xref: /netbsd-src/sys/arch/x86/acpi/acpi_cpu_md.c (revision b757af438b42b93f8c6571f026d8b8ef3eaf5fc9)
1 /* $NetBSD: acpi_cpu_md.c,v 1.71 2012/02/11 22:09:47 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 1.71 2012/02/11 22:09:47 jruoho Exp $");
31 
32 #include <sys/param.h>
33 #include <sys/bus.h>
34 #include <sys/cpufreq.h>
35 #include <sys/device.h>
36 #include <sys/kcore.h>
37 #include <sys/sysctl.h>
38 #include <sys/xcall.h>
39 
40 #include <x86/cpu.h>
41 #include <x86/cpufunc.h>
42 #include <x86/cputypes.h>
43 #include <x86/cpuvar.h>
44 #include <x86/cpu_msr.h>
45 #include <x86/machdep.h>
46 
47 #include <dev/acpi/acpica.h>
48 #include <dev/acpi/acpi_cpu.h>
49 
50 #include <dev/pci/pcivar.h>
51 #include <dev/pci/pcidevs.h>
52 
53 #include <machine/acpi_machdep.h>
54 
55 /*
56  * Intel IA32_MISC_ENABLE.
57  */
58 #define MSR_MISC_ENABLE_EST	__BIT(16)
59 #define MSR_MISC_ENABLE_TURBO	__BIT(38)
60 
61 /*
62  * AMD C1E.
63  */
64 #define MSR_CMPHALT		0xc0010055
65 
66 #define MSR_CMPHALT_SMI		__BIT(27)
67 #define MSR_CMPHALT_C1E		__BIT(28)
68 #define MSR_CMPHALT_BMSTS	__BIT(29)
69 
70 /*
71  * AMD families 10h, 11h, 12h, 14h, and 15h.
72  */
73 #define MSR_10H_LIMIT		0xc0010061
74 #define MSR_10H_CONTROL		0xc0010062
75 #define MSR_10H_STATUS		0xc0010063
76 #define MSR_10H_CONFIG		0xc0010064
77 
78 /*
79  * AMD family 0Fh.
80  */
81 #define MSR_0FH_CONTROL		0xc0010041
82 #define MSR_0FH_STATUS		0xc0010042
83 
84 #define MSR_0FH_STATUS_CFID	__BITS( 0,  5)
85 #define MSR_0FH_STATUS_CVID	__BITS(32, 36)
86 #define MSR_0FH_STATUS_PENDING	__BITS(31, 31)
87 
88 #define MSR_0FH_CONTROL_FID	__BITS( 0,  5)
89 #define MSR_0FH_CONTROL_VID	__BITS( 8, 12)
90 #define MSR_0FH_CONTROL_CHG	__BITS(16, 16)
91 #define MSR_0FH_CONTROL_CNT	__BITS(32, 51)
92 
93 #define ACPI_0FH_STATUS_FID	__BITS( 0,  5)
94 #define ACPI_0FH_STATUS_VID	__BITS( 6, 10)
95 
96 #define ACPI_0FH_CONTROL_FID	__BITS( 0,  5)
97 #define ACPI_0FH_CONTROL_VID	__BITS( 6, 10)
98 #define ACPI_0FH_CONTROL_VST	__BITS(11, 17)
99 #define ACPI_0FH_CONTROL_MVS	__BITS(18, 19)
100 #define ACPI_0FH_CONTROL_PLL	__BITS(20, 26)
101 #define ACPI_0FH_CONTROL_RVO	__BITS(28, 29)
102 #define ACPI_0FH_CONTROL_IRT	__BITS(30, 31)
103 
104 #define FID_TO_VCO_FID(fidd)	(((fid) < 8) ? (8 + ((fid) << 1)) : (fid))
105 
106 static char	  native_idle_text[16];
107 void		(*native_idle)(void) = NULL;
108 
109 static int	 acpicpu_md_quirk_piix4(const struct pci_attach_args *);
110 static void	 acpicpu_md_pstate_hwf_reset(void *, void *);
111 static int	 acpicpu_md_pstate_fidvid_get(struct acpicpu_softc *,
112                                               uint32_t *);
113 static int	 acpicpu_md_pstate_fidvid_set(struct acpicpu_pstate *);
114 static int	 acpicpu_md_pstate_fidvid_read(uint32_t *, uint32_t *);
115 static void	 acpicpu_md_pstate_fidvid_write(uint32_t, uint32_t,
116 					        uint32_t, uint32_t);
117 static int	 acpicpu_md_pstate_sysctl_init(void);
118 static int	 acpicpu_md_pstate_sysctl_get(SYSCTLFN_PROTO);
119 static int	 acpicpu_md_pstate_sysctl_set(SYSCTLFN_PROTO);
120 static int	 acpicpu_md_pstate_sysctl_all(SYSCTLFN_PROTO);
121 
122 extern struct acpicpu_softc **acpicpu_sc;
123 static struct sysctllog *acpicpu_log = NULL;
124 
125 struct cpu_info *
126 acpicpu_md_match(device_t parent, cfdata_t match, void *aux)
127 {
128 	struct cpufeature_attach_args *cfaa = aux;
129 
130 	if (strcmp(cfaa->name, "frequency") != 0)
131 		return NULL;
132 
133 	return cfaa->ci;
134 }
135 
136 struct cpu_info *
137 acpicpu_md_attach(device_t parent, device_t self, void *aux)
138 {
139 	struct cpufeature_attach_args *cfaa = aux;
140 
141 	return cfaa->ci;
142 }
143 
144 uint32_t
145 acpicpu_md_flags(void)
146 {
147 	struct cpu_info *ci = curcpu();
148 	struct pci_attach_args pa;
149 	uint32_t family, val = 0;
150 	uint32_t regs[4];
151 	uint64_t msr;
152 
153 	if (acpi_md_ncpus() == 1)
154 		val |= ACPICPU_FLAG_C_BM;
155 
156 	if ((ci->ci_feat_val[1] & CPUID2_MONITOR) != 0)
157 		val |= ACPICPU_FLAG_C_FFH;
158 
159 	/*
160 	 * By default, assume that the local APIC timer
161 	 * as well as TSC are stalled during C3 sleep.
162 	 */
163 	val |= ACPICPU_FLAG_C_APIC | ACPICPU_FLAG_C_TSC;
164 
165 	switch (cpu_vendor) {
166 
167 	case CPUVENDOR_IDT:
168 
169 		if ((ci->ci_feat_val[1] & CPUID2_EST) != 0)
170 			val |= ACPICPU_FLAG_P_FFH;
171 
172 		if ((ci->ci_feat_val[0] & CPUID_ACPI) != 0)
173 			val |= ACPICPU_FLAG_T_FFH;
174 
175 		break;
176 
177 	case CPUVENDOR_INTEL:
178 
179 		/*
180 		 * Bus master control and arbitration should be
181 		 * available on all supported Intel CPUs (to be
182 		 * sure, this is double-checked later from the
183 		 * firmware data). These flags imply that it is
184 		 * not necessary to flush caches before C3 state.
185 		 */
186 		val |= ACPICPU_FLAG_C_BM | ACPICPU_FLAG_C_ARB;
187 
188 		/*
189 		 * Check if we can use "native", MSR-based,
190 		 * access. If not, we have to resort to I/O.
191 		 */
192 		if ((ci->ci_feat_val[1] & CPUID2_EST) != 0)
193 			val |= ACPICPU_FLAG_P_FFH;
194 
195 		if ((ci->ci_feat_val[0] & CPUID_ACPI) != 0)
196 			val |= ACPICPU_FLAG_T_FFH;
197 
198 		/*
199 		 * Check whether MSR_APERF, MSR_MPERF, and Turbo
200 		 * Boost are available. Also see if we might have
201 		 * an invariant local APIC timer ("ARAT").
202 		 */
203 		if (cpuid_level >= 0x06) {
204 
205 			x86_cpuid(0x00000006, regs);
206 
207 			if ((regs[2] & CPUID_DSPM_HWF) != 0)
208 				val |= ACPICPU_FLAG_P_HWF;
209 
210 			if ((regs[0] & CPUID_DSPM_IDA) != 0)
211 				val |= ACPICPU_FLAG_P_TURBO;
212 
213 			if ((regs[0] & CPUID_DSPM_ARAT) != 0)
214 				val &= ~ACPICPU_FLAG_C_APIC;
215 		}
216 
217 		/*
218 		 * Detect whether TSC is invariant. If it is not,
219 		 * we keep the flag to note that TSC will not run
220 		 * at constant rate. Depending on the CPU, this may
221 		 * affect P- and T-state changes, but especially
222 		 * relevant are C-states; with variant TSC, states
223 		 * larger than C1 may completely stop the counter.
224 		 */
225 		x86_cpuid(0x80000000, regs);
226 
227 		if (regs[0] >= 0x80000007) {
228 
229 			x86_cpuid(0x80000007, regs);
230 
231 			if ((regs[3] & __BIT(8)) != 0)
232 				val &= ~ACPICPU_FLAG_C_TSC;
233 		}
234 
235 		break;
236 
237 	case CPUVENDOR_AMD:
238 
239 		x86_cpuid(0x80000000, regs);
240 
241 		if (regs[0] < 0x80000007)
242 			break;
243 
244 		x86_cpuid(0x80000007, regs);
245 
246 		family = CPUID2FAMILY(ci->ci_signature);
247 
248 		if (family == 0xf)
249 			family += CPUID2EXTFAMILY(ci->ci_signature);
250 
251     		switch (family) {
252 
253 		case 0x0f:
254 
255 			/*
256 			 * Evaluate support for the "FID/VID
257 			 * algorithm" also used by powernow(4).
258 			 */
259 			if ((regs[3] & CPUID_APM_FID) == 0)
260 				break;
261 
262 			if ((regs[3] & CPUID_APM_VID) == 0)
263 				break;
264 
265 			val |= ACPICPU_FLAG_P_FFH | ACPICPU_FLAG_P_FIDVID;
266 			break;
267 
268 		case 0x10:
269 		case 0x11:
270 
271 			if (rdmsr_safe(MSR_CMPHALT, &msr) != EFAULT)
272 				val |= ACPICPU_FLAG_C_C1E;
273 
274 			/* FALLTHROUGH */
275 
276 		case 0x12:
277 		case 0x14: /* AMD Fusion */
278 		case 0x15: /* AMD Bulldozer */
279 
280 			/*
281 			 * Like with Intel, detect invariant TSC,
282 			 * MSR-based P-states, and AMD's "turbo"
283 			 * (Core Performance Boost), respectively.
284 			 */
285 			if ((regs[3] & CPUID_APM_TSC) != 0)
286 				val &= ~ACPICPU_FLAG_C_TSC;
287 
288 			if ((regs[3] & CPUID_APM_HWP) != 0)
289 				val |= ACPICPU_FLAG_P_FFH;
290 
291 			if ((regs[3] & CPUID_APM_CPB) != 0)
292 				val |= ACPICPU_FLAG_P_TURBO;
293 
294 			/*
295 			 * Also check for APERF and MPERF,
296 			 * first available in the family 10h.
297 			 */
298 			if (cpuid_level >= 0x06) {
299 
300 				x86_cpuid(0x00000006, regs);
301 
302 				if ((regs[2] & CPUID_DSPM_HWF) != 0)
303 					val |= ACPICPU_FLAG_P_HWF;
304 			}
305 
306 			break;
307 		}
308 
309 		break;
310 	}
311 
312 	/*
313 	 * There are several erratums for PIIX4.
314 	 */
315 	if (pci_find_device(&pa, acpicpu_md_quirk_piix4) != 0)
316 		val |= ACPICPU_FLAG_PIIX4;
317 
318 	return val;
319 }
320 
321 static int
322 acpicpu_md_quirk_piix4(const struct pci_attach_args *pa)
323 {
324 
325 	/*
326 	 * XXX: The pci_find_device(9) function only
327 	 *	deals with attached devices. Change this
328 	 *	to use something like pci_device_foreach().
329 	 */
330 	if (PCI_VENDOR(pa->pa_id) != PCI_VENDOR_INTEL)
331 		return 0;
332 
333 	if (PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_INTEL_82371AB_ISA ||
334 	    PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_INTEL_82440MX_PMC)
335 		return 1;
336 
337 	return 0;
338 }
339 
340 void
341 acpicpu_md_quirk_c1e(void)
342 {
343 	const uint64_t c1e = MSR_CMPHALT_SMI | MSR_CMPHALT_C1E;
344 	uint64_t val;
345 
346 	val = rdmsr(MSR_CMPHALT);
347 
348 	if ((val & c1e) != 0)
349 		wrmsr(MSR_CMPHALT, val & ~c1e);
350 }
351 
352 int
353 acpicpu_md_cstate_start(struct acpicpu_softc *sc)
354 {
355 	const size_t size = sizeof(native_idle_text);
356 	struct acpicpu_cstate *cs;
357 	bool ipi = false;
358 	int i;
359 
360 	/*
361 	 * Save the cpu_idle(9) loop used by default.
362 	 */
363 	x86_cpu_idle_get(&native_idle, native_idle_text, size);
364 
365 	for (i = 0; i < ACPI_C_STATE_COUNT; i++) {
366 
367 		cs = &sc->sc_cstate[i];
368 
369 		if (cs->cs_method == ACPICPU_C_STATE_HALT) {
370 			ipi = true;
371 			break;
372 		}
373 	}
374 
375 	x86_cpu_idle_set(acpicpu_cstate_idle, "acpi", ipi);
376 
377 	return 0;
378 }
379 
380 int
381 acpicpu_md_cstate_stop(void)
382 {
383 	static char text[16];
384 	void (*func)(void);
385 	uint64_t xc;
386 	bool ipi;
387 
388 	x86_cpu_idle_get(&func, text, sizeof(text));
389 
390 	if (func == native_idle)
391 		return EALREADY;
392 
393 	ipi = (native_idle != x86_cpu_idle_halt) ? false : true;
394 	x86_cpu_idle_set(native_idle, native_idle_text, ipi);
395 
396 	/*
397 	 * Run a cross-call to ensure that all CPUs are
398 	 * out from the ACPI idle-loop before detachment.
399 	 */
400 	xc = xc_broadcast(0, (xcfunc_t)nullop, NULL, NULL);
401 	xc_wait(xc);
402 
403 	return 0;
404 }
405 
406 /*
407  * Called with interrupts enabled.
408  */
409 void
410 acpicpu_md_cstate_enter(int method, int state)
411 {
412 	struct cpu_info *ci = curcpu();
413 
414 	KASSERT(ci->ci_ilevel == IPL_NONE);
415 
416 	switch (method) {
417 
418 	case ACPICPU_C_STATE_FFH:
419 
420 		x86_monitor(&ci->ci_want_resched, 0, 0);
421 
422 		if (__predict_false(ci->ci_want_resched != 0))
423 			return;
424 
425 		x86_mwait((state - 1) << 4, 0);
426 		break;
427 
428 	case ACPICPU_C_STATE_HALT:
429 
430 		x86_disable_intr();
431 
432 		if (__predict_false(ci->ci_want_resched != 0)) {
433 			x86_enable_intr();
434 			return;
435 		}
436 
437 		x86_stihlt();
438 		break;
439 	}
440 }
441 
442 int
443 acpicpu_md_pstate_start(struct acpicpu_softc *sc)
444 {
445 	uint64_t xc, val;
446 
447 	switch (cpu_vendor) {
448 
449 	case CPUVENDOR_IDT:
450 	case CPUVENDOR_INTEL:
451 
452 		/*
453 		 * Make sure EST is enabled.
454 		 */
455 		if ((sc->sc_flags & ACPICPU_FLAG_P_FFH) != 0) {
456 
457 			val = rdmsr(MSR_MISC_ENABLE);
458 
459 			if ((val & MSR_MISC_ENABLE_EST) == 0) {
460 
461 				val |= MSR_MISC_ENABLE_EST;
462 				wrmsr(MSR_MISC_ENABLE, val);
463 				val = rdmsr(MSR_MISC_ENABLE);
464 
465 				if ((val & MSR_MISC_ENABLE_EST) == 0)
466 					return ENOTTY;
467 			}
468 		}
469 	}
470 
471 	/*
472 	 * Reset the APERF and MPERF counters.
473 	 */
474 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
475 		xc = xc_broadcast(0, acpicpu_md_pstate_hwf_reset, NULL, NULL);
476 		xc_wait(xc);
477 	}
478 
479 	return acpicpu_md_pstate_sysctl_init();
480 }
481 
482 int
483 acpicpu_md_pstate_stop(void)
484 {
485 
486 	if (acpicpu_log == NULL)
487 		return EALREADY;
488 
489 	sysctl_teardown(&acpicpu_log);
490 	acpicpu_log = NULL;
491 
492 	return 0;
493 }
494 
495 int
496 acpicpu_md_pstate_init(struct acpicpu_softc *sc)
497 {
498 	struct cpu_info *ci = sc->sc_ci;
499 	struct acpicpu_pstate *ps, msr;
500 	uint32_t family, i = 0;
501 
502 	(void)memset(&msr, 0, sizeof(struct acpicpu_pstate));
503 
504 	switch (cpu_vendor) {
505 
506 	case CPUVENDOR_IDT:
507 	case CPUVENDOR_INTEL:
508 
509 		/*
510 		 * If the so-called Turbo Boost is present,
511 		 * the P0-state is always the "turbo state".
512 		 * It is shown as the P1 frequency + 1 MHz.
513 		 *
514 		 * For discussion, see:
515 		 *
516 		 *	Intel Corporation: Intel Turbo Boost Technology
517 		 *	in Intel Core(tm) Microarchitectures (Nehalem)
518 		 *	Based Processors. White Paper, November 2008.
519 		 */
520 		if (sc->sc_pstate_count >= 2 &&
521 		   (sc->sc_flags & ACPICPU_FLAG_P_TURBO) != 0) {
522 
523 			ps = &sc->sc_pstate[0];
524 
525 			if (ps->ps_freq == sc->sc_pstate[1].ps_freq + 1)
526 				ps->ps_flags |= ACPICPU_FLAG_P_TURBO;
527 		}
528 
529 		msr.ps_control_addr = MSR_PERF_CTL;
530 		msr.ps_control_mask = __BITS(0, 15);
531 
532 		msr.ps_status_addr  = MSR_PERF_STATUS;
533 		msr.ps_status_mask  = __BITS(0, 15);
534 		break;
535 
536 	case CPUVENDOR_AMD:
537 
538 		if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
539 			msr.ps_flags |= ACPICPU_FLAG_P_FIDVID;
540 
541 		family = CPUID2FAMILY(ci->ci_signature);
542 
543 		if (family == 0xf)
544 			family += CPUID2EXTFAMILY(ci->ci_signature);
545 
546 		switch (family) {
547 
548 		case 0x0f:
549 			msr.ps_control_addr = MSR_0FH_CONTROL;
550 			msr.ps_status_addr  = MSR_0FH_STATUS;
551 			break;
552 
553 		case 0x10:
554 		case 0x11:
555 		case 0x12:
556 		case 0x14:
557 		case 0x15:
558 			msr.ps_control_addr = MSR_10H_CONTROL;
559 			msr.ps_control_mask = __BITS(0, 2);
560 
561 			msr.ps_status_addr  = MSR_10H_STATUS;
562 			msr.ps_status_mask  = __BITS(0, 2);
563 			break;
564 
565 		default:
566 			/*
567 			 * If we have an unknown AMD CPU, rely on XPSS.
568 			 */
569 			if ((sc->sc_flags & ACPICPU_FLAG_P_XPSS) == 0)
570 				return EOPNOTSUPP;
571 		}
572 
573 		break;
574 
575 	default:
576 		return ENODEV;
577 	}
578 
579 	/*
580 	 * Fill the P-state structures with MSR addresses that are
581 	 * known to be correct. If we do not know the addresses,
582 	 * leave the values intact. If a vendor uses XPSS, we do
583 	 * not necessarily need to do anything to support new CPUs.
584 	 */
585 	while (i < sc->sc_pstate_count) {
586 
587 		ps = &sc->sc_pstate[i];
588 
589 		if (msr.ps_flags != 0)
590 			ps->ps_flags |= msr.ps_flags;
591 
592 		if (msr.ps_status_addr != 0)
593 			ps->ps_status_addr = msr.ps_status_addr;
594 
595 		if (msr.ps_status_mask != 0)
596 			ps->ps_status_mask = msr.ps_status_mask;
597 
598 		if (msr.ps_control_addr != 0)
599 			ps->ps_control_addr = msr.ps_control_addr;
600 
601 		if (msr.ps_control_mask != 0)
602 			ps->ps_control_mask = msr.ps_control_mask;
603 
604 		i++;
605 	}
606 
607 	return 0;
608 }
609 
610 /*
611  * Read the IA32_APERF and IA32_MPERF counters. The first
612  * increments at the rate of the fixed maximum frequency
613  * configured during the boot, whereas APERF counts at the
614  * rate of the actual frequency. Note that the MSRs must be
615  * read without delay, and that only the ratio between
616  * IA32_APERF and IA32_MPERF is architecturally defined.
617  *
618  * The function thus returns the percentage of the actual
619  * frequency in terms of the maximum frequency of the calling
620  * CPU since the last call. A value zero implies an error.
621  *
622  * For further details, refer to:
623  *
624  *	Intel Corporation: Intel 64 and IA-32 Architectures
625  *	Software Developer's Manual. Section 13.2, Volume 3A:
626  *	System Programming Guide, Part 1. July, 2008.
627  *
628  *	Advanced Micro Devices: BIOS and Kernel Developer's
629  *	Guide (BKDG) for AMD Family 10h Processors. Section
630  *	2.4.5, Revision 3.48, April 2010.
631  */
632 uint8_t
633 acpicpu_md_pstate_hwf(struct cpu_info *ci)
634 {
635 	struct acpicpu_softc *sc;
636 	uint64_t aperf, mperf;
637 	uint8_t rv = 0;
638 
639 	sc = acpicpu_sc[ci->ci_acpiid];
640 
641 	if (__predict_false(sc == NULL))
642 		return 0;
643 
644 	if (__predict_false((sc->sc_flags & ACPICPU_FLAG_P_HWF) == 0))
645 		return 0;
646 
647 	aperf = sc->sc_pstate_aperf;
648 	mperf = sc->sc_pstate_mperf;
649 
650 	x86_disable_intr();
651 
652 	sc->sc_pstate_aperf = rdmsr(MSR_APERF);
653 	sc->sc_pstate_mperf = rdmsr(MSR_MPERF);
654 
655 	x86_enable_intr();
656 
657 	aperf = sc->sc_pstate_aperf - aperf;
658 	mperf = sc->sc_pstate_mperf - mperf;
659 
660 	if (__predict_true(mperf != 0))
661 		rv = (aperf * 100) / mperf;
662 
663 	return rv;
664 }
665 
666 static void
667 acpicpu_md_pstate_hwf_reset(void *arg1, void *arg2)
668 {
669 	struct cpu_info *ci = curcpu();
670 	struct acpicpu_softc *sc;
671 
672 	sc = acpicpu_sc[ci->ci_acpiid];
673 
674 	if (__predict_false(sc == NULL))
675 		return;
676 
677 	x86_disable_intr();
678 
679 	wrmsr(MSR_APERF, 0);
680 	wrmsr(MSR_MPERF, 0);
681 
682 	x86_enable_intr();
683 
684 	sc->sc_pstate_aperf = 0;
685 	sc->sc_pstate_mperf = 0;
686 }
687 
688 int
689 acpicpu_md_pstate_get(struct acpicpu_softc *sc, uint32_t *freq)
690 {
691 	struct acpicpu_pstate *ps = NULL;
692 	uint64_t val;
693 	uint32_t i;
694 
695 	if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
696 		return acpicpu_md_pstate_fidvid_get(sc, freq);
697 
698 	/*
699 	 * Pick any P-state for the status address.
700 	 */
701 	for (i = 0; i < sc->sc_pstate_count; i++) {
702 
703 		ps = &sc->sc_pstate[i];
704 
705 		if (__predict_true(ps->ps_freq != 0))
706 			break;
707 	}
708 
709 	if (__predict_false(ps == NULL))
710 		return ENODEV;
711 
712 	if (__predict_false(ps->ps_status_addr == 0))
713 		return EINVAL;
714 
715 	val = rdmsr(ps->ps_status_addr);
716 
717 	if (__predict_true(ps->ps_status_mask != 0))
718 		val = val & ps->ps_status_mask;
719 
720 	/*
721 	 * Search for the value from known P-states.
722 	 */
723 	for (i = 0; i < sc->sc_pstate_count; i++) {
724 
725 		ps = &sc->sc_pstate[i];
726 
727 		if (__predict_false(ps->ps_freq == 0))
728 			continue;
729 
730 		if (val == ps->ps_status) {
731 			*freq = ps->ps_freq;
732 			return 0;
733 		}
734 	}
735 
736 	/*
737 	 * If the value was not found, try APERF/MPERF.
738 	 * The state is P0 if the return value is 100 %.
739 	 */
740 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
741 
742 		KASSERT(sc->sc_pstate_count > 0);
743 		KASSERT(sc->sc_pstate[0].ps_freq != 0);
744 
745 		if (acpicpu_md_pstate_hwf(sc->sc_ci) == 100) {
746 			*freq = sc->sc_pstate[0].ps_freq;
747 			return 0;
748 		}
749 	}
750 
751 	return EIO;
752 }
753 
754 int
755 acpicpu_md_pstate_set(struct acpicpu_pstate *ps)
756 {
757 	uint64_t val = 0;
758 
759 	if (__predict_false(ps->ps_control_addr == 0))
760 		return EINVAL;
761 
762 	if ((ps->ps_flags & ACPICPU_FLAG_P_FIDVID) != 0)
763 		return acpicpu_md_pstate_fidvid_set(ps);
764 
765 	/*
766 	 * If the mask is set, do a read-modify-write.
767 	 */
768 	if (__predict_true(ps->ps_control_mask != 0)) {
769 		val = rdmsr(ps->ps_control_addr);
770 		val &= ~ps->ps_control_mask;
771 	}
772 
773 	val |= ps->ps_control;
774 
775 	wrmsr(ps->ps_control_addr, val);
776 	DELAY(ps->ps_latency);
777 
778 	return 0;
779 }
780 
781 static int
782 acpicpu_md_pstate_fidvid_get(struct acpicpu_softc *sc, uint32_t *freq)
783 {
784 	struct acpicpu_pstate *ps;
785 	uint32_t fid, i, vid;
786 	uint32_t cfid, cvid;
787 	int rv;
788 
789 	/*
790 	 * AMD family 0Fh needs special treatment.
791 	 * While it wants to use ACPI, it does not
792 	 * comply with the ACPI specifications.
793 	 */
794 	rv = acpicpu_md_pstate_fidvid_read(&cfid, &cvid);
795 
796 	if (rv != 0)
797 		return rv;
798 
799 	for (i = 0; i < sc->sc_pstate_count; i++) {
800 
801 		ps = &sc->sc_pstate[i];
802 
803 		if (__predict_false(ps->ps_freq == 0))
804 			continue;
805 
806 		fid = __SHIFTOUT(ps->ps_status, ACPI_0FH_STATUS_FID);
807 		vid = __SHIFTOUT(ps->ps_status, ACPI_0FH_STATUS_VID);
808 
809 		if (cfid == fid && cvid == vid) {
810 			*freq = ps->ps_freq;
811 			return 0;
812 		}
813 	}
814 
815 	return EIO;
816 }
817 
818 static int
819 acpicpu_md_pstate_fidvid_set(struct acpicpu_pstate *ps)
820 {
821 	const uint64_t ctrl = ps->ps_control;
822 	uint32_t cfid, cvid, fid, i, irt;
823 	uint32_t pll, vco_cfid, vco_fid;
824 	uint32_t val, vid, vst;
825 	int rv;
826 
827 	rv = acpicpu_md_pstate_fidvid_read(&cfid, &cvid);
828 
829 	if (rv != 0)
830 		return rv;
831 
832 	fid = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_FID);
833 	vid = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_VID);
834 	irt = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_IRT);
835 	vst = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_VST);
836 	pll = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_PLL);
837 
838 	vst = vst * 20;
839 	pll = pll * 1000 / 5;
840 	irt = 10 * __BIT(irt);
841 
842 	/*
843 	 * Phase 1.
844 	 */
845 	while (cvid > vid) {
846 
847 		val = 1 << __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_MVS);
848 		val = (val > cvid) ? 0 : cvid - val;
849 
850 		acpicpu_md_pstate_fidvid_write(cfid, val, 1, vst);
851 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
852 
853 		if (rv != 0)
854 			return rv;
855 	}
856 
857 	i = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_RVO);
858 
859 	for (; i > 0 && cvid > 0; --i) {
860 
861 		acpicpu_md_pstate_fidvid_write(cfid, cvid - 1, 1, vst);
862 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
863 
864 		if (rv != 0)
865 			return rv;
866 	}
867 
868 	/*
869 	 * Phase 2.
870 	 */
871 	if (cfid != fid) {
872 
873 		vco_fid  = FID_TO_VCO_FID(fid);
874 		vco_cfid = FID_TO_VCO_FID(cfid);
875 
876 		while (abs(vco_fid - vco_cfid) > 2) {
877 
878 			if (fid <= cfid)
879 				val = cfid - 2;
880 			else {
881 				val = (cfid > 6) ? cfid + 2 :
882 				    FID_TO_VCO_FID(cfid) + 2;
883 			}
884 
885 			acpicpu_md_pstate_fidvid_write(val, cvid, pll, irt);
886 			rv = acpicpu_md_pstate_fidvid_read(&cfid, NULL);
887 
888 			if (rv != 0)
889 				return rv;
890 
891 			vco_cfid = FID_TO_VCO_FID(cfid);
892 		}
893 
894 		acpicpu_md_pstate_fidvid_write(fid, cvid, pll, irt);
895 		rv = acpicpu_md_pstate_fidvid_read(&cfid, NULL);
896 
897 		if (rv != 0)
898 			return rv;
899 	}
900 
901 	/*
902 	 * Phase 3.
903 	 */
904 	if (cvid != vid) {
905 
906 		acpicpu_md_pstate_fidvid_write(cfid, vid, 1, vst);
907 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
908 
909 		if (rv != 0)
910 			return rv;
911 	}
912 
913 	return 0;
914 }
915 
916 static int
917 acpicpu_md_pstate_fidvid_read(uint32_t *cfid, uint32_t *cvid)
918 {
919 	int i = ACPICPU_P_STATE_RETRY * 100;
920 	uint64_t val;
921 
922 	do {
923 		val = rdmsr(MSR_0FH_STATUS);
924 
925 	} while (__SHIFTOUT(val, MSR_0FH_STATUS_PENDING) != 0 && --i >= 0);
926 
927 	if (i == 0)
928 		return EAGAIN;
929 
930 	if (cfid != NULL)
931 		*cfid = __SHIFTOUT(val, MSR_0FH_STATUS_CFID);
932 
933 	if (cvid != NULL)
934 		*cvid = __SHIFTOUT(val, MSR_0FH_STATUS_CVID);
935 
936 	return 0;
937 }
938 
939 static void
940 acpicpu_md_pstate_fidvid_write(uint32_t fid,
941     uint32_t vid, uint32_t cnt, uint32_t tmo)
942 {
943 	uint64_t val = 0;
944 
945 	val |= __SHIFTIN(fid, MSR_0FH_CONTROL_FID);
946 	val |= __SHIFTIN(vid, MSR_0FH_CONTROL_VID);
947 	val |= __SHIFTIN(cnt, MSR_0FH_CONTROL_CNT);
948 	val |= __SHIFTIN(0x1, MSR_0FH_CONTROL_CHG);
949 
950 	wrmsr(MSR_0FH_CONTROL, val);
951 	DELAY(tmo);
952 }
953 
954 int
955 acpicpu_md_tstate_get(struct acpicpu_softc *sc, uint32_t *percent)
956 {
957 	struct acpicpu_tstate *ts;
958 	uint64_t val;
959 	uint32_t i;
960 
961 	val = rdmsr(MSR_THERM_CONTROL);
962 
963 	for (i = 0; i < sc->sc_tstate_count; i++) {
964 
965 		ts = &sc->sc_tstate[i];
966 
967 		if (ts->ts_percent == 0)
968 			continue;
969 
970 		if (val == ts->ts_status) {
971 			*percent = ts->ts_percent;
972 			return 0;
973 		}
974 	}
975 
976 	return EIO;
977 }
978 
979 int
980 acpicpu_md_tstate_set(struct acpicpu_tstate *ts)
981 {
982 	uint64_t val;
983 	uint8_t i;
984 
985 	val = ts->ts_control;
986 	val = val & __BITS(1, 4);
987 
988 	wrmsr(MSR_THERM_CONTROL, val);
989 
990 	if (ts->ts_status == 0) {
991 		DELAY(ts->ts_latency);
992 		return 0;
993 	}
994 
995 	for (i = val = 0; i < ACPICPU_T_STATE_RETRY; i++) {
996 
997 		val = rdmsr(MSR_THERM_CONTROL);
998 
999 		if (val == ts->ts_status)
1000 			return 0;
1001 
1002 		DELAY(ts->ts_latency);
1003 	}
1004 
1005 	return EAGAIN;
1006 }
1007 
1008 /*
1009  * A kludge for backwards compatibility.
1010  */
1011 static int
1012 acpicpu_md_pstate_sysctl_init(void)
1013 {
1014 	const struct sysctlnode	*fnode, *mnode, *rnode;
1015 	const char *str;
1016 	int rv;
1017 
1018 	switch (cpu_vendor) {
1019 
1020 	case CPUVENDOR_IDT:
1021 	case CPUVENDOR_INTEL:
1022 		str = "est";
1023 		break;
1024 
1025 	case CPUVENDOR_AMD:
1026 		str = "powernow";
1027 		break;
1028 
1029 	default:
1030 		return ENODEV;
1031 	}
1032 
1033 
1034 	rv = sysctl_createv(&acpicpu_log, 0, NULL, &rnode,
1035 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "machdep", NULL,
1036 	    NULL, 0, NULL, 0, CTL_MACHDEP, CTL_EOL);
1037 
1038 	if (rv != 0)
1039 		goto fail;
1040 
1041 	rv = sysctl_createv(&acpicpu_log, 0, &rnode, &mnode,
1042 	    0, CTLTYPE_NODE, str, NULL,
1043 	    NULL, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1044 
1045 	if (rv != 0)
1046 		goto fail;
1047 
1048 	rv = sysctl_createv(&acpicpu_log, 0, &mnode, &fnode,
1049 	    0, CTLTYPE_NODE, "frequency", NULL,
1050 	    NULL, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1051 
1052 	if (rv != 0)
1053 		goto fail;
1054 
1055 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1056 	    CTLFLAG_READWRITE, CTLTYPE_INT, "target", NULL,
1057 	    acpicpu_md_pstate_sysctl_set, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1058 
1059 	if (rv != 0)
1060 		goto fail;
1061 
1062 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1063 	    CTLFLAG_READONLY, CTLTYPE_INT, "current", NULL,
1064 	    acpicpu_md_pstate_sysctl_get, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1065 
1066 	if (rv != 0)
1067 		goto fail;
1068 
1069 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1070 	    CTLFLAG_READONLY, CTLTYPE_STRING, "available", NULL,
1071 	    acpicpu_md_pstate_sysctl_all, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1072 
1073 	if (rv != 0)
1074 		goto fail;
1075 
1076 	return 0;
1077 
1078 fail:
1079 	if (acpicpu_log != NULL) {
1080 		sysctl_teardown(&acpicpu_log);
1081 		acpicpu_log = NULL;
1082 	}
1083 
1084 	return rv;
1085 }
1086 
1087 static int
1088 acpicpu_md_pstate_sysctl_get(SYSCTLFN_ARGS)
1089 {
1090 	struct sysctlnode node;
1091 	uint32_t freq;
1092 	int err;
1093 
1094 	freq = cpufreq_get(curcpu());
1095 
1096 	if (freq == 0)
1097 		return ENXIO;
1098 
1099 	node = *rnode;
1100 	node.sysctl_data = &freq;
1101 
1102 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1103 
1104 	if (err != 0 || newp == NULL)
1105 		return err;
1106 
1107 	return 0;
1108 }
1109 
1110 static int
1111 acpicpu_md_pstate_sysctl_set(SYSCTLFN_ARGS)
1112 {
1113 	struct sysctlnode node;
1114 	uint32_t freq;
1115 	int err;
1116 
1117 	freq = cpufreq_get(curcpu());
1118 
1119 	if (freq == 0)
1120 		return ENXIO;
1121 
1122 	node = *rnode;
1123 	node.sysctl_data = &freq;
1124 
1125 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1126 
1127 	if (err != 0 || newp == NULL)
1128 		return err;
1129 
1130 	cpufreq_set_all(freq);
1131 
1132 	return 0;
1133 }
1134 
1135 static int
1136 acpicpu_md_pstate_sysctl_all(SYSCTLFN_ARGS)
1137 {
1138 	struct cpu_info *ci = curcpu();
1139 	struct acpicpu_softc *sc;
1140 	struct sysctlnode node;
1141 	char buf[1024];
1142 	size_t len;
1143 	uint32_t i;
1144 	int err;
1145 
1146 	sc = acpicpu_sc[ci->ci_acpiid];
1147 
1148 	if (sc == NULL)
1149 		return ENXIO;
1150 
1151 	(void)memset(&buf, 0, sizeof(buf));
1152 
1153 	mutex_enter(&sc->sc_mtx);
1154 
1155 	for (len = 0, i = sc->sc_pstate_max; i < sc->sc_pstate_count; i++) {
1156 
1157 		if (sc->sc_pstate[i].ps_freq == 0)
1158 			continue;
1159 
1160 		len += snprintf(buf + len, sizeof(buf) - len, "%u%s",
1161 		    sc->sc_pstate[i].ps_freq,
1162 		    i < (sc->sc_pstate_count - 1) ? " " : "");
1163 	}
1164 
1165 	mutex_exit(&sc->sc_mtx);
1166 
1167 	node = *rnode;
1168 	node.sysctl_data = buf;
1169 
1170 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1171 
1172 	if (err != 0 || newp == NULL)
1173 		return err;
1174 
1175 	return 0;
1176 }
1177 
1178