xref: /netbsd-src/sys/arch/x86/acpi/acpi_cpu_md.c (revision d25ffa98a4bfca1fe272f3c182496ec9934faac7)
1 /* $NetBSD: acpi_cpu_md.c,v 1.61 2011/06/12 10:11:52 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 1.61 2011/06/12 10:11:52 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 	uint64_t xc;
377 	bool ipi;
378 
379 	ipi = (native_idle != x86_cpu_idle_halt) ? false : true;
380 	x86_cpu_idle_set(native_idle, native_idle_text, ipi);
381 
382 	/*
383 	 * Run a cross-call to ensure that all CPUs are
384 	 * out from the ACPI idle-loop before detachment.
385 	 */
386 	xc = xc_broadcast(0, (xcfunc_t)nullop, NULL, NULL);
387 	xc_wait(xc);
388 
389 	return 0;
390 }
391 
392 /*
393  * Called with interrupts disabled.
394  * Caller should enable interrupts after return.
395  */
396 void
397 acpicpu_md_cstate_enter(int method, int state)
398 {
399 	struct cpu_info *ci = curcpu();
400 
401 	switch (method) {
402 
403 	case ACPICPU_C_STATE_FFH:
404 
405 		x86_enable_intr();
406 		x86_monitor(&ci->ci_want_resched, 0, 0);
407 
408 		if (__predict_false(ci->ci_want_resched != 0))
409 			return;
410 
411 		x86_mwait((state - 1) << 4, 0);
412 		break;
413 
414 	case ACPICPU_C_STATE_HALT:
415 
416 		if (__predict_false(ci->ci_want_resched != 0))
417 			return;
418 
419 		x86_stihlt();
420 		break;
421 	}
422 }
423 
424 int
425 acpicpu_md_pstate_start(struct acpicpu_softc *sc)
426 {
427 	uint64_t xc;
428 
429 	/*
430 	 * Reset the APERF and MPERF counters.
431 	 */
432 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
433 		xc = xc_broadcast(0, acpicpu_md_pstate_hwf_reset, NULL, NULL);
434 		xc_wait(xc);
435 	}
436 
437 	return acpicpu_md_pstate_sysctl_init();
438 }
439 
440 int
441 acpicpu_md_pstate_stop(void)
442 {
443 	if (acpicpu_log != NULL)
444 		sysctl_teardown(&acpicpu_log);
445 
446 	return 0;
447 }
448 
449 int
450 acpicpu_md_pstate_init(struct acpicpu_softc *sc)
451 {
452 	struct cpu_info *ci = sc->sc_ci;
453 	struct acpicpu_pstate *ps, msr;
454 	uint32_t family, i = 0;
455 	uint64_t val;
456 
457 	(void)memset(&msr, 0, sizeof(struct acpicpu_pstate));
458 
459 	switch (cpu_vendor) {
460 
461 	case CPUVENDOR_IDT:
462 	case CPUVENDOR_INTEL:
463 
464 		/*
465 		 * Make sure EST is enabled.
466 		 */
467 		if ((sc->sc_flags & ACPICPU_FLAG_P_FFH) != 0) {
468 
469 			val = rdmsr(MSR_MISC_ENABLE);
470 
471 			if ((val & MSR_MISC_ENABLE_EST) == 0) {
472 
473 				val |= MSR_MISC_ENABLE_EST;
474 				wrmsr(MSR_MISC_ENABLE, val);
475 				val = rdmsr(MSR_MISC_ENABLE);
476 
477 				if ((val & MSR_MISC_ENABLE_EST) == 0)
478 					return ENOTTY;
479 			}
480 		}
481 
482 		/*
483 		 * If the so-called Turbo Boost is present,
484 		 * the P0-state is always the "turbo state".
485 		 * It is shown as the P1 frequency + 1 MHz.
486 		 *
487 		 * For discussion, see:
488 		 *
489 		 *	Intel Corporation: Intel Turbo Boost Technology
490 		 *	in Intel Core(tm) Microarchitectures (Nehalem)
491 		 *	Based Processors. White Paper, November 2008.
492 		 */
493 		if (sc->sc_pstate_count >= 2 &&
494 		   (sc->sc_flags & ACPICPU_FLAG_P_TURBO) != 0) {
495 
496 			ps = &sc->sc_pstate[0];
497 
498 			if (ps->ps_freq == sc->sc_pstate[1].ps_freq + 1)
499 				ps->ps_flags |= ACPICPU_FLAG_P_TURBO;
500 		}
501 
502 		msr.ps_control_addr = MSR_PERF_CTL;
503 		msr.ps_control_mask = __BITS(0, 15);
504 
505 		msr.ps_status_addr  = MSR_PERF_STATUS;
506 		msr.ps_status_mask  = __BITS(0, 15);
507 		break;
508 
509 	case CPUVENDOR_AMD:
510 
511 		if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
512 			msr.ps_flags |= ACPICPU_FLAG_P_FIDVID;
513 
514 		family = CPUID2FAMILY(ci->ci_signature);
515 
516 		if (family == 0xf)
517 			family += CPUID2EXTFAMILY(ci->ci_signature);
518 
519 		switch (family) {
520 
521 		case 0x0f:
522 			msr.ps_control_addr = MSR_0FH_CONTROL;
523 			msr.ps_status_addr  = MSR_0FH_STATUS;
524 			break;
525 
526 		case 0x10:
527 		case 0x11:
528 		case 0x14: /* AMD Fusion */
529 			msr.ps_control_addr = MSR_10H_CONTROL;
530 			msr.ps_control_mask = __BITS(0, 2);
531 
532 			msr.ps_status_addr  = MSR_10H_STATUS;
533 			msr.ps_status_mask  = __BITS(0, 2);
534 			break;
535 
536 		default:
537 			/*
538 			 * If we have an unknown AMD CPU, rely on XPSS.
539 			 */
540 			if ((sc->sc_flags & ACPICPU_FLAG_P_XPSS) == 0)
541 				return EOPNOTSUPP;
542 		}
543 
544 		break;
545 
546 	default:
547 		return ENODEV;
548 	}
549 
550 	/*
551 	 * Fill the P-state structures with MSR addresses that are
552 	 * known to be correct. If we do not know the addresses,
553 	 * leave the values intact. If a vendor uses XPSS, we do
554 	 * not necessarily need to do anything to support new CPUs.
555 	 */
556 	while (i < sc->sc_pstate_count) {
557 
558 		ps = &sc->sc_pstate[i];
559 
560 		if (msr.ps_flags != 0)
561 			ps->ps_flags |= msr.ps_flags;
562 
563 		if (msr.ps_status_addr != 0)
564 			ps->ps_status_addr = msr.ps_status_addr;
565 
566 		if (msr.ps_status_mask != 0)
567 			ps->ps_status_mask = msr.ps_status_mask;
568 
569 		if (msr.ps_control_addr != 0)
570 			ps->ps_control_addr = msr.ps_control_addr;
571 
572 		if (msr.ps_control_mask != 0)
573 			ps->ps_control_mask = msr.ps_control_mask;
574 
575 		i++;
576 	}
577 
578 	return 0;
579 }
580 
581 /*
582  * Read the IA32_APERF and IA32_MPERF counters. The first
583  * increments at the rate of the fixed maximum frequency
584  * configured during the boot, whereas APERF counts at the
585  * rate of the actual frequency. Note that the MSRs must be
586  * read without delay, and that only the ratio between
587  * IA32_APERF and IA32_MPERF is architecturally defined.
588  *
589  * The function thus returns the percentage of the actual
590  * frequency in terms of the maximum frequency of the calling
591  * CPU since the last call. A value zero implies an error.
592  *
593  * For further details, refer to:
594  *
595  *	Intel Corporation: Intel 64 and IA-32 Architectures
596  *	Software Developer's Manual. Section 13.2, Volume 3A:
597  *	System Programming Guide, Part 1. July, 2008.
598  *
599  *	Advanced Micro Devices: BIOS and Kernel Developer's
600  *	Guide (BKDG) for AMD Family 10h Processors. Section
601  *	2.4.5, Revision 3.48, April 2010.
602  */
603 uint8_t
604 acpicpu_md_pstate_hwf(struct cpu_info *ci)
605 {
606 	struct acpicpu_softc *sc;
607 	uint64_t aperf, mperf;
608 	uint8_t rv = 0;
609 
610 	sc = acpicpu_sc[ci->ci_acpiid];
611 
612 	if (__predict_false(sc == NULL))
613 		return 0;
614 
615 	if (__predict_false((sc->sc_flags & ACPICPU_FLAG_P_HWF) == 0))
616 		return 0;
617 
618 	aperf = sc->sc_pstate_aperf;
619 	mperf = sc->sc_pstate_mperf;
620 
621 	x86_disable_intr();
622 
623 	sc->sc_pstate_aperf = rdmsr(MSR_APERF);
624 	sc->sc_pstate_mperf = rdmsr(MSR_MPERF);
625 
626 	x86_enable_intr();
627 
628 	aperf = sc->sc_pstate_aperf - aperf;
629 	mperf = sc->sc_pstate_mperf - mperf;
630 
631 	if (__predict_true(mperf != 0))
632 		rv = (aperf * 100) / mperf;
633 
634 	return rv;
635 }
636 
637 static void
638 acpicpu_md_pstate_hwf_reset(void *arg1, void *arg2)
639 {
640 	struct cpu_info *ci = curcpu();
641 	struct acpicpu_softc *sc;
642 
643 	sc = acpicpu_sc[ci->ci_acpiid];
644 
645 	if (__predict_false(sc == NULL))
646 		return;
647 
648 	x86_disable_intr();
649 
650 	wrmsr(MSR_APERF, 0);
651 	wrmsr(MSR_MPERF, 0);
652 
653 	x86_enable_intr();
654 
655 	sc->sc_pstate_aperf = 0;
656 	sc->sc_pstate_mperf = 0;
657 }
658 
659 int
660 acpicpu_md_pstate_get(struct acpicpu_softc *sc, uint32_t *freq)
661 {
662 	struct acpicpu_pstate *ps = NULL;
663 	uint64_t val;
664 	uint32_t i;
665 
666 	if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
667 		return acpicpu_md_pstate_fidvid_get(sc, freq);
668 
669 	/*
670 	 * Pick any P-state for the status address.
671 	*/
672 	for (i = 0; i < sc->sc_pstate_count; i++) {
673 
674 		ps = &sc->sc_pstate[i];
675 
676 		if (__predict_true(ps->ps_freq != 0))
677 			break;
678 	}
679 
680 	if (__predict_false(ps == NULL))
681 		return ENODEV;
682 
683 	if (__predict_false(ps->ps_status_addr == 0))
684 		return EINVAL;
685 
686 	val = rdmsr(ps->ps_status_addr);
687 
688 	if (__predict_true(ps->ps_status_mask != 0))
689 		val = val & ps->ps_status_mask;
690 
691 	/*
692 	 * Search for the value from known P-states.
693 	 */
694 	for (i = 0; i < sc->sc_pstate_count; i++) {
695 
696 		ps = &sc->sc_pstate[i];
697 
698 		if (__predict_false(ps->ps_freq == 0))
699 			continue;
700 
701 		if (val == ps->ps_status) {
702 			*freq = ps->ps_freq;
703 			return 0;
704 		}
705 	}
706 
707 	/*
708 	 * If the value was not found, try APERF/MPERF.
709 	 * The state is P0 if the return value is 100 %.
710 	 */
711 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
712 
713 		if (acpicpu_md_pstate_hwf(sc->sc_ci) == 100) {
714 			*freq = sc->sc_pstate[0].ps_freq;
715 			return 0;
716 		}
717 	}
718 
719 	return EIO;
720 }
721 
722 int
723 acpicpu_md_pstate_set(struct acpicpu_pstate *ps)
724 {
725 	uint64_t val = 0;
726 
727 	if (__predict_false(ps->ps_control_addr == 0))
728 		return EINVAL;
729 
730 	if ((ps->ps_flags & ACPICPU_FLAG_P_FIDVID) != 0)
731 		return acpicpu_md_pstate_fidvid_set(ps);
732 
733 	/*
734 	 * If the mask is set, do a read-modify-write.
735 	 */
736 	if (__predict_true(ps->ps_control_mask != 0)) {
737 		val = rdmsr(ps->ps_control_addr);
738 		val &= ~ps->ps_control_mask;
739 	}
740 
741 	val |= ps->ps_control;
742 
743 	wrmsr(ps->ps_control_addr, val);
744 	DELAY(ps->ps_latency);
745 
746 	return 0;
747 }
748 
749 static int
750 acpicpu_md_pstate_fidvid_get(struct acpicpu_softc *sc, uint32_t *freq)
751 {
752 	struct acpicpu_pstate *ps;
753 	uint32_t fid, i, vid;
754 	uint32_t cfid, cvid;
755 	int rv;
756 
757 	/*
758 	 * AMD family 0Fh needs special treatment.
759 	 * While it wants to use ACPI, it does not
760 	 * comply with the ACPI specifications.
761 	 */
762 	rv = acpicpu_md_pstate_fidvid_read(&cfid, &cvid);
763 
764 	if (rv != 0)
765 		return rv;
766 
767 	for (i = 0; i < sc->sc_pstate_count; i++) {
768 
769 		ps = &sc->sc_pstate[i];
770 
771 		if (__predict_false(ps->ps_freq == 0))
772 			continue;
773 
774 		fid = __SHIFTOUT(ps->ps_status, ACPI_0FH_STATUS_FID);
775 		vid = __SHIFTOUT(ps->ps_status, ACPI_0FH_STATUS_VID);
776 
777 		if (cfid == fid && cvid == vid) {
778 			*freq = ps->ps_freq;
779 			return 0;
780 		}
781 	}
782 
783 	return EIO;
784 }
785 
786 static int
787 acpicpu_md_pstate_fidvid_set(struct acpicpu_pstate *ps)
788 {
789 	const uint64_t ctrl = ps->ps_control;
790 	uint32_t cfid, cvid, fid, i, irt;
791 	uint32_t pll, vco_cfid, vco_fid;
792 	uint32_t val, vid, vst;
793 	int rv;
794 
795 	rv = acpicpu_md_pstate_fidvid_read(&cfid, &cvid);
796 
797 	if (rv != 0)
798 		return rv;
799 
800 	fid = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_FID);
801 	vid = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_VID);
802 	irt = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_IRT);
803 	vst = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_VST);
804 	pll = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_PLL);
805 
806 	vst = vst * 20;
807 	pll = pll * 1000 / 5;
808 	irt = 10 * __BIT(irt);
809 
810 	/*
811 	 * Phase 1.
812 	 */
813 	while (cvid > vid) {
814 
815 		val = 1 << __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_MVS);
816 		val = (val > cvid) ? 0 : cvid - val;
817 
818 		acpicpu_md_pstate_fidvid_write(cfid, val, 1, vst);
819 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
820 
821 		if (rv != 0)
822 			return rv;
823 	}
824 
825 	i = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_RVO);
826 
827 	for (; i > 0 && cvid > 0; --i) {
828 
829 		acpicpu_md_pstate_fidvid_write(cfid, cvid - 1, 1, vst);
830 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
831 
832 		if (rv != 0)
833 			return rv;
834 	}
835 
836 	/*
837 	 * Phase 2.
838 	 */
839 	if (cfid != fid) {
840 
841 		vco_fid  = FID_TO_VCO_FID(fid);
842 		vco_cfid = FID_TO_VCO_FID(cfid);
843 
844 		while (abs(vco_fid - vco_cfid) > 2) {
845 
846 			if (fid <= cfid)
847 				val = cfid - 2;
848 			else {
849 				val = (cfid > 6) ? cfid + 2 :
850 				    FID_TO_VCO_FID(cfid) + 2;
851 			}
852 
853 			acpicpu_md_pstate_fidvid_write(val, cvid, pll, irt);
854 			rv = acpicpu_md_pstate_fidvid_read(&cfid, NULL);
855 
856 			if (rv != 0)
857 				return rv;
858 
859 			vco_cfid = FID_TO_VCO_FID(cfid);
860 		}
861 
862 		acpicpu_md_pstate_fidvid_write(fid, cvid, pll, irt);
863 		rv = acpicpu_md_pstate_fidvid_read(&cfid, NULL);
864 
865 		if (rv != 0)
866 			return rv;
867 	}
868 
869 	/*
870 	 * Phase 3.
871 	 */
872 	if (cvid != vid) {
873 
874 		acpicpu_md_pstate_fidvid_write(cfid, vid, 1, vst);
875 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
876 
877 		if (rv != 0)
878 			return rv;
879 	}
880 
881 	return 0;
882 }
883 
884 static int
885 acpicpu_md_pstate_fidvid_read(uint32_t *cfid, uint32_t *cvid)
886 {
887 	int i = ACPICPU_P_STATE_RETRY * 100;
888 	uint64_t val;
889 
890 	do {
891 		val = rdmsr(MSR_0FH_STATUS);
892 
893 	} while (__SHIFTOUT(val, MSR_0FH_STATUS_PENDING) != 0 && --i >= 0);
894 
895 	if (i == 0)
896 		return EAGAIN;
897 
898 	if (cfid != NULL)
899 		*cfid = __SHIFTOUT(val, MSR_0FH_STATUS_CFID);
900 
901 	if (cvid != NULL)
902 		*cvid = __SHIFTOUT(val, MSR_0FH_STATUS_CVID);
903 
904 	return 0;
905 }
906 
907 static void
908 acpicpu_md_pstate_fidvid_write(uint32_t fid,
909     uint32_t vid, uint32_t cnt, uint32_t tmo)
910 {
911 	uint64_t val = 0;
912 
913 	val |= __SHIFTIN(fid, MSR_0FH_CONTROL_FID);
914 	val |= __SHIFTIN(vid, MSR_0FH_CONTROL_VID);
915 	val |= __SHIFTIN(cnt, MSR_0FH_CONTROL_CNT);
916 	val |= __SHIFTIN(0x1, MSR_0FH_CONTROL_CHG);
917 
918 	wrmsr(MSR_0FH_CONTROL, val);
919 	DELAY(tmo);
920 }
921 
922 int
923 acpicpu_md_tstate_get(struct acpicpu_softc *sc, uint32_t *percent)
924 {
925 	struct acpicpu_tstate *ts;
926 	uint64_t val;
927 	uint32_t i;
928 
929 	val = rdmsr(MSR_THERM_CONTROL);
930 
931 	for (i = 0; i < sc->sc_tstate_count; i++) {
932 
933 		ts = &sc->sc_tstate[i];
934 
935 		if (ts->ts_percent == 0)
936 			continue;
937 
938 		if (val == ts->ts_status) {
939 			*percent = ts->ts_percent;
940 			return 0;
941 		}
942 	}
943 
944 	return EIO;
945 }
946 
947 int
948 acpicpu_md_tstate_set(struct acpicpu_tstate *ts)
949 {
950 	uint64_t val;
951 	uint8_t i;
952 
953 	val = ts->ts_control;
954 	val = val & __BITS(1, 4);
955 
956 	wrmsr(MSR_THERM_CONTROL, val);
957 
958 	if (ts->ts_status == 0) {
959 		DELAY(ts->ts_latency);
960 		return 0;
961 	}
962 
963 	for (i = val = 0; i < ACPICPU_T_STATE_RETRY; i++) {
964 
965 		val = rdmsr(MSR_THERM_CONTROL);
966 
967 		if (val == ts->ts_status)
968 			return 0;
969 
970 		DELAY(ts->ts_latency);
971 	}
972 
973 	return EAGAIN;
974 }
975 
976 /*
977  * A kludge for backwards compatibility.
978  */
979 static int
980 acpicpu_md_pstate_sysctl_init(void)
981 {
982 	const struct sysctlnode	*fnode, *mnode, *rnode;
983 	const char *str;
984 	int rv;
985 
986 	switch (cpu_vendor) {
987 
988 	case CPUVENDOR_IDT:
989 	case CPUVENDOR_INTEL:
990 		str = "est";
991 		break;
992 
993 	case CPUVENDOR_AMD:
994 		str = "powernow";
995 		break;
996 
997 	default:
998 		return ENODEV;
999 	}
1000 
1001 
1002 	rv = sysctl_createv(&acpicpu_log, 0, NULL, &rnode,
1003 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "machdep", NULL,
1004 	    NULL, 0, NULL, 0, CTL_MACHDEP, CTL_EOL);
1005 
1006 	if (rv != 0)
1007 		goto fail;
1008 
1009 	rv = sysctl_createv(&acpicpu_log, 0, &rnode, &mnode,
1010 	    0, CTLTYPE_NODE, str, NULL,
1011 	    NULL, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1012 
1013 	if (rv != 0)
1014 		goto fail;
1015 
1016 	rv = sysctl_createv(&acpicpu_log, 0, &mnode, &fnode,
1017 	    0, CTLTYPE_NODE, "frequency", NULL,
1018 	    NULL, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1019 
1020 	if (rv != 0)
1021 		goto fail;
1022 
1023 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1024 	    CTLFLAG_READWRITE, CTLTYPE_INT, "target", NULL,
1025 	    acpicpu_md_pstate_sysctl_set, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1026 
1027 	if (rv != 0)
1028 		goto fail;
1029 
1030 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1031 	    CTLFLAG_READONLY, CTLTYPE_INT, "current", NULL,
1032 	    acpicpu_md_pstate_sysctl_get, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1033 
1034 	if (rv != 0)
1035 		goto fail;
1036 
1037 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1038 	    CTLFLAG_READONLY, CTLTYPE_STRING, "available", NULL,
1039 	    acpicpu_md_pstate_sysctl_all, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1040 
1041 	if (rv != 0)
1042 		goto fail;
1043 
1044 	return 0;
1045 
1046 fail:
1047 	if (acpicpu_log != NULL) {
1048 		sysctl_teardown(&acpicpu_log);
1049 		acpicpu_log = NULL;
1050 	}
1051 
1052 	return rv;
1053 }
1054 
1055 static int
1056 acpicpu_md_pstate_sysctl_get(SYSCTLFN_ARGS)
1057 {
1058 	struct cpu_info *ci = curcpu();
1059 	struct sysctlnode node;
1060 	uint32_t freq;
1061 	int err;
1062 
1063 	err = acpicpu_pstate_get(ci, &freq);
1064 
1065 	if (err != 0)
1066 		return err;
1067 
1068 	node = *rnode;
1069 	node.sysctl_data = &freq;
1070 
1071 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1072 
1073 	if (err != 0 || newp == NULL)
1074 		return err;
1075 
1076 	return 0;
1077 }
1078 
1079 static int
1080 acpicpu_md_pstate_sysctl_set(SYSCTLFN_ARGS)
1081 {
1082 	struct cpu_info *ci = curcpu();
1083 	struct sysctlnode node;
1084 	uint32_t freq;
1085 	int err;
1086 
1087 	err = acpicpu_pstate_get(ci, &freq);
1088 
1089 	if (err != 0)
1090 		return err;
1091 
1092 	node = *rnode;
1093 	node.sysctl_data = &freq;
1094 
1095 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1096 
1097 	if (err != 0 || newp == NULL)
1098 		return err;
1099 
1100 	acpicpu_pstate_set(ci, freq);
1101 
1102 	return 0;
1103 }
1104 
1105 static int
1106 acpicpu_md_pstate_sysctl_all(SYSCTLFN_ARGS)
1107 {
1108 	struct cpu_info *ci = curcpu();
1109 	struct acpicpu_softc *sc;
1110 	struct sysctlnode node;
1111 	char buf[1024];
1112 	size_t len;
1113 	uint32_t i;
1114 	int err;
1115 
1116 	sc = acpicpu_sc[ci->ci_acpiid];
1117 
1118 	if (sc == NULL)
1119 		return ENXIO;
1120 
1121 	(void)memset(&buf, 0, sizeof(buf));
1122 
1123 	mutex_enter(&sc->sc_mtx);
1124 
1125 	for (len = 0, i = sc->sc_pstate_max; i < sc->sc_pstate_count; i++) {
1126 
1127 		if (sc->sc_pstate[i].ps_freq == 0)
1128 			continue;
1129 
1130 		len += snprintf(buf + len, sizeof(buf) - len, "%u%s",
1131 		    sc->sc_pstate[i].ps_freq,
1132 		    i < (sc->sc_pstate_count - 1) ? " " : "");
1133 	}
1134 
1135 	mutex_exit(&sc->sc_mtx);
1136 
1137 	node = *rnode;
1138 	node.sysctl_data = buf;
1139 
1140 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1141 
1142 	if (err != 0 || newp == NULL)
1143 		return err;
1144 
1145 	return 0;
1146 }
1147 
1148