 128 | ||
109 | if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED)) | 129 | if (nmi_watchdog == NMI_NONE || nmi_watchdog == NMI_DISABLED) |
110 | return 0; | 130 | return 0; |
111 | 131 | ||
112 | if (!atomic_read(&nmi_active)) | 132 | if (!atomic_read(&nmi_active)) |
@@ -126,7 +146,7 @@ int __init check_nmi_watchdog(void) | |||
126 | for_each_possible_cpu(cpu) | 146 | for_each_possible_cpu(cpu) |
127 | prev_nmi_count[cpu] = get_nmi_count(cpu); | 147 | prev_nmi_count[cpu] = get_nmi_count(cpu); |
128 | local_irq_enable(); | 148 | local_irq_enable(); |
129 | mdelay((20*1000)/nmi_hz); // wait 20 ticks | 149 | mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */ |
130 | 150 | ||
131 | for_each_online_cpu(cpu) { | 151 | for_each_online_cpu(cpu) { |
132 | if (!per_cpu(wd_enabled, cpu)) | 152 | if (!per_cpu(wd_enabled, cpu)) |
@@ -149,16 +169,20 @@ int __init check_nmi_watchdog(void) | |||
149 | } | 169 | } |
150 | printk("OK.\n"); | 170 | printk("OK.\n"); |
151 | 171 | ||
152 | /* now that we know it works we can reduce NMI frequency to | 172 | /* |
153 | something more reasonable; makes a difference in some configs */ | 173 | * now that we know it works we can reduce NMI frequency to |
174 | * something more reasonable; makes a difference in some configs | ||
175 | */ | ||
154 | if (nmi_watchdog == NMI_LOCAL_APIC) | 176 | if (nmi_watchdog == NMI_LOCAL_APIC) |
155 | nmi_hz = lapic_adjust_nmi_hz(1); | 177 | nmi_hz = lapic_adjust_nmi_hz(1); |
156 | 178 | ||
157 | kfree(prev_nmi_count); | 179 | kfree(prev_nmi_count); |
158 | return 0; | 180 | return 0; |
181 | |||
159 | error: | 182 | error: |
183 | #ifdef CONFIG_X86_32 | ||
160 | timer_ack = !cpu_has_tsc; | 184 | timer_ack = !cpu_has_tsc; |
161 | 185 | #endif | |
162 | return -1; | 186 | return -1; |
163 | } | 187 | } |
164 | 188 | ||
@@ -176,18 +200,17 @@ static int __init setup_nmi_watchdog(char *str) | |||
176 | 200 | ||
177 | get_option(&str, &nmi); | 201 | get_option(&str, &nmi); |
178 | 202 | ||
179 | if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE)) | 203 | if (nmi >= NMI_INVALID || nmi < NMI_NONE) |
180 | return 0; | 204 | return 0; |
181 | 205 | ||
182 | nmi_watchdog = nmi; | 206 | nmi_watchdog = nmi; |
183 | return 1; | 207 | return 1; |
184 | } | 208 | } |
185 | |||
186 | __setup("nmi_watchdog=", setup_nmi_watchdog); | 209 | __setup("nmi_watchdog=", setup_nmi_watchdog); |
187 | 210 | ||
188 | 211 | /* | |
189 | /* Suspend/resume support */ | 212 | * Suspend/resume support |
190 | 213 | */ | |
191 | #ifdef CONFIG_PM | 214 | #ifdef CONFIG_PM |
192 | 215 | ||
193 | static int nmi_pm_active; /* nmi_active before suspend */ | 216 | static int nmi_pm_active; /* nmi_active before suspend */ |
@@ -211,7 +234,6 @@ static int lapic_nmi_resume(struct sys_device *dev) | |||
211 | return 0; | 234 | return 0; |
212 | } | 235 | } |
213 | 236 | ||
214 | |||
215 | static struct sysdev_class nmi_sysclass = { | 237 | static struct sysdev_class nmi_sysclass = { |
216 | .name = "lapic_nmi", | 238 | .name = "lapic_nmi", |
217 | .resume = lapic_nmi_resume, | 239 | .resume = lapic_nmi_resume, |
@@ -227,7 +249,8 @@ static int __init init_lapic_nmi_sysfs(void) | |||
227 | { | 249 | { |
228 | int error; | 250 | int error; |
229 | 251 | ||
230 | /* should really be a BUG_ON but b/c this is an | 252 | /* |
253 | * should really be a BUG_ON but b/c this is an | ||
231 | * init call, it just doesn't work. -dcz | 254 | * init call, it just doesn't work. -dcz |
232 | */ | 255 | */ |
233 | if (nmi_watchdog != NMI_LOCAL_APIC) | 256 | if (nmi_watchdog != NMI_LOCAL_APIC) |
@@ -241,6 +264,7 @@ static int __init init_lapic_nmi_sysfs(void) | |||
241 | error = sysdev_register(&device_lapic_nmi); | 264 | error = sysdev_register(&device_lapic_nmi); |
242 | return error; | 265 | return error; |
243 | } | 266 | } |
267 | |||
244 | /* must come after the local APIC's device_initcall() */ | 268 | /* must come after the local APIC's device_initcall() */ |
245 | late_initcall(init_lapic_nmi_sysfs); | 269 | late_initcall(init_lapic_nmi_sysfs); |
246 | 270 | ||
@@ -281,12 +305,13 @@ void setup_apic_nmi_watchdog(void *unused) | |||
281 | 305 | ||
282 | /* cheap hack to support suspend/resume */ | 306 | /* cheap hack to support suspend/resume */ |
283 | /* if cpu0 is not active neither should the other cpus */ | 307 | /* if cpu0 is not active neither should the other cpus */ |
284 | if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0)) | 308 | if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0) |
285 | return; | 309 | return; |
286 | 310 | ||
287 | switch (nmi_watchdog) { | 311 | switch (nmi_watchdog) { |
288 | case NMI_LOCAL_APIC: | 312 | case NMI_LOCAL_APIC: |
289 | __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */ | 313 | /* enable it before to avoid race with handler */ |
314 | __get_cpu_var(wd_enabled) = 1; | ||
290 | if (lapic_watchdog_init(nmi_hz) < 0) { | 315 | if (lapic_watchdog_init(nmi_hz) < 0) { |
291 | __get_cpu_var(wd_enabled) = 0; | 316 | __get_cpu_var(wd_enabled) = 0; |
292 | return; | 317 | return; |
@@ -301,9 +326,9 @@ void setup_apic_nmi_watchdog(void *unused) | |||
301 | void stop_apic_nmi_watchdog(void *unused) | 326 | void stop_apic_nmi_watchdog(void *unused) |
302 | { | 327 | { |
303 | /* only support LOCAL and IO APICs for now */ | 328 | /* only support LOCAL and IO APICs for now */ |
304 | if ((nmi_watchdog != NMI_LOCAL_APIC) && | 329 | if (nmi_watchdog != NMI_LOCAL_APIC && |
305 | (nmi_watchdog != NMI_IO_APIC)) | 330 | nmi_watchdog != NMI_IO_APIC) |
306 | return; | 331 | return; |
307 | if (__get_cpu_var(wd_enabled) == 0) | 332 | if (__get_cpu_var(wd_enabled) == 0) |
308 | return; | 333 | return; |
309 | if (nmi_watchdog == NMI_LOCAL_APIC) | 334 | if (nmi_watchdog == NMI_LOCAL_APIC) |
@@ -323,8 +348,7 @@ void stop_apic_nmi_watchdog(void *unused) | |||
323 | * since NMIs don't listen to _any_ locks, we have to be extremely | 348 | * since NMIs don't listen to _any_ locks, we have to be extremely |
324 | * careful not to rely on unsafe variables. The printk might lock | 349 | * careful not to rely on unsafe variables. The printk might lock |
325 | * up though, so we have to break up any console locks first ... | 350 | * up though, so we have to break up any console locks first ... |
326 | * [when there will be more tty-related locks, break them up | 351 | * [when there will be more tty-related locks, break them up here too!] |
327 | * here too!] | ||
328 | */ | 352 | */ |
329 | 353 | ||
330 | static DEFINE_PER_CPU(unsigned, last_irq_sum); | 354 | static DEFINE_PER_CPU(unsigned, last_irq_sum); |
@@ -357,7 +381,6 @@ EXPORT_SYMBOL(touch_nmi_watchdog); | |||
357 | notrace __kprobes int | 381 | notrace __kprobes int |
358 | nmi_watchdog_tick(struct pt_regs *regs, unsigned reason) | 382 | nmi_watchdog_tick(struct pt_regs *regs, unsigned reason) |
359 | { | 383 | { |
360 | |||
361 | /* | 384 | /* |
362 | * Since current_thread_info()-> is always on the stack, and we | 385 | * Since current_thread_info()-> is always on the stack, and we |
363 | * always switch the stack NMI-atomically, it's safe to use | 386 | * always switch the stack NMI-atomically, it's safe to use |
@@ -403,7 +426,7 @@ nmi_watchdog_tick(struct pt_regs *regs, unsigned reason) | |||
403 | * wait a few IRQs (5 seconds) before doing the oops ... | 426 | * wait a few IRQs (5 seconds) before doing the oops ... |
404 | */ | 427 | */ |
405 | local_inc(&__get_cpu_var(alert_counter)); | 428 | local_inc(&__get_cpu_var(alert_counter)); |
406 | if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz) | 429 | if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz) |
407 | /* | 430 | /* |
408 | * die_nmi will return ONLY if NOTIFY_STOP happens.. | 431 | * die_nmi will return ONLY if NOTIFY_STOP happens.. |
409 | */ | 432 | */ |
@@ -413,6 +436,7 @@ nmi_watchdog_tick(struct pt_regs *regs, unsigned reason) | |||
413 | __get_cpu_var(last_irq_sum) = sum; | 436 | __get_cpu_var(last_irq_sum) = sum; |
414 | local_set(&__get_cpu_var(alert_counter), 0); | 437 | local_set(&__get_cpu_var(alert_counter), 0); |
415 | } | 438 | } |
439 | |||
416 | /* see if the nmi watchdog went off */ | 440 | /* see if the nmi watchdog went off */ |
417 | if (!__get_cpu_var(wd_enabled)) | 441 | if (!__get_cpu_var(wd_enabled)) |
418 | return rc; | 442 | return rc; |
@@ -421,7 +445,8 @@ nmi_watchdog_tick(struct pt_regs *regs, unsigned reason) | |||
421 | rc |= lapic_wd_event(nmi_hz); | 445 | rc |= lapic_wd_event(nmi_hz); |
422 | break; | 446 | break; |
423 | case NMI_IO_APIC: | 447 | case NMI_IO_APIC: |
424 | /* don't know how to accurately check for this. | 448 | /* |
449 | * don't know how to accurately check for this. | ||
425 | * just assume it was a watchdog timer interrupt | 450 | * just assume it was a watchdog timer interrupt |
426 | * This matches the old behaviour. | 451 | * This matches the old behaviour. |
427 | */ | 452 | */ |
@@ -458,7 +483,8 @@ int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file, | |||
458 | return 0; | 483 | return 0; |
459 | 484 | ||
460 | if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) { | 485 | if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) { |
461 | printk( KERN_WARNING "NMI watchdog is permanently disabled\n"); | 486 | printk(KERN_WARNING |
487 | "NMI watchdog is permanently disabled\n"); | ||
462 | return -EIO; | 488 | return -EIO; |
463 | } | 489 | } |
464 | 490 | ||
@@ -471,14 +497,14 @@ int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file, | |||
471 | else | 497 | else |
472 | disable_lapic_nmi_watchdog(); | 498 | disable_lapic_nmi_watchdog(); |
473 | } else { | 499 | } else { |
474 | printk( KERN_WARNING | 500 | printk(KERN_WARNING |
475 | "NMI watchdog doesn't know what hardware to touch\n"); | 501 | "NMI watchdog doesn't know what hardware to touch\n"); |
476 | return -EIO; | 502 | return -EIO; |
477 | } | 503 | } |
478 | return 0; | 504 | return 0; |
479 | } | 505 | } |
480 | 506 | ||
481 | #endif | 507 | #endif /* CONFIG_SYSCTL */ |
482 | 508 | ||
483 | int do_nmi_callback(struct pt_regs *regs, int cpu) | 509 | int do_nmi_callback(struct pt_regs *regs, int cpu) |
484 | { | 510 | { |
@@ -504,3 +530,4 @@ void __trigger_all_cpu_backtrace(void) | |||
504 | 530 | ||
505 | EXPORT_SYMBOL(nmi_active); | 531 | EXPORT_SYMBOL(nmi_active); |
506 | EXPORT_SYMBOL(nmi_watchdog); | 532 | EXPORT_SYMBOL(nmi_watchdog); |
533 | |||
diff --git a/arch/x86/kernel/nmi_64.c b/arch/x86/kernel/nmi_64.c deleted file mode 100644 index c6802447262c..000000000000 --- a/arch/x86/kernel/nmi_64.c +++ /dev/null | |||
@@ -1,477 +0,0 @@ | |||
1 | /* | ||
2 | * NMI watchdog support on APIC systems | ||
3 | * | ||
4 | * Started by Ingo Molnar <mingo@redhat.com> | ||
5 | * | ||
6 | * Fixes: | ||
7 | * Mikael Pettersson : AMD K7 support for local APIC NMI watchdog. | ||
8 | * Mikael Pettersson : Power Management for local APIC NMI watchdog. | ||
9 | * Pavel Machek and | ||
10 | * Mikael Pettersson : PM converted to driver model. Disable/enable API. | ||
11 | */ | ||
12 | |||
13 | #include <linux/nmi.h> | ||
14 | #include <linux/mm.h> | ||
15 | #include <linux/delay.h> | ||
16 | #include <linux/interrupt.h> | ||
17 | #include <linux/module.h> | ||
18 | #include <linux/sysdev.h> | ||
19 | #include <linux/sysctl.h> | ||
20 | #include <linux/kprobes.h> | ||
21 | #include <linux/cpumask.h> | ||
22 | #include <linux/kdebug.h> | ||
23 | |||
24 | #include <asm/smp.h> | ||
25 | #include <asm/nmi.h> | ||
26 | #include <asm/proto.h> | ||
27 | #include <asm/mce.h> | ||
28 | |||
29 | #include <mach_traps.h> | ||
30 | |||
31 | int unknown_nmi_panic; | ||
32 | int nmi_watchdog_enabled; | ||
33 | |||
34 | static cpumask_t backtrace_mask = CPU_MASK_NONE; | ||
35 | |||
36 | /* nmi_active: | ||
37 | * >0: the lapic NMI watchdog is active, but can be disabled | ||
38 | * <0: the lapic NMI watchdog has not been set up, and cannot | ||
39 | * be enabled | ||
40 | * 0: the lapic NMI watchdog is disabled, but can be enabled | ||
41 | */ | ||
42 | atomic_t nmi_active = ATOMIC_INIT(0); /* oprofile uses this */ | ||
43 | static int panic_on_timeout; | ||
44 | |||
45 | unsigned int nmi_watchdog = NMI_DEFAULT; | ||
46 | static unsigned int nmi_hz = HZ; | ||
47 | |||
48 | static DEFINE_PER_CPU(short, wd_enabled); | ||
49 | |||
50 | static int endflag __initdata = 0; | ||
51 | |||
52 | static inline unsigned int get_nmi_count(int cpu) | ||
53 | { | ||
54 | return cpu_pda(cpu)->__nmi_count; | ||
55 | } | ||
56 | |||
57 | static inline int mce_in_progress(void) | ||
58 | { | ||
59 | #ifdef CONFIG_X86_MCE | ||
60 | return atomic_read(&mce_entry) > 0; | ||
61 | #endif | ||
62 | return 0; | ||
63 | } | ||
64 | |||
65 | /* | ||
66 | * Take the local apic timer and PIT/HPET into account. We don't | ||
67 | * know which one is active, when we have highres/dyntick on | ||
68 | */ | ||
69 | static inline unsigned int get_timer_irqs(int cpu) | ||
70 | { | ||
71 | return read_pda(apic_timer_irqs) + read_pda(irq0_irqs); | ||
72 | } | ||
73 | |||
74 | /* Run after command line and cpu_init init, but before all other checks */ | ||
75 | void nmi_watchdog_default(void) | ||
76 | { | ||
77 | if (nmi_watchdog != NMI_DEFAULT) | ||
78 | return; | ||
79 | nmi_watchdog = NMI_NONE; | ||
80 | } | ||
81 | |||
82 | #ifdef CONFIG_SMP | ||
83 | /* The performance counters used by NMI_LOCAL_APIC don't trigger when | ||
84 | * the CPU is idle. To make sure the NMI watchdog really ticks on all | ||
85 | * CPUs during the test make them busy. | ||
86 | */ | ||
87 | static __init void nmi_cpu_busy(void *data) | ||
88 | { | ||
89 | local_irq_enable_in_hardirq(); | ||
90 | /* Intentionally don't use cpu_relax here. This is | ||
91 | to make sure that the performance counter really ticks, | ||
92 | even if there is a simulator or similar that catches the | ||
93 | pause instruction. On a real HT machine this is fine because | ||
94 | all other CPUs are busy with "useless" delay loops and don't | ||
95 | care if they get somewhat less cycles. */ | ||
96 | while (endflag == 0) | ||
97 | mb(); | ||
98 | } | ||
99 | #endif | ||
100 | |||
101 | int __init check_nmi_watchdog(void) | ||
102 | { | ||
103 | unsigned int *prev_nmi_count; | ||
104 | int cpu; | ||
105 | |||
106 | if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED)) | ||
107 | return 0; | ||
108 | |||
109 | if (!atomic_read(&nmi_active)) | ||
110 | return 0; | ||
111 | |||
112 | prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL); | ||
113 | if (!prev_nmi_count) | ||
114 | return -1; | ||
115 | |||
116 | printk(KERN_INFO "Testing NMI watchdog ... "); | ||
117 | |||
118 | #ifdef CONFIG_SMP | ||
119 | if (nmi_watchdog == NMI_LOCAL_APIC) | ||
120 | smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0); | ||
121 | #endif | ||
122 | |||
123 | for_each_possible_cpu(cpu) | ||
124 | prev_nmi_count[cpu] = get_nmi_count(cpu); | ||
125 | local_irq_enable(); | ||
126 | mdelay((20*1000)/nmi_hz); // wait 20 ticks | ||
127 | |||
128 | for_each_online_cpu(cpu) { | ||
129 | if (!per_cpu(wd_enabled, cpu)) | ||
130 | continue; | ||
131 | if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5) { | ||
132 | printk(KERN_WARNING "WARNING: CPU#%d: NMI " | ||
133 | "appears to be stuck (%d->%d)!\n", | ||
134 | cpu, | ||
135 | prev_nmi_count[cpu], | ||
136 | get_nmi_count(cpu)); | ||
137 | per_cpu(wd_enabled, cpu) = 0; | ||
138 | atomic_dec(&nmi_active); | ||
139 | } | ||
140 | } | ||
141 | endflag = 1; | ||
142 | if (!atomic_read(&nmi_active)) { | ||
143 | kfree(prev_nmi_count); | ||
144 | atomic_set(&nmi_active, -1); | ||
145 | return -1; | ||
146 | } | ||
147 | printk("OK.\n"); | ||
148 | |||
149 | /* now that we know it works we can reduce NMI frequency to | ||
150 | something more reasonable; makes a difference in some configs */ | ||
151 | if (nmi_watchdog == NMI_LOCAL_APIC) | ||
152 | nmi_hz = lapic_adjust_nmi_hz(1); | ||
153 | |||
154 | kfree(prev_nmi_count); | ||
155 | return 0; | ||
156 | } | ||
157 | |||
158 | static int __init setup_nmi_watchdog(char *str) | ||
159 | { | ||
160 | int nmi; | ||
161 | |||
162 | if (!strncmp(str,"panic",5)) { | ||
163 | panic_on_timeout = 1; | ||
164 | str = strchr(str, ','); | ||
165 | if (!str) | ||
166 | return 1; | ||
167 | ++str; | ||
168 | } | ||
169 | |||
170 | get_option(&str, &nmi); | ||
171 | |||
172 | if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE)) | ||
173 | return 0; | ||
174 | |||
175 | nmi_watchdog = nmi; | ||
176 | return 1; | ||
177 | } | ||
178 | |||
179 | __setup("nmi_watchdog=", setup_nmi_watchdog); | ||
180 | |||
181 | #ifdef CONFIG_PM | ||
182 | |||
183 | static int nmi_pm_active; /* nmi_active before suspend */ | ||
184 | |||
185 | static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state) | ||
186 | { | ||
187 | /* only CPU0 goes here, other CPUs should be offline */ | ||
188 | nmi_pm_active = atomic_read(&nmi_active); | ||
189 | stop_apic_nmi_watchdog(NULL); | ||
190 | BUG_ON(atomic_read(&nmi_active) != 0); | ||
191 | return 0; | ||
192 | } | ||
193 | |||
194 | static int lapic_nmi_resume(struct sys_device *dev) | ||
195 | { | ||
196 | /* only CPU0 goes here, other CPUs should be offline */ | ||
197 | if (nmi_pm_active > 0) { | ||
198 | setup_apic_nmi_watchdog(NULL); | ||
199 | touch_nmi_watchdog(); | ||
200 | } | ||
201 | return 0; | ||
202 | } | ||
203 | |||
204 | static struct sysdev_class nmi_sysclass = { | ||
205 | .name = "lapic_nmi", | ||
206 | .resume = lapic_nmi_resume, | ||
207 | .suspend = lapic_nmi_suspend, | ||
208 | }; | ||
209 | |||
210 | static struct sys_device device_lapic_nmi = { | ||
211 | .id = 0, | ||
212 | .cls = &nmi_sysclass, | ||
213 | }; | ||
214 | |||
215 | static int __init init_lapic_nmi_sysfs(void) | ||
216 | { | ||
217 | int error; | ||
218 | |||
219 | /* should really be a BUG_ON but b/c this is an | ||
220 | * init call, it just doesn't work. -dcz | ||
221 | */ | ||
222 | if (nmi_watchdog != NMI_LOCAL_APIC) | ||
223 | return 0; | ||
224 | |||
225 | if (atomic_read(&nmi_active) < 0) | ||
226 | return 0; | ||
227 | |||
228 | error = sysdev_class_register(&nmi_sysclass); | ||
229 | if (!error) | ||
230 | error = sysdev_register(&device_lapic_nmi); | ||
231 | return error; | ||
232 | } | ||
233 | /* must come after the local APIC's device_initcall() */ | ||
234 | late_initcall(init_lapic_nmi_sysfs); | ||
235 | |||
236 | #endif /* CONFIG_PM */ | ||
237 | |||
238 | static void __acpi_nmi_enable(void *__unused) | ||
239 | { | ||
240 | apic_write_around(APIC_LVT0, APIC_DM_NMI); | ||
241 | } | ||
242 | |||
243 | /* | ||
244 | * Enable timer based NMIs on all CPUs: | ||
245 | */ | ||
246 | void acpi_nmi_enable(void) | ||
247 | { | ||
248 | if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC) | ||
249 | on_each_cpu(__acpi_nmi_enable, NULL, 0, 1); | ||
250 | } | ||
251 | |||
252 | static void __acpi_nmi_disable(void *__unused) | ||
253 | { | ||
254 | apic_write_around(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED); | ||
255 | } | ||
256 | |||
257 | /* | ||
258 | * Disable timer based NMIs on all CPUs: | ||
259 | */ | ||
260 | void acpi_nmi_disable(void) | ||
261 | { | ||
262 | if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC) | ||
263 | on_each_cpu(__acpi_nmi_disable, NULL, 0, 1); | ||
264 | } | ||
265 | |||
266 | void setup_apic_nmi_watchdog(void *unused) | ||
267 | { | ||
268 | if (__get_cpu_var(wd_enabled)) | ||
269 | return; | ||
270 | |||
271 | /* cheap hack to support suspend/resume */ | ||
272 | /* if cpu0 is not active neither should the other cpus */ | ||
273 | if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0)) | ||
274 | return; | ||
275 | |||
276 | switch (nmi_watchdog) { | ||
277 | case NMI_LOCAL_APIC: | ||
278 | __get_cpu_var(wd_enabled) = 1; | ||
279 | if (lapic_watchdog_init(nmi_hz) < 0) { | ||
280 | __get_cpu_var(wd_enabled) = 0; | ||
281 | return; | ||
282 | } | ||
283 | /* FALL THROUGH */ | ||
284 | case NMI_IO_APIC: | ||
285 | __get_cpu_var(wd_enabled) = 1; | ||
286 | atomic_inc(&nmi_active); | ||
287 | } | ||
288 | } | ||
289 | |||
290 | void stop_apic_nmi_watchdog(void *unused) | ||
291 | { | ||
292 | /* only support LOCAL and IO APICs for now */ | ||
293 | if ((nmi_watchdog != NMI_LOCAL_APIC) && | ||
294 | (nmi_watchdog != NMI_IO_APIC)) | ||
295 | return; | ||
296 | if (__get_cpu_var(wd_enabled) == 0) | ||
297 | return; | ||
298 | if (nmi_watchdog == NMI_LOCAL_APIC) | ||
299 | lapic_watchdog_stop(); | ||
300 | __get_cpu_var(wd_enabled) = 0; | ||
301 | atomic_dec(&nmi_active); | ||
302 | } | ||
303 | |||
304 | /* | ||
305 | * the best way to detect whether a CPU has a 'hard lockup' problem | ||
306 | * is to check it's local APIC timer IRQ counts. If they are not | ||
307 | * changing then that CPU has some problem. | ||
308 | * | ||
309 | * as these watchdog NMI IRQs are generated on every CPU, we only | ||
310 | * have to check the current processor. | ||
311 | */ | ||
312 | |||
313 | static DEFINE_PER_CPU(unsigned, last_irq_sum); | ||
314 | static DEFINE_PER_CPU(local_t, alert_counter); | ||
315 | static DEFINE_PER_CPU(int, nmi_touch); | ||
316 | |||
317 | void touch_nmi_watchdog(void) | ||
318 | { | ||
319 | if (nmi_watchdog > 0) { | ||
320 | unsigned cpu; | ||
321 | |||
322 | /* | ||
323 | * Tell other CPUs to reset their alert counters. We cannot | ||
324 | * do it ourselves because the alert count increase is not | ||
325 | * atomic. | ||
326 | */ | ||
327 | for_each_present_cpu(cpu) { | ||
328 | if (per_cpu(nmi_touch, cpu) != 1) | ||
329 | per_cpu(nmi_touch, cpu) = 1; | ||
330 | } | ||
331 | } | ||
332 | |||
333 | touch_softlockup_watchdog(); | ||
334 | } | ||
335 | EXPORT_SYMBOL(touch_nmi_watchdog); | ||
336 | |||
337 | notrace __kprobes int | ||
338 | nmi_watchdog_tick(struct pt_regs *regs, unsigned reason) | ||
339 | { | ||
340 | unsigned int sum; | ||
341 | int touched = 0; | ||
342 | int cpu = smp_processor_id(); | ||
343 | int rc = 0; | ||
344 | |||
345 | /* check for other users first */ | ||
346 | if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT) | ||
347 | == NOTIFY_STOP) { | ||
348 | rc = 1; | ||
349 | touched = 1; | ||
350 | } | ||
351 | |||
352 | sum = get_timer_irqs(cpu); | ||
353 | |||
354 | if (__get_cpu_var(nmi_touch)) { | ||
355 | __get_cpu_var(nmi_touch) = 0; | ||
356 | touched = 1; | ||
357 | } | ||
358 | |||
359 | if (cpu_isset(cpu, backtrace_mask)) { | ||
360 | static DEFINE_SPINLOCK(lock); /* Serialise the printks */ | ||
361 | |||
362 | spin_lock(&lock); | ||
363 | printk("NMI backtrace for cpu %d\n", cpu); | ||
364 | dump_stack(); | ||
365 | spin_unlock(&lock); | ||
366 | cpu_clear(cpu, backtrace_mask); | ||
367 | } | ||
368 | |||
369 | if (mce_in_progress()) | ||
370 | touched = 1; | ||
371 | |||
372 | /* if the apic timer isn't firing, this cpu isn't doing much */ | ||
373 | if (!touched && __get_cpu_var(last_irq_sum) == sum) { | ||
374 | /* | ||
375 | * Ayiee, looks like this CPU is stuck ... | ||
376 | * wait a few IRQs (5 seconds) before doing the oops ... | ||
377 | */ | ||
378 | local_inc(&__get_cpu_var(alert_counter)); | ||
379 | if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz) | ||
380 | die_nmi("NMI Watchdog detected LOCKUP", | ||
381 | regs, panic_on_timeout); | ||
382 | } else { | ||
383 | __get_cpu_var(last_irq_sum) = sum; | ||
384 | local_set(&__get_cpu_var(alert_counter), 0); | ||
385 | } | ||
386 | |||
387 | /* see if the nmi watchdog went off */ | ||
388 | if (!__get_cpu_var(wd_enabled)) | ||
389 | return rc; | ||
390 | switch (nmi_watchdog) { | ||
391 | case NMI_LOCAL_APIC: | ||
392 | rc |= lapic_wd_event(nmi_hz); | ||
393 | break; | ||
394 | case NMI_IO_APIC: | ||
395 | /* don't know how to accurately check for this. | ||
396 | * just assume it was a watchdog timer interrupt | ||
397 | * This matches the old behaviour. | ||
398 | */ | ||
399 | rc = 1; | ||
400 | break; | ||
401 | } | ||
402 | return rc; | ||
403 | } | ||
404 | |||
405 | #ifdef CONFIG_SYSCTL | ||
406 | |||
407 | static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu) | ||
408 | { | ||
409 | unsigned char reason = get_nmi_reason(); | ||
410 | char buf[64]; | ||
411 | |||
412 | sprintf(buf, "NMI received for unknown reason %02x\n", reason); | ||
413 | die_nmi(buf, regs, 1); /* Always panic here */ | ||
414 | return 0; | ||
415 | } | ||
416 | |||
417 | /* | ||
418 | * proc handler for /proc/sys/kernel/nmi | ||
419 | */ | ||
420 | int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file, | ||
421 | void __user *buffer, size_t *length, loff_t *ppos) | ||
422 | { | ||
423 | int old_state; | ||
424 | |||
425 | nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0; | ||
426 | old_state = nmi_watchdog_enabled; | ||
427 | proc_dointvec(table, write, file, buffer, length, ppos); | ||
428 | if (!!old_state == !!nmi_watchdog_enabled) | ||
429 | return 0; | ||
430 | |||
431 | if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) { | ||
432 | printk( KERN_WARNING "NMI watchdog is permanently disabled\n"); | ||
433 | return -EIO; | ||
434 | } | ||
435 | |||
436 | /* if nmi_watchdog is not set yet, then set it */ | ||
437 | nmi_watchdog_default(); | ||
438 | |||
439 | if (nmi_watchdog == NMI_LOCAL_APIC) { | ||
440 | if (nmi_watchdog_enabled) | ||
441 | enable_lapic_nmi_watchdog(); | ||
442 | else | ||
443 | disable_lapic_nmi_watchdog(); | ||
444 | } else { | ||
445 | printk( KERN_WARNING | ||
446 | "NMI watchdog doesn't know what hardware to touch\n"); | ||
447 | return -EIO; | ||
448 | } | ||
449 | return 0; | ||
450 | } | ||
451 | |||
452 | #endif | ||
453 | |||
454 | int do_nmi_callback(struct pt_regs *regs, int cpu) | ||
455 | { | ||
456 | #ifdef CONFIG_SYSCTL | ||
457 | if (unknown_nmi_panic) | ||
458 | return unknown_nmi_panic_callback(regs, cpu); | ||
459 | #endif | ||
460 | return 0; | ||
461 | } | ||
462 | |||
463 | void __trigger_all_cpu_backtrace(void) | ||
464 | { | ||
465 | int i; | ||
466 | |||
467 | backtrace_mask = cpu_online_map; | ||
468 | /* Wait for up to 10 seconds for all CPUs to do the backtrace */ | ||
469 | for (i = 0; i < 10 * 1000; i++) { | ||
470 | if (cpus_empty(backtrace_mask)) | ||
471 | break; | ||
472 | mdelay(1); | ||
473 | } | ||
474 | } | ||
475 | |||
476 | EXPORT_SYMBOL(nmi_active); | ||
477 | EXPORT_SYMBOL(nmi_watchdog); | ||