xref: /openbmc/linux/arch/x86/kernel/cpu/perfctr-watchdog.c (revision 96de0e252cedffad61b3cb5e05662c591898e69a)
1 /* local apic based NMI watchdog for various CPUs.
2    This file also handles reservation of performance counters for coordination
3    with other users (like oprofile).
4 
5    Note that these events normally don't tick when the CPU idles. This means
6    the frequency varies with CPU load.
7 
8    Original code for K7/P6 written by Keith Owens */
9 
10 #include <linux/percpu.h>
11 #include <linux/module.h>
12 #include <linux/kernel.h>
13 #include <linux/bitops.h>
14 #include <linux/smp.h>
15 #include <linux/nmi.h>
16 #include <asm/apic.h>
17 #include <asm/intel_arch_perfmon.h>
18 
19 struct nmi_watchdog_ctlblk {
20 	unsigned int cccr_msr;
21 	unsigned int perfctr_msr;  /* the MSR to reset in NMI handler */
22 	unsigned int evntsel_msr;  /* the MSR to select the events to handle */
23 };
24 
25 /* Interface defining a CPU specific perfctr watchdog */
26 struct wd_ops {
27 	int (*reserve)(void);
28 	void (*unreserve)(void);
29 	int (*setup)(unsigned nmi_hz);
30 	void (*rearm)(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz);
31 	void (*stop)(void);
32 	unsigned perfctr;
33 	unsigned evntsel;
34 	u64 checkbit;
35 };
36 
37 static const struct wd_ops *wd_ops;
38 
39 /* this number is calculated from Intel's MSR_P4_CRU_ESCR5 register and it's
40  * offset from MSR_P4_BSU_ESCR0.  It will be the max for all platforms (for now)
41  */
42 #define NMI_MAX_COUNTER_BITS 66
43 
44 /* perfctr_nmi_owner tracks the ownership of the perfctr registers:
45  * evtsel_nmi_owner tracks the ownership of the event selection
46  * - different performance counters/ event selection may be reserved for
47  *   different subsystems this reservation system just tries to coordinate
48  *   things a little
49  */
50 static DECLARE_BITMAP(perfctr_nmi_owner, NMI_MAX_COUNTER_BITS);
51 static DECLARE_BITMAP(evntsel_nmi_owner, NMI_MAX_COUNTER_BITS);
52 
53 static DEFINE_PER_CPU(struct nmi_watchdog_ctlblk, nmi_watchdog_ctlblk);
54 
55 /* converts an msr to an appropriate reservation bit */
56 static inline unsigned int nmi_perfctr_msr_to_bit(unsigned int msr)
57 {
58 	/* returns the bit offset of the performance counter register */
59 	switch (boot_cpu_data.x86_vendor) {
60 	case X86_VENDOR_AMD:
61 		return (msr - MSR_K7_PERFCTR0);
62 	case X86_VENDOR_INTEL:
63 		if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
64 			return (msr - MSR_ARCH_PERFMON_PERFCTR0);
65 
66 		switch (boot_cpu_data.x86) {
67 		case 6:
68 			return (msr - MSR_P6_PERFCTR0);
69 		case 15:
70 			return (msr - MSR_P4_BPU_PERFCTR0);
71 		}
72 	}
73 	return 0;
74 }
75 
76 /* converts an msr to an appropriate reservation bit */
77 /* returns the bit offset of the event selection register */
78 static inline unsigned int nmi_evntsel_msr_to_bit(unsigned int msr)
79 {
80 	/* returns the bit offset of the event selection register */
81 	switch (boot_cpu_data.x86_vendor) {
82 	case X86_VENDOR_AMD:
83 		return (msr - MSR_K7_EVNTSEL0);
84 	case X86_VENDOR_INTEL:
85 		if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
86 			return (msr - MSR_ARCH_PERFMON_EVENTSEL0);
87 
88 		switch (boot_cpu_data.x86) {
89 		case 6:
90 			return (msr - MSR_P6_EVNTSEL0);
91 		case 15:
92 			return (msr - MSR_P4_BSU_ESCR0);
93 		}
94 	}
95 	return 0;
96 
97 }
98 
99 /* checks for a bit availability (hack for oprofile) */
100 int avail_to_resrv_perfctr_nmi_bit(unsigned int counter)
101 {
102 	BUG_ON(counter > NMI_MAX_COUNTER_BITS);
103 
104 	return (!test_bit(counter, perfctr_nmi_owner));
105 }
106 
107 /* checks the an msr for availability */
108 int avail_to_resrv_perfctr_nmi(unsigned int msr)
109 {
110 	unsigned int counter;
111 
112 	counter = nmi_perfctr_msr_to_bit(msr);
113 	BUG_ON(counter > NMI_MAX_COUNTER_BITS);
114 
115 	return (!test_bit(counter, perfctr_nmi_owner));
116 }
117 
118 int reserve_perfctr_nmi(unsigned int msr)
119 {
120 	unsigned int counter;
121 
122 	counter = nmi_perfctr_msr_to_bit(msr);
123 	BUG_ON(counter > NMI_MAX_COUNTER_BITS);
124 
125 	if (!test_and_set_bit(counter, perfctr_nmi_owner))
126 		return 1;
127 	return 0;
128 }
129 
130 void release_perfctr_nmi(unsigned int msr)
131 {
132 	unsigned int counter;
133 
134 	counter = nmi_perfctr_msr_to_bit(msr);
135 	BUG_ON(counter > NMI_MAX_COUNTER_BITS);
136 
137 	clear_bit(counter, perfctr_nmi_owner);
138 }
139 
140 int reserve_evntsel_nmi(unsigned int msr)
141 {
142 	unsigned int counter;
143 
144 	counter = nmi_evntsel_msr_to_bit(msr);
145 	BUG_ON(counter > NMI_MAX_COUNTER_BITS);
146 
147 	if (!test_and_set_bit(counter, evntsel_nmi_owner))
148 		return 1;
149 	return 0;
150 }
151 
152 void release_evntsel_nmi(unsigned int msr)
153 {
154 	unsigned int counter;
155 
156 	counter = nmi_evntsel_msr_to_bit(msr);
157 	BUG_ON(counter > NMI_MAX_COUNTER_BITS);
158 
159 	clear_bit(counter, evntsel_nmi_owner);
160 }
161 
162 EXPORT_SYMBOL(avail_to_resrv_perfctr_nmi);
163 EXPORT_SYMBOL(avail_to_resrv_perfctr_nmi_bit);
164 EXPORT_SYMBOL(reserve_perfctr_nmi);
165 EXPORT_SYMBOL(release_perfctr_nmi);
166 EXPORT_SYMBOL(reserve_evntsel_nmi);
167 EXPORT_SYMBOL(release_evntsel_nmi);
168 
169 void disable_lapic_nmi_watchdog(void)
170 {
171 	BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
172 
173 	if (atomic_read(&nmi_active) <= 0)
174 		return;
175 
176 	on_each_cpu(stop_apic_nmi_watchdog, NULL, 0, 1);
177 	wd_ops->unreserve();
178 
179 	BUG_ON(atomic_read(&nmi_active) != 0);
180 }
181 
182 void enable_lapic_nmi_watchdog(void)
183 {
184 	BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
185 
186 	/* are we already enabled */
187 	if (atomic_read(&nmi_active) != 0)
188 		return;
189 
190 	/* are we lapic aware */
191 	if (!wd_ops)
192 		return;
193 	if (!wd_ops->reserve()) {
194 		printk(KERN_ERR "NMI watchdog: cannot reserve perfctrs\n");
195 		return;
196 	}
197 
198 	on_each_cpu(setup_apic_nmi_watchdog, NULL, 0, 1);
199 	touch_nmi_watchdog();
200 }
201 
202 /*
203  * Activate the NMI watchdog via the local APIC.
204  */
205 
206 static unsigned int adjust_for_32bit_ctr(unsigned int hz)
207 {
208 	u64 counter_val;
209 	unsigned int retval = hz;
210 
211 	/*
212 	 * On Intel CPUs with P6/ARCH_PERFMON only 32 bits in the counter
213 	 * are writable, with higher bits sign extending from bit 31.
214 	 * So, we can only program the counter with 31 bit values and
215 	 * 32nd bit should be 1, for 33.. to be 1.
216 	 * Find the appropriate nmi_hz
217 	 */
218 	counter_val = (u64)cpu_khz * 1000;
219 	do_div(counter_val, retval);
220  	if (counter_val > 0x7fffffffULL) {
221 		u64 count = (u64)cpu_khz * 1000;
222 		do_div(count, 0x7fffffffUL);
223 		retval = count + 1;
224 	}
225 	return retval;
226 }
227 
228 static void
229 write_watchdog_counter(unsigned int perfctr_msr, const char *descr, unsigned nmi_hz)
230 {
231 	u64 count = (u64)cpu_khz * 1000;
232 
233 	do_div(count, nmi_hz);
234 	if(descr)
235 		Dprintk("setting %s to -0x%08Lx\n", descr, count);
236 	wrmsrl(perfctr_msr, 0 - count);
237 }
238 
239 static void write_watchdog_counter32(unsigned int perfctr_msr,
240 		const char *descr, unsigned nmi_hz)
241 {
242 	u64 count = (u64)cpu_khz * 1000;
243 
244 	do_div(count, nmi_hz);
245 	if(descr)
246 		Dprintk("setting %s to -0x%08Lx\n", descr, count);
247 	wrmsr(perfctr_msr, (u32)(-count), 0);
248 }
249 
250 /* AMD K7/K8/Family10h/Family11h support. AMD keeps this interface
251    nicely stable so there is not much variety */
252 
253 #define K7_EVNTSEL_ENABLE	(1 << 22)
254 #define K7_EVNTSEL_INT		(1 << 20)
255 #define K7_EVNTSEL_OS		(1 << 17)
256 #define K7_EVNTSEL_USR		(1 << 16)
257 #define K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING	0x76
258 #define K7_NMI_EVENT		K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING
259 
260 static int setup_k7_watchdog(unsigned nmi_hz)
261 {
262 	unsigned int perfctr_msr, evntsel_msr;
263 	unsigned int evntsel;
264 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
265 
266 	perfctr_msr = wd_ops->perfctr;
267 	evntsel_msr = wd_ops->evntsel;
268 
269 	wrmsrl(perfctr_msr, 0UL);
270 
271 	evntsel = K7_EVNTSEL_INT
272 		| K7_EVNTSEL_OS
273 		| K7_EVNTSEL_USR
274 		| K7_NMI_EVENT;
275 
276 	/* setup the timer */
277 	wrmsr(evntsel_msr, evntsel, 0);
278 	write_watchdog_counter(perfctr_msr, "K7_PERFCTR0",nmi_hz);
279 	apic_write(APIC_LVTPC, APIC_DM_NMI);
280 	evntsel |= K7_EVNTSEL_ENABLE;
281 	wrmsr(evntsel_msr, evntsel, 0);
282 
283 	wd->perfctr_msr = perfctr_msr;
284 	wd->evntsel_msr = evntsel_msr;
285 	wd->cccr_msr = 0;  //unused
286 	return 1;
287 }
288 
289 static void single_msr_stop_watchdog(void)
290 {
291 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
292 
293 	wrmsr(wd->evntsel_msr, 0, 0);
294 }
295 
296 static int single_msr_reserve(void)
297 {
298 	if (!reserve_perfctr_nmi(wd_ops->perfctr))
299 		return 0;
300 
301 	if (!reserve_evntsel_nmi(wd_ops->evntsel)) {
302 		release_perfctr_nmi(wd_ops->perfctr);
303 		return 0;
304 	}
305 	return 1;
306 }
307 
308 static void single_msr_unreserve(void)
309 {
310 	release_evntsel_nmi(wd_ops->evntsel);
311 	release_perfctr_nmi(wd_ops->perfctr);
312 }
313 
314 static void single_msr_rearm(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz)
315 {
316 	/* start the cycle over again */
317 	write_watchdog_counter(wd->perfctr_msr, NULL, nmi_hz);
318 }
319 
320 static const struct wd_ops k7_wd_ops = {
321 	.reserve = single_msr_reserve,
322 	.unreserve = single_msr_unreserve,
323 	.setup = setup_k7_watchdog,
324 	.rearm = single_msr_rearm,
325 	.stop = single_msr_stop_watchdog,
326 	.perfctr = MSR_K7_PERFCTR0,
327 	.evntsel = MSR_K7_EVNTSEL0,
328 	.checkbit = 1ULL<<47,
329 };
330 
331 /* Intel Model 6 (PPro+,P2,P3,P-M,Core1) */
332 
333 #define P6_EVNTSEL0_ENABLE	(1 << 22)
334 #define P6_EVNTSEL_INT		(1 << 20)
335 #define P6_EVNTSEL_OS		(1 << 17)
336 #define P6_EVNTSEL_USR		(1 << 16)
337 #define P6_EVENT_CPU_CLOCKS_NOT_HALTED	0x79
338 #define P6_NMI_EVENT		P6_EVENT_CPU_CLOCKS_NOT_HALTED
339 
340 static int setup_p6_watchdog(unsigned nmi_hz)
341 {
342 	unsigned int perfctr_msr, evntsel_msr;
343 	unsigned int evntsel;
344 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
345 
346 	perfctr_msr = wd_ops->perfctr;
347 	evntsel_msr = wd_ops->evntsel;
348 
349 	/* KVM doesn't implement this MSR */
350 	if (wrmsr_safe(perfctr_msr, 0, 0) < 0)
351 		return 0;
352 
353 	evntsel = P6_EVNTSEL_INT
354 		| P6_EVNTSEL_OS
355 		| P6_EVNTSEL_USR
356 		| P6_NMI_EVENT;
357 
358 	/* setup the timer */
359 	wrmsr(evntsel_msr, evntsel, 0);
360 	nmi_hz = adjust_for_32bit_ctr(nmi_hz);
361 	write_watchdog_counter32(perfctr_msr, "P6_PERFCTR0",nmi_hz);
362 	apic_write(APIC_LVTPC, APIC_DM_NMI);
363 	evntsel |= P6_EVNTSEL0_ENABLE;
364 	wrmsr(evntsel_msr, evntsel, 0);
365 
366 	wd->perfctr_msr = perfctr_msr;
367 	wd->evntsel_msr = evntsel_msr;
368 	wd->cccr_msr = 0;  //unused
369 	return 1;
370 }
371 
372 static void p6_rearm(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz)
373 {
374 	/* P6 based Pentium M need to re-unmask
375 	 * the apic vector but it doesn't hurt
376 	 * other P6 variant.
377 	 * ArchPerfom/Core Duo also needs this */
378 	apic_write(APIC_LVTPC, APIC_DM_NMI);
379 	/* P6/ARCH_PERFMON has 32 bit counter write */
380 	write_watchdog_counter32(wd->perfctr_msr, NULL,nmi_hz);
381 }
382 
383 static const struct wd_ops p6_wd_ops = {
384 	.reserve = single_msr_reserve,
385 	.unreserve = single_msr_unreserve,
386 	.setup = setup_p6_watchdog,
387 	.rearm = p6_rearm,
388 	.stop = single_msr_stop_watchdog,
389 	.perfctr = MSR_P6_PERFCTR0,
390 	.evntsel = MSR_P6_EVNTSEL0,
391 	.checkbit = 1ULL<<39,
392 };
393 
394 /* Intel P4 performance counters. By far the most complicated of all. */
395 
396 #define MSR_P4_MISC_ENABLE_PERF_AVAIL	(1<<7)
397 #define P4_ESCR_EVENT_SELECT(N)	((N)<<25)
398 #define P4_ESCR_OS		(1<<3)
399 #define P4_ESCR_USR		(1<<2)
400 #define P4_CCCR_OVF_PMI0	(1<<26)
401 #define P4_CCCR_OVF_PMI1	(1<<27)
402 #define P4_CCCR_THRESHOLD(N)	((N)<<20)
403 #define P4_CCCR_COMPLEMENT	(1<<19)
404 #define P4_CCCR_COMPARE		(1<<18)
405 #define P4_CCCR_REQUIRED	(3<<16)
406 #define P4_CCCR_ESCR_SELECT(N)	((N)<<13)
407 #define P4_CCCR_ENABLE		(1<<12)
408 #define P4_CCCR_OVF 		(1<<31)
409 
410 /* Set up IQ_COUNTER0 to behave like a clock, by having IQ_CCCR0 filter
411    CRU_ESCR0 (with any non-null event selector) through a complemented
412    max threshold. [IA32-Vol3, Section 14.9.9] */
413 
414 static int setup_p4_watchdog(unsigned nmi_hz)
415 {
416 	unsigned int perfctr_msr, evntsel_msr, cccr_msr;
417 	unsigned int evntsel, cccr_val;
418 	unsigned int misc_enable, dummy;
419 	unsigned int ht_num;
420 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
421 
422 	rdmsr(MSR_IA32_MISC_ENABLE, misc_enable, dummy);
423 	if (!(misc_enable & MSR_P4_MISC_ENABLE_PERF_AVAIL))
424 		return 0;
425 
426 #ifdef CONFIG_SMP
427 	/* detect which hyperthread we are on */
428 	if (smp_num_siblings == 2) {
429 		unsigned int ebx, apicid;
430 
431         	ebx = cpuid_ebx(1);
432 	        apicid = (ebx >> 24) & 0xff;
433         	ht_num = apicid & 1;
434 	} else
435 #endif
436 		ht_num = 0;
437 
438 	/* performance counters are shared resources
439 	 * assign each hyperthread its own set
440 	 * (re-use the ESCR0 register, seems safe
441 	 * and keeps the cccr_val the same)
442 	 */
443 	if (!ht_num) {
444 		/* logical cpu 0 */
445 		perfctr_msr = MSR_P4_IQ_PERFCTR0;
446 		evntsel_msr = MSR_P4_CRU_ESCR0;
447 		cccr_msr = MSR_P4_IQ_CCCR0;
448 		cccr_val = P4_CCCR_OVF_PMI0 | P4_CCCR_ESCR_SELECT(4);
449 	} else {
450 		/* logical cpu 1 */
451 		perfctr_msr = MSR_P4_IQ_PERFCTR1;
452 		evntsel_msr = MSR_P4_CRU_ESCR0;
453 		cccr_msr = MSR_P4_IQ_CCCR1;
454 		cccr_val = P4_CCCR_OVF_PMI1 | P4_CCCR_ESCR_SELECT(4);
455 	}
456 
457 	evntsel = P4_ESCR_EVENT_SELECT(0x3F)
458 	 	| P4_ESCR_OS
459 		| P4_ESCR_USR;
460 
461 	cccr_val |= P4_CCCR_THRESHOLD(15)
462 		 | P4_CCCR_COMPLEMENT
463 		 | P4_CCCR_COMPARE
464 		 | P4_CCCR_REQUIRED;
465 
466 	wrmsr(evntsel_msr, evntsel, 0);
467 	wrmsr(cccr_msr, cccr_val, 0);
468 	write_watchdog_counter(perfctr_msr, "P4_IQ_COUNTER0", nmi_hz);
469 	apic_write(APIC_LVTPC, APIC_DM_NMI);
470 	cccr_val |= P4_CCCR_ENABLE;
471 	wrmsr(cccr_msr, cccr_val, 0);
472 	wd->perfctr_msr = perfctr_msr;
473 	wd->evntsel_msr = evntsel_msr;
474 	wd->cccr_msr = cccr_msr;
475 	return 1;
476 }
477 
478 static void stop_p4_watchdog(void)
479 {
480 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
481 	wrmsr(wd->cccr_msr, 0, 0);
482 	wrmsr(wd->evntsel_msr, 0, 0);
483 }
484 
485 static int p4_reserve(void)
486 {
487 	if (!reserve_perfctr_nmi(MSR_P4_IQ_PERFCTR0))
488 		return 0;
489 #ifdef CONFIG_SMP
490 	if (smp_num_siblings > 1 && !reserve_perfctr_nmi(MSR_P4_IQ_PERFCTR1))
491 		goto fail1;
492 #endif
493 	if (!reserve_evntsel_nmi(MSR_P4_CRU_ESCR0))
494 		goto fail2;
495 	/* RED-PEN why is ESCR1 not reserved here? */
496 	return 1;
497  fail2:
498 #ifdef CONFIG_SMP
499 	if (smp_num_siblings > 1)
500 		release_perfctr_nmi(MSR_P4_IQ_PERFCTR1);
501  fail1:
502 #endif
503 	release_perfctr_nmi(MSR_P4_IQ_PERFCTR0);
504 	return 0;
505 }
506 
507 static void p4_unreserve(void)
508 {
509 #ifdef CONFIG_SMP
510 	if (smp_num_siblings > 1)
511 		release_perfctr_nmi(MSR_P4_IQ_PERFCTR1);
512 #endif
513 	release_evntsel_nmi(MSR_P4_CRU_ESCR0);
514 	release_perfctr_nmi(MSR_P4_IQ_PERFCTR0);
515 }
516 
517 static void p4_rearm(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz)
518 {
519 	unsigned dummy;
520 	/*
521  	 * P4 quirks:
522 	 * - An overflown perfctr will assert its interrupt
523 	 *   until the OVF flag in its CCCR is cleared.
524 	 * - LVTPC is masked on interrupt and must be
525 	 *   unmasked by the LVTPC handler.
526 	 */
527 	rdmsrl(wd->cccr_msr, dummy);
528 	dummy &= ~P4_CCCR_OVF;
529 	wrmsrl(wd->cccr_msr, dummy);
530 	apic_write(APIC_LVTPC, APIC_DM_NMI);
531 	/* start the cycle over again */
532 	write_watchdog_counter(wd->perfctr_msr, NULL, nmi_hz);
533 }
534 
535 static const struct wd_ops p4_wd_ops = {
536 	.reserve = p4_reserve,
537 	.unreserve = p4_unreserve,
538 	.setup = setup_p4_watchdog,
539 	.rearm = p4_rearm,
540 	.stop = stop_p4_watchdog,
541 	/* RED-PEN this is wrong for the other sibling */
542 	.perfctr = MSR_P4_BPU_PERFCTR0,
543 	.evntsel = MSR_P4_BSU_ESCR0,
544 	.checkbit = 1ULL<<39,
545 };
546 
547 /* Watchdog using the Intel architected PerfMon. Used for Core2 and hopefully
548    all future Intel CPUs. */
549 
550 #define ARCH_PERFMON_NMI_EVENT_SEL	ARCH_PERFMON_UNHALTED_CORE_CYCLES_SEL
551 #define ARCH_PERFMON_NMI_EVENT_UMASK	ARCH_PERFMON_UNHALTED_CORE_CYCLES_UMASK
552 
553 static struct wd_ops intel_arch_wd_ops;
554 
555 static int setup_intel_arch_watchdog(unsigned nmi_hz)
556 {
557 	unsigned int ebx;
558 	union cpuid10_eax eax;
559 	unsigned int unused;
560 	unsigned int perfctr_msr, evntsel_msr;
561 	unsigned int evntsel;
562 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
563 
564 	/*
565 	 * Check whether the Architectural PerfMon supports
566 	 * Unhalted Core Cycles Event or not.
567 	 * NOTE: Corresponding bit = 0 in ebx indicates event present.
568 	 */
569 	cpuid(10, &(eax.full), &ebx, &unused, &unused);
570 	if ((eax.split.mask_length < (ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX+1)) ||
571 	    (ebx & ARCH_PERFMON_UNHALTED_CORE_CYCLES_PRESENT))
572 		return 0;
573 
574 	perfctr_msr = wd_ops->perfctr;
575 	evntsel_msr = wd_ops->evntsel;
576 
577 	wrmsrl(perfctr_msr, 0UL);
578 
579 	evntsel = ARCH_PERFMON_EVENTSEL_INT
580 		| ARCH_PERFMON_EVENTSEL_OS
581 		| ARCH_PERFMON_EVENTSEL_USR
582 		| ARCH_PERFMON_NMI_EVENT_SEL
583 		| ARCH_PERFMON_NMI_EVENT_UMASK;
584 
585 	/* setup the timer */
586 	wrmsr(evntsel_msr, evntsel, 0);
587 	nmi_hz = adjust_for_32bit_ctr(nmi_hz);
588 	write_watchdog_counter32(perfctr_msr, "INTEL_ARCH_PERFCTR0", nmi_hz);
589 	apic_write(APIC_LVTPC, APIC_DM_NMI);
590 	evntsel |= ARCH_PERFMON_EVENTSEL0_ENABLE;
591 	wrmsr(evntsel_msr, evntsel, 0);
592 
593 	wd->perfctr_msr = perfctr_msr;
594 	wd->evntsel_msr = evntsel_msr;
595 	wd->cccr_msr = 0;  //unused
596 	intel_arch_wd_ops.checkbit = 1ULL << (eax.split.bit_width - 1);
597 	return 1;
598 }
599 
600 static struct wd_ops intel_arch_wd_ops __read_mostly = {
601 	.reserve = single_msr_reserve,
602 	.unreserve = single_msr_unreserve,
603 	.setup = setup_intel_arch_watchdog,
604 	.rearm = p6_rearm,
605 	.stop = single_msr_stop_watchdog,
606 	.perfctr = MSR_ARCH_PERFMON_PERFCTR1,
607 	.evntsel = MSR_ARCH_PERFMON_EVENTSEL1,
608 };
609 
610 static struct wd_ops coreduo_wd_ops = {
611 	.reserve = single_msr_reserve,
612 	.unreserve = single_msr_unreserve,
613 	.setup = setup_intel_arch_watchdog,
614 	.rearm = p6_rearm,
615 	.stop = single_msr_stop_watchdog,
616 	.perfctr = MSR_ARCH_PERFMON_PERFCTR0,
617 	.evntsel = MSR_ARCH_PERFMON_EVENTSEL0,
618 };
619 
620 static void probe_nmi_watchdog(void)
621 {
622 	switch (boot_cpu_data.x86_vendor) {
623 	case X86_VENDOR_AMD:
624 		if (boot_cpu_data.x86 != 6 && boot_cpu_data.x86 != 15 &&
625 		    boot_cpu_data.x86 != 16)
626 			return;
627 		wd_ops = &k7_wd_ops;
628 		break;
629 	case X86_VENDOR_INTEL:
630 		/* Work around Core Duo (Yonah) errata AE49 where perfctr1
631 		   doesn't have a working enable bit. */
632 		if (boot_cpu_data.x86 == 6 && boot_cpu_data.x86_model == 14) {
633 			wd_ops = &coreduo_wd_ops;
634 			break;
635 		}
636 		if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
637 			wd_ops = &intel_arch_wd_ops;
638 			break;
639 		}
640 		switch (boot_cpu_data.x86) {
641 		case 6:
642 			if (boot_cpu_data.x86_model > 0xd)
643 				return;
644 
645 			wd_ops = &p6_wd_ops;
646 			break;
647 		case 15:
648 			if (boot_cpu_data.x86_model > 0x4)
649 				return;
650 
651 			wd_ops = &p4_wd_ops;
652 			break;
653 		default:
654 			return;
655 		}
656 		break;
657 	}
658 }
659 
660 /* Interface to nmi.c */
661 
662 int lapic_watchdog_init(unsigned nmi_hz)
663 {
664 	if (!wd_ops) {
665 		probe_nmi_watchdog();
666 		if (!wd_ops)
667 			return -1;
668 
669 		if (!wd_ops->reserve()) {
670 			printk(KERN_ERR
671 				"NMI watchdog: cannot reserve perfctrs\n");
672 			return -1;
673 		}
674 	}
675 
676 	if (!(wd_ops->setup(nmi_hz))) {
677 		printk(KERN_ERR "Cannot setup NMI watchdog on CPU %d\n",
678 		       raw_smp_processor_id());
679 		return -1;
680 	}
681 
682 	return 0;
683 }
684 
685 void lapic_watchdog_stop(void)
686 {
687 	if (wd_ops)
688 		wd_ops->stop();
689 }
690 
691 unsigned lapic_adjust_nmi_hz(unsigned hz)
692 {
693 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
694 	if (wd->perfctr_msr == MSR_P6_PERFCTR0 ||
695 	    wd->perfctr_msr == MSR_ARCH_PERFMON_PERFCTR1)
696 		hz = adjust_for_32bit_ctr(hz);
697 	return hz;
698 }
699 
700 int lapic_wd_event(unsigned nmi_hz)
701 {
702 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
703 	u64 ctr;
704 	rdmsrl(wd->perfctr_msr, ctr);
705 	if (ctr & wd_ops->checkbit) { /* perfctr still running? */
706 		return 0;
707 	}
708 	wd_ops->rearm(wd, nmi_hz);
709 	return 1;
710 }
711 
712 int lapic_watchdog_ok(void)
713 {
714 	return wd_ops != NULL;
715 }
716