xref: /netbsd-src/sys/arch/x86/acpi/acpi_cpu_md.c (revision 33881f779a77dce6440bdc44610d94de75bebefe)
1 /* $NetBSD: acpi_cpu_md.c,v 1.82 2020/03/14 13:50:46 ad 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 1.82 2020/03/14 13:50:46 ad 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/machdep.h>
45 #include <x86/x86/tsc.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 #ifdef ACPICPU_ENABLE_C3
107 static char	  native_idle_text[16];
108 void		(*native_idle)(void) = NULL;
109 #endif
110 
111 static int	 acpicpu_md_quirk_piix4(const struct pci_attach_args *);
112 static void	 acpicpu_md_pstate_hwf_reset(void *, void *);
113 static int	 acpicpu_md_pstate_fidvid_get(struct acpicpu_softc *,
114                                               uint32_t *);
115 static int	 acpicpu_md_pstate_fidvid_set(struct acpicpu_pstate *);
116 static int	 acpicpu_md_pstate_fidvid_read(uint32_t *, uint32_t *);
117 static void	 acpicpu_md_pstate_fidvid_write(uint32_t, uint32_t,
118 					        uint32_t, uint32_t);
119 static int	 acpicpu_md_pstate_sysctl_init(void);
120 static int	 acpicpu_md_pstate_sysctl_get(SYSCTLFN_PROTO);
121 static int	 acpicpu_md_pstate_sysctl_set(SYSCTLFN_PROTO);
122 static int	 acpicpu_md_pstate_sysctl_all(SYSCTLFN_PROTO);
123 
124 extern struct acpicpu_softc **acpicpu_sc;
125 static struct sysctllog *acpicpu_log = NULL;
126 
127 struct cpu_info *
128 acpicpu_md_match(device_t parent, cfdata_t match, void *aux)
129 {
130 	struct cpufeature_attach_args *cfaa = aux;
131 
132 	if (strcmp(cfaa->name, "frequency") != 0)
133 		return NULL;
134 
135 	return cfaa->ci;
136 }
137 
138 struct cpu_info *
139 acpicpu_md_attach(device_t parent, device_t self, void *aux)
140 {
141 	struct cpufeature_attach_args *cfaa = aux;
142 
143 	return cfaa->ci;
144 }
145 
146 uint32_t
147 acpicpu_md_flags(void)
148 {
149 	struct cpu_info *ci = curcpu();
150 	struct pci_attach_args pa;
151 	uint32_t family, val = 0;
152 	uint32_t regs[4];
153 	uint64_t msr;
154 
155 	if (acpi_md_ncpus() == 1)
156 		val |= ACPICPU_FLAG_C_BM;
157 
158 	if ((ci->ci_feat_val[1] & CPUID2_MONITOR) != 0)
159 		val |= ACPICPU_FLAG_C_FFH;
160 
161 	/*
162 	 * By default, assume that the local APIC timer
163 	 * as well as TSC are stalled during C3 sleep.
164 	 */
165 	val |= ACPICPU_FLAG_C_APIC | ACPICPU_FLAG_C_TSC;
166 
167 	/*
168 	 * Detect whether TSC is invariant. If it is not, we keep the flag to
169 	 * note that TSC will not run at constant rate. Depending on the CPU,
170 	 * this may affect P- and T-state changes, but especially relevant
171 	 * are C-states; with variant TSC, states larger than C1 may
172 	 * completely stop the counter.
173 	 */
174 	if (tsc_is_invariant())
175 		val &= ~ACPICPU_FLAG_C_TSC;
176 
177 	switch (cpu_vendor) {
178 
179 	case CPUVENDOR_IDT:
180 
181 		if ((ci->ci_feat_val[1] & CPUID2_EST) != 0)
182 			val |= ACPICPU_FLAG_P_FFH;
183 
184 		if ((ci->ci_feat_val[0] & CPUID_ACPI) != 0)
185 			val |= ACPICPU_FLAG_T_FFH;
186 
187 		break;
188 
189 	case CPUVENDOR_INTEL:
190 
191 		/*
192 		 * Bus master control and arbitration should be
193 		 * available on all supported Intel CPUs (to be
194 		 * sure, this is double-checked later from the
195 		 * firmware data). These flags imply that it is
196 		 * not necessary to flush caches before C3 state.
197 		 */
198 		val |= ACPICPU_FLAG_C_BM | ACPICPU_FLAG_C_ARB;
199 
200 		/*
201 		 * Check if we can use "native", MSR-based,
202 		 * access. If not, we have to resort to I/O.
203 		 */
204 		if ((ci->ci_feat_val[1] & CPUID2_EST) != 0)
205 			val |= ACPICPU_FLAG_P_FFH;
206 
207 		if ((ci->ci_feat_val[0] & CPUID_ACPI) != 0)
208 			val |= ACPICPU_FLAG_T_FFH;
209 
210 		/*
211 		 * Check whether MSR_APERF, MSR_MPERF, and Turbo
212 		 * Boost are available. Also see if we might have
213 		 * an invariant local APIC timer ("ARAT").
214 		 */
215 		if (cpuid_level >= 0x06) {
216 
217 			x86_cpuid(0x00000006, regs);
218 
219 			if ((regs[2] & CPUID_DSPM_HWF) != 0)
220 				val |= ACPICPU_FLAG_P_HWF;
221 
222 			if ((regs[0] & CPUID_DSPM_IDA) != 0)
223 				val |= ACPICPU_FLAG_P_TURBO;
224 
225 			if ((regs[0] & CPUID_DSPM_ARAT) != 0)
226 				val &= ~ACPICPU_FLAG_C_APIC;
227 
228 		}
229 
230 		break;
231 
232 	case CPUVENDOR_AMD:
233 
234 		x86_cpuid(0x80000000, regs);
235 
236 		if (regs[0] < 0x80000007)
237 			break;
238 
239 		x86_cpuid(0x80000007, regs);
240 
241 		family = CPUID_TO_FAMILY(ci->ci_signature);
242 
243     		switch (family) {
244 
245 		case 0x0f:
246 
247 			/*
248 			 * Disable C1E if present.
249 			 */
250 			if (rdmsr_safe(MSR_CMPHALT, &msr) != EFAULT)
251 				val |= ACPICPU_FLAG_C_C1E;
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 
269 			/*
270 			 * Disable C1E if present.
271 			 */
272 			if (rdmsr_safe(MSR_CMPHALT, &msr) != EFAULT)
273 				val |= ACPICPU_FLAG_C_C1E;
274 
275 			/* FALLTHROUGH */
276 
277 		case 0x12:
278 		case 0x14: /* AMD Fusion */
279 		case 0x15: /* AMD Bulldozer */
280 
281 			/*
282 			 * Like with Intel, detect MSR-based P-states,
283 			 * and AMD's "turbo" (Core Performance Boost),
284 			 * respectively.
285 			 */
286 			if ((regs[3] & CPUID_APM_HWP) != 0)
287 				val |= ACPICPU_FLAG_P_FFH;
288 
289 			if ((regs[3] & CPUID_APM_CPB) != 0)
290 				val |= ACPICPU_FLAG_P_TURBO;
291 
292 			/*
293 			 * Also check for APERF and MPERF,
294 			 * first available in the family 10h.
295 			 */
296 			if (cpuid_level >= 0x06) {
297 
298 				x86_cpuid(0x00000006, regs);
299 
300 				if ((regs[2] & CPUID_DSPM_HWF) != 0)
301 					val |= ACPICPU_FLAG_P_HWF;
302 			}
303 
304 			break;
305 		}
306 
307 		break;
308 	}
309 
310 	/*
311 	 * There are several erratums for PIIX4.
312 	 */
313 	if (pci_find_device(&pa, acpicpu_md_quirk_piix4) != 0)
314 		val |= ACPICPU_FLAG_PIIX4;
315 
316 	return val;
317 }
318 
319 static int
320 acpicpu_md_quirk_piix4(const struct pci_attach_args *pa)
321 {
322 
323 	/*
324 	 * XXX: The pci_find_device(9) function only
325 	 *	deals with attached devices. Change this
326 	 *	to use something like pci_device_foreach().
327 	 */
328 	if (PCI_VENDOR(pa->pa_id) != PCI_VENDOR_INTEL)
329 		return 0;
330 
331 	if (PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_INTEL_82371AB_ISA ||
332 	    PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_INTEL_82440MX_PMC)
333 		return 1;
334 
335 	return 0;
336 }
337 
338 void
339 acpicpu_md_quirk_c1e(void)
340 {
341 	const uint64_t c1e = MSR_CMPHALT_SMI | MSR_CMPHALT_C1E;
342 	uint64_t val;
343 
344 	val = rdmsr(MSR_CMPHALT);
345 
346 	if ((val & c1e) != 0)
347 		wrmsr(MSR_CMPHALT, val & ~c1e);
348 }
349 
350 int
351 acpicpu_md_cstate_start(struct acpicpu_softc *sc)
352 {
353 #ifdef ACPICPU_ENABLE_C3
354 	/*
355 	 * XXX There are performance problems with the ACPI idle loop, and
356 	 * it does not enter deep sleep.  Once those are resolved it'll be
357 	 * re-enabled.
358 	 */
359 	const size_t size = sizeof(native_idle_text);
360 	struct acpicpu_cstate *cs;
361 	bool ipi = false;
362 	int i;
363 
364 	/*
365 	 * Save the cpu_idle(9) loop used by default.
366 	 */
367 	x86_cpu_idle_get(&native_idle, native_idle_text, size);
368 
369 	for (i = 0; i < ACPI_C_STATE_COUNT; i++) {
370 
371 		cs = &sc->sc_cstate[i];
372 
373 		if (cs->cs_method == ACPICPU_C_STATE_HALT) {
374 			ipi = true;
375 			break;
376 		}
377 	}
378 
379 	x86_cpu_idle_set(acpicpu_cstate_idle, "acpi", ipi);
380 #endif	/* ACPICPU_ENABLE_C3 */
381 
382 	return 0;
383 }
384 
385 int
386 acpicpu_md_cstate_stop(void)
387 {
388 #ifdef ACPICPU_ENABLE_C3
389 	/*
390 	 * XXX There are performance problems with the ACPI idle loop, and
391 	 * it does not enter deep sleep.  Once those are resolved it'll be
392 	 * re-enabled.
393 	 */
394 	static char text[16];
395 	void (*func)(void);
396 	bool ipi;
397 
398 	x86_cpu_idle_get(&func, text, sizeof(text));
399 
400 	if (func == native_idle)
401 		return EALREADY;
402 
403 	ipi = (native_idle != x86_cpu_idle_halt) ? false : true;
404 	x86_cpu_idle_set(native_idle, native_idle_text, ipi);
405 
406 	/*
407 	 * Run a cross-call to ensure that all CPUs are
408 	 * out from the ACPI idle-loop before detachment.
409 	 */
410 	xc_barrier(0);
411 #endif	/* ACPICPU_ENABLE_C3 */
412 
413 	return 0;
414 }
415 
416 /*
417  * Called with interrupts enabled.
418  */
419 void __nocsan
420 acpicpu_md_cstate_enter(int method, int state)
421 {
422 	struct cpu_info *ci = curcpu();
423 
424 	KASSERT(ci->ci_ilevel == IPL_NONE);
425 
426 	switch (method) {
427 
428 	case ACPICPU_C_STATE_FFH:
429 
430 		x86_monitor(&ci->ci_want_resched, 0, 0);
431 
432 		if (__predict_false(ci->ci_want_resched != 0))
433 			return;
434 
435 		x86_mwait((state - 1) << 4, 0);
436 		break;
437 
438 	case ACPICPU_C_STATE_HALT:
439 
440 		x86_disable_intr();
441 
442 		if (__predict_false(ci->ci_want_resched != 0)) {
443 			x86_enable_intr();
444 			return;
445 		}
446 
447 		x86_stihlt();
448 		break;
449 	}
450 }
451 
452 int
453 acpicpu_md_pstate_start(struct acpicpu_softc *sc)
454 {
455 	uint64_t xc, val;
456 
457 	switch (cpu_vendor) {
458 
459 	case CPUVENDOR_IDT:
460 	case CPUVENDOR_INTEL:
461 
462 		/*
463 		 * Make sure EST is enabled.
464 		 */
465 		if ((sc->sc_flags & ACPICPU_FLAG_P_FFH) != 0) {
466 
467 			val = rdmsr(MSR_MISC_ENABLE);
468 
469 			if ((val & MSR_MISC_ENABLE_EST) == 0) {
470 
471 				val |= MSR_MISC_ENABLE_EST;
472 				wrmsr(MSR_MISC_ENABLE, val);
473 				val = rdmsr(MSR_MISC_ENABLE);
474 
475 				if ((val & MSR_MISC_ENABLE_EST) == 0)
476 					return ENOTTY;
477 			}
478 		}
479 	}
480 
481 	/*
482 	 * Reset the APERF and MPERF counters.
483 	 */
484 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
485 		xc = xc_broadcast(0, acpicpu_md_pstate_hwf_reset, NULL, NULL);
486 		xc_wait(xc);
487 	}
488 
489 	return acpicpu_md_pstate_sysctl_init();
490 }
491 
492 int
493 acpicpu_md_pstate_stop(void)
494 {
495 
496 	if (acpicpu_log == NULL)
497 		return EALREADY;
498 
499 	sysctl_teardown(&acpicpu_log);
500 	acpicpu_log = NULL;
501 
502 	return 0;
503 }
504 
505 int
506 acpicpu_md_pstate_init(struct acpicpu_softc *sc)
507 {
508 	struct cpu_info *ci = sc->sc_ci;
509 	struct acpicpu_pstate *ps, msr;
510 	uint32_t family, i = 0;
511 
512 	(void)memset(&msr, 0, sizeof(struct acpicpu_pstate));
513 
514 	switch (cpu_vendor) {
515 
516 	case CPUVENDOR_IDT:
517 	case CPUVENDOR_INTEL:
518 
519 		/*
520 		 * If the so-called Turbo Boost is present,
521 		 * the P0-state is always the "turbo state".
522 		 * It is shown as the P1 frequency + 1 MHz.
523 		 *
524 		 * For discussion, see:
525 		 *
526 		 *	Intel Corporation: Intel Turbo Boost Technology
527 		 *	in Intel Core(tm) Microarchitectures (Nehalem)
528 		 *	Based Processors. White Paper, November 2008.
529 		 */
530 		if (sc->sc_pstate_count >= 2 &&
531 		   (sc->sc_flags & ACPICPU_FLAG_P_TURBO) != 0) {
532 
533 			ps = &sc->sc_pstate[0];
534 
535 			if (ps->ps_freq == sc->sc_pstate[1].ps_freq + 1)
536 				ps->ps_flags |= ACPICPU_FLAG_P_TURBO;
537 		}
538 
539 		msr.ps_control_addr = MSR_PERF_CTL;
540 		msr.ps_control_mask = __BITS(0, 15);
541 
542 		msr.ps_status_addr  = MSR_PERF_STATUS;
543 		msr.ps_status_mask  = __BITS(0, 15);
544 		break;
545 
546 	case CPUVENDOR_AMD:
547 
548 		if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
549 			msr.ps_flags |= ACPICPU_FLAG_P_FIDVID;
550 
551 		family = CPUID_TO_FAMILY(ci->ci_signature);
552 
553 		switch (family) {
554 
555 		case 0x0f:
556 			msr.ps_control_addr = MSR_0FH_CONTROL;
557 			msr.ps_status_addr  = MSR_0FH_STATUS;
558 			break;
559 
560 		case 0x10:
561 		case 0x11:
562 		case 0x12:
563 		case 0x14:
564 		case 0x15:
565 			msr.ps_control_addr = MSR_10H_CONTROL;
566 			msr.ps_control_mask = __BITS(0, 2);
567 
568 			msr.ps_status_addr  = MSR_10H_STATUS;
569 			msr.ps_status_mask  = __BITS(0, 2);
570 			break;
571 
572 		default:
573 			/*
574 			 * If we have an unknown AMD CPU, rely on XPSS.
575 			 */
576 			if ((sc->sc_flags & ACPICPU_FLAG_P_XPSS) == 0)
577 				return EOPNOTSUPP;
578 		}
579 
580 		break;
581 
582 	default:
583 		return ENODEV;
584 	}
585 
586 	/*
587 	 * Fill the P-state structures with MSR addresses that are
588 	 * known to be correct. If we do not know the addresses,
589 	 * leave the values intact. If a vendor uses XPSS, we do
590 	 * not necessarily need to do anything to support new CPUs.
591 	 */
592 	while (i < sc->sc_pstate_count) {
593 
594 		ps = &sc->sc_pstate[i];
595 
596 		if (msr.ps_flags != 0)
597 			ps->ps_flags |= msr.ps_flags;
598 
599 		if (msr.ps_status_addr != 0)
600 			ps->ps_status_addr = msr.ps_status_addr;
601 
602 		if (msr.ps_status_mask != 0)
603 			ps->ps_status_mask = msr.ps_status_mask;
604 
605 		if (msr.ps_control_addr != 0)
606 			ps->ps_control_addr = msr.ps_control_addr;
607 
608 		if (msr.ps_control_mask != 0)
609 			ps->ps_control_mask = msr.ps_control_mask;
610 
611 		i++;
612 	}
613 
614 	return 0;
615 }
616 
617 /*
618  * Read the IA32_APERF and IA32_MPERF counters. The first
619  * increments at the rate of the fixed maximum frequency
620  * configured during the boot, whereas APERF counts at the
621  * rate of the actual frequency. Note that the MSRs must be
622  * read without delay, and that only the ratio between
623  * IA32_APERF and IA32_MPERF is architecturally defined.
624  *
625  * The function thus returns the percentage of the actual
626  * frequency in terms of the maximum frequency of the calling
627  * CPU since the last call. A value zero implies an error.
628  *
629  * For further details, refer to:
630  *
631  *	Intel Corporation: Intel 64 and IA-32 Architectures
632  *	Software Developer's Manual. Section 13.2, Volume 3A:
633  *	System Programming Guide, Part 1. July, 2008.
634  *
635  *	Advanced Micro Devices: BIOS and Kernel Developer's
636  *	Guide (BKDG) for AMD Family 10h Processors. Section
637  *	2.4.5, Revision 3.48, April 2010.
638  */
639 uint8_t
640 acpicpu_md_pstate_hwf(struct cpu_info *ci)
641 {
642 	struct acpicpu_softc *sc;
643 	uint64_t aperf, mperf;
644 	uint8_t rv = 0;
645 
646 	sc = acpicpu_sc[ci->ci_acpiid];
647 
648 	if (__predict_false(sc == NULL))
649 		return 0;
650 
651 	if (__predict_false((sc->sc_flags & ACPICPU_FLAG_P_HWF) == 0))
652 		return 0;
653 
654 	aperf = sc->sc_pstate_aperf;
655 	mperf = sc->sc_pstate_mperf;
656 
657 	x86_disable_intr();
658 
659 	sc->sc_pstate_aperf = rdmsr(MSR_APERF);
660 	sc->sc_pstate_mperf = rdmsr(MSR_MPERF);
661 
662 	x86_enable_intr();
663 
664 	aperf = sc->sc_pstate_aperf - aperf;
665 	mperf = sc->sc_pstate_mperf - mperf;
666 
667 	if (__predict_true(mperf != 0))
668 		rv = (aperf * 100) / mperf;
669 
670 	return rv;
671 }
672 
673 static void
674 acpicpu_md_pstate_hwf_reset(void *arg1, void *arg2)
675 {
676 	struct cpu_info *ci = curcpu();
677 	struct acpicpu_softc *sc;
678 
679 	sc = acpicpu_sc[ci->ci_acpiid];
680 
681 	if (__predict_false(sc == NULL))
682 		return;
683 
684 	x86_disable_intr();
685 
686 	wrmsr(MSR_APERF, 0);
687 	wrmsr(MSR_MPERF, 0);
688 
689 	x86_enable_intr();
690 
691 	sc->sc_pstate_aperf = 0;
692 	sc->sc_pstate_mperf = 0;
693 }
694 
695 int
696 acpicpu_md_pstate_get(struct acpicpu_softc *sc, uint32_t *freq)
697 {
698 	struct acpicpu_pstate *ps = NULL;
699 	uint64_t val;
700 	uint32_t i;
701 
702 	if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
703 		return acpicpu_md_pstate_fidvid_get(sc, freq);
704 
705 	/*
706 	 * Pick any P-state for the status address.
707 	 */
708 	for (i = 0; i < sc->sc_pstate_count; i++) {
709 
710 		ps = &sc->sc_pstate[i];
711 
712 		if (__predict_true(ps->ps_freq != 0))
713 			break;
714 	}
715 
716 	if (__predict_false(ps == NULL))
717 		return ENODEV;
718 
719 	if (__predict_false(ps->ps_status_addr == 0))
720 		return EINVAL;
721 
722 	val = rdmsr(ps->ps_status_addr);
723 
724 	if (__predict_true(ps->ps_status_mask != 0))
725 		val = val & ps->ps_status_mask;
726 
727 	/*
728 	 * Search for the value from known P-states.
729 	 */
730 	for (i = 0; i < sc->sc_pstate_count; i++) {
731 
732 		ps = &sc->sc_pstate[i];
733 
734 		if (__predict_false(ps->ps_freq == 0))
735 			continue;
736 
737 		if (val == ps->ps_status) {
738 			*freq = ps->ps_freq;
739 			return 0;
740 		}
741 	}
742 
743 	/*
744 	 * If the value was not found, try APERF/MPERF.
745 	 * The state is P0 if the return value is 100 %.
746 	 */
747 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
748 
749 		KASSERT(sc->sc_pstate_count > 0);
750 		KASSERT(sc->sc_pstate[0].ps_freq != 0);
751 
752 		if (acpicpu_md_pstate_hwf(sc->sc_ci) == 100) {
753 			*freq = sc->sc_pstate[0].ps_freq;
754 			return 0;
755 		}
756 	}
757 
758 	return EIO;
759 }
760 
761 int
762 acpicpu_md_pstate_set(struct acpicpu_pstate *ps)
763 {
764 	uint64_t val = 0;
765 
766 	if (__predict_false(ps->ps_control_addr == 0))
767 		return EINVAL;
768 
769 	if ((ps->ps_flags & ACPICPU_FLAG_P_FIDVID) != 0)
770 		return acpicpu_md_pstate_fidvid_set(ps);
771 
772 	/*
773 	 * If the mask is set, do a read-modify-write.
774 	 */
775 	if (__predict_true(ps->ps_control_mask != 0)) {
776 		val = rdmsr(ps->ps_control_addr);
777 		val &= ~ps->ps_control_mask;
778 	}
779 
780 	val |= ps->ps_control;
781 
782 	wrmsr(ps->ps_control_addr, val);
783 	DELAY(ps->ps_latency);
784 
785 	return 0;
786 }
787 
788 static int
789 acpicpu_md_pstate_fidvid_get(struct acpicpu_softc *sc, uint32_t *freq)
790 {
791 	struct acpicpu_pstate *ps;
792 	uint32_t fid, i, vid;
793 	uint32_t cfid, cvid;
794 	int rv;
795 
796 	/*
797 	 * AMD family 0Fh needs special treatment.
798 	 * While it wants to use ACPI, it does not
799 	 * comply with the ACPI specifications.
800 	 */
801 	rv = acpicpu_md_pstate_fidvid_read(&cfid, &cvid);
802 
803 	if (rv != 0)
804 		return rv;
805 
806 	for (i = 0; i < sc->sc_pstate_count; i++) {
807 
808 		ps = &sc->sc_pstate[i];
809 
810 		if (__predict_false(ps->ps_freq == 0))
811 			continue;
812 
813 		fid = __SHIFTOUT(ps->ps_status, ACPI_0FH_STATUS_FID);
814 		vid = __SHIFTOUT(ps->ps_status, ACPI_0FH_STATUS_VID);
815 
816 		if (cfid == fid && cvid == vid) {
817 			*freq = ps->ps_freq;
818 			return 0;
819 		}
820 	}
821 
822 	return EIO;
823 }
824 
825 static int
826 acpicpu_md_pstate_fidvid_set(struct acpicpu_pstate *ps)
827 {
828 	const uint64_t ctrl = ps->ps_control;
829 	uint32_t cfid, cvid, fid, i, irt;
830 	uint32_t pll, vco_cfid, vco_fid;
831 	uint32_t val, vid, vst;
832 	int rv;
833 
834 	rv = acpicpu_md_pstate_fidvid_read(&cfid, &cvid);
835 
836 	if (rv != 0)
837 		return rv;
838 
839 	fid = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_FID);
840 	vid = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_VID);
841 	irt = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_IRT);
842 	vst = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_VST);
843 	pll = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_PLL);
844 
845 	vst = vst * 20;
846 	pll = pll * 1000 / 5;
847 	irt = 10 * __BIT(irt);
848 
849 	/*
850 	 * Phase 1.
851 	 */
852 	while (cvid > vid) {
853 
854 		val = 1 << __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_MVS);
855 		val = (val > cvid) ? 0 : cvid - val;
856 
857 		acpicpu_md_pstate_fidvid_write(cfid, val, 1, vst);
858 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
859 
860 		if (rv != 0)
861 			return rv;
862 	}
863 
864 	i = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_RVO);
865 
866 	for (; i > 0 && cvid > 0; --i) {
867 
868 		acpicpu_md_pstate_fidvid_write(cfid, cvid - 1, 1, vst);
869 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
870 
871 		if (rv != 0)
872 			return rv;
873 	}
874 
875 	/*
876 	 * Phase 2.
877 	 */
878 	if (cfid != fid) {
879 
880 		vco_fid  = FID_TO_VCO_FID(fid);
881 		vco_cfid = FID_TO_VCO_FID(cfid);
882 
883 		while (abs(vco_fid - vco_cfid) > 2) {
884 
885 			if (fid <= cfid)
886 				val = cfid - 2;
887 			else {
888 				val = (cfid > 6) ? cfid + 2 :
889 				    FID_TO_VCO_FID(cfid) + 2;
890 			}
891 
892 			acpicpu_md_pstate_fidvid_write(val, cvid, pll, irt);
893 			rv = acpicpu_md_pstate_fidvid_read(&cfid, NULL);
894 
895 			if (rv != 0)
896 				return rv;
897 
898 			vco_cfid = FID_TO_VCO_FID(cfid);
899 		}
900 
901 		acpicpu_md_pstate_fidvid_write(fid, cvid, pll, irt);
902 		rv = acpicpu_md_pstate_fidvid_read(&cfid, NULL);
903 
904 		if (rv != 0)
905 			return rv;
906 	}
907 
908 	/*
909 	 * Phase 3.
910 	 */
911 	if (cvid != vid) {
912 
913 		acpicpu_md_pstate_fidvid_write(cfid, vid, 1, vst);
914 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
915 
916 		if (rv != 0)
917 			return rv;
918 	}
919 
920 	return 0;
921 }
922 
923 static int
924 acpicpu_md_pstate_fidvid_read(uint32_t *cfid, uint32_t *cvid)
925 {
926 	int i = ACPICPU_P_STATE_RETRY * 100;
927 	uint64_t val;
928 
929 	do {
930 		val = rdmsr(MSR_0FH_STATUS);
931 
932 	} while (__SHIFTOUT(val, MSR_0FH_STATUS_PENDING) != 0 && --i >= 0);
933 
934 	if (i == 0)
935 		return EAGAIN;
936 
937 	if (cfid != NULL)
938 		*cfid = __SHIFTOUT(val, MSR_0FH_STATUS_CFID);
939 
940 	if (cvid != NULL)
941 		*cvid = __SHIFTOUT(val, MSR_0FH_STATUS_CVID);
942 
943 	return 0;
944 }
945 
946 static void
947 acpicpu_md_pstate_fidvid_write(uint32_t fid,
948     uint32_t vid, uint32_t cnt, uint32_t tmo)
949 {
950 	uint64_t val = 0;
951 
952 	val |= __SHIFTIN(fid, MSR_0FH_CONTROL_FID);
953 	val |= __SHIFTIN(vid, MSR_0FH_CONTROL_VID);
954 	val |= __SHIFTIN(cnt, MSR_0FH_CONTROL_CNT);
955 	val |= __SHIFTIN(0x1, MSR_0FH_CONTROL_CHG);
956 
957 	wrmsr(MSR_0FH_CONTROL, val);
958 	DELAY(tmo);
959 }
960 
961 int
962 acpicpu_md_tstate_get(struct acpicpu_softc *sc, uint32_t *percent)
963 {
964 	struct acpicpu_tstate *ts;
965 	uint64_t val;
966 	uint32_t i;
967 
968 	val = rdmsr(MSR_THERM_CONTROL);
969 
970 	for (i = 0; i < sc->sc_tstate_count; i++) {
971 
972 		ts = &sc->sc_tstate[i];
973 
974 		if (ts->ts_percent == 0)
975 			continue;
976 
977 		if (val == ts->ts_status) {
978 			*percent = ts->ts_percent;
979 			return 0;
980 		}
981 	}
982 
983 	return EIO;
984 }
985 
986 int
987 acpicpu_md_tstate_set(struct acpicpu_tstate *ts)
988 {
989 	uint64_t val;
990 	uint8_t i;
991 
992 	val = ts->ts_control;
993 	val = val & __BITS(0, 4);
994 
995 	wrmsr(MSR_THERM_CONTROL, val);
996 
997 	if (ts->ts_status == 0) {
998 		DELAY(ts->ts_latency);
999 		return 0;
1000 	}
1001 
1002 	for (i = val = 0; i < ACPICPU_T_STATE_RETRY; i++) {
1003 
1004 		val = rdmsr(MSR_THERM_CONTROL);
1005 
1006 		if (val == ts->ts_status)
1007 			return 0;
1008 
1009 		DELAY(ts->ts_latency);
1010 	}
1011 
1012 	return EAGAIN;
1013 }
1014 
1015 /*
1016  * A kludge for backwards compatibility.
1017  */
1018 static int
1019 acpicpu_md_pstate_sysctl_init(void)
1020 {
1021 	const struct sysctlnode	*fnode, *mnode, *rnode;
1022 	const char *str;
1023 	int rv;
1024 
1025 	switch (cpu_vendor) {
1026 
1027 	case CPUVENDOR_IDT:
1028 	case CPUVENDOR_INTEL:
1029 		str = "est";
1030 		break;
1031 
1032 	case CPUVENDOR_AMD:
1033 		str = "powernow";
1034 		break;
1035 
1036 	default:
1037 		return ENODEV;
1038 	}
1039 
1040 
1041 	rv = sysctl_createv(&acpicpu_log, 0, NULL, &rnode,
1042 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "machdep", NULL,
1043 	    NULL, 0, NULL, 0, CTL_MACHDEP, CTL_EOL);
1044 
1045 	if (rv != 0)
1046 		goto fail;
1047 
1048 	rv = sysctl_createv(&acpicpu_log, 0, &rnode, &mnode,
1049 	    0, CTLTYPE_NODE, str, 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, &mnode, &fnode,
1056 	    0, CTLTYPE_NODE, "frequency", NULL,
1057 	    NULL, 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_READWRITE, CTLTYPE_INT, "target", NULL,
1064 	    acpicpu_md_pstate_sysctl_set, 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_INT, "current", NULL,
1071 	    acpicpu_md_pstate_sysctl_get, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1072 
1073 	if (rv != 0)
1074 		goto fail;
1075 
1076 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1077 	    CTLFLAG_READONLY, CTLTYPE_STRING, "available", NULL,
1078 	    acpicpu_md_pstate_sysctl_all, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1079 
1080 	if (rv != 0)
1081 		goto fail;
1082 
1083 	return 0;
1084 
1085 fail:
1086 	if (acpicpu_log != NULL) {
1087 		sysctl_teardown(&acpicpu_log);
1088 		acpicpu_log = NULL;
1089 	}
1090 
1091 	return rv;
1092 }
1093 
1094 static int
1095 acpicpu_md_pstate_sysctl_get(SYSCTLFN_ARGS)
1096 {
1097 	struct sysctlnode node;
1098 	uint32_t freq;
1099 	int err;
1100 
1101 	freq = cpufreq_get(curcpu());
1102 
1103 	if (freq == 0)
1104 		return ENXIO;
1105 
1106 	node = *rnode;
1107 	node.sysctl_data = &freq;
1108 
1109 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1110 
1111 	if (err != 0 || newp == NULL)
1112 		return err;
1113 
1114 	return 0;
1115 }
1116 
1117 static int
1118 acpicpu_md_pstate_sysctl_set(SYSCTLFN_ARGS)
1119 {
1120 	struct sysctlnode node;
1121 	uint32_t freq;
1122 	int err;
1123 
1124 	freq = cpufreq_get(curcpu());
1125 
1126 	if (freq == 0)
1127 		return ENXIO;
1128 
1129 	node = *rnode;
1130 	node.sysctl_data = &freq;
1131 
1132 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1133 
1134 	if (err != 0 || newp == NULL)
1135 		return err;
1136 
1137 	cpufreq_set_all(freq);
1138 
1139 	return 0;
1140 }
1141 
1142 static int
1143 acpicpu_md_pstate_sysctl_all(SYSCTLFN_ARGS)
1144 {
1145 	struct cpu_info *ci = curcpu();
1146 	struct acpicpu_softc *sc;
1147 	struct sysctlnode node;
1148 	char buf[1024];
1149 	size_t len;
1150 	uint32_t i;
1151 	int err;
1152 
1153 	sc = acpicpu_sc[ci->ci_acpiid];
1154 
1155 	if (sc == NULL)
1156 		return ENXIO;
1157 
1158 	(void)memset(&buf, 0, sizeof(buf));
1159 
1160 	mutex_enter(&sc->sc_mtx);
1161 
1162 	for (len = 0, i = sc->sc_pstate_max; i < sc->sc_pstate_count; i++) {
1163 
1164 		if (sc->sc_pstate[i].ps_freq == 0)
1165 			continue;
1166 
1167 		if (len >= sizeof(buf))
1168 			break;
1169 		len += snprintf(buf + len, sizeof(buf) - len, "%u%s",
1170 		    sc->sc_pstate[i].ps_freq,
1171 		    i < (sc->sc_pstate_count - 1) ? " " : "");
1172 	}
1173 
1174 	mutex_exit(&sc->sc_mtx);
1175 
1176 	node = *rnode;
1177 	node.sysctl_data = buf;
1178 
1179 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1180 
1181 	if (err != 0 || newp == NULL)
1182 		return err;
1183 
1184 	return 0;
1185 }
1186 
1187