xref: /netbsd-src/sys/arch/x86/acpi/acpi_cpu_md.c (revision 1b9578b8c2c1f848eeb16dabbfd7d1f0d9fdefbd)
1 /* $NetBSD: acpi_cpu_md.c,v 1.63 2011/06/23 08:10:35 jruoho Exp $ */
2 
3 /*-
4  * Copyright (c) 2010, 2011 Jukka Ruohonen <jruohonen@iki.fi>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 #include <sys/cdefs.h>
30 __KERNEL_RCSID(0, "$NetBSD: acpi_cpu_md.c,v 1.63 2011/06/23 08:10:35 jruoho Exp $");
31 
32 #include <sys/param.h>
33 #include <sys/bus.h>
34 #include <sys/device.h>
35 #include <sys/kcore.h>
36 #include <sys/sysctl.h>
37 #include <sys/xcall.h>
38 
39 #include <x86/cpu.h>
40 #include <x86/cpufunc.h>
41 #include <x86/cputypes.h>
42 #include <x86/cpuvar.h>
43 #include <x86/cpu_msr.h>
44 #include <x86/machdep.h>
45 
46 #include <dev/acpi/acpica.h>
47 #include <dev/acpi/acpi_cpu.h>
48 
49 #include <dev/pci/pcivar.h>
50 #include <dev/pci/pcidevs.h>
51 
52 #include <machine/acpi_machdep.h>
53 
54 /*
55  * Intel IA32_MISC_ENABLE.
56  */
57 #define MSR_MISC_ENABLE_EST	__BIT(16)
58 #define MSR_MISC_ENABLE_TURBO	__BIT(38)
59 
60 /*
61  * AMD C1E.
62  */
63 #define MSR_CMPHALT		0xc0010055
64 
65 #define MSR_CMPHALT_SMI		__BIT(27)
66 #define MSR_CMPHALT_C1E		__BIT(28)
67 #define MSR_CMPHALT_BMSTS	__BIT(29)
68 
69 /*
70  * AMD families 10h, 11h, and 14h
71  */
72 #define MSR_10H_LIMIT		0xc0010061
73 #define MSR_10H_CONTROL		0xc0010062
74 #define MSR_10H_STATUS		0xc0010063
75 #define MSR_10H_CONFIG		0xc0010064
76 
77 /*
78  * AMD family 0Fh.
79  */
80 #define MSR_0FH_CONTROL		0xc0010041
81 #define MSR_0FH_STATUS		0xc0010042
82 
83 #define MSR_0FH_STATUS_CFID	__BITS( 0,  5)
84 #define MSR_0FH_STATUS_CVID	__BITS(32, 36)
85 #define MSR_0FH_STATUS_PENDING	__BITS(31, 31)
86 
87 #define MSR_0FH_CONTROL_FID	__BITS( 0,  5)
88 #define MSR_0FH_CONTROL_VID	__BITS( 8, 12)
89 #define MSR_0FH_CONTROL_CHG	__BITS(16, 16)
90 #define MSR_0FH_CONTROL_CNT	__BITS(32, 51)
91 
92 #define ACPI_0FH_STATUS_FID	__BITS( 0,  5)
93 #define ACPI_0FH_STATUS_VID	__BITS( 6, 10)
94 
95 #define ACPI_0FH_CONTROL_FID	__BITS( 0,  5)
96 #define ACPI_0FH_CONTROL_VID	__BITS( 6, 10)
97 #define ACPI_0FH_CONTROL_VST	__BITS(11, 17)
98 #define ACPI_0FH_CONTROL_MVS	__BITS(18, 19)
99 #define ACPI_0FH_CONTROL_PLL	__BITS(20, 26)
100 #define ACPI_0FH_CONTROL_RVO	__BITS(28, 29)
101 #define ACPI_0FH_CONTROL_IRT	__BITS(30, 31)
102 
103 #define FID_TO_VCO_FID(fidd)	(((fid) < 8) ? (8 + ((fid) << 1)) : (fid))
104 
105 static char	  native_idle_text[16];
106 void		(*native_idle)(void) = NULL;
107 
108 static int	 acpicpu_md_quirk_piix4(const struct pci_attach_args *);
109 static void	 acpicpu_md_pstate_hwf_reset(void *, void *);
110 static int	 acpicpu_md_pstate_fidvid_get(struct acpicpu_softc *,
111                                               uint32_t *);
112 static int	 acpicpu_md_pstate_fidvid_set(struct acpicpu_pstate *);
113 static int	 acpicpu_md_pstate_fidvid_read(uint32_t *, uint32_t *);
114 static void	 acpicpu_md_pstate_fidvid_write(uint32_t, uint32_t,
115 					        uint32_t, uint32_t);
116 static int	 acpicpu_md_pstate_sysctl_init(void);
117 static int	 acpicpu_md_pstate_sysctl_get(SYSCTLFN_PROTO);
118 static int	 acpicpu_md_pstate_sysctl_set(SYSCTLFN_PROTO);
119 static int	 acpicpu_md_pstate_sysctl_all(SYSCTLFN_PROTO);
120 
121 extern struct acpicpu_softc **acpicpu_sc;
122 static struct sysctllog *acpicpu_log = NULL;
123 
124 struct cpu_info *
125 acpicpu_md_match(device_t parent, cfdata_t match, void *aux)
126 {
127 	struct cpufeature_attach_args *cfaa = aux;
128 
129 	if (strcmp(cfaa->name, "frequency") != 0)
130 		return NULL;
131 
132 	return cfaa->ci;
133 }
134 
135 struct cpu_info *
136 acpicpu_md_attach(device_t parent, device_t self, void *aux)
137 {
138 	struct cpufeature_attach_args *cfaa = aux;
139 
140 	return cfaa->ci;
141 }
142 
143 uint32_t
144 acpicpu_md_flags(void)
145 {
146 	struct cpu_info *ci = curcpu();
147 	struct pci_attach_args pa;
148 	uint32_t family, val = 0;
149 	uint32_t regs[4];
150 
151 	if (acpi_md_ncpus() == 1)
152 		val |= ACPICPU_FLAG_C_BM;
153 
154 	if ((ci->ci_feat_val[1] & CPUID2_MONITOR) != 0)
155 		val |= ACPICPU_FLAG_C_FFH;
156 
157 	/*
158 	 * By default, assume that the local APIC timer
159 	 * as well as TSC are stalled during C3 sleep.
160 	 */
161 	val |= ACPICPU_FLAG_C_APIC | ACPICPU_FLAG_C_TSC;
162 
163 	switch (cpu_vendor) {
164 
165 	case CPUVENDOR_IDT:
166 
167 		if ((ci->ci_feat_val[1] & CPUID2_EST) != 0)
168 			val |= ACPICPU_FLAG_P_FFH;
169 
170 		if ((ci->ci_feat_val[0] & CPUID_ACPI) != 0)
171 			val |= ACPICPU_FLAG_T_FFH;
172 
173 		break;
174 
175 	case CPUVENDOR_INTEL:
176 
177 		/*
178 		 * Bus master control and arbitration should be
179 		 * available on all supported Intel CPUs (to be
180 		 * sure, this is double-checked later from the
181 		 * firmware data). These flags imply that it is
182 		 * not necessary to flush caches before C3 state.
183 		 */
184 		val |= ACPICPU_FLAG_C_BM | ACPICPU_FLAG_C_ARB;
185 
186 		/*
187 		 * Check if we can use "native", MSR-based,
188 		 * access. If not, we have to resort to I/O.
189 		 */
190 		if ((ci->ci_feat_val[1] & CPUID2_EST) != 0)
191 			val |= ACPICPU_FLAG_P_FFH;
192 
193 		if ((ci->ci_feat_val[0] & CPUID_ACPI) != 0)
194 			val |= ACPICPU_FLAG_T_FFH;
195 
196 		/*
197 		 * Check whether MSR_APERF, MSR_MPERF, and Turbo
198 		 * Boost are available. Also see if we might have
199 		 * an invariant local APIC timer ("ARAT").
200 		 */
201 		if (cpuid_level >= 0x06) {
202 
203 			x86_cpuid(0x00000006, regs);
204 
205 			if ((regs[2] & CPUID_DSPM_HWF) != 0)
206 				val |= ACPICPU_FLAG_P_HWF;
207 
208 			if ((regs[0] & CPUID_DSPM_IDA) != 0)
209 				val |= ACPICPU_FLAG_P_TURBO;
210 
211 			if ((regs[0] & CPUID_DSPM_ARAT) != 0)
212 				val &= ~ACPICPU_FLAG_C_APIC;
213 		}
214 
215 		/*
216 		 * Detect whether TSC is invariant. If it is not,
217 		 * we keep the flag to note that TSC will not run
218 		 * at constant rate. Depending on the CPU, this may
219 		 * affect P- and T-state changes, but especially
220 		 * relevant are C-states; with variant TSC, states
221 		 * larger than C1 may completely stop the counter.
222 		 */
223 		x86_cpuid(0x80000000, regs);
224 
225 		if (regs[0] >= 0x80000007) {
226 
227 			x86_cpuid(0x80000007, regs);
228 
229 			if ((regs[3] & __BIT(8)) != 0)
230 				val &= ~ACPICPU_FLAG_C_TSC;
231 		}
232 
233 		break;
234 
235 	case CPUVENDOR_AMD:
236 
237 		x86_cpuid(0x80000000, regs);
238 
239 		if (regs[0] < 0x80000007)
240 			break;
241 
242 		x86_cpuid(0x80000007, regs);
243 
244 		family = CPUID2FAMILY(ci->ci_signature);
245 
246 		if (family == 0xf)
247 			family += CPUID2EXTFAMILY(ci->ci_signature);
248 
249     		switch (family) {
250 
251 		case 0x0f:
252 
253 			/*
254 			 * Evaluate support for the "FID/VID
255 			 * algorithm" also used by powernow(4).
256 			 */
257 			if ((regs[3] & CPUID_APM_FID) == 0)
258 				break;
259 
260 			if ((regs[3] & CPUID_APM_VID) == 0)
261 				break;
262 
263 			val |= ACPICPU_FLAG_P_FFH | ACPICPU_FLAG_P_FIDVID;
264 			break;
265 
266 		case 0x10:
267 		case 0x11:
268 			val |= ACPICPU_FLAG_C_C1E;
269 			/* FALLTHROUGH */
270 
271 		case 0x14: /* AMD Fusion */
272 
273 			/*
274 			 * Like with Intel, detect invariant TSC,
275 			 * MSR-based P-states, and AMD's "turbo"
276 			 * (Core Performance Boost), respectively.
277 			 */
278 			if ((regs[3] & CPUID_APM_TSC) != 0)
279 				val &= ~ACPICPU_FLAG_C_TSC;
280 
281 			if ((regs[3] & CPUID_APM_HWP) != 0)
282 				val |= ACPICPU_FLAG_P_FFH;
283 
284 			if ((regs[3] & CPUID_APM_CPB) != 0)
285 				val |= ACPICPU_FLAG_P_TURBO;
286 
287 			/*
288 			 * Also check for APERF and MPERF,
289 			 * first available in the family 10h.
290 			 */
291 			if (cpuid_level >= 0x06) {
292 
293 				x86_cpuid(0x00000006, regs);
294 
295 				if ((regs[2] & CPUID_DSPM_HWF) != 0)
296 					val |= ACPICPU_FLAG_P_HWF;
297 			}
298 
299 			break;
300 		}
301 
302 		break;
303 	}
304 
305 	/*
306 	 * There are several erratums for PIIX4.
307 	 */
308 	if (pci_find_device(&pa, acpicpu_md_quirk_piix4) != 0)
309 		val |= ACPICPU_FLAG_PIIX4;
310 
311 	return val;
312 }
313 
314 static int
315 acpicpu_md_quirk_piix4(const struct pci_attach_args *pa)
316 {
317 
318 	/*
319 	 * XXX: The pci_find_device(9) function only
320 	 *	deals with attached devices. Change this
321 	 *	to use something like pci_device_foreach().
322 	 */
323 	if (PCI_VENDOR(pa->pa_id) != PCI_VENDOR_INTEL)
324 		return 0;
325 
326 	if (PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_INTEL_82371AB_ISA ||
327 	    PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_INTEL_82440MX_PMC)
328 		return 1;
329 
330 	return 0;
331 }
332 
333 void
334 acpicpu_md_quirk_c1e(void)
335 {
336 	const uint64_t c1e = MSR_CMPHALT_SMI | MSR_CMPHALT_C1E;
337 	uint64_t val;
338 
339 	val = rdmsr(MSR_CMPHALT);
340 
341 	if ((val & c1e) != 0)
342 		wrmsr(MSR_CMPHALT, val & ~c1e);
343 }
344 
345 int
346 acpicpu_md_cstate_start(struct acpicpu_softc *sc)
347 {
348 	const size_t size = sizeof(native_idle_text);
349 	struct acpicpu_cstate *cs;
350 	bool ipi = false;
351 	int i;
352 
353 	/*
354 	 * Save the cpu_idle(9) loop used by default.
355 	 */
356 	x86_cpu_idle_get(&native_idle, native_idle_text, size);
357 
358 	for (i = 0; i < ACPI_C_STATE_COUNT; i++) {
359 
360 		cs = &sc->sc_cstate[i];
361 
362 		if (cs->cs_method == ACPICPU_C_STATE_HALT) {
363 			ipi = true;
364 			break;
365 		}
366 	}
367 
368 	x86_cpu_idle_set(acpicpu_cstate_idle, "acpi", ipi);
369 
370 	return 0;
371 }
372 
373 int
374 acpicpu_md_cstate_stop(void)
375 {
376 	static char text[16];
377 	void (*func)(void);
378 	uint64_t xc;
379 	bool ipi;
380 
381 	x86_cpu_idle_get(&func, text, sizeof(text));
382 
383 	if (func == native_idle)
384 		return EALREADY;
385 
386 	ipi = (native_idle != x86_cpu_idle_halt) ? false : true;
387 	x86_cpu_idle_set(native_idle, native_idle_text, ipi);
388 
389 	/*
390 	 * Run a cross-call to ensure that all CPUs are
391 	 * out from the ACPI idle-loop before detachment.
392 	 */
393 	xc = xc_broadcast(0, (xcfunc_t)nullop, NULL, NULL);
394 	xc_wait(xc);
395 
396 	return 0;
397 }
398 
399 /*
400  * Called with interrupts disabled.
401  * Caller should enable interrupts after return.
402  */
403 void
404 acpicpu_md_cstate_enter(int method, int state)
405 {
406 	struct cpu_info *ci = curcpu();
407 
408 	switch (method) {
409 
410 	case ACPICPU_C_STATE_FFH:
411 
412 		x86_enable_intr();
413 		x86_monitor(&ci->ci_want_resched, 0, 0);
414 
415 		if (__predict_false(ci->ci_want_resched != 0))
416 			return;
417 
418 		x86_mwait((state - 1) << 4, 0);
419 		break;
420 
421 	case ACPICPU_C_STATE_HALT:
422 
423 		if (__predict_false(ci->ci_want_resched != 0))
424 			return;
425 
426 		x86_stihlt();
427 		break;
428 	}
429 }
430 
431 int
432 acpicpu_md_pstate_start(struct acpicpu_softc *sc)
433 {
434 	uint64_t xc, val;
435 
436 	switch (cpu_vendor) {
437 
438 	case CPUVENDOR_IDT:
439 	case CPUVENDOR_INTEL:
440 
441 		/*
442 		 * Make sure EST is enabled.
443 		 */
444 		if ((sc->sc_flags & ACPICPU_FLAG_P_FFH) != 0) {
445 
446 			val = rdmsr(MSR_MISC_ENABLE);
447 
448 			if ((val & MSR_MISC_ENABLE_EST) == 0) {
449 
450 				val |= MSR_MISC_ENABLE_EST;
451 				wrmsr(MSR_MISC_ENABLE, val);
452 				val = rdmsr(MSR_MISC_ENABLE);
453 
454 				if ((val & MSR_MISC_ENABLE_EST) == 0)
455 					return ENOTTY;
456 			}
457 		}
458 	}
459 
460 	/*
461 	 * Reset the APERF and MPERF counters.
462 	 */
463 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
464 		xc = xc_broadcast(0, acpicpu_md_pstate_hwf_reset, NULL, NULL);
465 		xc_wait(xc);
466 	}
467 
468 	return acpicpu_md_pstate_sysctl_init();
469 }
470 
471 int
472 acpicpu_md_pstate_stop(void)
473 {
474 
475 	if (acpicpu_log == NULL)
476 		return EALREADY;
477 
478 	sysctl_teardown(&acpicpu_log);
479 	acpicpu_log = NULL;
480 
481 	return 0;
482 }
483 
484 int
485 acpicpu_md_pstate_init(struct acpicpu_softc *sc)
486 {
487 	struct cpu_info *ci = sc->sc_ci;
488 	struct acpicpu_pstate *ps, msr;
489 	uint32_t family, i = 0;
490 
491 	(void)memset(&msr, 0, sizeof(struct acpicpu_pstate));
492 
493 	switch (cpu_vendor) {
494 
495 	case CPUVENDOR_IDT:
496 	case CPUVENDOR_INTEL:
497 
498 		/*
499 		 * If the so-called Turbo Boost is present,
500 		 * the P0-state is always the "turbo state".
501 		 * It is shown as the P1 frequency + 1 MHz.
502 		 *
503 		 * For discussion, see:
504 		 *
505 		 *	Intel Corporation: Intel Turbo Boost Technology
506 		 *	in Intel Core(tm) Microarchitectures (Nehalem)
507 		 *	Based Processors. White Paper, November 2008.
508 		 */
509 		if (sc->sc_pstate_count >= 2 &&
510 		   (sc->sc_flags & ACPICPU_FLAG_P_TURBO) != 0) {
511 
512 			ps = &sc->sc_pstate[0];
513 
514 			if (ps->ps_freq == sc->sc_pstate[1].ps_freq + 1)
515 				ps->ps_flags |= ACPICPU_FLAG_P_TURBO;
516 		}
517 
518 		msr.ps_control_addr = MSR_PERF_CTL;
519 		msr.ps_control_mask = __BITS(0, 15);
520 
521 		msr.ps_status_addr  = MSR_PERF_STATUS;
522 		msr.ps_status_mask  = __BITS(0, 15);
523 		break;
524 
525 	case CPUVENDOR_AMD:
526 
527 		if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
528 			msr.ps_flags |= ACPICPU_FLAG_P_FIDVID;
529 
530 		family = CPUID2FAMILY(ci->ci_signature);
531 
532 		if (family == 0xf)
533 			family += CPUID2EXTFAMILY(ci->ci_signature);
534 
535 		switch (family) {
536 
537 		case 0x0f:
538 			msr.ps_control_addr = MSR_0FH_CONTROL;
539 			msr.ps_status_addr  = MSR_0FH_STATUS;
540 			break;
541 
542 		case 0x10:
543 		case 0x11:
544 		case 0x14: /* AMD Fusion */
545 			msr.ps_control_addr = MSR_10H_CONTROL;
546 			msr.ps_control_mask = __BITS(0, 2);
547 
548 			msr.ps_status_addr  = MSR_10H_STATUS;
549 			msr.ps_status_mask  = __BITS(0, 2);
550 			break;
551 
552 		default:
553 			/*
554 			 * If we have an unknown AMD CPU, rely on XPSS.
555 			 */
556 			if ((sc->sc_flags & ACPICPU_FLAG_P_XPSS) == 0)
557 				return EOPNOTSUPP;
558 		}
559 
560 		break;
561 
562 	default:
563 		return ENODEV;
564 	}
565 
566 	/*
567 	 * Fill the P-state structures with MSR addresses that are
568 	 * known to be correct. If we do not know the addresses,
569 	 * leave the values intact. If a vendor uses XPSS, we do
570 	 * not necessarily need to do anything to support new CPUs.
571 	 */
572 	while (i < sc->sc_pstate_count) {
573 
574 		ps = &sc->sc_pstate[i];
575 
576 		if (msr.ps_flags != 0)
577 			ps->ps_flags |= msr.ps_flags;
578 
579 		if (msr.ps_status_addr != 0)
580 			ps->ps_status_addr = msr.ps_status_addr;
581 
582 		if (msr.ps_status_mask != 0)
583 			ps->ps_status_mask = msr.ps_status_mask;
584 
585 		if (msr.ps_control_addr != 0)
586 			ps->ps_control_addr = msr.ps_control_addr;
587 
588 		if (msr.ps_control_mask != 0)
589 			ps->ps_control_mask = msr.ps_control_mask;
590 
591 		i++;
592 	}
593 
594 	return 0;
595 }
596 
597 /*
598  * Read the IA32_APERF and IA32_MPERF counters. The first
599  * increments at the rate of the fixed maximum frequency
600  * configured during the boot, whereas APERF counts at the
601  * rate of the actual frequency. Note that the MSRs must be
602  * read without delay, and that only the ratio between
603  * IA32_APERF and IA32_MPERF is architecturally defined.
604  *
605  * The function thus returns the percentage of the actual
606  * frequency in terms of the maximum frequency of the calling
607  * CPU since the last call. A value zero implies an error.
608  *
609  * For further details, refer to:
610  *
611  *	Intel Corporation: Intel 64 and IA-32 Architectures
612  *	Software Developer's Manual. Section 13.2, Volume 3A:
613  *	System Programming Guide, Part 1. July, 2008.
614  *
615  *	Advanced Micro Devices: BIOS and Kernel Developer's
616  *	Guide (BKDG) for AMD Family 10h Processors. Section
617  *	2.4.5, Revision 3.48, April 2010.
618  */
619 uint8_t
620 acpicpu_md_pstate_hwf(struct cpu_info *ci)
621 {
622 	struct acpicpu_softc *sc;
623 	uint64_t aperf, mperf;
624 	uint8_t rv = 0;
625 
626 	sc = acpicpu_sc[ci->ci_acpiid];
627 
628 	if (__predict_false(sc == NULL))
629 		return 0;
630 
631 	if (__predict_false((sc->sc_flags & ACPICPU_FLAG_P_HWF) == 0))
632 		return 0;
633 
634 	aperf = sc->sc_pstate_aperf;
635 	mperf = sc->sc_pstate_mperf;
636 
637 	x86_disable_intr();
638 
639 	sc->sc_pstate_aperf = rdmsr(MSR_APERF);
640 	sc->sc_pstate_mperf = rdmsr(MSR_MPERF);
641 
642 	x86_enable_intr();
643 
644 	aperf = sc->sc_pstate_aperf - aperf;
645 	mperf = sc->sc_pstate_mperf - mperf;
646 
647 	if (__predict_true(mperf != 0))
648 		rv = (aperf * 100) / mperf;
649 
650 	return rv;
651 }
652 
653 static void
654 acpicpu_md_pstate_hwf_reset(void *arg1, void *arg2)
655 {
656 	struct cpu_info *ci = curcpu();
657 	struct acpicpu_softc *sc;
658 
659 	sc = acpicpu_sc[ci->ci_acpiid];
660 
661 	if (__predict_false(sc == NULL))
662 		return;
663 
664 	x86_disable_intr();
665 
666 	wrmsr(MSR_APERF, 0);
667 	wrmsr(MSR_MPERF, 0);
668 
669 	x86_enable_intr();
670 
671 	sc->sc_pstate_aperf = 0;
672 	sc->sc_pstate_mperf = 0;
673 }
674 
675 int
676 acpicpu_md_pstate_get(struct acpicpu_softc *sc, uint32_t *freq)
677 {
678 	struct acpicpu_pstate *ps = NULL;
679 	uint64_t val;
680 	uint32_t i;
681 
682 	if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
683 		return acpicpu_md_pstate_fidvid_get(sc, freq);
684 
685 	/*
686 	 * Pick any P-state for the status address.
687 	*/
688 	for (i = 0; i < sc->sc_pstate_count; i++) {
689 
690 		ps = &sc->sc_pstate[i];
691 
692 		if (__predict_true(ps->ps_freq != 0))
693 			break;
694 	}
695 
696 	if (__predict_false(ps == NULL))
697 		return ENODEV;
698 
699 	if (__predict_false(ps->ps_status_addr == 0))
700 		return EINVAL;
701 
702 	val = rdmsr(ps->ps_status_addr);
703 
704 	if (__predict_true(ps->ps_status_mask != 0))
705 		val = val & ps->ps_status_mask;
706 
707 	/*
708 	 * Search for the value from known P-states.
709 	 */
710 	for (i = 0; i < sc->sc_pstate_count; i++) {
711 
712 		ps = &sc->sc_pstate[i];
713 
714 		if (__predict_false(ps->ps_freq == 0))
715 			continue;
716 
717 		if (val == ps->ps_status) {
718 			*freq = ps->ps_freq;
719 			return 0;
720 		}
721 	}
722 
723 	/*
724 	 * If the value was not found, try APERF/MPERF.
725 	 * The state is P0 if the return value is 100 %.
726 	 */
727 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
728 
729 		if (acpicpu_md_pstate_hwf(sc->sc_ci) == 100) {
730 			*freq = sc->sc_pstate[0].ps_freq;
731 			return 0;
732 		}
733 	}
734 
735 	return EIO;
736 }
737 
738 int
739 acpicpu_md_pstate_set(struct acpicpu_pstate *ps)
740 {
741 	uint64_t val = 0;
742 
743 	if (__predict_false(ps->ps_control_addr == 0))
744 		return EINVAL;
745 
746 	if ((ps->ps_flags & ACPICPU_FLAG_P_FIDVID) != 0)
747 		return acpicpu_md_pstate_fidvid_set(ps);
748 
749 	/*
750 	 * If the mask is set, do a read-modify-write.
751 	 */
752 	if (__predict_true(ps->ps_control_mask != 0)) {
753 		val = rdmsr(ps->ps_control_addr);
754 		val &= ~ps->ps_control_mask;
755 	}
756 
757 	val |= ps->ps_control;
758 
759 	wrmsr(ps->ps_control_addr, val);
760 	DELAY(ps->ps_latency);
761 
762 	return 0;
763 }
764 
765 static int
766 acpicpu_md_pstate_fidvid_get(struct acpicpu_softc *sc, uint32_t *freq)
767 {
768 	struct acpicpu_pstate *ps;
769 	uint32_t fid, i, vid;
770 	uint32_t cfid, cvid;
771 	int rv;
772 
773 	/*
774 	 * AMD family 0Fh needs special treatment.
775 	 * While it wants to use ACPI, it does not
776 	 * comply with the ACPI specifications.
777 	 */
778 	rv = acpicpu_md_pstate_fidvid_read(&cfid, &cvid);
779 
780 	if (rv != 0)
781 		return rv;
782 
783 	for (i = 0; i < sc->sc_pstate_count; i++) {
784 
785 		ps = &sc->sc_pstate[i];
786 
787 		if (__predict_false(ps->ps_freq == 0))
788 			continue;
789 
790 		fid = __SHIFTOUT(ps->ps_status, ACPI_0FH_STATUS_FID);
791 		vid = __SHIFTOUT(ps->ps_status, ACPI_0FH_STATUS_VID);
792 
793 		if (cfid == fid && cvid == vid) {
794 			*freq = ps->ps_freq;
795 			return 0;
796 		}
797 	}
798 
799 	return EIO;
800 }
801 
802 static int
803 acpicpu_md_pstate_fidvid_set(struct acpicpu_pstate *ps)
804 {
805 	const uint64_t ctrl = ps->ps_control;
806 	uint32_t cfid, cvid, fid, i, irt;
807 	uint32_t pll, vco_cfid, vco_fid;
808 	uint32_t val, vid, vst;
809 	int rv;
810 
811 	rv = acpicpu_md_pstate_fidvid_read(&cfid, &cvid);
812 
813 	if (rv != 0)
814 		return rv;
815 
816 	fid = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_FID);
817 	vid = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_VID);
818 	irt = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_IRT);
819 	vst = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_VST);
820 	pll = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_PLL);
821 
822 	vst = vst * 20;
823 	pll = pll * 1000 / 5;
824 	irt = 10 * __BIT(irt);
825 
826 	/*
827 	 * Phase 1.
828 	 */
829 	while (cvid > vid) {
830 
831 		val = 1 << __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_MVS);
832 		val = (val > cvid) ? 0 : cvid - val;
833 
834 		acpicpu_md_pstate_fidvid_write(cfid, val, 1, vst);
835 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
836 
837 		if (rv != 0)
838 			return rv;
839 	}
840 
841 	i = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_RVO);
842 
843 	for (; i > 0 && cvid > 0; --i) {
844 
845 		acpicpu_md_pstate_fidvid_write(cfid, cvid - 1, 1, vst);
846 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
847 
848 		if (rv != 0)
849 			return rv;
850 	}
851 
852 	/*
853 	 * Phase 2.
854 	 */
855 	if (cfid != fid) {
856 
857 		vco_fid  = FID_TO_VCO_FID(fid);
858 		vco_cfid = FID_TO_VCO_FID(cfid);
859 
860 		while (abs(vco_fid - vco_cfid) > 2) {
861 
862 			if (fid <= cfid)
863 				val = cfid - 2;
864 			else {
865 				val = (cfid > 6) ? cfid + 2 :
866 				    FID_TO_VCO_FID(cfid) + 2;
867 			}
868 
869 			acpicpu_md_pstate_fidvid_write(val, cvid, pll, irt);
870 			rv = acpicpu_md_pstate_fidvid_read(&cfid, NULL);
871 
872 			if (rv != 0)
873 				return rv;
874 
875 			vco_cfid = FID_TO_VCO_FID(cfid);
876 		}
877 
878 		acpicpu_md_pstate_fidvid_write(fid, cvid, pll, irt);
879 		rv = acpicpu_md_pstate_fidvid_read(&cfid, NULL);
880 
881 		if (rv != 0)
882 			return rv;
883 	}
884 
885 	/*
886 	 * Phase 3.
887 	 */
888 	if (cvid != vid) {
889 
890 		acpicpu_md_pstate_fidvid_write(cfid, vid, 1, vst);
891 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
892 
893 		if (rv != 0)
894 			return rv;
895 	}
896 
897 	return 0;
898 }
899 
900 static int
901 acpicpu_md_pstate_fidvid_read(uint32_t *cfid, uint32_t *cvid)
902 {
903 	int i = ACPICPU_P_STATE_RETRY * 100;
904 	uint64_t val;
905 
906 	do {
907 		val = rdmsr(MSR_0FH_STATUS);
908 
909 	} while (__SHIFTOUT(val, MSR_0FH_STATUS_PENDING) != 0 && --i >= 0);
910 
911 	if (i == 0)
912 		return EAGAIN;
913 
914 	if (cfid != NULL)
915 		*cfid = __SHIFTOUT(val, MSR_0FH_STATUS_CFID);
916 
917 	if (cvid != NULL)
918 		*cvid = __SHIFTOUT(val, MSR_0FH_STATUS_CVID);
919 
920 	return 0;
921 }
922 
923 static void
924 acpicpu_md_pstate_fidvid_write(uint32_t fid,
925     uint32_t vid, uint32_t cnt, uint32_t tmo)
926 {
927 	uint64_t val = 0;
928 
929 	val |= __SHIFTIN(fid, MSR_0FH_CONTROL_FID);
930 	val |= __SHIFTIN(vid, MSR_0FH_CONTROL_VID);
931 	val |= __SHIFTIN(cnt, MSR_0FH_CONTROL_CNT);
932 	val |= __SHIFTIN(0x1, MSR_0FH_CONTROL_CHG);
933 
934 	wrmsr(MSR_0FH_CONTROL, val);
935 	DELAY(tmo);
936 }
937 
938 int
939 acpicpu_md_tstate_get(struct acpicpu_softc *sc, uint32_t *percent)
940 {
941 	struct acpicpu_tstate *ts;
942 	uint64_t val;
943 	uint32_t i;
944 
945 	val = rdmsr(MSR_THERM_CONTROL);
946 
947 	for (i = 0; i < sc->sc_tstate_count; i++) {
948 
949 		ts = &sc->sc_tstate[i];
950 
951 		if (ts->ts_percent == 0)
952 			continue;
953 
954 		if (val == ts->ts_status) {
955 			*percent = ts->ts_percent;
956 			return 0;
957 		}
958 	}
959 
960 	return EIO;
961 }
962 
963 int
964 acpicpu_md_tstate_set(struct acpicpu_tstate *ts)
965 {
966 	uint64_t val;
967 	uint8_t i;
968 
969 	val = ts->ts_control;
970 	val = val & __BITS(1, 4);
971 
972 	wrmsr(MSR_THERM_CONTROL, val);
973 
974 	if (ts->ts_status == 0) {
975 		DELAY(ts->ts_latency);
976 		return 0;
977 	}
978 
979 	for (i = val = 0; i < ACPICPU_T_STATE_RETRY; i++) {
980 
981 		val = rdmsr(MSR_THERM_CONTROL);
982 
983 		if (val == ts->ts_status)
984 			return 0;
985 
986 		DELAY(ts->ts_latency);
987 	}
988 
989 	return EAGAIN;
990 }
991 
992 /*
993  * A kludge for backwards compatibility.
994  */
995 static int
996 acpicpu_md_pstate_sysctl_init(void)
997 {
998 	const struct sysctlnode	*fnode, *mnode, *rnode;
999 	const char *str;
1000 	int rv;
1001 
1002 	switch (cpu_vendor) {
1003 
1004 	case CPUVENDOR_IDT:
1005 	case CPUVENDOR_INTEL:
1006 		str = "est";
1007 		break;
1008 
1009 	case CPUVENDOR_AMD:
1010 		str = "powernow";
1011 		break;
1012 
1013 	default:
1014 		return ENODEV;
1015 	}
1016 
1017 
1018 	rv = sysctl_createv(&acpicpu_log, 0, NULL, &rnode,
1019 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "machdep", NULL,
1020 	    NULL, 0, NULL, 0, CTL_MACHDEP, CTL_EOL);
1021 
1022 	if (rv != 0)
1023 		goto fail;
1024 
1025 	rv = sysctl_createv(&acpicpu_log, 0, &rnode, &mnode,
1026 	    0, CTLTYPE_NODE, str, NULL,
1027 	    NULL, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1028 
1029 	if (rv != 0)
1030 		goto fail;
1031 
1032 	rv = sysctl_createv(&acpicpu_log, 0, &mnode, &fnode,
1033 	    0, CTLTYPE_NODE, "frequency", NULL,
1034 	    NULL, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1035 
1036 	if (rv != 0)
1037 		goto fail;
1038 
1039 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1040 	    CTLFLAG_READWRITE, CTLTYPE_INT, "target", NULL,
1041 	    acpicpu_md_pstate_sysctl_set, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1042 
1043 	if (rv != 0)
1044 		goto fail;
1045 
1046 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1047 	    CTLFLAG_READONLY, CTLTYPE_INT, "current", NULL,
1048 	    acpicpu_md_pstate_sysctl_get, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1049 
1050 	if (rv != 0)
1051 		goto fail;
1052 
1053 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1054 	    CTLFLAG_READONLY, CTLTYPE_STRING, "available", NULL,
1055 	    acpicpu_md_pstate_sysctl_all, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1056 
1057 	if (rv != 0)
1058 		goto fail;
1059 
1060 	return 0;
1061 
1062 fail:
1063 	if (acpicpu_log != NULL) {
1064 		sysctl_teardown(&acpicpu_log);
1065 		acpicpu_log = NULL;
1066 	}
1067 
1068 	return rv;
1069 }
1070 
1071 static int
1072 acpicpu_md_pstate_sysctl_get(SYSCTLFN_ARGS)
1073 {
1074 	struct cpu_info *ci = curcpu();
1075 	struct sysctlnode node;
1076 	uint32_t freq;
1077 	int err;
1078 
1079 	err = acpicpu_pstate_get(ci, &freq);
1080 
1081 	if (err != 0)
1082 		return err;
1083 
1084 	node = *rnode;
1085 	node.sysctl_data = &freq;
1086 
1087 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1088 
1089 	if (err != 0 || newp == NULL)
1090 		return err;
1091 
1092 	return 0;
1093 }
1094 
1095 static int
1096 acpicpu_md_pstate_sysctl_set(SYSCTLFN_ARGS)
1097 {
1098 	struct cpu_info *ci = curcpu();
1099 	struct sysctlnode node;
1100 	uint32_t freq;
1101 	int err;
1102 
1103 	err = acpicpu_pstate_get(ci, &freq);
1104 
1105 	if (err != 0)
1106 		return err;
1107 
1108 	node = *rnode;
1109 	node.sysctl_data = &freq;
1110 
1111 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1112 
1113 	if (err != 0 || newp == NULL)
1114 		return err;
1115 
1116 	acpicpu_pstate_set(ci, freq);
1117 
1118 	return 0;
1119 }
1120 
1121 static int
1122 acpicpu_md_pstate_sysctl_all(SYSCTLFN_ARGS)
1123 {
1124 	struct cpu_info *ci = curcpu();
1125 	struct acpicpu_softc *sc;
1126 	struct sysctlnode node;
1127 	char buf[1024];
1128 	size_t len;
1129 	uint32_t i;
1130 	int err;
1131 
1132 	sc = acpicpu_sc[ci->ci_acpiid];
1133 
1134 	if (sc == NULL)
1135 		return ENXIO;
1136 
1137 	(void)memset(&buf, 0, sizeof(buf));
1138 
1139 	mutex_enter(&sc->sc_mtx);
1140 
1141 	for (len = 0, i = sc->sc_pstate_max; i < sc->sc_pstate_count; i++) {
1142 
1143 		if (sc->sc_pstate[i].ps_freq == 0)
1144 			continue;
1145 
1146 		len += snprintf(buf + len, sizeof(buf) - len, "%u%s",
1147 		    sc->sc_pstate[i].ps_freq,
1148 		    i < (sc->sc_pstate_count - 1) ? " " : "");
1149 	}
1150 
1151 	mutex_exit(&sc->sc_mtx);
1152 
1153 	node = *rnode;
1154 	node.sysctl_data = buf;
1155 
1156 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1157 
1158 	if (err != 0 || newp == NULL)
1159 		return err;
1160 
1161 	return 0;
1162 }
1163 
1164