xref: /netbsd-src/sys/arch/x86/acpi/acpi_cpu_md.c (revision 46f5119e40af2e51998f686b2fdcc76b5488f7f3)
1 /* $NetBSD: acpi_cpu_md.c,v 1.58 2011/04/04 20:37:55 dyoung 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 1.58 2011/04/04 20:37:55 dyoung 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_cap(void)
145 {
146 	struct cpu_info *ci = curcpu();
147 	uint32_t regs[4];
148 	uint32_t val = 0;
149 
150 	if (cpu_vendor != CPUVENDOR_IDT &&
151 	    cpu_vendor != CPUVENDOR_INTEL)
152 		return val;
153 
154 	/*
155 	 * Basic SMP C-states (required for e.g. _CST).
156 	 */
157 	val |= ACPICPU_PDC_C_C1PT | ACPICPU_PDC_C_C2C3;
158 
159 	/*
160 	 * Claim to support dependency coordination.
161 	 */
162 	val |= ACPICPU_PDC_P_SW | ACPICPU_PDC_C_SW | ACPICPU_PDC_T_SW;
163 
164         /*
165 	 * If MONITOR/MWAIT is available, announce
166 	 * support for native instructions in all C-states.
167 	 */
168         if ((ci->ci_feat_val[1] & CPUID2_MONITOR) != 0)
169 		val |= ACPICPU_PDC_C_C1_FFH | ACPICPU_PDC_C_C2C3_FFH;
170 
171 	/*
172 	 * Set native P- and T-states, if available.
173 	 */
174         if ((ci->ci_feat_val[1] & CPUID2_EST) != 0)
175 		val |= ACPICPU_PDC_P_FFH;
176 
177 	if ((ci->ci_feat_val[0] & CPUID_ACPI) != 0)
178 		val |= ACPICPU_PDC_T_FFH;
179 
180 	/*
181 	 * Declare support for APERF and MPERF.
182 	 */
183 	if (cpuid_level >= 0x06) {
184 
185 		x86_cpuid(0x00000006, regs);
186 
187 		if ((regs[2] & CPUID_DSPM_HWF) != 0)
188 			val |= ACPICPU_PDC_P_HWF;
189 	}
190 
191 	return val;
192 }
193 
194 uint32_t
195 acpicpu_md_flags(void)
196 {
197 	struct cpu_info *ci = curcpu();
198 	struct pci_attach_args pa;
199 	uint32_t family, val = 0;
200 	uint32_t regs[4];
201 
202 	if (acpi_md_ncpus() == 1)
203 		val |= ACPICPU_FLAG_C_BM;
204 
205 	if ((ci->ci_feat_val[1] & CPUID2_MONITOR) != 0)
206 		val |= ACPICPU_FLAG_C_FFH;
207 
208 	/*
209 	 * By default, assume that the local APIC timer
210 	 * as well as TSC are stalled during C3 sleep.
211 	 */
212 	val |= ACPICPU_FLAG_C_APIC | ACPICPU_FLAG_C_TSC;
213 
214 	switch (cpu_vendor) {
215 
216 	case CPUVENDOR_IDT:
217 
218 		if ((ci->ci_feat_val[1] & CPUID2_EST) != 0)
219 			val |= ACPICPU_FLAG_P_FFH;
220 
221 		if ((ci->ci_feat_val[0] & CPUID_ACPI) != 0)
222 			val |= ACPICPU_FLAG_T_FFH;
223 
224 		break;
225 
226 	case CPUVENDOR_INTEL:
227 
228 		/*
229 		 * Bus master control and arbitration should be
230 		 * available on all supported Intel CPUs (to be
231 		 * sure, this is double-checked later from the
232 		 * firmware data). These flags imply that it is
233 		 * not necessary to flush caches before C3 state.
234 		 */
235 		val |= ACPICPU_FLAG_C_BM | ACPICPU_FLAG_C_ARB;
236 
237 		/*
238 		 * Check if we can use "native", MSR-based,
239 		 * access. If not, we have to resort to I/O.
240 		 */
241 		if ((ci->ci_feat_val[1] & CPUID2_EST) != 0)
242 			val |= ACPICPU_FLAG_P_FFH;
243 
244 		if ((ci->ci_feat_val[0] & CPUID_ACPI) != 0)
245 			val |= ACPICPU_FLAG_T_FFH;
246 
247 		/*
248 		 * Check whether MSR_APERF, MSR_MPERF, and Turbo
249 		 * Boost are available. Also see if we might have
250 		 * an invariant local APIC timer ("ARAT").
251 		 */
252 		if (cpuid_level >= 0x06) {
253 
254 			x86_cpuid(0x00000006, regs);
255 
256 			if ((regs[2] & CPUID_DSPM_HWF) != 0)
257 				val |= ACPICPU_FLAG_P_HWF;
258 
259 			if ((regs[0] & CPUID_DSPM_IDA) != 0)
260 				val |= ACPICPU_FLAG_P_TURBO;
261 
262 			if ((regs[0] & CPUID_DSPM_ARAT) != 0)
263 				val &= ~ACPICPU_FLAG_C_APIC;
264 		}
265 
266 		/*
267 		 * Detect whether TSC is invariant. If it is not,
268 		 * we keep the flag to note that TSC will not run
269 		 * at constant rate. Depending on the CPU, this may
270 		 * affect P- and T-state changes, but especially
271 		 * relevant are C-states; with variant TSC, states
272 		 * larger than C1 may completely stop the counter.
273 		 */
274 		x86_cpuid(0x80000000, regs);
275 
276 		if (regs[0] >= 0x80000007) {
277 
278 			x86_cpuid(0x80000007, regs);
279 
280 			if ((regs[3] & __BIT(8)) != 0)
281 				val &= ~ACPICPU_FLAG_C_TSC;
282 		}
283 
284 		break;
285 
286 	case CPUVENDOR_AMD:
287 
288 		x86_cpuid(0x80000000, regs);
289 
290 		if (regs[0] < 0x80000007)
291 			break;
292 
293 		x86_cpuid(0x80000007, regs);
294 
295 		family = CPUID2FAMILY(ci->ci_signature);
296 
297 		if (family == 0xf)
298 			family += CPUID2EXTFAMILY(ci->ci_signature);
299 
300     		switch (family) {
301 
302 		case 0x0f:
303 
304 			/*
305 			 * Evaluate support for the "FID/VID
306 			 * algorithm" also used by powernow(4).
307 			 */
308 			if ((regs[3] & CPUID_APM_FID) == 0)
309 				break;
310 
311 			if ((regs[3] & CPUID_APM_VID) == 0)
312 				break;
313 
314 			val |= ACPICPU_FLAG_P_FFH | ACPICPU_FLAG_P_FIDVID;
315 			break;
316 
317 		case 0x10:
318 		case 0x11:
319 			val |= ACPICPU_FLAG_C_C1E;
320 			/* FALLTHROUGH */
321 
322 		case 0x14: /* AMD Fusion */
323 
324 			/*
325 			 * Like with Intel, detect invariant TSC,
326 			 * MSR-based P-states, and AMD's "turbo"
327 			 * (Core Performance Boost), respectively.
328 			 */
329 			if ((regs[3] & CPUID_APM_TSC) != 0)
330 				val &= ~ACPICPU_FLAG_C_TSC;
331 
332 			if ((regs[3] & CPUID_APM_HWP) != 0)
333 				val |= ACPICPU_FLAG_P_FFH;
334 
335 			if ((regs[3] & CPUID_APM_CPB) != 0)
336 				val |= ACPICPU_FLAG_P_TURBO;
337 
338 			/*
339 			 * Also check for APERF and MPERF,
340 			 * first available in the family 10h.
341 			 */
342 			if (cpuid_level >= 0x06) {
343 
344 				x86_cpuid(0x00000006, regs);
345 
346 				if ((regs[2] & CPUID_DSPM_HWF) != 0)
347 					val |= ACPICPU_FLAG_P_HWF;
348 			}
349 
350 			break;
351 		}
352 
353 		break;
354 	}
355 
356 	/*
357 	 * There are several erratums for PIIX4.
358 	 */
359 	if (pci_find_device(&pa, acpicpu_md_quirk_piix4) != 0)
360 		val |= ACPICPU_FLAG_PIIX4;
361 
362 	return val;
363 }
364 
365 static int
366 acpicpu_md_quirk_piix4(const struct pci_attach_args *pa)
367 {
368 
369 	/*
370 	 * XXX: The pci_find_device(9) function only
371 	 *	deals with attached devices. Change this
372 	 *	to use something like pci_device_foreach().
373 	 */
374 	if (PCI_VENDOR(pa->pa_id) != PCI_VENDOR_INTEL)
375 		return 0;
376 
377 	if (PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_INTEL_82371AB_ISA ||
378 	    PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_INTEL_82440MX_PMC)
379 		return 1;
380 
381 	return 0;
382 }
383 
384 void
385 acpicpu_md_quirk_c1e(void)
386 {
387 	const uint64_t c1e = MSR_CMPHALT_SMI | MSR_CMPHALT_C1E;
388 	uint64_t val;
389 
390 	val = rdmsr(MSR_CMPHALT);
391 
392 	if ((val & c1e) != 0)
393 		wrmsr(MSR_CMPHALT, val & ~c1e);
394 }
395 
396 int
397 acpicpu_md_cstate_start(struct acpicpu_softc *sc)
398 {
399 	const size_t size = sizeof(native_idle_text);
400 	struct acpicpu_cstate *cs;
401 	bool ipi = false;
402 	int i;
403 
404 	/*
405 	 * Save the cpu_idle(9) loop used by default.
406 	 */
407 	x86_cpu_idle_get(&native_idle, native_idle_text, size);
408 
409 	for (i = 0; i < ACPI_C_STATE_COUNT; i++) {
410 
411 		cs = &sc->sc_cstate[i];
412 
413 		if (cs->cs_method == ACPICPU_C_STATE_HALT) {
414 			ipi = true;
415 			break;
416 		}
417 	}
418 
419 	x86_cpu_idle_set(acpicpu_cstate_idle, "acpi", ipi);
420 
421 	return 0;
422 }
423 
424 int
425 acpicpu_md_cstate_stop(void)
426 {
427 	uint64_t xc;
428 	bool ipi;
429 
430 	ipi = (native_idle != x86_cpu_idle_halt) ? false : true;
431 	x86_cpu_idle_set(native_idle, native_idle_text, ipi);
432 
433 	/*
434 	 * Run a cross-call to ensure that all CPUs are
435 	 * out from the ACPI idle-loop before detachment.
436 	 */
437 	xc = xc_broadcast(0, (xcfunc_t)nullop, NULL, NULL);
438 	xc_wait(xc);
439 
440 	return 0;
441 }
442 
443 /*
444  * Called with interrupts disabled.
445  * Caller should enable interrupts after return.
446  */
447 void
448 acpicpu_md_cstate_enter(int method, int state)
449 {
450 	struct cpu_info *ci = curcpu();
451 
452 	switch (method) {
453 
454 	case ACPICPU_C_STATE_FFH:
455 
456 		x86_enable_intr();
457 		x86_monitor(&ci->ci_want_resched, 0, 0);
458 
459 		if (__predict_false(ci->ci_want_resched != 0))
460 			return;
461 
462 		x86_mwait((state - 1) << 4, 0);
463 		break;
464 
465 	case ACPICPU_C_STATE_HALT:
466 
467 		if (__predict_false(ci->ci_want_resched != 0))
468 			return;
469 
470 		x86_stihlt();
471 		break;
472 	}
473 }
474 
475 int
476 acpicpu_md_pstate_start(struct acpicpu_softc *sc)
477 {
478 	uint64_t xc;
479 
480 	/*
481 	 * Reset the APERF and MPERF counters.
482 	 */
483 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
484 		xc = xc_broadcast(0, acpicpu_md_pstate_hwf_reset, NULL, NULL);
485 		xc_wait(xc);
486 	}
487 
488 	return acpicpu_md_pstate_sysctl_init();
489 }
490 
491 int
492 acpicpu_md_pstate_stop(void)
493 {
494 	if (acpicpu_log != NULL)
495 		sysctl_teardown(&acpicpu_log);
496 
497 	return 0;
498 }
499 
500 int
501 acpicpu_md_pstate_init(struct acpicpu_softc *sc)
502 {
503 	struct cpu_info *ci = sc->sc_ci;
504 	struct acpicpu_pstate *ps, msr;
505 	uint32_t family, i = 0;
506 	uint64_t val;
507 
508 	(void)memset(&msr, 0, sizeof(struct acpicpu_pstate));
509 
510 	switch (cpu_vendor) {
511 
512 	case CPUVENDOR_IDT:
513 	case CPUVENDOR_INTEL:
514 
515 		/*
516 		 * Make sure EST is enabled.
517 		 */
518 		if ((sc->sc_flags & ACPICPU_FLAG_P_FFH) != 0) {
519 
520 			val = rdmsr(MSR_MISC_ENABLE);
521 
522 			if ((val & MSR_MISC_ENABLE_EST) == 0) {
523 
524 				val |= MSR_MISC_ENABLE_EST;
525 				wrmsr(MSR_MISC_ENABLE, val);
526 				val = rdmsr(MSR_MISC_ENABLE);
527 
528 				if ((val & MSR_MISC_ENABLE_EST) == 0)
529 					return ENOTTY;
530 			}
531 		}
532 
533 		/*
534 		 * If the so-called Turbo Boost is present,
535 		 * the P0-state is always the "turbo state".
536 		 * It is shown as the P1 frequency + 1 MHz.
537 		 *
538 		 * For discussion, see:
539 		 *
540 		 *	Intel Corporation: Intel Turbo Boost Technology
541 		 *	in Intel Core(tm) Microarchitectures (Nehalem)
542 		 *	Based Processors. White Paper, November 2008.
543 		 */
544 		if (sc->sc_pstate_count >= 2 &&
545 		   (sc->sc_flags & ACPICPU_FLAG_P_TURBO) != 0) {
546 
547 			ps = &sc->sc_pstate[0];
548 
549 			if (ps->ps_freq == sc->sc_pstate[1].ps_freq + 1)
550 				ps->ps_flags |= ACPICPU_FLAG_P_TURBO;
551 		}
552 
553 		msr.ps_control_addr = MSR_PERF_CTL;
554 		msr.ps_control_mask = __BITS(0, 15);
555 
556 		msr.ps_status_addr  = MSR_PERF_STATUS;
557 		msr.ps_status_mask  = __BITS(0, 15);
558 		break;
559 
560 	case CPUVENDOR_AMD:
561 
562 		if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
563 			msr.ps_flags |= ACPICPU_FLAG_P_FIDVID;
564 
565 		family = CPUID2FAMILY(ci->ci_signature);
566 
567 		if (family == 0xf)
568 			family += CPUID2EXTFAMILY(ci->ci_signature);
569 
570 		switch (family) {
571 
572 		case 0x0f:
573 			msr.ps_control_addr = MSR_0FH_CONTROL;
574 			msr.ps_status_addr  = MSR_0FH_STATUS;
575 			break;
576 
577 		case 0x10:
578 		case 0x11:
579 		case 0x14: /* AMD Fusion */
580 			msr.ps_control_addr = MSR_10H_CONTROL;
581 			msr.ps_control_mask = __BITS(0, 2);
582 
583 			msr.ps_status_addr  = MSR_10H_STATUS;
584 			msr.ps_status_mask  = __BITS(0, 2);
585 			break;
586 
587 		default:
588 			/*
589 			 * If we have an unknown AMD CPU, rely on XPSS.
590 			 */
591 			if ((sc->sc_flags & ACPICPU_FLAG_P_XPSS) == 0)
592 				return EOPNOTSUPP;
593 		}
594 
595 		break;
596 
597 	default:
598 		return ENODEV;
599 	}
600 
601 	/*
602 	 * Fill the P-state structures with MSR addresses that are
603 	 * known to be correct. If we do not know the addresses,
604 	 * leave the values intact. If a vendor uses XPSS, we do
605 	 * not necessarily need to do anything to support new CPUs.
606 	 */
607 	while (i < sc->sc_pstate_count) {
608 
609 		ps = &sc->sc_pstate[i];
610 
611 		if (msr.ps_flags != 0)
612 			ps->ps_flags |= msr.ps_flags;
613 
614 		if (msr.ps_status_addr != 0)
615 			ps->ps_status_addr = msr.ps_status_addr;
616 
617 		if (msr.ps_status_mask != 0)
618 			ps->ps_status_mask = msr.ps_status_mask;
619 
620 		if (msr.ps_control_addr != 0)
621 			ps->ps_control_addr = msr.ps_control_addr;
622 
623 		if (msr.ps_control_mask != 0)
624 			ps->ps_control_mask = msr.ps_control_mask;
625 
626 		i++;
627 	}
628 
629 	return 0;
630 }
631 
632 /*
633  * Read the IA32_APERF and IA32_MPERF counters. The first
634  * increments at the rate of the fixed maximum frequency
635  * configured during the boot, whereas APERF counts at the
636  * rate of the actual frequency. Note that the MSRs must be
637  * read without delay, and that only the ratio between
638  * IA32_APERF and IA32_MPERF is architecturally defined.
639  *
640  * The function thus returns the percentage of the actual
641  * frequency in terms of the maximum frequency of the calling
642  * CPU since the last call. A value zero implies an error.
643  *
644  * For further details, refer to:
645  *
646  *	Intel Corporation: Intel 64 and IA-32 Architectures
647  *	Software Developer's Manual. Section 13.2, Volume 3A:
648  *	System Programming Guide, Part 1. July, 2008.
649  *
650  *	Advanced Micro Devices: BIOS and Kernel Developer's
651  *	Guide (BKDG) for AMD Family 10h Processors. Section
652  *	2.4.5, Revision 3.48, April 2010.
653  */
654 uint8_t
655 acpicpu_md_pstate_hwf(struct cpu_info *ci)
656 {
657 	struct acpicpu_softc *sc;
658 	uint64_t aperf, mperf;
659 	uint8_t rv = 0;
660 
661 	sc = acpicpu_sc[ci->ci_acpiid];
662 
663 	if (__predict_false(sc == NULL))
664 		return 0;
665 
666 	if (__predict_false((sc->sc_flags & ACPICPU_FLAG_P_HWF) == 0))
667 		return 0;
668 
669 	aperf = sc->sc_pstate_aperf;
670 	mperf = sc->sc_pstate_mperf;
671 
672 	x86_disable_intr();
673 
674 	sc->sc_pstate_aperf = rdmsr(MSR_APERF);
675 	sc->sc_pstate_mperf = rdmsr(MSR_MPERF);
676 
677 	x86_enable_intr();
678 
679 	aperf = sc->sc_pstate_aperf - aperf;
680 	mperf = sc->sc_pstate_mperf - mperf;
681 
682 	if (__predict_true(mperf != 0))
683 		rv = (aperf * 100) / mperf;
684 
685 	return rv;
686 }
687 
688 static void
689 acpicpu_md_pstate_hwf_reset(void *arg1, void *arg2)
690 {
691 	struct cpu_info *ci = curcpu();
692 	struct acpicpu_softc *sc;
693 
694 	sc = acpicpu_sc[ci->ci_acpiid];
695 
696 	if (__predict_false(sc == NULL))
697 		return;
698 
699 	x86_disable_intr();
700 
701 	wrmsr(MSR_APERF, 0);
702 	wrmsr(MSR_MPERF, 0);
703 
704 	x86_enable_intr();
705 
706 	sc->sc_pstate_aperf = 0;
707 	sc->sc_pstate_mperf = 0;
708 }
709 
710 int
711 acpicpu_md_pstate_get(struct acpicpu_softc *sc, uint32_t *freq)
712 {
713 	struct acpicpu_pstate *ps = NULL;
714 	uint64_t val;
715 	uint32_t i;
716 
717 	if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
718 		return acpicpu_md_pstate_fidvid_get(sc, freq);
719 
720 	/*
721 	 * Pick any P-state for the status address.
722 	*/
723 	for (i = 0; i < sc->sc_pstate_count; i++) {
724 
725 		ps = &sc->sc_pstate[i];
726 
727 		if (__predict_true(ps->ps_freq != 0))
728 			break;
729 	}
730 
731 	if (__predict_false(ps == NULL))
732 		return ENODEV;
733 
734 	if (__predict_false(ps->ps_status_addr == 0))
735 		return EINVAL;
736 
737 	val = rdmsr(ps->ps_status_addr);
738 
739 	if (__predict_true(ps->ps_status_mask != 0))
740 		val = val & ps->ps_status_mask;
741 
742 	/*
743 	 * Search for the value from known P-states.
744 	 */
745 	for (i = 0; i < sc->sc_pstate_count; i++) {
746 
747 		ps = &sc->sc_pstate[i];
748 
749 		if (__predict_false(ps->ps_freq == 0))
750 			continue;
751 
752 		if (val == ps->ps_status) {
753 			*freq = ps->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 	if (cfid != fid || cvid != vid)
921 		return EIO;
922 
923 	return 0;
924 }
925 
926 static int
927 acpicpu_md_pstate_fidvid_read(uint32_t *cfid, uint32_t *cvid)
928 {
929 	int i = ACPICPU_P_STATE_RETRY * 100;
930 	uint64_t val;
931 
932 	do {
933 		val = rdmsr(MSR_0FH_STATUS);
934 
935 	} while (__SHIFTOUT(val, MSR_0FH_STATUS_PENDING) != 0 && --i >= 0);
936 
937 	if (i == 0)
938 		return EAGAIN;
939 
940 	if (cfid != NULL)
941 		*cfid = __SHIFTOUT(val, MSR_0FH_STATUS_CFID);
942 
943 	if (cvid != NULL)
944 		*cvid = __SHIFTOUT(val, MSR_0FH_STATUS_CVID);
945 
946 	return 0;
947 }
948 
949 static void
950 acpicpu_md_pstate_fidvid_write(uint32_t fid,
951     uint32_t vid, uint32_t cnt, uint32_t tmo)
952 {
953 	uint64_t val = 0;
954 
955 	val |= __SHIFTIN(fid, MSR_0FH_CONTROL_FID);
956 	val |= __SHIFTIN(vid, MSR_0FH_CONTROL_VID);
957 	val |= __SHIFTIN(cnt, MSR_0FH_CONTROL_CNT);
958 	val |= __SHIFTIN(0x1, MSR_0FH_CONTROL_CHG);
959 
960 	wrmsr(MSR_0FH_CONTROL, val);
961 	DELAY(tmo);
962 }
963 
964 int
965 acpicpu_md_tstate_get(struct acpicpu_softc *sc, uint32_t *percent)
966 {
967 	struct acpicpu_tstate *ts;
968 	uint64_t val;
969 	uint32_t i;
970 
971 	val = rdmsr(MSR_THERM_CONTROL);
972 
973 	for (i = 0; i < sc->sc_tstate_count; i++) {
974 
975 		ts = &sc->sc_tstate[i];
976 
977 		if (ts->ts_percent == 0)
978 			continue;
979 
980 		if (val == ts->ts_status) {
981 			*percent = ts->ts_percent;
982 			return 0;
983 		}
984 	}
985 
986 	return EIO;
987 }
988 
989 int
990 acpicpu_md_tstate_set(struct acpicpu_tstate *ts)
991 {
992 	uint64_t val;
993 	uint8_t i;
994 
995 	val = ts->ts_control;
996 	val = val & __BITS(1, 4);
997 
998 	wrmsr(MSR_THERM_CONTROL, val);
999 
1000 	if (ts->ts_status == 0) {
1001 		DELAY(ts->ts_latency);
1002 		return 0;
1003 	}
1004 
1005 	for (i = val = 0; i < ACPICPU_T_STATE_RETRY; i++) {
1006 
1007 		val = rdmsr(MSR_THERM_CONTROL);
1008 
1009 		if (val == ts->ts_status)
1010 			return 0;
1011 
1012 		DELAY(ts->ts_latency);
1013 	}
1014 
1015 	return EAGAIN;
1016 }
1017 
1018 /*
1019  * A kludge for backwards compatibility.
1020  */
1021 static int
1022 acpicpu_md_pstate_sysctl_init(void)
1023 {
1024 	const struct sysctlnode	*fnode, *mnode, *rnode;
1025 	const char *str;
1026 	int rv;
1027 
1028 	switch (cpu_vendor) {
1029 
1030 	case CPUVENDOR_IDT:
1031 	case CPUVENDOR_INTEL:
1032 		str = "est";
1033 		break;
1034 
1035 	case CPUVENDOR_AMD:
1036 		str = "powernow";
1037 		break;
1038 
1039 	default:
1040 		return ENODEV;
1041 	}
1042 
1043 
1044 	rv = sysctl_createv(&acpicpu_log, 0, NULL, &rnode,
1045 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "machdep", NULL,
1046 	    NULL, 0, NULL, 0, CTL_MACHDEP, CTL_EOL);
1047 
1048 	if (rv != 0)
1049 		goto fail;
1050 
1051 	rv = sysctl_createv(&acpicpu_log, 0, &rnode, &mnode,
1052 	    0, CTLTYPE_NODE, str, NULL,
1053 	    NULL, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1054 
1055 	if (rv != 0)
1056 		goto fail;
1057 
1058 	rv = sysctl_createv(&acpicpu_log, 0, &mnode, &fnode,
1059 	    0, CTLTYPE_NODE, "frequency", NULL,
1060 	    NULL, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1061 
1062 	if (rv != 0)
1063 		goto fail;
1064 
1065 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1066 	    CTLFLAG_READWRITE, CTLTYPE_INT, "target", NULL,
1067 	    acpicpu_md_pstate_sysctl_set, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1068 
1069 	if (rv != 0)
1070 		goto fail;
1071 
1072 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1073 	    CTLFLAG_READONLY, CTLTYPE_INT, "current", NULL,
1074 	    acpicpu_md_pstate_sysctl_get, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1075 
1076 	if (rv != 0)
1077 		goto fail;
1078 
1079 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1080 	    CTLFLAG_READONLY, CTLTYPE_STRING, "available", NULL,
1081 	    acpicpu_md_pstate_sysctl_all, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1082 
1083 	if (rv != 0)
1084 		goto fail;
1085 
1086 	return 0;
1087 
1088 fail:
1089 	if (acpicpu_log != NULL) {
1090 		sysctl_teardown(&acpicpu_log);
1091 		acpicpu_log = NULL;
1092 	}
1093 
1094 	return rv;
1095 }
1096 
1097 static int
1098 acpicpu_md_pstate_sysctl_get(SYSCTLFN_ARGS)
1099 {
1100 	struct cpu_info *ci = curcpu();
1101 	struct sysctlnode node;
1102 	uint32_t freq;
1103 	int err;
1104 
1105 	err = acpicpu_pstate_get(ci, &freq);
1106 
1107 	if (err != 0)
1108 		return err;
1109 
1110 	node = *rnode;
1111 	node.sysctl_data = &freq;
1112 
1113 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1114 
1115 	if (err != 0 || newp == NULL)
1116 		return err;
1117 
1118 	return 0;
1119 }
1120 
1121 static int
1122 acpicpu_md_pstate_sysctl_set(SYSCTLFN_ARGS)
1123 {
1124 	struct cpu_info *ci = curcpu();
1125 	struct sysctlnode node;
1126 	uint32_t freq;
1127 	int err;
1128 
1129 	err = acpicpu_pstate_get(ci, &freq);
1130 
1131 	if (err != 0)
1132 		return err;
1133 
1134 	node = *rnode;
1135 	node.sysctl_data = &freq;
1136 
1137 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1138 
1139 	if (err != 0 || newp == NULL)
1140 		return err;
1141 
1142 	acpicpu_pstate_set(ci, freq);
1143 
1144 	return 0;
1145 }
1146 
1147 static int
1148 acpicpu_md_pstate_sysctl_all(SYSCTLFN_ARGS)
1149 {
1150 	struct cpu_info *ci = curcpu();
1151 	struct acpicpu_softc *sc;
1152 	struct sysctlnode node;
1153 	char buf[1024];
1154 	size_t len;
1155 	uint32_t i;
1156 	int err;
1157 
1158 	sc = acpicpu_sc[ci->ci_acpiid];
1159 
1160 	if (sc == NULL)
1161 		return ENXIO;
1162 
1163 	(void)memset(&buf, 0, sizeof(buf));
1164 
1165 	mutex_enter(&sc->sc_mtx);
1166 
1167 	for (len = 0, i = sc->sc_pstate_max; i < sc->sc_pstate_count; i++) {
1168 
1169 		if (sc->sc_pstate[i].ps_freq == 0)
1170 			continue;
1171 
1172 		len += snprintf(buf + len, sizeof(buf) - len, "%u%s",
1173 		    sc->sc_pstate[i].ps_freq,
1174 		    i < (sc->sc_pstate_count - 1) ? " " : "");
1175 	}
1176 
1177 	mutex_exit(&sc->sc_mtx);
1178 
1179 	node = *rnode;
1180 	node.sysctl_data = buf;
1181 
1182 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1183 
1184 	if (err != 0 || newp == NULL)
1185 		return err;
1186 
1187 	return 0;
1188 }
1189 
1190