xref: /netbsd-src/sys/arch/x86/acpi/acpi_cpu_md.c (revision 946379e7b37692fc43f68eb0d1c10daa0a7f3b6c)
1 /* $NetBSD: acpi_cpu_md.c,v 1.77 2014/04/17 16:01:24 christos 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 1.77 2014/04/17 16:01:24 christos 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 #include <x86/x86/tsc.h>
47 
48 #include <dev/acpi/acpica.h>
49 #include <dev/acpi/acpi_cpu.h>
50 
51 #include <dev/pci/pcivar.h>
52 #include <dev/pci/pcidevs.h>
53 
54 #include <machine/acpi_machdep.h>
55 
56 /*
57  * Intel IA32_MISC_ENABLE.
58  */
59 #define MSR_MISC_ENABLE_EST	__BIT(16)
60 #define MSR_MISC_ENABLE_TURBO	__BIT(38)
61 
62 /*
63  * AMD C1E.
64  */
65 #define MSR_CMPHALT		0xc0010055
66 
67 #define MSR_CMPHALT_SMI		__BIT(27)
68 #define MSR_CMPHALT_C1E		__BIT(28)
69 #define MSR_CMPHALT_BMSTS	__BIT(29)
70 
71 /*
72  * AMD families 10h, 11h, 12h, 14h, and 15h.
73  */
74 #define MSR_10H_LIMIT		0xc0010061
75 #define MSR_10H_CONTROL		0xc0010062
76 #define MSR_10H_STATUS		0xc0010063
77 #define MSR_10H_CONFIG		0xc0010064
78 
79 /*
80  * AMD family 0Fh.
81  */
82 #define MSR_0FH_CONTROL		0xc0010041
83 #define MSR_0FH_STATUS		0xc0010042
84 
85 #define MSR_0FH_STATUS_CFID	__BITS( 0,  5)
86 #define MSR_0FH_STATUS_CVID	__BITS(32, 36)
87 #define MSR_0FH_STATUS_PENDING	__BITS(31, 31)
88 
89 #define MSR_0FH_CONTROL_FID	__BITS( 0,  5)
90 #define MSR_0FH_CONTROL_VID	__BITS( 8, 12)
91 #define MSR_0FH_CONTROL_CHG	__BITS(16, 16)
92 #define MSR_0FH_CONTROL_CNT	__BITS(32, 51)
93 
94 #define ACPI_0FH_STATUS_FID	__BITS( 0,  5)
95 #define ACPI_0FH_STATUS_VID	__BITS( 6, 10)
96 
97 #define ACPI_0FH_CONTROL_FID	__BITS( 0,  5)
98 #define ACPI_0FH_CONTROL_VID	__BITS( 6, 10)
99 #define ACPI_0FH_CONTROL_VST	__BITS(11, 17)
100 #define ACPI_0FH_CONTROL_MVS	__BITS(18, 19)
101 #define ACPI_0FH_CONTROL_PLL	__BITS(20, 26)
102 #define ACPI_0FH_CONTROL_RVO	__BITS(28, 29)
103 #define ACPI_0FH_CONTROL_IRT	__BITS(30, 31)
104 
105 #define FID_TO_VCO_FID(fidd)	(((fid) < 8) ? (8 + ((fid) << 1)) : (fid))
106 
107 static char	  native_idle_text[16];
108 void		(*native_idle)(void) = NULL;
109 
110 static int	 acpicpu_md_quirk_piix4(const struct pci_attach_args *);
111 static void	 acpicpu_md_pstate_hwf_reset(void *, void *);
112 static int	 acpicpu_md_pstate_fidvid_get(struct acpicpu_softc *,
113                                               uint32_t *);
114 static int	 acpicpu_md_pstate_fidvid_set(struct acpicpu_pstate *);
115 static int	 acpicpu_md_pstate_fidvid_read(uint32_t *, uint32_t *);
116 static void	 acpicpu_md_pstate_fidvid_write(uint32_t, uint32_t,
117 					        uint32_t, uint32_t);
118 static int	 acpicpu_md_pstate_sysctl_init(void);
119 static int	 acpicpu_md_pstate_sysctl_get(SYSCTLFN_PROTO);
120 static int	 acpicpu_md_pstate_sysctl_set(SYSCTLFN_PROTO);
121 static int	 acpicpu_md_pstate_sysctl_all(SYSCTLFN_PROTO);
122 
123 extern struct acpicpu_softc **acpicpu_sc;
124 static struct sysctllog *acpicpu_log = NULL;
125 
126 struct cpu_info *
127 acpicpu_md_match(device_t parent, cfdata_t match, void *aux)
128 {
129 	struct cpufeature_attach_args *cfaa = aux;
130 
131 	if (strcmp(cfaa->name, "frequency") != 0)
132 		return NULL;
133 
134 	return cfaa->ci;
135 }
136 
137 struct cpu_info *
138 acpicpu_md_attach(device_t parent, device_t self, void *aux)
139 {
140 	struct cpufeature_attach_args *cfaa = aux;
141 
142 	return cfaa->ci;
143 }
144 
145 uint32_t
146 acpicpu_md_flags(void)
147 {
148 	struct cpu_info *ci = curcpu();
149 	struct pci_attach_args pa;
150 	uint32_t family, val = 0;
151 	uint32_t regs[4];
152 	uint64_t msr;
153 
154 	if (acpi_md_ncpus() == 1)
155 		val |= ACPICPU_FLAG_C_BM;
156 
157 	if ((ci->ci_feat_val[1] & CPUID2_MONITOR) != 0)
158 		val |= ACPICPU_FLAG_C_FFH;
159 
160 	/*
161 	 * By default, assume that the local APIC timer
162 	 * as well as TSC are stalled during C3 sleep.
163 	 */
164 	val |= ACPICPU_FLAG_C_APIC | ACPICPU_FLAG_C_TSC;
165 
166 	/*
167 	 * Detect whether TSC is invariant. If it is not, we keep the flag to
168 	 * note that TSC will not run at constant rate. Depending on the CPU,
169 	 * this may affect P- and T-state changes, but especially relevant
170 	 * are C-states; with variant TSC, states larger than C1 may
171 	 * completely stop the counter.
172 	 */
173 	if (tsc_is_invariant())
174 		val &= ~ACPICPU_FLAG_C_TSC;
175 
176 	switch (cpu_vendor) {
177 
178 	case CPUVENDOR_IDT:
179 
180 		if ((ci->ci_feat_val[1] & CPUID2_EST) != 0)
181 			val |= ACPICPU_FLAG_P_FFH;
182 
183 		if ((ci->ci_feat_val[0] & CPUID_ACPI) != 0)
184 			val |= ACPICPU_FLAG_T_FFH;
185 
186 		break;
187 
188 	case CPUVENDOR_INTEL:
189 
190 		/*
191 		 * Bus master control and arbitration should be
192 		 * available on all supported Intel CPUs (to be
193 		 * sure, this is double-checked later from the
194 		 * firmware data). These flags imply that it is
195 		 * not necessary to flush caches before C3 state.
196 		 */
197 		val |= ACPICPU_FLAG_C_BM | ACPICPU_FLAG_C_ARB;
198 
199 		/*
200 		 * Check if we can use "native", MSR-based,
201 		 * access. If not, we have to resort to I/O.
202 		 */
203 		if ((ci->ci_feat_val[1] & CPUID2_EST) != 0)
204 			val |= ACPICPU_FLAG_P_FFH;
205 
206 		if ((ci->ci_feat_val[0] & CPUID_ACPI) != 0)
207 			val |= ACPICPU_FLAG_T_FFH;
208 
209 		/*
210 		 * Check whether MSR_APERF, MSR_MPERF, and Turbo
211 		 * Boost are available. Also see if we might have
212 		 * an invariant local APIC timer ("ARAT").
213 		 */
214 		if (cpuid_level >= 0x06) {
215 
216 			x86_cpuid(0x00000006, regs);
217 
218 			if ((regs[2] & CPUID_DSPM_HWF) != 0)
219 				val |= ACPICPU_FLAG_P_HWF;
220 
221 			if ((regs[0] & CPUID_DSPM_IDA) != 0)
222 				val |= ACPICPU_FLAG_P_TURBO;
223 
224 			if ((regs[0] & CPUID_DSPM_ARAT) != 0)
225 				val &= ~ACPICPU_FLAG_C_APIC;
226 		}
227 
228 		break;
229 
230 	case CPUVENDOR_AMD:
231 
232 		x86_cpuid(0x80000000, regs);
233 
234 		if (regs[0] < 0x80000007)
235 			break;
236 
237 		x86_cpuid(0x80000007, regs);
238 
239 		family = CPUID_TO_FAMILY(ci->ci_signature);
240 
241     		switch (family) {
242 
243 		case 0x0f:
244 
245 			/*
246 			 * Disable C1E if present.
247 			 */
248 			if (rdmsr_safe(MSR_CMPHALT, &msr) != EFAULT)
249 				val |= ACPICPU_FLAG_C_C1E;
250 
251 			/*
252 			 * Evaluate support for the "FID/VID
253 			 * algorithm" also used by powernow(4).
254 			 */
255 			if ((regs[3] & CPUID_APM_FID) == 0)
256 				break;
257 
258 			if ((regs[3] & CPUID_APM_VID) == 0)
259 				break;
260 
261 			val |= ACPICPU_FLAG_P_FFH | ACPICPU_FLAG_P_FIDVID;
262 			break;
263 
264 		case 0x10:
265 		case 0x11:
266 
267 			/*
268 			 * Disable C1E if present.
269 			 */
270 			if (rdmsr_safe(MSR_CMPHALT, &msr) != EFAULT)
271 				val |= ACPICPU_FLAG_C_C1E;
272 
273 			/* FALLTHROUGH */
274 
275 		case 0x12:
276 		case 0x14: /* AMD Fusion */
277 		case 0x15: /* AMD Bulldozer */
278 
279 			/*
280 			 * Like with Intel, detect MSR-based P-states,
281 			 * and AMD's "turbo" (Core Performance Boost),
282 			 * respectively.
283 			 */
284 			if ((regs[3] & CPUID_APM_HWP) != 0)
285 				val |= ACPICPU_FLAG_P_FFH;
286 
287 			if ((regs[3] & CPUID_APM_CPB) != 0)
288 				val |= ACPICPU_FLAG_P_TURBO;
289 
290 			/*
291 			 * Also check for APERF and MPERF,
292 			 * first available in the family 10h.
293 			 */
294 			if (cpuid_level >= 0x06) {
295 
296 				x86_cpuid(0x00000006, regs);
297 
298 				if ((regs[2] & CPUID_DSPM_HWF) != 0)
299 					val |= ACPICPU_FLAG_P_HWF;
300 			}
301 
302 			break;
303 		}
304 
305 		break;
306 	}
307 
308 	/*
309 	 * There are several erratums for PIIX4.
310 	 */
311 	if (pci_find_device(&pa, acpicpu_md_quirk_piix4) != 0)
312 		val |= ACPICPU_FLAG_PIIX4;
313 
314 	return val;
315 }
316 
317 static int
318 acpicpu_md_quirk_piix4(const struct pci_attach_args *pa)
319 {
320 
321 	/*
322 	 * XXX: The pci_find_device(9) function only
323 	 *	deals with attached devices. Change this
324 	 *	to use something like pci_device_foreach().
325 	 */
326 	if (PCI_VENDOR(pa->pa_id) != PCI_VENDOR_INTEL)
327 		return 0;
328 
329 	if (PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_INTEL_82371AB_ISA ||
330 	    PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_INTEL_82440MX_PMC)
331 		return 1;
332 
333 	return 0;
334 }
335 
336 void
337 acpicpu_md_quirk_c1e(void)
338 {
339 	const uint64_t c1e = MSR_CMPHALT_SMI | MSR_CMPHALT_C1E;
340 	uint64_t val;
341 
342 	val = rdmsr(MSR_CMPHALT);
343 
344 	if ((val & c1e) != 0)
345 		wrmsr(MSR_CMPHALT, val & ~c1e);
346 }
347 
348 int
349 acpicpu_md_cstate_start(struct acpicpu_softc *sc)
350 {
351 	const size_t size = sizeof(native_idle_text);
352 	struct acpicpu_cstate *cs;
353 	bool ipi = false;
354 	int i;
355 
356 	/*
357 	 * Save the cpu_idle(9) loop used by default.
358 	 */
359 	x86_cpu_idle_get(&native_idle, native_idle_text, size);
360 
361 	for (i = 0; i < ACPI_C_STATE_COUNT; i++) {
362 
363 		cs = &sc->sc_cstate[i];
364 
365 		if (cs->cs_method == ACPICPU_C_STATE_HALT) {
366 			ipi = true;
367 			break;
368 		}
369 	}
370 
371 	x86_cpu_idle_set(acpicpu_cstate_idle, "acpi", ipi);
372 
373 	return 0;
374 }
375 
376 int
377 acpicpu_md_cstate_stop(void)
378 {
379 	static char text[16];
380 	void (*func)(void);
381 	uint64_t xc;
382 	bool ipi;
383 
384 	x86_cpu_idle_get(&func, text, sizeof(text));
385 
386 	if (func == native_idle)
387 		return EALREADY;
388 
389 	ipi = (native_idle != x86_cpu_idle_halt) ? false : true;
390 	x86_cpu_idle_set(native_idle, native_idle_text, ipi);
391 
392 	/*
393 	 * Run a cross-call to ensure that all CPUs are
394 	 * out from the ACPI idle-loop before detachment.
395 	 */
396 	xc = xc_broadcast(0, (xcfunc_t)nullop, NULL, NULL);
397 	xc_wait(xc);
398 
399 	return 0;
400 }
401 
402 /*
403  * Called with interrupts enabled.
404  */
405 void
406 acpicpu_md_cstate_enter(int method, int state)
407 {
408 	struct cpu_info *ci = curcpu();
409 
410 	KASSERT(ci->ci_ilevel == IPL_NONE);
411 
412 	switch (method) {
413 
414 	case ACPICPU_C_STATE_FFH:
415 
416 		x86_monitor(&ci->ci_want_resched, 0, 0);
417 
418 		if (__predict_false(ci->ci_want_resched != 0))
419 			return;
420 
421 		x86_mwait((state - 1) << 4, 0);
422 		break;
423 
424 	case ACPICPU_C_STATE_HALT:
425 
426 		x86_disable_intr();
427 
428 		if (__predict_false(ci->ci_want_resched != 0)) {
429 			x86_enable_intr();
430 			return;
431 		}
432 
433 		x86_stihlt();
434 		break;
435 	}
436 }
437 
438 int
439 acpicpu_md_pstate_start(struct acpicpu_softc *sc)
440 {
441 	uint64_t xc, val;
442 
443 	switch (cpu_vendor) {
444 
445 	case CPUVENDOR_IDT:
446 	case CPUVENDOR_INTEL:
447 
448 		/*
449 		 * Make sure EST is enabled.
450 		 */
451 		if ((sc->sc_flags & ACPICPU_FLAG_P_FFH) != 0) {
452 
453 			val = rdmsr(MSR_MISC_ENABLE);
454 
455 			if ((val & MSR_MISC_ENABLE_EST) == 0) {
456 
457 				val |= MSR_MISC_ENABLE_EST;
458 				wrmsr(MSR_MISC_ENABLE, val);
459 				val = rdmsr(MSR_MISC_ENABLE);
460 
461 				if ((val & MSR_MISC_ENABLE_EST) == 0)
462 					return ENOTTY;
463 			}
464 		}
465 	}
466 
467 	/*
468 	 * Reset the APERF and MPERF counters.
469 	 */
470 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
471 		xc = xc_broadcast(0, acpicpu_md_pstate_hwf_reset, NULL, NULL);
472 		xc_wait(xc);
473 	}
474 
475 	return acpicpu_md_pstate_sysctl_init();
476 }
477 
478 int
479 acpicpu_md_pstate_stop(void)
480 {
481 
482 	if (acpicpu_log == NULL)
483 		return EALREADY;
484 
485 	sysctl_teardown(&acpicpu_log);
486 	acpicpu_log = NULL;
487 
488 	return 0;
489 }
490 
491 int
492 acpicpu_md_pstate_init(struct acpicpu_softc *sc)
493 {
494 	struct cpu_info *ci = sc->sc_ci;
495 	struct acpicpu_pstate *ps, msr;
496 	uint32_t family, i = 0;
497 
498 	(void)memset(&msr, 0, sizeof(struct acpicpu_pstate));
499 
500 	switch (cpu_vendor) {
501 
502 	case CPUVENDOR_IDT:
503 	case CPUVENDOR_INTEL:
504 
505 		/*
506 		 * If the so-called Turbo Boost is present,
507 		 * the P0-state is always the "turbo state".
508 		 * It is shown as the P1 frequency + 1 MHz.
509 		 *
510 		 * For discussion, see:
511 		 *
512 		 *	Intel Corporation: Intel Turbo Boost Technology
513 		 *	in Intel Core(tm) Microarchitectures (Nehalem)
514 		 *	Based Processors. White Paper, November 2008.
515 		 */
516 		if (sc->sc_pstate_count >= 2 &&
517 		   (sc->sc_flags & ACPICPU_FLAG_P_TURBO) != 0) {
518 
519 			ps = &sc->sc_pstate[0];
520 
521 			if (ps->ps_freq == sc->sc_pstate[1].ps_freq + 1)
522 				ps->ps_flags |= ACPICPU_FLAG_P_TURBO;
523 		}
524 
525 		msr.ps_control_addr = MSR_PERF_CTL;
526 		msr.ps_control_mask = __BITS(0, 15);
527 
528 		msr.ps_status_addr  = MSR_PERF_STATUS;
529 		msr.ps_status_mask  = __BITS(0, 15);
530 		break;
531 
532 	case CPUVENDOR_AMD:
533 
534 		if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
535 			msr.ps_flags |= ACPICPU_FLAG_P_FIDVID;
536 
537 		family = CPUID_TO_FAMILY(ci->ci_signature);
538 
539 		switch (family) {
540 
541 		case 0x0f:
542 			msr.ps_control_addr = MSR_0FH_CONTROL;
543 			msr.ps_status_addr  = MSR_0FH_STATUS;
544 			break;
545 
546 		case 0x10:
547 		case 0x11:
548 		case 0x12:
549 		case 0x14:
550 		case 0x15:
551 			msr.ps_control_addr = MSR_10H_CONTROL;
552 			msr.ps_control_mask = __BITS(0, 2);
553 
554 			msr.ps_status_addr  = MSR_10H_STATUS;
555 			msr.ps_status_mask  = __BITS(0, 2);
556 			break;
557 
558 		default:
559 			/*
560 			 * If we have an unknown AMD CPU, rely on XPSS.
561 			 */
562 			if ((sc->sc_flags & ACPICPU_FLAG_P_XPSS) == 0)
563 				return EOPNOTSUPP;
564 		}
565 
566 		break;
567 
568 	default:
569 		return ENODEV;
570 	}
571 
572 	/*
573 	 * Fill the P-state structures with MSR addresses that are
574 	 * known to be correct. If we do not know the addresses,
575 	 * leave the values intact. If a vendor uses XPSS, we do
576 	 * not necessarily need to do anything to support new CPUs.
577 	 */
578 	while (i < sc->sc_pstate_count) {
579 
580 		ps = &sc->sc_pstate[i];
581 
582 		if (msr.ps_flags != 0)
583 			ps->ps_flags |= msr.ps_flags;
584 
585 		if (msr.ps_status_addr != 0)
586 			ps->ps_status_addr = msr.ps_status_addr;
587 
588 		if (msr.ps_status_mask != 0)
589 			ps->ps_status_mask = msr.ps_status_mask;
590 
591 		if (msr.ps_control_addr != 0)
592 			ps->ps_control_addr = msr.ps_control_addr;
593 
594 		if (msr.ps_control_mask != 0)
595 			ps->ps_control_mask = msr.ps_control_mask;
596 
597 		i++;
598 	}
599 
600 	return 0;
601 }
602 
603 /*
604  * Read the IA32_APERF and IA32_MPERF counters. The first
605  * increments at the rate of the fixed maximum frequency
606  * configured during the boot, whereas APERF counts at the
607  * rate of the actual frequency. Note that the MSRs must be
608  * read without delay, and that only the ratio between
609  * IA32_APERF and IA32_MPERF is architecturally defined.
610  *
611  * The function thus returns the percentage of the actual
612  * frequency in terms of the maximum frequency of the calling
613  * CPU since the last call. A value zero implies an error.
614  *
615  * For further details, refer to:
616  *
617  *	Intel Corporation: Intel 64 and IA-32 Architectures
618  *	Software Developer's Manual. Section 13.2, Volume 3A:
619  *	System Programming Guide, Part 1. July, 2008.
620  *
621  *	Advanced Micro Devices: BIOS and Kernel Developer's
622  *	Guide (BKDG) for AMD Family 10h Processors. Section
623  *	2.4.5, Revision 3.48, April 2010.
624  */
625 uint8_t
626 acpicpu_md_pstate_hwf(struct cpu_info *ci)
627 {
628 	struct acpicpu_softc *sc;
629 	uint64_t aperf, mperf;
630 	uint8_t rv = 0;
631 
632 	sc = acpicpu_sc[ci->ci_acpiid];
633 
634 	if (__predict_false(sc == NULL))
635 		return 0;
636 
637 	if (__predict_false((sc->sc_flags & ACPICPU_FLAG_P_HWF) == 0))
638 		return 0;
639 
640 	aperf = sc->sc_pstate_aperf;
641 	mperf = sc->sc_pstate_mperf;
642 
643 	x86_disable_intr();
644 
645 	sc->sc_pstate_aperf = rdmsr(MSR_APERF);
646 	sc->sc_pstate_mperf = rdmsr(MSR_MPERF);
647 
648 	x86_enable_intr();
649 
650 	aperf = sc->sc_pstate_aperf - aperf;
651 	mperf = sc->sc_pstate_mperf - mperf;
652 
653 	if (__predict_true(mperf != 0))
654 		rv = (aperf * 100) / mperf;
655 
656 	return rv;
657 }
658 
659 static void
660 acpicpu_md_pstate_hwf_reset(void *arg1, void *arg2)
661 {
662 	struct cpu_info *ci = curcpu();
663 	struct acpicpu_softc *sc;
664 
665 	sc = acpicpu_sc[ci->ci_acpiid];
666 
667 	if (__predict_false(sc == NULL))
668 		return;
669 
670 	x86_disable_intr();
671 
672 	wrmsr(MSR_APERF, 0);
673 	wrmsr(MSR_MPERF, 0);
674 
675 	x86_enable_intr();
676 
677 	sc->sc_pstate_aperf = 0;
678 	sc->sc_pstate_mperf = 0;
679 }
680 
681 int
682 acpicpu_md_pstate_get(struct acpicpu_softc *sc, uint32_t *freq)
683 {
684 	struct acpicpu_pstate *ps = NULL;
685 	uint64_t val;
686 	uint32_t i;
687 
688 	if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
689 		return acpicpu_md_pstate_fidvid_get(sc, freq);
690 
691 	/*
692 	 * Pick any P-state for the status address.
693 	 */
694 	for (i = 0; i < sc->sc_pstate_count; i++) {
695 
696 		ps = &sc->sc_pstate[i];
697 
698 		if (__predict_true(ps->ps_freq != 0))
699 			break;
700 	}
701 
702 	if (__predict_false(ps == NULL))
703 		return ENODEV;
704 
705 	if (__predict_false(ps->ps_status_addr == 0))
706 		return EINVAL;
707 
708 	val = rdmsr(ps->ps_status_addr);
709 
710 	if (__predict_true(ps->ps_status_mask != 0))
711 		val = val & ps->ps_status_mask;
712 
713 	/*
714 	 * Search for the value from known P-states.
715 	 */
716 	for (i = 0; i < sc->sc_pstate_count; i++) {
717 
718 		ps = &sc->sc_pstate[i];
719 
720 		if (__predict_false(ps->ps_freq == 0))
721 			continue;
722 
723 		if (val == ps->ps_status) {
724 			*freq = ps->ps_freq;
725 			return 0;
726 		}
727 	}
728 
729 	/*
730 	 * If the value was not found, try APERF/MPERF.
731 	 * The state is P0 if the return value is 100 %.
732 	 */
733 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
734 
735 		KASSERT(sc->sc_pstate_count > 0);
736 		KASSERT(sc->sc_pstate[0].ps_freq != 0);
737 
738 		if (acpicpu_md_pstate_hwf(sc->sc_ci) == 100) {
739 			*freq = sc->sc_pstate[0].ps_freq;
740 			return 0;
741 		}
742 	}
743 
744 	return EIO;
745 }
746 
747 int
748 acpicpu_md_pstate_set(struct acpicpu_pstate *ps)
749 {
750 	uint64_t val = 0;
751 
752 	if (__predict_false(ps->ps_control_addr == 0))
753 		return EINVAL;
754 
755 	if ((ps->ps_flags & ACPICPU_FLAG_P_FIDVID) != 0)
756 		return acpicpu_md_pstate_fidvid_set(ps);
757 
758 	/*
759 	 * If the mask is set, do a read-modify-write.
760 	 */
761 	if (__predict_true(ps->ps_control_mask != 0)) {
762 		val = rdmsr(ps->ps_control_addr);
763 		val &= ~ps->ps_control_mask;
764 	}
765 
766 	val |= ps->ps_control;
767 
768 	wrmsr(ps->ps_control_addr, val);
769 	DELAY(ps->ps_latency);
770 
771 	return 0;
772 }
773 
774 static int
775 acpicpu_md_pstate_fidvid_get(struct acpicpu_softc *sc, uint32_t *freq)
776 {
777 	struct acpicpu_pstate *ps;
778 	uint32_t fid, i, vid;
779 	uint32_t cfid, cvid;
780 	int rv;
781 
782 	/*
783 	 * AMD family 0Fh needs special treatment.
784 	 * While it wants to use ACPI, it does not
785 	 * comply with the ACPI specifications.
786 	 */
787 	rv = acpicpu_md_pstate_fidvid_read(&cfid, &cvid);
788 
789 	if (rv != 0)
790 		return rv;
791 
792 	for (i = 0; i < sc->sc_pstate_count; i++) {
793 
794 		ps = &sc->sc_pstate[i];
795 
796 		if (__predict_false(ps->ps_freq == 0))
797 			continue;
798 
799 		fid = __SHIFTOUT(ps->ps_status, ACPI_0FH_STATUS_FID);
800 		vid = __SHIFTOUT(ps->ps_status, ACPI_0FH_STATUS_VID);
801 
802 		if (cfid == fid && cvid == vid) {
803 			*freq = ps->ps_freq;
804 			return 0;
805 		}
806 	}
807 
808 	return EIO;
809 }
810 
811 static int
812 acpicpu_md_pstate_fidvid_set(struct acpicpu_pstate *ps)
813 {
814 	const uint64_t ctrl = ps->ps_control;
815 	uint32_t cfid, cvid, fid, i, irt;
816 	uint32_t pll, vco_cfid, vco_fid;
817 	uint32_t val, vid, vst;
818 	int rv;
819 
820 	rv = acpicpu_md_pstate_fidvid_read(&cfid, &cvid);
821 
822 	if (rv != 0)
823 		return rv;
824 
825 	fid = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_FID);
826 	vid = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_VID);
827 	irt = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_IRT);
828 	vst = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_VST);
829 	pll = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_PLL);
830 
831 	vst = vst * 20;
832 	pll = pll * 1000 / 5;
833 	irt = 10 * __BIT(irt);
834 
835 	/*
836 	 * Phase 1.
837 	 */
838 	while (cvid > vid) {
839 
840 		val = 1 << __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_MVS);
841 		val = (val > cvid) ? 0 : cvid - val;
842 
843 		acpicpu_md_pstate_fidvid_write(cfid, val, 1, vst);
844 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
845 
846 		if (rv != 0)
847 			return rv;
848 	}
849 
850 	i = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_RVO);
851 
852 	for (; i > 0 && cvid > 0; --i) {
853 
854 		acpicpu_md_pstate_fidvid_write(cfid, cvid - 1, 1, vst);
855 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
856 
857 		if (rv != 0)
858 			return rv;
859 	}
860 
861 	/*
862 	 * Phase 2.
863 	 */
864 	if (cfid != fid) {
865 
866 		vco_fid  = FID_TO_VCO_FID(fid);
867 		vco_cfid = FID_TO_VCO_FID(cfid);
868 
869 		while (abs(vco_fid - vco_cfid) > 2) {
870 
871 			if (fid <= cfid)
872 				val = cfid - 2;
873 			else {
874 				val = (cfid > 6) ? cfid + 2 :
875 				    FID_TO_VCO_FID(cfid) + 2;
876 			}
877 
878 			acpicpu_md_pstate_fidvid_write(val, cvid, pll, irt);
879 			rv = acpicpu_md_pstate_fidvid_read(&cfid, NULL);
880 
881 			if (rv != 0)
882 				return rv;
883 
884 			vco_cfid = FID_TO_VCO_FID(cfid);
885 		}
886 
887 		acpicpu_md_pstate_fidvid_write(fid, cvid, pll, irt);
888 		rv = acpicpu_md_pstate_fidvid_read(&cfid, NULL);
889 
890 		if (rv != 0)
891 			return rv;
892 	}
893 
894 	/*
895 	 * Phase 3.
896 	 */
897 	if (cvid != vid) {
898 
899 		acpicpu_md_pstate_fidvid_write(cfid, vid, 1, vst);
900 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
901 
902 		if (rv != 0)
903 			return rv;
904 	}
905 
906 	return 0;
907 }
908 
909 static int
910 acpicpu_md_pstate_fidvid_read(uint32_t *cfid, uint32_t *cvid)
911 {
912 	int i = ACPICPU_P_STATE_RETRY * 100;
913 	uint64_t val;
914 
915 	do {
916 		val = rdmsr(MSR_0FH_STATUS);
917 
918 	} while (__SHIFTOUT(val, MSR_0FH_STATUS_PENDING) != 0 && --i >= 0);
919 
920 	if (i == 0)
921 		return EAGAIN;
922 
923 	if (cfid != NULL)
924 		*cfid = __SHIFTOUT(val, MSR_0FH_STATUS_CFID);
925 
926 	if (cvid != NULL)
927 		*cvid = __SHIFTOUT(val, MSR_0FH_STATUS_CVID);
928 
929 	return 0;
930 }
931 
932 static void
933 acpicpu_md_pstate_fidvid_write(uint32_t fid,
934     uint32_t vid, uint32_t cnt, uint32_t tmo)
935 {
936 	uint64_t val = 0;
937 
938 	val |= __SHIFTIN(fid, MSR_0FH_CONTROL_FID);
939 	val |= __SHIFTIN(vid, MSR_0FH_CONTROL_VID);
940 	val |= __SHIFTIN(cnt, MSR_0FH_CONTROL_CNT);
941 	val |= __SHIFTIN(0x1, MSR_0FH_CONTROL_CHG);
942 
943 	wrmsr(MSR_0FH_CONTROL, val);
944 	DELAY(tmo);
945 }
946 
947 int
948 acpicpu_md_tstate_get(struct acpicpu_softc *sc, uint32_t *percent)
949 {
950 	struct acpicpu_tstate *ts;
951 	uint64_t val;
952 	uint32_t i;
953 
954 	val = rdmsr(MSR_THERM_CONTROL);
955 
956 	for (i = 0; i < sc->sc_tstate_count; i++) {
957 
958 		ts = &sc->sc_tstate[i];
959 
960 		if (ts->ts_percent == 0)
961 			continue;
962 
963 		if (val == ts->ts_status) {
964 			*percent = ts->ts_percent;
965 			return 0;
966 		}
967 	}
968 
969 	return EIO;
970 }
971 
972 int
973 acpicpu_md_tstate_set(struct acpicpu_tstate *ts)
974 {
975 	uint64_t val;
976 	uint8_t i;
977 
978 	val = ts->ts_control;
979 	val = val & __BITS(0, 4);
980 
981 	wrmsr(MSR_THERM_CONTROL, val);
982 
983 	if (ts->ts_status == 0) {
984 		DELAY(ts->ts_latency);
985 		return 0;
986 	}
987 
988 	for (i = val = 0; i < ACPICPU_T_STATE_RETRY; i++) {
989 
990 		val = rdmsr(MSR_THERM_CONTROL);
991 
992 		if (val == ts->ts_status)
993 			return 0;
994 
995 		DELAY(ts->ts_latency);
996 	}
997 
998 	return EAGAIN;
999 }
1000 
1001 /*
1002  * A kludge for backwards compatibility.
1003  */
1004 static int
1005 acpicpu_md_pstate_sysctl_init(void)
1006 {
1007 	const struct sysctlnode	*fnode, *mnode, *rnode;
1008 	const char *str;
1009 	int rv;
1010 
1011 	switch (cpu_vendor) {
1012 
1013 	case CPUVENDOR_IDT:
1014 	case CPUVENDOR_INTEL:
1015 		str = "est";
1016 		break;
1017 
1018 	case CPUVENDOR_AMD:
1019 		str = "powernow";
1020 		break;
1021 
1022 	default:
1023 		return ENODEV;
1024 	}
1025 
1026 
1027 	rv = sysctl_createv(&acpicpu_log, 0, NULL, &rnode,
1028 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "machdep", NULL,
1029 	    NULL, 0, NULL, 0, CTL_MACHDEP, CTL_EOL);
1030 
1031 	if (rv != 0)
1032 		goto fail;
1033 
1034 	rv = sysctl_createv(&acpicpu_log, 0, &rnode, &mnode,
1035 	    0, CTLTYPE_NODE, str, NULL,
1036 	    NULL, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1037 
1038 	if (rv != 0)
1039 		goto fail;
1040 
1041 	rv = sysctl_createv(&acpicpu_log, 0, &mnode, &fnode,
1042 	    0, CTLTYPE_NODE, "frequency", 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, &fnode, &rnode,
1049 	    CTLFLAG_READWRITE, CTLTYPE_INT, "target", NULL,
1050 	    acpicpu_md_pstate_sysctl_set, 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_READONLY, CTLTYPE_INT, "current", NULL,
1057 	    acpicpu_md_pstate_sysctl_get, 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_STRING, "available", NULL,
1064 	    acpicpu_md_pstate_sysctl_all, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1065 
1066 	if (rv != 0)
1067 		goto fail;
1068 
1069 	return 0;
1070 
1071 fail:
1072 	if (acpicpu_log != NULL) {
1073 		sysctl_teardown(&acpicpu_log);
1074 		acpicpu_log = NULL;
1075 	}
1076 
1077 	return rv;
1078 }
1079 
1080 static int
1081 acpicpu_md_pstate_sysctl_get(SYSCTLFN_ARGS)
1082 {
1083 	struct sysctlnode node;
1084 	uint32_t freq;
1085 	int err;
1086 
1087 	freq = cpufreq_get(curcpu());
1088 
1089 	if (freq == 0)
1090 		return ENXIO;
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 	return 0;
1101 }
1102 
1103 static int
1104 acpicpu_md_pstate_sysctl_set(SYSCTLFN_ARGS)
1105 {
1106 	struct sysctlnode node;
1107 	uint32_t freq;
1108 	int err;
1109 
1110 	freq = cpufreq_get(curcpu());
1111 
1112 	if (freq == 0)
1113 		return ENXIO;
1114 
1115 	node = *rnode;
1116 	node.sysctl_data = &freq;
1117 
1118 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1119 
1120 	if (err != 0 || newp == NULL)
1121 		return err;
1122 
1123 	cpufreq_set_all(freq);
1124 
1125 	return 0;
1126 }
1127 
1128 static int
1129 acpicpu_md_pstate_sysctl_all(SYSCTLFN_ARGS)
1130 {
1131 	struct cpu_info *ci = curcpu();
1132 	struct acpicpu_softc *sc;
1133 	struct sysctlnode node;
1134 	char buf[1024];
1135 	size_t len;
1136 	uint32_t i;
1137 	int err;
1138 
1139 	sc = acpicpu_sc[ci->ci_acpiid];
1140 
1141 	if (sc == NULL)
1142 		return ENXIO;
1143 
1144 	(void)memset(&buf, 0, sizeof(buf));
1145 
1146 	mutex_enter(&sc->sc_mtx);
1147 
1148 	for (len = 0, i = sc->sc_pstate_max; i < sc->sc_pstate_count; i++) {
1149 
1150 		if (sc->sc_pstate[i].ps_freq == 0)
1151 			continue;
1152 
1153 		if (len >= sizeof(buf))
1154 			break;
1155 		len += snprintf(buf + len, sizeof(buf) - len, "%u%s",
1156 		    sc->sc_pstate[i].ps_freq,
1157 		    i < (sc->sc_pstate_count - 1) ? " " : "");
1158 	}
1159 
1160 	mutex_exit(&sc->sc_mtx);
1161 
1162 	node = *rnode;
1163 	node.sysctl_data = buf;
1164 
1165 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1166 
1167 	if (err != 0 || newp == NULL)
1168 		return err;
1169 
1170 	return 0;
1171 }
1172 
1173