1 /*
2  * local apic based NMI watchdog for various CPUs.
3  *
4  * This file also handles reservation of performance counters for coordination
5  * with other users (like oprofile).
6  *
7  * Note that these events normally don't tick when the CPU idles. This means
8  * the frequency varies with CPU load.
9  *
10  * Original code for K7/P6 written by Keith Owens
11  *
12  */
13 
14 #include <linux/percpu.h>
15 #include <linux/module.h>
16 #include <linux/kernel.h>
17 #include <linux/bitops.h>
18 #include <linux/smp.h>
19 #include <linux/nmi.h>
20 #include <asm/apic.h>
21 #include <asm/intel_arch_perfmon.h>
22 
23 struct nmi_watchdog_ctlblk {
24 	unsigned int cccr_msr;
25 	unsigned int perfctr_msr;  /* the MSR to reset in NMI handler */
26 	unsigned int evntsel_msr;  /* the MSR to select the events to handle */
27 };
28 
29 /* Interface defining a CPU specific perfctr watchdog */
30 struct wd_ops {
31 	int (*reserve)(void);
32 	void (*unreserve)(void);
33 	int (*setup)(unsigned nmi_hz);
34 	void (*rearm)(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz);
35 	void (*stop)(void);
36 	unsigned perfctr;
37 	unsigned evntsel;
38 	u64 checkbit;
39 };
40 
41 static const struct wd_ops *wd_ops;
42 
43 /*
44  * this number is calculated from Intel's MSR_P4_CRU_ESCR5 register and it's
45  * offset from MSR_P4_BSU_ESCR0.
46  *
47  * It will be the max for all platforms (for now)
48  */
49 #define NMI_MAX_COUNTER_BITS 66
50 
51 /*
52  * perfctr_nmi_owner tracks the ownership of the perfctr registers:
53  * evtsel_nmi_owner tracks the ownership of the event selection
54  * - different performance counters/ event selection may be reserved for
55  *   different subsystems this reservation system just tries to coordinate
56  *   things a little
57  */
58 static DECLARE_BITMAP(perfctr_nmi_owner, NMI_MAX_COUNTER_BITS);
59 static DECLARE_BITMAP(evntsel_nmi_owner, NMI_MAX_COUNTER_BITS);
60 
61 static DEFINE_PER_CPU(struct nmi_watchdog_ctlblk, nmi_watchdog_ctlblk);
62 
63 /* converts an msr to an appropriate reservation bit */
64 static inline unsigned int nmi_perfctr_msr_to_bit(unsigned int msr)
65 {
66 	/* returns the bit offset of the performance counter register */
67 	switch (boot_cpu_data.x86_vendor) {
68 	case X86_VENDOR_AMD:
69 		return (msr - MSR_K7_PERFCTR0);
70 	case X86_VENDOR_INTEL:
71 		if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
72 			return (msr - MSR_ARCH_PERFMON_PERFCTR0);
73 
74 		switch (boot_cpu_data.x86) {
75 		case 6:
76 			return (msr - MSR_P6_PERFCTR0);
77 		case 15:
78 			return (msr - MSR_P4_BPU_PERFCTR0);
79 		}
80 	}
81 	return 0;
82 }
83 
84 /*
85  * converts an msr to an appropriate reservation bit
86  * returns the bit offset of the event selection register
87  */
88 static inline unsigned int nmi_evntsel_msr_to_bit(unsigned int msr)
89 {
90 	/* returns the bit offset of the event selection register */
91 	switch (boot_cpu_data.x86_vendor) {
92 	case X86_VENDOR_AMD:
93 		return (msr - MSR_K7_EVNTSEL0);
94 	case X86_VENDOR_INTEL:
95 		if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
96 			return (msr - MSR_ARCH_PERFMON_EVENTSEL0);
97 
98 		switch (boot_cpu_data.x86) {
99 		case 6:
100 			return (msr - MSR_P6_EVNTSEL0);
101 		case 15:
102 			return (msr - MSR_P4_BSU_ESCR0);
103 		}
104 	}
105 	return 0;
106 
107 }
108 
109 /* checks for a bit availability (hack for oprofile) */
110 int avail_to_resrv_perfctr_nmi_bit(unsigned int counter)
111 {
112 	BUG_ON(counter > NMI_MAX_COUNTER_BITS);
113 
114 	return (!test_bit(counter, perfctr_nmi_owner));
115 }
116 
117 /* checks the an msr for availability */
118 int avail_to_resrv_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 	return (!test_bit(counter, perfctr_nmi_owner));
126 }
127 EXPORT_SYMBOL(avail_to_resrv_perfctr_nmi_bit);
128 
129 int reserve_perfctr_nmi(unsigned int msr)
130 {
131 	unsigned int counter;
132 
133 	counter = nmi_perfctr_msr_to_bit(msr);
134 	/* register not managed by the allocator? */
135 	if (counter > NMI_MAX_COUNTER_BITS)
136 		return 1;
137 
138 	if (!test_and_set_bit(counter, perfctr_nmi_owner))
139 		return 1;
140 	return 0;
141 }
142 EXPORT_SYMBOL(reserve_perfctr_nmi);
143 
144 void release_perfctr_nmi(unsigned int msr)
145 {
146 	unsigned int counter;
147 
148 	counter = nmi_perfctr_msr_to_bit(msr);
149 	/* register not managed by the allocator? */
150 	if (counter > NMI_MAX_COUNTER_BITS)
151 		return;
152 
153 	clear_bit(counter, perfctr_nmi_owner);
154 }
155 EXPORT_SYMBOL(release_perfctr_nmi);
156 
157 int reserve_evntsel_nmi(unsigned int msr)
158 {
159 	unsigned int counter;
160 
161 	counter = nmi_evntsel_msr_to_bit(msr);
162 	/* register not managed by the allocator? */
163 	if (counter > NMI_MAX_COUNTER_BITS)
164 		return 1;
165 
166 	if (!test_and_set_bit(counter, evntsel_nmi_owner))
167 		return 1;
168 	return 0;
169 }
170 EXPORT_SYMBOL(reserve_evntsel_nmi);
171 
172 void release_evntsel_nmi(unsigned int msr)
173 {
174 	unsigned int counter;
175 
176 	counter = nmi_evntsel_msr_to_bit(msr);
177 	/* register not managed by the allocator? */
178 	if (counter > NMI_MAX_COUNTER_BITS)
179 		return;
180 
181 	clear_bit(counter, evntsel_nmi_owner);
182 }
183 EXPORT_SYMBOL(release_evntsel_nmi);
184 
185 void disable_lapic_nmi_watchdog(void)
186 {
187 	BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
188 
189 	if (atomic_read(&nmi_active) <= 0)
190 		return;
191 
192 	on_each_cpu(stop_apic_nmi_watchdog, NULL, 1);
193 
194 	if (wd_ops)
195 		wd_ops->unreserve();
196 
197 	BUG_ON(atomic_read(&nmi_active) != 0);
198 }
199 
200 void enable_lapic_nmi_watchdog(void)
201 {
202 	BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
203 
204 	/* are we already enabled */
205 	if (atomic_read(&nmi_active) != 0)
206 		return;
207 
208 	/* are we lapic aware */
209 	if (!wd_ops)
210 		return;
211 	if (!wd_ops->reserve()) {
212 		printk(KERN_ERR "NMI watchdog: cannot reserve perfctrs\n");
213 		return;
214 	}
215 
216 	on_each_cpu(setup_apic_nmi_watchdog, NULL, 1);
217 	touch_nmi_watchdog();
218 }
219 
220 /*
221  * Activate the NMI watchdog via the local APIC.
222  */
223 
224 static unsigned int adjust_for_32bit_ctr(unsigned int hz)
225 {
226 	u64 counter_val;
227 	unsigned int retval = hz;
228 
229 	/*
230 	 * On Intel CPUs with P6/ARCH_PERFMON only 32 bits in the counter
231 	 * are writable, with higher bits sign extending from bit 31.
232 	 * So, we can only program the counter with 31 bit values and
233 	 * 32nd bit should be 1, for 33.. to be 1.
234 	 * Find the appropriate nmi_hz
235 	 */
236 	counter_val = (u64)cpu_khz * 1000;
237 	do_div(counter_val, retval);
238  	if (counter_val > 0x7fffffffULL) {
239 		u64 count = (u64)cpu_khz * 1000;
240 		do_div(count, 0x7fffffffUL);
241 		retval = count + 1;
242 	}
243 	return retval;
244 }
245 
246 static void write_watchdog_counter(unsigned int perfctr_msr,
247 				const char *descr, unsigned nmi_hz)
248 {
249 	u64 count = (u64)cpu_khz * 1000;
250 
251 	do_div(count, nmi_hz);
252 	if(descr)
253 		pr_debug("setting %s to -0x%08Lx\n", descr, count);
254 	wrmsrl(perfctr_msr, 0 - count);
255 }
256 
257 static void write_watchdog_counter32(unsigned int perfctr_msr,
258 				const char *descr, unsigned nmi_hz)
259 {
260 	u64 count = (u64)cpu_khz * 1000;
261 
262 	do_div(count, nmi_hz);
263 	if(descr)
264 		pr_debug("setting %s to -0x%08Lx\n", descr, count);
265 	wrmsr(perfctr_msr, (u32)(-count), 0);
266 }
267 
268 /*
269  * AMD K7/K8/Family10h/Family11h support.
270  * AMD keeps this interface nicely stable so there is not much variety
271  */
272 #define K7_EVNTSEL_ENABLE	(1 << 22)
273 #define K7_EVNTSEL_INT		(1 << 20)
274 #define K7_EVNTSEL_OS		(1 << 17)
275 #define K7_EVNTSEL_USR		(1 << 16)
276 #define K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING	0x76
277 #define K7_NMI_EVENT		K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING
278 
279 static int setup_k7_watchdog(unsigned nmi_hz)
280 {
281 	unsigned int perfctr_msr, evntsel_msr;
282 	unsigned int evntsel;
283 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
284 
285 	perfctr_msr = wd_ops->perfctr;
286 	evntsel_msr = wd_ops->evntsel;
287 
288 	wrmsrl(perfctr_msr, 0UL);
289 
290 	evntsel = K7_EVNTSEL_INT
291 		| K7_EVNTSEL_OS
292 		| K7_EVNTSEL_USR
293 		| K7_NMI_EVENT;
294 
295 	/* setup the timer */
296 	wrmsr(evntsel_msr, evntsel, 0);
297 	write_watchdog_counter(perfctr_msr, "K7_PERFCTR0",nmi_hz);
298 
299 	/* initialize the wd struct before enabling */
300 	wd->perfctr_msr = perfctr_msr;
301 	wd->evntsel_msr = evntsel_msr;
302 	wd->cccr_msr = 0;  /* unused */
303 
304 	/* ok, everything is initialized, announce that we're set */
305 	cpu_nmi_set_wd_enabled();
306 
307 	apic_write(APIC_LVTPC, APIC_DM_NMI);
308 	evntsel |= K7_EVNTSEL_ENABLE;
309 	wrmsr(evntsel_msr, evntsel, 0);
310 
311 	return 1;
312 }
313 
314 static void single_msr_stop_watchdog(void)
315 {
316 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
317 
318 	wrmsr(wd->evntsel_msr, 0, 0);
319 }
320 
321 static int single_msr_reserve(void)
322 {
323 	if (!reserve_perfctr_nmi(wd_ops->perfctr))
324 		return 0;
325 
326 	if (!reserve_evntsel_nmi(wd_ops->evntsel)) {
327 		release_perfctr_nmi(wd_ops->perfctr);
328 		return 0;
329 	}
330 	return 1;
331 }
332 
333 static void single_msr_unreserve(void)
334 {
335 	release_evntsel_nmi(wd_ops->evntsel);
336 	release_perfctr_nmi(wd_ops->perfctr);
337 }
338 
339 static void single_msr_rearm(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz)
340 {
341 	/* start the cycle over again */
342 	write_watchdog_counter(wd->perfctr_msr, NULL, nmi_hz);
343 }
344 
345 static const struct wd_ops k7_wd_ops = {
346 	.reserve	= single_msr_reserve,
347 	.unreserve	= single_msr_unreserve,
348 	.setup		= setup_k7_watchdog,
349 	.rearm		= single_msr_rearm,
350 	.stop		= single_msr_stop_watchdog,
351 	.perfctr	= MSR_K7_PERFCTR0,
352 	.evntsel	= MSR_K7_EVNTSEL0,
353 	.checkbit	= 1ULL << 47,
354 };
355 
356 /*
357  * Intel Model 6 (PPro+,P2,P3,P-M,Core1)
358  */
359 #define P6_EVNTSEL0_ENABLE	(1 << 22)
360 #define P6_EVNTSEL_INT		(1 << 20)
361 #define P6_EVNTSEL_OS		(1 << 17)
362 #define P6_EVNTSEL_USR		(1 << 16)
363 #define P6_EVENT_CPU_CLOCKS_NOT_HALTED	0x79
364 #define P6_NMI_EVENT		P6_EVENT_CPU_CLOCKS_NOT_HALTED
365 
366 static int setup_p6_watchdog(unsigned nmi_hz)
367 {
368 	unsigned int perfctr_msr, evntsel_msr;
369 	unsigned int evntsel;
370 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
371 
372 	perfctr_msr = wd_ops->perfctr;
373 	evntsel_msr = wd_ops->evntsel;
374 
375 	/* KVM doesn't implement this MSR */
376 	if (wrmsr_safe(perfctr_msr, 0, 0) < 0)
377 		return 0;
378 
379 	evntsel = P6_EVNTSEL_INT
380 		| P6_EVNTSEL_OS
381 		| P6_EVNTSEL_USR
382 		| P6_NMI_EVENT;
383 
384 	/* setup the timer */
385 	wrmsr(evntsel_msr, evntsel, 0);
386 	nmi_hz = adjust_for_32bit_ctr(nmi_hz);
387 	write_watchdog_counter32(perfctr_msr, "P6_PERFCTR0",nmi_hz);
388 
389 	/* initialize the wd struct before enabling */
390 	wd->perfctr_msr = perfctr_msr;
391 	wd->evntsel_msr = evntsel_msr;
392 	wd->cccr_msr = 0;  /* unused */
393 
394 	/* ok, everything is initialized, announce that we're set */
395 	cpu_nmi_set_wd_enabled();
396 
397 	apic_write(APIC_LVTPC, APIC_DM_NMI);
398 	evntsel |= P6_EVNTSEL0_ENABLE;
399 	wrmsr(evntsel_msr, evntsel, 0);
400 
401 	return 1;
402 }
403 
404 static void p6_rearm(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz)
405 {
406 	/*
407 	 * P6 based Pentium M need to re-unmask
408 	 * the apic vector but it doesn't hurt
409 	 * other P6 variant.
410 	 * ArchPerfom/Core Duo also needs this
411 	 */
412 	apic_write(APIC_LVTPC, APIC_DM_NMI);
413 
414 	/* P6/ARCH_PERFMON has 32 bit counter write */
415 	write_watchdog_counter32(wd->perfctr_msr, NULL,nmi_hz);
416 }
417 
418 static const struct wd_ops p6_wd_ops = {
419 	.reserve	= single_msr_reserve,
420 	.unreserve	= single_msr_unreserve,
421 	.setup		= setup_p6_watchdog,
422 	.rearm		= p6_rearm,
423 	.stop		= single_msr_stop_watchdog,
424 	.perfctr	= MSR_P6_PERFCTR0,
425 	.evntsel	= MSR_P6_EVNTSEL0,
426 	.checkbit	= 1ULL << 39,
427 };
428 
429 /*
430  * Intel P4 performance counters.
431  * By far the most complicated of all.
432  */
433 #define MSR_P4_MISC_ENABLE_PERF_AVAIL	(1 << 7)
434 #define P4_ESCR_EVENT_SELECT(N)	((N) << 25)
435 #define P4_ESCR_OS		(1 << 3)
436 #define P4_ESCR_USR		(1 << 2)
437 #define P4_CCCR_OVF_PMI0	(1 << 26)
438 #define P4_CCCR_OVF_PMI1	(1 << 27)
439 #define P4_CCCR_THRESHOLD(N)	((N) << 20)
440 #define P4_CCCR_COMPLEMENT	(1 << 19)
441 #define P4_CCCR_COMPARE		(1 << 18)
442 #define P4_CCCR_REQUIRED	(3 << 16)
443 #define P4_CCCR_ESCR_SELECT(N)	((N) << 13)
444 #define P4_CCCR_ENABLE		(1 << 12)
445 #define P4_CCCR_OVF 		(1 << 31)
446 
447 #define P4_CONTROLS 18
448 static unsigned int p4_controls[18] = {
449 	MSR_P4_BPU_CCCR0,
450 	MSR_P4_BPU_CCCR1,
451 	MSR_P4_BPU_CCCR2,
452 	MSR_P4_BPU_CCCR3,
453 	MSR_P4_MS_CCCR0,
454 	MSR_P4_MS_CCCR1,
455 	MSR_P4_MS_CCCR2,
456 	MSR_P4_MS_CCCR3,
457 	MSR_P4_FLAME_CCCR0,
458 	MSR_P4_FLAME_CCCR1,
459 	MSR_P4_FLAME_CCCR2,
460 	MSR_P4_FLAME_CCCR3,
461 	MSR_P4_IQ_CCCR0,
462 	MSR_P4_IQ_CCCR1,
463 	MSR_P4_IQ_CCCR2,
464 	MSR_P4_IQ_CCCR3,
465 	MSR_P4_IQ_CCCR4,
466 	MSR_P4_IQ_CCCR5,
467 };
468 /*
469  * Set up IQ_COUNTER0 to behave like a clock, by having IQ_CCCR0 filter
470  * CRU_ESCR0 (with any non-null event selector) through a complemented
471  * max threshold. [IA32-Vol3, Section 14.9.9]
472  */
473 static int setup_p4_watchdog(unsigned nmi_hz)
474 {
475 	unsigned int perfctr_msr, evntsel_msr, cccr_msr;
476 	unsigned int evntsel, cccr_val;
477 	unsigned int misc_enable, dummy;
478 	unsigned int ht_num;
479 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
480 
481 	rdmsr(MSR_IA32_MISC_ENABLE, misc_enable, dummy);
482 	if (!(misc_enable & MSR_P4_MISC_ENABLE_PERF_AVAIL))
483 		return 0;
484 
485 #ifdef CONFIG_SMP
486 	/* detect which hyperthread we are on */
487 	if (smp_num_siblings == 2) {
488 		unsigned int ebx, apicid;
489 
490         	ebx = cpuid_ebx(1);
491 	        apicid = (ebx >> 24) & 0xff;
492         	ht_num = apicid & 1;
493 	} else
494 #endif
495 		ht_num = 0;
496 
497 	/*
498 	 * performance counters are shared resources
499 	 * assign each hyperthread its own set
500 	 * (re-use the ESCR0 register, seems safe
501 	 * and keeps the cccr_val the same)
502 	 */
503 	if (!ht_num) {
504 		/* logical cpu 0 */
505 		perfctr_msr = MSR_P4_IQ_PERFCTR0;
506 		evntsel_msr = MSR_P4_CRU_ESCR0;
507 		cccr_msr = MSR_P4_IQ_CCCR0;
508 		cccr_val = P4_CCCR_OVF_PMI0 | P4_CCCR_ESCR_SELECT(4);
509 
510 		/*
511 		 * If we're on the kdump kernel or other situation, we may
512 		 * still have other performance counter registers set to
513 		 * interrupt and they'll keep interrupting forever because
514 		 * of the P4_CCCR_OVF quirk. So we need to ACK all the
515 		 * pending interrupts and disable all the registers here,
516 		 * before reenabling the NMI delivery. Refer to p4_rearm()
517 		 * about the P4_CCCR_OVF quirk.
518 		 */
519 		if (reset_devices) {
520 			unsigned int low, high;
521 			int i;
522 
523 			for (i = 0; i < P4_CONTROLS; i++) {
524 				rdmsr(p4_controls[i], low, high);
525 				low &= ~(P4_CCCR_ENABLE | P4_CCCR_OVF);
526 				wrmsr(p4_controls[i], low, high);
527 			}
528 		}
529 	} else {
530 		/* logical cpu 1 */
531 		perfctr_msr = MSR_P4_IQ_PERFCTR1;
532 		evntsel_msr = MSR_P4_CRU_ESCR0;
533 		cccr_msr = MSR_P4_IQ_CCCR1;
534 
535 		/* Pentium 4 D processors don't support P4_CCCR_OVF_PMI1 */
536 		if (boot_cpu_data.x86_model == 4 && boot_cpu_data.x86_mask == 4)
537 			cccr_val = P4_CCCR_OVF_PMI0;
538 		else
539 			cccr_val = P4_CCCR_OVF_PMI1;
540 		cccr_val |= P4_CCCR_ESCR_SELECT(4);
541 	}
542 
543 	evntsel = P4_ESCR_EVENT_SELECT(0x3F)
544 	 	| P4_ESCR_OS
545 		| P4_ESCR_USR;
546 
547 	cccr_val |= P4_CCCR_THRESHOLD(15)
548 		 | P4_CCCR_COMPLEMENT
549 		 | P4_CCCR_COMPARE
550 		 | P4_CCCR_REQUIRED;
551 
552 	wrmsr(evntsel_msr, evntsel, 0);
553 	wrmsr(cccr_msr, cccr_val, 0);
554 	write_watchdog_counter(perfctr_msr, "P4_IQ_COUNTER0", nmi_hz);
555 
556 	wd->perfctr_msr = perfctr_msr;
557 	wd->evntsel_msr = evntsel_msr;
558 	wd->cccr_msr = cccr_msr;
559 
560 	/* ok, everything is initialized, announce that we're set */
561 	cpu_nmi_set_wd_enabled();
562 
563 	apic_write(APIC_LVTPC, APIC_DM_NMI);
564 	cccr_val |= P4_CCCR_ENABLE;
565 	wrmsr(cccr_msr, cccr_val, 0);
566 	return 1;
567 }
568 
569 static void stop_p4_watchdog(void)
570 {
571 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
572 	wrmsr(wd->cccr_msr, 0, 0);
573 	wrmsr(wd->evntsel_msr, 0, 0);
574 }
575 
576 static int p4_reserve(void)
577 {
578 	if (!reserve_perfctr_nmi(MSR_P4_IQ_PERFCTR0))
579 		return 0;
580 #ifdef CONFIG_SMP
581 	if (smp_num_siblings > 1 && !reserve_perfctr_nmi(MSR_P4_IQ_PERFCTR1))
582 		goto fail1;
583 #endif
584 	if (!reserve_evntsel_nmi(MSR_P4_CRU_ESCR0))
585 		goto fail2;
586 	/* RED-PEN why is ESCR1 not reserved here? */
587 	return 1;
588  fail2:
589 #ifdef CONFIG_SMP
590 	if (smp_num_siblings > 1)
591 		release_perfctr_nmi(MSR_P4_IQ_PERFCTR1);
592  fail1:
593 #endif
594 	release_perfctr_nmi(MSR_P4_IQ_PERFCTR0);
595 	return 0;
596 }
597 
598 static void p4_unreserve(void)
599 {
600 #ifdef CONFIG_SMP
601 	if (smp_num_siblings > 1)
602 		release_perfctr_nmi(MSR_P4_IQ_PERFCTR1);
603 #endif
604 	release_evntsel_nmi(MSR_P4_CRU_ESCR0);
605 	release_perfctr_nmi(MSR_P4_IQ_PERFCTR0);
606 }
607 
608 static void p4_rearm(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz)
609 {
610 	unsigned dummy;
611 	/*
612  	 * P4 quirks:
613 	 * - An overflown perfctr will assert its interrupt
614 	 *   until the OVF flag in its CCCR is cleared.
615 	 * - LVTPC is masked on interrupt and must be
616 	 *   unmasked by the LVTPC handler.
617 	 */
618 	rdmsrl(wd->cccr_msr, dummy);
619 	dummy &= ~P4_CCCR_OVF;
620 	wrmsrl(wd->cccr_msr, dummy);
621 	apic_write(APIC_LVTPC, APIC_DM_NMI);
622 	/* start the cycle over again */
623 	write_watchdog_counter(wd->perfctr_msr, NULL, nmi_hz);
624 }
625 
626 static const struct wd_ops p4_wd_ops = {
627 	.reserve	= p4_reserve,
628 	.unreserve	= p4_unreserve,
629 	.setup		= setup_p4_watchdog,
630 	.rearm		= p4_rearm,
631 	.stop		= stop_p4_watchdog,
632 	/* RED-PEN this is wrong for the other sibling */
633 	.perfctr	= MSR_P4_BPU_PERFCTR0,
634 	.evntsel	= MSR_P4_BSU_ESCR0,
635 	.checkbit	= 1ULL << 39,
636 };
637 
638 /*
639  * Watchdog using the Intel architected PerfMon.
640  * Used for Core2 and hopefully all future Intel CPUs.
641  */
642 #define ARCH_PERFMON_NMI_EVENT_SEL	ARCH_PERFMON_UNHALTED_CORE_CYCLES_SEL
643 #define ARCH_PERFMON_NMI_EVENT_UMASK	ARCH_PERFMON_UNHALTED_CORE_CYCLES_UMASK
644 
645 static struct wd_ops intel_arch_wd_ops;
646 
647 static int setup_intel_arch_watchdog(unsigned nmi_hz)
648 {
649 	unsigned int ebx;
650 	union cpuid10_eax eax;
651 	unsigned int unused;
652 	unsigned int perfctr_msr, evntsel_msr;
653 	unsigned int evntsel;
654 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
655 
656 	/*
657 	 * Check whether the Architectural PerfMon supports
658 	 * Unhalted Core Cycles Event or not.
659 	 * NOTE: Corresponding bit = 0 in ebx indicates event present.
660 	 */
661 	cpuid(10, &(eax.full), &ebx, &unused, &unused);
662 	if ((eax.split.mask_length < (ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX+1)) ||
663 	    (ebx & ARCH_PERFMON_UNHALTED_CORE_CYCLES_PRESENT))
664 		return 0;
665 
666 	perfctr_msr = wd_ops->perfctr;
667 	evntsel_msr = wd_ops->evntsel;
668 
669 	wrmsrl(perfctr_msr, 0UL);
670 
671 	evntsel = ARCH_PERFMON_EVENTSEL_INT
672 		| ARCH_PERFMON_EVENTSEL_OS
673 		| ARCH_PERFMON_EVENTSEL_USR
674 		| ARCH_PERFMON_NMI_EVENT_SEL
675 		| ARCH_PERFMON_NMI_EVENT_UMASK;
676 
677 	/* setup the timer */
678 	wrmsr(evntsel_msr, evntsel, 0);
679 	nmi_hz = adjust_for_32bit_ctr(nmi_hz);
680 	write_watchdog_counter32(perfctr_msr, "INTEL_ARCH_PERFCTR0", nmi_hz);
681 
682 	wd->perfctr_msr = perfctr_msr;
683 	wd->evntsel_msr = evntsel_msr;
684 	wd->cccr_msr = 0;  /* unused */
685 
686 	/* ok, everything is initialized, announce that we're set */
687 	cpu_nmi_set_wd_enabled();
688 
689 	apic_write(APIC_LVTPC, APIC_DM_NMI);
690 	evntsel |= ARCH_PERFMON_EVENTSEL0_ENABLE;
691 	wrmsr(evntsel_msr, evntsel, 0);
692 	intel_arch_wd_ops.checkbit = 1ULL << (eax.split.bit_width - 1);
693 	return 1;
694 }
695 
696 static struct wd_ops intel_arch_wd_ops __read_mostly = {
697 	.reserve	= single_msr_reserve,
698 	.unreserve	= single_msr_unreserve,
699 	.setup		= setup_intel_arch_watchdog,
700 	.rearm		= p6_rearm,
701 	.stop		= single_msr_stop_watchdog,
702 	.perfctr	= MSR_ARCH_PERFMON_PERFCTR1,
703 	.evntsel	= MSR_ARCH_PERFMON_EVENTSEL1,
704 };
705 
706 static void probe_nmi_watchdog(void)
707 {
708 	switch (boot_cpu_data.x86_vendor) {
709 	case X86_VENDOR_AMD:
710 		if (boot_cpu_data.x86 != 6 && boot_cpu_data.x86 != 15 &&
711 		    boot_cpu_data.x86 != 16)
712 			return;
713 		wd_ops = &k7_wd_ops;
714 		break;
715 	case X86_VENDOR_INTEL:
716 		/*
717 		 * Work around Core Duo (Yonah) errata AE49 where perfctr1
718 		 * doesn't have a working enable bit.
719 		 */
720 		if (boot_cpu_data.x86 == 6 && boot_cpu_data.x86_model == 14) {
721 			intel_arch_wd_ops.perfctr = MSR_ARCH_PERFMON_PERFCTR0;
722 			intel_arch_wd_ops.evntsel = MSR_ARCH_PERFMON_EVENTSEL0;
723 		}
724 		if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
725 			wd_ops = &intel_arch_wd_ops;
726 			break;
727 		}
728 		switch (boot_cpu_data.x86) {
729 		case 6:
730 			if (boot_cpu_data.x86_model > 13)
731 				return;
732 
733 			wd_ops = &p6_wd_ops;
734 			break;
735 		case 15:
736 			wd_ops = &p4_wd_ops;
737 			break;
738 		default:
739 			return;
740 		}
741 		break;
742 	}
743 }
744 
745 /* Interface to nmi.c */
746 
747 int lapic_watchdog_init(unsigned nmi_hz)
748 {
749 	if (!wd_ops) {
750 		probe_nmi_watchdog();
751 		if (!wd_ops) {
752 			printk(KERN_INFO "NMI watchdog: CPU not supported\n");
753 			return -1;
754 		}
755 
756 		if (!wd_ops->reserve()) {
757 			printk(KERN_ERR
758 				"NMI watchdog: cannot reserve perfctrs\n");
759 			return -1;
760 		}
761 	}
762 
763 	if (!(wd_ops->setup(nmi_hz))) {
764 		printk(KERN_ERR "Cannot setup NMI watchdog on CPU %d\n",
765 		       raw_smp_processor_id());
766 		return -1;
767 	}
768 
769 	return 0;
770 }
771 
772 void lapic_watchdog_stop(void)
773 {
774 	if (wd_ops)
775 		wd_ops->stop();
776 }
777 
778 unsigned lapic_adjust_nmi_hz(unsigned hz)
779 {
780 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
781 	if (wd->perfctr_msr == MSR_P6_PERFCTR0 ||
782 	    wd->perfctr_msr == MSR_ARCH_PERFMON_PERFCTR1)
783 		hz = adjust_for_32bit_ctr(hz);
784 	return hz;
785 }
786 
787 int lapic_wd_event(unsigned nmi_hz)
788 {
789 	struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
790 	u64 ctr;
791 
792 	rdmsrl(wd->perfctr_msr, ctr);
793 	if (ctr & wd_ops->checkbit) /* perfctr still running? */
794 		return 0;
795 
796 	wd_ops->rearm(wd, nmi_hz);
797 	return 1;
798 }
799 
800 int lapic_watchdog_ok(void)
801 {
802 	return wd_ops != NULL;
803 }
804