[PATCH] x86: Temporarily revert parts of the Core 2 nmi nmi watchdog support
[linux-2.6] / arch / x86_64 / kernel / nmi.c
1 /*
2  *  linux/arch/x86_64/nmi.c
3  *
4  *  NMI watchdog support on APIC systems
5  *
6  *  Started by Ingo Molnar <mingo@redhat.com>
7  *
8  *  Fixes:
9  *  Mikael Pettersson   : AMD K7 support for local APIC NMI watchdog.
10  *  Mikael Pettersson   : Power Management for local APIC NMI watchdog.
11  *  Pavel Machek and
12  *  Mikael Pettersson   : PM converted to driver model. Disable/enable API.
13  */
14
15 #include <linux/mm.h>
16 #include <linux/delay.h>
17 #include <linux/interrupt.h>
18 #include <linux/module.h>
19 #include <linux/sysdev.h>
20 #include <linux/nmi.h>
21 #include <linux/sysctl.h>
22 #include <linux/kprobes.h>
23
24 #include <asm/smp.h>
25 #include <asm/nmi.h>
26 #include <asm/proto.h>
27 #include <asm/kdebug.h>
28 #include <asm/mce.h>
29
30 /*
31  * lapic_nmi_owner tracks the ownership of the lapic NMI hardware:
32  * - it may be reserved by some other driver, or not
33  * - when not reserved by some other driver, it may be used for
34  *   the NMI watchdog, or not
35  *
36  * This is maintained separately from nmi_active because the NMI
37  * watchdog may also be driven from the I/O APIC timer.
38  */
39 static DEFINE_SPINLOCK(lapic_nmi_owner_lock);
40 static unsigned int lapic_nmi_owner;
41 #define LAPIC_NMI_WATCHDOG      (1<<0)
42 #define LAPIC_NMI_RESERVED      (1<<1)
43
44 /* nmi_active:
45  * +1: the lapic NMI watchdog is active, but can be disabled
46  *  0: the lapic NMI watchdog has not been set up, and cannot
47  *     be enabled
48  * -1: the lapic NMI watchdog is disabled, but can be enabled
49  */
50 int nmi_active;         /* oprofile uses this */
51 int panic_on_timeout;
52
53 unsigned int nmi_watchdog = NMI_DEFAULT;
54 static unsigned int nmi_hz = HZ;
55 static unsigned int nmi_perfctr_msr;    /* the MSR to reset in NMI handler */
56 static unsigned int nmi_p4_cccr_val;
57
58 /* Note that these events don't tick when the CPU idles. This means
59    the frequency varies with CPU load. */
60
61 #define K7_EVNTSEL_ENABLE       (1 << 22)
62 #define K7_EVNTSEL_INT          (1 << 20)
63 #define K7_EVNTSEL_OS           (1 << 17)
64 #define K7_EVNTSEL_USR          (1 << 16)
65 #define K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING    0x76
66 #define K7_NMI_EVENT            K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING
67
68 #define MSR_P4_MISC_ENABLE      0x1A0
69 #define MSR_P4_MISC_ENABLE_PERF_AVAIL   (1<<7)
70 #define MSR_P4_MISC_ENABLE_PEBS_UNAVAIL (1<<12)
71 #define MSR_P4_PERFCTR0         0x300
72 #define MSR_P4_CCCR0            0x360
73 #define P4_ESCR_EVENT_SELECT(N) ((N)<<25)
74 #define P4_ESCR_OS              (1<<3)
75 #define P4_ESCR_USR             (1<<2)
76 #define P4_CCCR_OVF_PMI0        (1<<26)
77 #define P4_CCCR_OVF_PMI1        (1<<27)
78 #define P4_CCCR_THRESHOLD(N)    ((N)<<20)
79 #define P4_CCCR_COMPLEMENT      (1<<19)
80 #define P4_CCCR_COMPARE         (1<<18)
81 #define P4_CCCR_REQUIRED        (3<<16)
82 #define P4_CCCR_ESCR_SELECT(N)  ((N)<<13)
83 #define P4_CCCR_ENABLE          (1<<12)
84 /* Set up IQ_COUNTER0 to behave like a clock, by having IQ_CCCR0 filter
85    CRU_ESCR0 (with any non-null event selector) through a complemented
86    max threshold. [IA32-Vol3, Section 14.9.9] */
87 #define MSR_P4_IQ_COUNTER0      0x30C
88 #define P4_NMI_CRU_ESCR0        (P4_ESCR_EVENT_SELECT(0x3F)|P4_ESCR_OS|P4_ESCR_USR)
89 #define P4_NMI_IQ_CCCR0 \
90         (P4_CCCR_OVF_PMI0|P4_CCCR_THRESHOLD(15)|P4_CCCR_COMPLEMENT|     \
91          P4_CCCR_COMPARE|P4_CCCR_REQUIRED|P4_CCCR_ESCR_SELECT(4)|P4_CCCR_ENABLE)
92
93 static __cpuinit inline int nmi_known_cpu(void)
94 {
95         switch (boot_cpu_data.x86_vendor) {
96         case X86_VENDOR_AMD:
97                 return boot_cpu_data.x86 == 15;
98         case X86_VENDOR_INTEL:
99                 return boot_cpu_data.x86 == 15;
100         }
101         return 0;
102 }
103
104 /* Run after command line and cpu_init init, but before all other checks */
105 void __cpuinit nmi_watchdog_default(void)
106 {
107         if (nmi_watchdog != NMI_DEFAULT)
108                 return;
109         if (nmi_known_cpu())
110                 nmi_watchdog = NMI_LOCAL_APIC;
111         else
112                 nmi_watchdog = NMI_IO_APIC;
113 }
114
115 #ifdef CONFIG_SMP
116 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
117  * the CPU is idle. To make sure the NMI watchdog really ticks on all
118  * CPUs during the test make them busy.
119  */
120 static __init void nmi_cpu_busy(void *data)
121 {
122         volatile int *endflag = data;
123         local_irq_enable_in_hardirq();
124         /* Intentionally don't use cpu_relax here. This is
125            to make sure that the performance counter really ticks,
126            even if there is a simulator or similar that catches the
127            pause instruction. On a real HT machine this is fine because
128            all other CPUs are busy with "useless" delay loops and don't
129            care if they get somewhat less cycles. */
130         while (*endflag == 0)
131                 barrier();
132 }
133 #endif
134
135 int __init check_nmi_watchdog (void)
136 {
137         volatile int endflag = 0;
138         int *counts;
139         int cpu;
140
141         counts = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
142         if (!counts)
143                 return -1;
144
145         printk(KERN_INFO "testing NMI watchdog ... ");
146
147 #ifdef CONFIG_SMP
148         if (nmi_watchdog == NMI_LOCAL_APIC)
149                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
150 #endif
151
152         for (cpu = 0; cpu < NR_CPUS; cpu++)
153                 counts[cpu] = cpu_pda(cpu)->__nmi_count;
154         local_irq_enable();
155         mdelay((10*1000)/nmi_hz); // wait 10 ticks
156
157         for_each_online_cpu(cpu) {
158                 if (cpu_pda(cpu)->__nmi_count - counts[cpu] <= 5) {
159                         endflag = 1;
160                         printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n",
161                                cpu,
162                                counts[cpu],
163                                cpu_pda(cpu)->__nmi_count);
164                         nmi_active = 0;
165                         lapic_nmi_owner &= ~LAPIC_NMI_WATCHDOG;
166                         nmi_perfctr_msr = 0;
167                         kfree(counts);
168                         return -1;
169                 }
170         }
171         endflag = 1;
172         printk("OK.\n");
173
174         /* now that we know it works we can reduce NMI frequency to
175            something more reasonable; makes a difference in some configs */
176         if (nmi_watchdog == NMI_LOCAL_APIC)
177                 nmi_hz = 1;
178
179         kfree(counts);
180         return 0;
181 }
182
183 int __init setup_nmi_watchdog(char *str)
184 {
185         int nmi;
186
187         if (!strncmp(str,"panic",5)) {
188                 panic_on_timeout = 1;
189                 str = strchr(str, ',');
190                 if (!str)
191                         return 1;
192                 ++str;
193         }
194
195         get_option(&str, &nmi);
196
197         if (nmi >= NMI_INVALID)
198                 return 0;
199         nmi_watchdog = nmi;
200         return 1;
201 }
202
203 __setup("nmi_watchdog=", setup_nmi_watchdog);
204
205 static void disable_lapic_nmi_watchdog(void)
206 {
207         if (nmi_active <= 0)
208                 return;
209         switch (boot_cpu_data.x86_vendor) {
210         case X86_VENDOR_AMD:
211                 wrmsr(MSR_K7_EVNTSEL0, 0, 0);
212                 break;
213         case X86_VENDOR_INTEL:
214                 if (boot_cpu_data.x86 == 15) {
215                         wrmsr(MSR_P4_IQ_CCCR0, 0, 0);
216                         wrmsr(MSR_P4_CRU_ESCR0, 0, 0);
217                 }
218                 break;
219         }
220         nmi_active = -1;
221         /* tell do_nmi() and others that we're not active any more */
222         nmi_watchdog = 0;
223 }
224
225 static void enable_lapic_nmi_watchdog(void)
226 {
227         if (nmi_active < 0) {
228                 nmi_watchdog = NMI_LOCAL_APIC;
229                 touch_nmi_watchdog();
230                 setup_apic_nmi_watchdog();
231         }
232 }
233
234 int reserve_lapic_nmi(void)
235 {
236         unsigned int old_owner;
237
238         spin_lock(&lapic_nmi_owner_lock);
239         old_owner = lapic_nmi_owner;
240         lapic_nmi_owner |= LAPIC_NMI_RESERVED;
241         spin_unlock(&lapic_nmi_owner_lock);
242         if (old_owner & LAPIC_NMI_RESERVED)
243                 return -EBUSY;
244         if (old_owner & LAPIC_NMI_WATCHDOG)
245                 disable_lapic_nmi_watchdog();
246         return 0;
247 }
248
249 void release_lapic_nmi(void)
250 {
251         unsigned int new_owner;
252
253         spin_lock(&lapic_nmi_owner_lock);
254         new_owner = lapic_nmi_owner & ~LAPIC_NMI_RESERVED;
255         lapic_nmi_owner = new_owner;
256         spin_unlock(&lapic_nmi_owner_lock);
257         if (new_owner & LAPIC_NMI_WATCHDOG)
258                 enable_lapic_nmi_watchdog();
259 }
260
261 void disable_timer_nmi_watchdog(void)
262 {
263         if ((nmi_watchdog != NMI_IO_APIC) || (nmi_active <= 0))
264                 return;
265
266         disable_irq(0);
267         unset_nmi_callback();
268         nmi_active = -1;
269         nmi_watchdog = NMI_NONE;
270 }
271
272 void enable_timer_nmi_watchdog(void)
273 {
274         if (nmi_active < 0) {
275                 nmi_watchdog = NMI_IO_APIC;
276                 touch_nmi_watchdog();
277                 nmi_active = 1;
278                 enable_irq(0);
279         }
280 }
281
282 #ifdef CONFIG_PM
283
284 static int nmi_pm_active; /* nmi_active before suspend */
285
286 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
287 {
288         nmi_pm_active = nmi_active;
289         disable_lapic_nmi_watchdog();
290         return 0;
291 }
292
293 static int lapic_nmi_resume(struct sys_device *dev)
294 {
295         if (nmi_pm_active > 0)
296         enable_lapic_nmi_watchdog();
297         return 0;
298 }
299
300 static struct sysdev_class nmi_sysclass = {
301         set_kset_name("lapic_nmi"),
302         .resume         = lapic_nmi_resume,
303         .suspend        = lapic_nmi_suspend,
304 };
305
306 static struct sys_device device_lapic_nmi = {
307         .id             = 0,
308         .cls    = &nmi_sysclass,
309 };
310
311 static int __init init_lapic_nmi_sysfs(void)
312 {
313         int error;
314
315         if (nmi_active == 0 || nmi_watchdog != NMI_LOCAL_APIC)
316                 return 0;
317
318         error = sysdev_class_register(&nmi_sysclass);
319         if (!error)
320                 error = sysdev_register(&device_lapic_nmi);
321         return error;
322 }
323 /* must come after the local APIC's device_initcall() */
324 late_initcall(init_lapic_nmi_sysfs);
325
326 #endif  /* CONFIG_PM */
327
328 /*
329  * Activate the NMI watchdog via the local APIC.
330  * Original code written by Keith Owens.
331  */
332
333 static void clear_msr_range(unsigned int base, unsigned int n)
334 {
335         unsigned int i;
336
337         for(i = 0; i < n; ++i)
338                 wrmsr(base+i, 0, 0);
339 }
340
341 static void setup_k7_watchdog(void)
342 {
343         int i;
344         unsigned int evntsel;
345
346         nmi_perfctr_msr = MSR_K7_PERFCTR0;
347
348         for(i = 0; i < 4; ++i) {
349                 /* Simulator may not support it */
350                 if (checking_wrmsrl(MSR_K7_EVNTSEL0+i, 0UL)) {
351                         nmi_perfctr_msr = 0;
352                         return;
353                 }
354                 wrmsrl(MSR_K7_PERFCTR0+i, 0UL);
355         }
356
357         evntsel = K7_EVNTSEL_INT
358                 | K7_EVNTSEL_OS
359                 | K7_EVNTSEL_USR
360                 | K7_NMI_EVENT;
361
362         wrmsr(MSR_K7_EVNTSEL0, evntsel, 0);
363         wrmsrl(MSR_K7_PERFCTR0, -((u64)cpu_khz * 1000 / nmi_hz));
364         apic_write(APIC_LVTPC, APIC_DM_NMI);
365         evntsel |= K7_EVNTSEL_ENABLE;
366         wrmsr(MSR_K7_EVNTSEL0, evntsel, 0);
367 }
368
369
370 static int setup_p4_watchdog(void)
371 {
372         unsigned int misc_enable, dummy;
373
374         rdmsr(MSR_P4_MISC_ENABLE, misc_enable, dummy);
375         if (!(misc_enable & MSR_P4_MISC_ENABLE_PERF_AVAIL))
376                 return 0;
377
378         nmi_perfctr_msr = MSR_P4_IQ_COUNTER0;
379         nmi_p4_cccr_val = P4_NMI_IQ_CCCR0;
380 #ifdef CONFIG_SMP
381         if (smp_num_siblings == 2)
382                 nmi_p4_cccr_val |= P4_CCCR_OVF_PMI1;
383 #endif
384
385         if (!(misc_enable & MSR_P4_MISC_ENABLE_PEBS_UNAVAIL))
386                 clear_msr_range(0x3F1, 2);
387         /* MSR 0x3F0 seems to have a default value of 0xFC00, but current
388            docs doesn't fully define it, so leave it alone for now. */
389         if (boot_cpu_data.x86_model >= 0x3) {
390                 /* MSR_P4_IQ_ESCR0/1 (0x3ba/0x3bb) removed */
391                 clear_msr_range(0x3A0, 26);
392                 clear_msr_range(0x3BC, 3);
393         } else {
394                 clear_msr_range(0x3A0, 31);
395         }
396         clear_msr_range(0x3C0, 6);
397         clear_msr_range(0x3C8, 6);
398         clear_msr_range(0x3E0, 2);
399         clear_msr_range(MSR_P4_CCCR0, 18);
400         clear_msr_range(MSR_P4_PERFCTR0, 18);
401
402         wrmsr(MSR_P4_CRU_ESCR0, P4_NMI_CRU_ESCR0, 0);
403         wrmsr(MSR_P4_IQ_CCCR0, P4_NMI_IQ_CCCR0 & ~P4_CCCR_ENABLE, 0);
404         Dprintk("setting P4_IQ_COUNTER0 to 0x%08lx\n", -(cpu_khz * 1000UL / nmi_hz));
405         wrmsrl(MSR_P4_IQ_COUNTER0, -((u64)cpu_khz * 1000 / nmi_hz));
406         apic_write(APIC_LVTPC, APIC_DM_NMI);
407         wrmsr(MSR_P4_IQ_CCCR0, nmi_p4_cccr_val, 0);
408         return 1;
409 }
410
411 void setup_apic_nmi_watchdog(void)
412 {
413         switch (boot_cpu_data.x86_vendor) {
414         case X86_VENDOR_AMD:
415                 if (boot_cpu_data.x86 != 15)
416                         return;
417                 if (strstr(boot_cpu_data.x86_model_id, "Screwdriver"))
418                         return;
419                 setup_k7_watchdog();
420                 break;
421         case X86_VENDOR_INTEL:
422                 if (boot_cpu_data.x86 != 15)
423                         return;
424                 if (!setup_p4_watchdog())
425                         return;
426                 break;
427
428         default:
429                 return;
430         }
431         lapic_nmi_owner = LAPIC_NMI_WATCHDOG;
432         nmi_active = 1;
433 }
434
435 /*
436  * the best way to detect whether a CPU has a 'hard lockup' problem
437  * is to check it's local APIC timer IRQ counts. If they are not
438  * changing then that CPU has some problem.
439  *
440  * as these watchdog NMI IRQs are generated on every CPU, we only
441  * have to check the current processor.
442  */
443
444 static DEFINE_PER_CPU(unsigned, last_irq_sum);
445 static DEFINE_PER_CPU(local_t, alert_counter);
446 static DEFINE_PER_CPU(int, nmi_touch);
447
448 void touch_nmi_watchdog (void)
449 {
450         if (nmi_watchdog > 0) {
451                 unsigned cpu;
452
453                 /*
454                  * Tell other CPUs to reset their alert counters. We cannot
455                  * do it ourselves because the alert count increase is not
456                  * atomic.
457                  */
458                 for_each_present_cpu (cpu)
459                         per_cpu(nmi_touch, cpu) = 1;
460         }
461
462         touch_softlockup_watchdog();
463 }
464
465 void __kprobes nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
466 {
467         int sum;
468         int touched = 0;
469
470         sum = read_pda(apic_timer_irqs);
471         if (__get_cpu_var(nmi_touch)) {
472                 __get_cpu_var(nmi_touch) = 0;
473                 touched = 1;
474         }
475 #ifdef CONFIG_X86_MCE
476         /* Could check oops_in_progress here too, but it's safer
477            not too */
478         if (atomic_read(&mce_entry) > 0)
479                 touched = 1;
480 #endif
481         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
482                 /*
483                  * Ayiee, looks like this CPU is stuck ...
484                  * wait a few IRQs (5 seconds) before doing the oops ...
485                  */
486                 local_inc(&__get_cpu_var(alert_counter));
487                 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz) {
488                         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
489                                                         == NOTIFY_STOP) {
490                                 local_set(&__get_cpu_var(alert_counter), 0);
491                                 return;
492                         }
493                         die_nmi("NMI Watchdog detected LOCKUP on CPU %d\n", regs);
494                 }
495         } else {
496                 __get_cpu_var(last_irq_sum) = sum;
497                 local_set(&__get_cpu_var(alert_counter), 0);
498         }
499         if (nmi_perfctr_msr) {
500                 if (nmi_perfctr_msr == MSR_P4_IQ_COUNTER0) {
501                         /*
502                          * P4 quirks:
503                          * - An overflown perfctr will assert its interrupt
504                          *   until the OVF flag in its CCCR is cleared.
505                          * - LVTPC is masked on interrupt and must be
506                          *   unmasked by the LVTPC handler.
507                          */
508                         wrmsr(MSR_P4_IQ_CCCR0, nmi_p4_cccr_val, 0);
509                         apic_write(APIC_LVTPC, APIC_DM_NMI);
510                 }
511                 wrmsrl(nmi_perfctr_msr, -((u64)cpu_khz * 1000 / nmi_hz));
512         }
513 }
514
515 static __kprobes int dummy_nmi_callback(struct pt_regs * regs, int cpu)
516 {
517         return 0;
518 }
519  
520 static nmi_callback_t nmi_callback = dummy_nmi_callback;
521  
522 asmlinkage __kprobes void do_nmi(struct pt_regs * regs, long error_code)
523 {
524         int cpu = safe_smp_processor_id();
525
526         nmi_enter();
527         add_pda(__nmi_count,1);
528         if (!rcu_dereference(nmi_callback)(regs, cpu))
529                 default_do_nmi(regs);
530         nmi_exit();
531 }
532
533 void set_nmi_callback(nmi_callback_t callback)
534 {
535         vmalloc_sync_all();
536         rcu_assign_pointer(nmi_callback, callback);
537 }
538 EXPORT_SYMBOL_GPL(set_nmi_callback);
539
540 void unset_nmi_callback(void)
541 {
542         nmi_callback = dummy_nmi_callback;
543 }
544 EXPORT_SYMBOL_GPL(unset_nmi_callback);
545
546 #ifdef CONFIG_SYSCTL
547
548 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
549 {
550         unsigned char reason = get_nmi_reason();
551         char buf[64];
552
553         if (!(reason & 0xc0)) {
554                 sprintf(buf, "NMI received for unknown reason %02x\n", reason);
555                 die_nmi(buf,regs);
556         }
557         return 0;
558 }
559
560 /*
561  * proc handler for /proc/sys/kernel/unknown_nmi_panic
562  */
563 int proc_unknown_nmi_panic(struct ctl_table *table, int write, struct file *file,
564                         void __user *buffer, size_t *length, loff_t *ppos)
565 {
566         int old_state;
567
568         old_state = unknown_nmi_panic;
569         proc_dointvec(table, write, file, buffer, length, ppos);
570         if (!!old_state == !!unknown_nmi_panic)
571                 return 0;
572
573         if (unknown_nmi_panic) {
574                 if (reserve_lapic_nmi() < 0) {
575                         unknown_nmi_panic = 0;
576                         return -EBUSY;
577                 } else {
578                         set_nmi_callback(unknown_nmi_panic_callback);
579                 }
580         } else {
581                 release_lapic_nmi();
582                 unset_nmi_callback();
583         }
584         return 0;
585 }
586
587 #endif
588
589 EXPORT_SYMBOL(nmi_active);
590 EXPORT_SYMBOL(nmi_watchdog);
591 EXPORT_SYMBOL(reserve_lapic_nmi);
592 EXPORT_SYMBOL(release_lapic_nmi);
593 EXPORT_SYMBOL(disable_timer_nmi_watchdog);
594 EXPORT_SYMBOL(enable_timer_nmi_watchdog);
595 EXPORT_SYMBOL(touch_nmi_watchdog);