xref: /netbsd-src/sys/arch/x86/acpi/acpi_cpu_md.c (revision bdc22b2e01993381dcefeff2bc9b56ca75a4235c)
1 /* $NetBSD: acpi_cpu_md.c,v 1.78 2016/12/08 11:31:12 nat 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 1.78 2016/12/08 11:31:12 nat 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 
229 		break;
230 
231 	case CPUVENDOR_AMD:
232 
233 		x86_cpuid(0x80000000, regs);
234 
235 		if (regs[0] < 0x80000007)
236 			break;
237 
238 		x86_cpuid(0x80000007, regs);
239 
240 		family = CPUID_TO_FAMILY(ci->ci_signature);
241 
242     		switch (family) {
243 
244 		case 0x0f:
245 
246 			/*
247 			 * Disable C1E if present.
248 			 */
249 			if (rdmsr_safe(MSR_CMPHALT, &msr) != EFAULT)
250 				val |= ACPICPU_FLAG_C_C1E;
251 
252 			/*
253 			 * Evaluate support for the "FID/VID
254 			 * algorithm" also used by powernow(4).
255 			 */
256 			if ((regs[3] & CPUID_APM_FID) == 0)
257 				break;
258 
259 			if ((regs[3] & CPUID_APM_VID) == 0)
260 				break;
261 
262 			val |= ACPICPU_FLAG_P_FFH | ACPICPU_FLAG_P_FIDVID;
263 			break;
264 
265 		case 0x10:
266 		case 0x11:
267 
268 			/*
269 			 * Disable C1E if present.
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 MSR-based P-states,
282 			 * and AMD's "turbo" (Core Performance Boost),
283 			 * respectively.
284 			 */
285 			if ((regs[3] & CPUID_APM_HWP) != 0)
286 				val |= ACPICPU_FLAG_P_FFH;
287 
288 			if ((regs[3] & CPUID_APM_CPB) != 0)
289 				val |= ACPICPU_FLAG_P_TURBO;
290 
291 			/*
292 			 * Also check for APERF and MPERF,
293 			 * first available in the family 10h.
294 			 */
295 			if (cpuid_level >= 0x06) {
296 
297 				x86_cpuid(0x00000006, regs);
298 
299 				if ((regs[2] & CPUID_DSPM_HWF) != 0)
300 					val |= ACPICPU_FLAG_P_HWF;
301 			}
302 
303 			break;
304 		}
305 
306 		break;
307 	}
308 
309 	/*
310 	 * There are several erratums for PIIX4.
311 	 */
312 	if (pci_find_device(&pa, acpicpu_md_quirk_piix4) != 0)
313 		val |= ACPICPU_FLAG_PIIX4;
314 
315 	return val;
316 }
317 
318 static int
319 acpicpu_md_quirk_piix4(const struct pci_attach_args *pa)
320 {
321 
322 	/*
323 	 * XXX: The pci_find_device(9) function only
324 	 *	deals with attached devices. Change this
325 	 *	to use something like pci_device_foreach().
326 	 */
327 	if (PCI_VENDOR(pa->pa_id) != PCI_VENDOR_INTEL)
328 		return 0;
329 
330 	if (PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_INTEL_82371AB_ISA ||
331 	    PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_INTEL_82440MX_PMC)
332 		return 1;
333 
334 	return 0;
335 }
336 
337 void
338 acpicpu_md_quirk_c1e(void)
339 {
340 	const uint64_t c1e = MSR_CMPHALT_SMI | MSR_CMPHALT_C1E;
341 	uint64_t val;
342 
343 	val = rdmsr(MSR_CMPHALT);
344 
345 	if ((val & c1e) != 0)
346 		wrmsr(MSR_CMPHALT, val & ~c1e);
347 }
348 
349 int
350 acpicpu_md_cstate_start(struct acpicpu_softc *sc)
351 {
352 	const size_t size = sizeof(native_idle_text);
353 	struct acpicpu_cstate *cs;
354 	bool ipi = false;
355 	int i;
356 
357 	/*
358 	 * Save the cpu_idle(9) loop used by default.
359 	 */
360 	x86_cpu_idle_get(&native_idle, native_idle_text, size);
361 
362 	for (i = 0; i < ACPI_C_STATE_COUNT; i++) {
363 
364 		cs = &sc->sc_cstate[i];
365 
366 		if (cs->cs_method == ACPICPU_C_STATE_HALT) {
367 			ipi = true;
368 			break;
369 		}
370 	}
371 
372 	x86_cpu_idle_set(acpicpu_cstate_idle, "acpi", ipi);
373 
374 	return 0;
375 }
376 
377 int
378 acpicpu_md_cstate_stop(void)
379 {
380 	static char text[16];
381 	void (*func)(void);
382 	uint64_t xc;
383 	bool ipi;
384 
385 	x86_cpu_idle_get(&func, text, sizeof(text));
386 
387 	if (func == native_idle)
388 		return EALREADY;
389 
390 	ipi = (native_idle != x86_cpu_idle_halt) ? false : true;
391 	x86_cpu_idle_set(native_idle, native_idle_text, ipi);
392 
393 	/*
394 	 * Run a cross-call to ensure that all CPUs are
395 	 * out from the ACPI idle-loop before detachment.
396 	 */
397 	xc = xc_broadcast(0, (xcfunc_t)nullop, NULL, NULL);
398 	xc_wait(xc);
399 
400 	return 0;
401 }
402 
403 /*
404  * Called with interrupts enabled.
405  */
406 void
407 acpicpu_md_cstate_enter(int method, int state)
408 {
409 	struct cpu_info *ci = curcpu();
410 
411 	KASSERT(ci->ci_ilevel == IPL_NONE);
412 
413 	switch (method) {
414 
415 	case ACPICPU_C_STATE_FFH:
416 
417 		x86_monitor(&ci->ci_want_resched, 0, 0);
418 
419 		if (__predict_false(ci->ci_want_resched != 0))
420 			return;
421 
422 		x86_mwait((state - 1) << 4, 0);
423 		break;
424 
425 	case ACPICPU_C_STATE_HALT:
426 
427 		x86_disable_intr();
428 
429 		if (__predict_false(ci->ci_want_resched != 0)) {
430 			x86_enable_intr();
431 			return;
432 		}
433 
434 		x86_stihlt();
435 		break;
436 	}
437 }
438 
439 int
440 acpicpu_md_pstate_start(struct acpicpu_softc *sc)
441 {
442 	uint64_t xc, val;
443 
444 	switch (cpu_vendor) {
445 
446 	case CPUVENDOR_IDT:
447 	case CPUVENDOR_INTEL:
448 
449 		/*
450 		 * Make sure EST is enabled.
451 		 */
452 		if ((sc->sc_flags & ACPICPU_FLAG_P_FFH) != 0) {
453 
454 			val = rdmsr(MSR_MISC_ENABLE);
455 
456 			if ((val & MSR_MISC_ENABLE_EST) == 0) {
457 
458 				val |= MSR_MISC_ENABLE_EST;
459 				wrmsr(MSR_MISC_ENABLE, val);
460 				val = rdmsr(MSR_MISC_ENABLE);
461 
462 				if ((val & MSR_MISC_ENABLE_EST) == 0)
463 					return ENOTTY;
464 			}
465 		}
466 	}
467 
468 	/*
469 	 * Reset the APERF and MPERF counters.
470 	 */
471 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
472 		xc = xc_broadcast(0, acpicpu_md_pstate_hwf_reset, NULL, NULL);
473 		xc_wait(xc);
474 	}
475 
476 	return acpicpu_md_pstate_sysctl_init();
477 }
478 
479 int
480 acpicpu_md_pstate_stop(void)
481 {
482 
483 	if (acpicpu_log == NULL)
484 		return EALREADY;
485 
486 	sysctl_teardown(&acpicpu_log);
487 	acpicpu_log = NULL;
488 
489 	return 0;
490 }
491 
492 int
493 acpicpu_md_pstate_init(struct acpicpu_softc *sc)
494 {
495 	struct cpu_info *ci = sc->sc_ci;
496 	struct acpicpu_pstate *ps, msr;
497 	uint32_t family, i = 0;
498 
499 	(void)memset(&msr, 0, sizeof(struct acpicpu_pstate));
500 
501 	switch (cpu_vendor) {
502 
503 	case CPUVENDOR_IDT:
504 	case CPUVENDOR_INTEL:
505 
506 		/*
507 		 * If the so-called Turbo Boost is present,
508 		 * the P0-state is always the "turbo state".
509 		 * It is shown as the P1 frequency + 1 MHz.
510 		 *
511 		 * For discussion, see:
512 		 *
513 		 *	Intel Corporation: Intel Turbo Boost Technology
514 		 *	in Intel Core(tm) Microarchitectures (Nehalem)
515 		 *	Based Processors. White Paper, November 2008.
516 		 */
517 		if (sc->sc_pstate_count >= 2 &&
518 		   (sc->sc_flags & ACPICPU_FLAG_P_TURBO) != 0) {
519 
520 			ps = &sc->sc_pstate[0];
521 
522 			if (ps->ps_freq == sc->sc_pstate[1].ps_freq + 1)
523 				ps->ps_flags |= ACPICPU_FLAG_P_TURBO;
524 		}
525 
526 		msr.ps_control_addr = MSR_PERF_CTL;
527 		msr.ps_control_mask = __BITS(0, 15);
528 
529 		msr.ps_status_addr  = MSR_PERF_STATUS;
530 		msr.ps_status_mask  = __BITS(0, 15);
531 		break;
532 
533 	case CPUVENDOR_AMD:
534 
535 		if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
536 			msr.ps_flags |= ACPICPU_FLAG_P_FIDVID;
537 
538 		family = CPUID_TO_FAMILY(ci->ci_signature);
539 
540 		switch (family) {
541 
542 		case 0x0f:
543 			msr.ps_control_addr = MSR_0FH_CONTROL;
544 			msr.ps_status_addr  = MSR_0FH_STATUS;
545 			break;
546 
547 		case 0x10:
548 		case 0x11:
549 		case 0x12:
550 		case 0x14:
551 		case 0x15:
552 			msr.ps_control_addr = MSR_10H_CONTROL;
553 			msr.ps_control_mask = __BITS(0, 2);
554 
555 			msr.ps_status_addr  = MSR_10H_STATUS;
556 			msr.ps_status_mask  = __BITS(0, 2);
557 			break;
558 
559 		default:
560 			/*
561 			 * If we have an unknown AMD CPU, rely on XPSS.
562 			 */
563 			if ((sc->sc_flags & ACPICPU_FLAG_P_XPSS) == 0)
564 				return EOPNOTSUPP;
565 		}
566 
567 		break;
568 
569 	default:
570 		return ENODEV;
571 	}
572 
573 	/*
574 	 * Fill the P-state structures with MSR addresses that are
575 	 * known to be correct. If we do not know the addresses,
576 	 * leave the values intact. If a vendor uses XPSS, we do
577 	 * not necessarily need to do anything to support new CPUs.
578 	 */
579 	while (i < sc->sc_pstate_count) {
580 
581 		ps = &sc->sc_pstate[i];
582 
583 		if (msr.ps_flags != 0)
584 			ps->ps_flags |= msr.ps_flags;
585 
586 		if (msr.ps_status_addr != 0)
587 			ps->ps_status_addr = msr.ps_status_addr;
588 
589 		if (msr.ps_status_mask != 0)
590 			ps->ps_status_mask = msr.ps_status_mask;
591 
592 		if (msr.ps_control_addr != 0)
593 			ps->ps_control_addr = msr.ps_control_addr;
594 
595 		if (msr.ps_control_mask != 0)
596 			ps->ps_control_mask = msr.ps_control_mask;
597 
598 		i++;
599 	}
600 
601 	return 0;
602 }
603 
604 /*
605  * Read the IA32_APERF and IA32_MPERF counters. The first
606  * increments at the rate of the fixed maximum frequency
607  * configured during the boot, whereas APERF counts at the
608  * rate of the actual frequency. Note that the MSRs must be
609  * read without delay, and that only the ratio between
610  * IA32_APERF and IA32_MPERF is architecturally defined.
611  *
612  * The function thus returns the percentage of the actual
613  * frequency in terms of the maximum frequency of the calling
614  * CPU since the last call. A value zero implies an error.
615  *
616  * For further details, refer to:
617  *
618  *	Intel Corporation: Intel 64 and IA-32 Architectures
619  *	Software Developer's Manual. Section 13.2, Volume 3A:
620  *	System Programming Guide, Part 1. July, 2008.
621  *
622  *	Advanced Micro Devices: BIOS and Kernel Developer's
623  *	Guide (BKDG) for AMD Family 10h Processors. Section
624  *	2.4.5, Revision 3.48, April 2010.
625  */
626 uint8_t
627 acpicpu_md_pstate_hwf(struct cpu_info *ci)
628 {
629 	struct acpicpu_softc *sc;
630 	uint64_t aperf, mperf;
631 	uint8_t rv = 0;
632 
633 	sc = acpicpu_sc[ci->ci_acpiid];
634 
635 	if (__predict_false(sc == NULL))
636 		return 0;
637 
638 	if (__predict_false((sc->sc_flags & ACPICPU_FLAG_P_HWF) == 0))
639 		return 0;
640 
641 	aperf = sc->sc_pstate_aperf;
642 	mperf = sc->sc_pstate_mperf;
643 
644 	x86_disable_intr();
645 
646 	sc->sc_pstate_aperf = rdmsr(MSR_APERF);
647 	sc->sc_pstate_mperf = rdmsr(MSR_MPERF);
648 
649 	x86_enable_intr();
650 
651 	aperf = sc->sc_pstate_aperf - aperf;
652 	mperf = sc->sc_pstate_mperf - mperf;
653 
654 	if (__predict_true(mperf != 0))
655 		rv = (aperf * 100) / mperf;
656 
657 	return rv;
658 }
659 
660 static void
661 acpicpu_md_pstate_hwf_reset(void *arg1, void *arg2)
662 {
663 	struct cpu_info *ci = curcpu();
664 	struct acpicpu_softc *sc;
665 
666 	sc = acpicpu_sc[ci->ci_acpiid];
667 
668 	if (__predict_false(sc == NULL))
669 		return;
670 
671 	x86_disable_intr();
672 
673 	wrmsr(MSR_APERF, 0);
674 	wrmsr(MSR_MPERF, 0);
675 
676 	x86_enable_intr();
677 
678 	sc->sc_pstate_aperf = 0;
679 	sc->sc_pstate_mperf = 0;
680 }
681 
682 int
683 acpicpu_md_pstate_get(struct acpicpu_softc *sc, uint32_t *freq)
684 {
685 	struct acpicpu_pstate *ps = NULL;
686 	uint64_t val;
687 	uint32_t i;
688 
689 	if ((sc->sc_flags & ACPICPU_FLAG_P_FIDVID) != 0)
690 		return acpicpu_md_pstate_fidvid_get(sc, freq);
691 
692 	/*
693 	 * Pick any P-state for the status address.
694 	 */
695 	for (i = 0; i < sc->sc_pstate_count; i++) {
696 
697 		ps = &sc->sc_pstate[i];
698 
699 		if (__predict_true(ps->ps_freq != 0))
700 			break;
701 	}
702 
703 	if (__predict_false(ps == NULL))
704 		return ENODEV;
705 
706 	if (__predict_false(ps->ps_status_addr == 0))
707 		return EINVAL;
708 
709 	val = rdmsr(ps->ps_status_addr);
710 
711 	if (__predict_true(ps->ps_status_mask != 0))
712 		val = val & ps->ps_status_mask;
713 
714 	/*
715 	 * Search for the value from known P-states.
716 	 */
717 	for (i = 0; i < sc->sc_pstate_count; i++) {
718 
719 		ps = &sc->sc_pstate[i];
720 
721 		if (__predict_false(ps->ps_freq == 0))
722 			continue;
723 
724 		if (val == ps->ps_status) {
725 			*freq = ps->ps_freq;
726 			return 0;
727 		}
728 	}
729 
730 	/*
731 	 * If the value was not found, try APERF/MPERF.
732 	 * The state is P0 if the return value is 100 %.
733 	 */
734 	if ((sc->sc_flags & ACPICPU_FLAG_P_HWF) != 0) {
735 
736 		KASSERT(sc->sc_pstate_count > 0);
737 		KASSERT(sc->sc_pstate[0].ps_freq != 0);
738 
739 		if (acpicpu_md_pstate_hwf(sc->sc_ci) == 100) {
740 			*freq = sc->sc_pstate[0].ps_freq;
741 			return 0;
742 		}
743 	}
744 
745 	return EIO;
746 }
747 
748 int
749 acpicpu_md_pstate_set(struct acpicpu_pstate *ps)
750 {
751 	uint64_t val = 0;
752 
753 	if (__predict_false(ps->ps_control_addr == 0))
754 		return EINVAL;
755 
756 	if ((ps->ps_flags & ACPICPU_FLAG_P_FIDVID) != 0)
757 		return acpicpu_md_pstate_fidvid_set(ps);
758 
759 	/*
760 	 * If the mask is set, do a read-modify-write.
761 	 */
762 	if (__predict_true(ps->ps_control_mask != 0)) {
763 		val = rdmsr(ps->ps_control_addr);
764 		val &= ~ps->ps_control_mask;
765 	}
766 
767 	val |= ps->ps_control;
768 
769 	wrmsr(ps->ps_control_addr, val);
770 	DELAY(ps->ps_latency);
771 
772 	return 0;
773 }
774 
775 static int
776 acpicpu_md_pstate_fidvid_get(struct acpicpu_softc *sc, uint32_t *freq)
777 {
778 	struct acpicpu_pstate *ps;
779 	uint32_t fid, i, vid;
780 	uint32_t cfid, cvid;
781 	int rv;
782 
783 	/*
784 	 * AMD family 0Fh needs special treatment.
785 	 * While it wants to use ACPI, it does not
786 	 * comply with the ACPI specifications.
787 	 */
788 	rv = acpicpu_md_pstate_fidvid_read(&cfid, &cvid);
789 
790 	if (rv != 0)
791 		return rv;
792 
793 	for (i = 0; i < sc->sc_pstate_count; i++) {
794 
795 		ps = &sc->sc_pstate[i];
796 
797 		if (__predict_false(ps->ps_freq == 0))
798 			continue;
799 
800 		fid = __SHIFTOUT(ps->ps_status, ACPI_0FH_STATUS_FID);
801 		vid = __SHIFTOUT(ps->ps_status, ACPI_0FH_STATUS_VID);
802 
803 		if (cfid == fid && cvid == vid) {
804 			*freq = ps->ps_freq;
805 			return 0;
806 		}
807 	}
808 
809 	return EIO;
810 }
811 
812 static int
813 acpicpu_md_pstate_fidvid_set(struct acpicpu_pstate *ps)
814 {
815 	const uint64_t ctrl = ps->ps_control;
816 	uint32_t cfid, cvid, fid, i, irt;
817 	uint32_t pll, vco_cfid, vco_fid;
818 	uint32_t val, vid, vst;
819 	int rv;
820 
821 	rv = acpicpu_md_pstate_fidvid_read(&cfid, &cvid);
822 
823 	if (rv != 0)
824 		return rv;
825 
826 	fid = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_FID);
827 	vid = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_VID);
828 	irt = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_IRT);
829 	vst = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_VST);
830 	pll = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_PLL);
831 
832 	vst = vst * 20;
833 	pll = pll * 1000 / 5;
834 	irt = 10 * __BIT(irt);
835 
836 	/*
837 	 * Phase 1.
838 	 */
839 	while (cvid > vid) {
840 
841 		val = 1 << __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_MVS);
842 		val = (val > cvid) ? 0 : cvid - val;
843 
844 		acpicpu_md_pstate_fidvid_write(cfid, val, 1, vst);
845 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
846 
847 		if (rv != 0)
848 			return rv;
849 	}
850 
851 	i = __SHIFTOUT(ctrl, ACPI_0FH_CONTROL_RVO);
852 
853 	for (; i > 0 && cvid > 0; --i) {
854 
855 		acpicpu_md_pstate_fidvid_write(cfid, cvid - 1, 1, vst);
856 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
857 
858 		if (rv != 0)
859 			return rv;
860 	}
861 
862 	/*
863 	 * Phase 2.
864 	 */
865 	if (cfid != fid) {
866 
867 		vco_fid  = FID_TO_VCO_FID(fid);
868 		vco_cfid = FID_TO_VCO_FID(cfid);
869 
870 		while (abs(vco_fid - vco_cfid) > 2) {
871 
872 			if (fid <= cfid)
873 				val = cfid - 2;
874 			else {
875 				val = (cfid > 6) ? cfid + 2 :
876 				    FID_TO_VCO_FID(cfid) + 2;
877 			}
878 
879 			acpicpu_md_pstate_fidvid_write(val, cvid, pll, irt);
880 			rv = acpicpu_md_pstate_fidvid_read(&cfid, NULL);
881 
882 			if (rv != 0)
883 				return rv;
884 
885 			vco_cfid = FID_TO_VCO_FID(cfid);
886 		}
887 
888 		acpicpu_md_pstate_fidvid_write(fid, cvid, pll, irt);
889 		rv = acpicpu_md_pstate_fidvid_read(&cfid, NULL);
890 
891 		if (rv != 0)
892 			return rv;
893 	}
894 
895 	/*
896 	 * Phase 3.
897 	 */
898 	if (cvid != vid) {
899 
900 		acpicpu_md_pstate_fidvid_write(cfid, vid, 1, vst);
901 		rv = acpicpu_md_pstate_fidvid_read(NULL, &cvid);
902 
903 		if (rv != 0)
904 			return rv;
905 	}
906 
907 	return 0;
908 }
909 
910 static int
911 acpicpu_md_pstate_fidvid_read(uint32_t *cfid, uint32_t *cvid)
912 {
913 	int i = ACPICPU_P_STATE_RETRY * 100;
914 	uint64_t val;
915 
916 	do {
917 		val = rdmsr(MSR_0FH_STATUS);
918 
919 	} while (__SHIFTOUT(val, MSR_0FH_STATUS_PENDING) != 0 && --i >= 0);
920 
921 	if (i == 0)
922 		return EAGAIN;
923 
924 	if (cfid != NULL)
925 		*cfid = __SHIFTOUT(val, MSR_0FH_STATUS_CFID);
926 
927 	if (cvid != NULL)
928 		*cvid = __SHIFTOUT(val, MSR_0FH_STATUS_CVID);
929 
930 	return 0;
931 }
932 
933 static void
934 acpicpu_md_pstate_fidvid_write(uint32_t fid,
935     uint32_t vid, uint32_t cnt, uint32_t tmo)
936 {
937 	uint64_t val = 0;
938 
939 	val |= __SHIFTIN(fid, MSR_0FH_CONTROL_FID);
940 	val |= __SHIFTIN(vid, MSR_0FH_CONTROL_VID);
941 	val |= __SHIFTIN(cnt, MSR_0FH_CONTROL_CNT);
942 	val |= __SHIFTIN(0x1, MSR_0FH_CONTROL_CHG);
943 
944 	wrmsr(MSR_0FH_CONTROL, val);
945 	DELAY(tmo);
946 }
947 
948 int
949 acpicpu_md_tstate_get(struct acpicpu_softc *sc, uint32_t *percent)
950 {
951 	struct acpicpu_tstate *ts;
952 	uint64_t val;
953 	uint32_t i;
954 
955 	val = rdmsr(MSR_THERM_CONTROL);
956 
957 	for (i = 0; i < sc->sc_tstate_count; i++) {
958 
959 		ts = &sc->sc_tstate[i];
960 
961 		if (ts->ts_percent == 0)
962 			continue;
963 
964 		if (val == ts->ts_status) {
965 			*percent = ts->ts_percent;
966 			return 0;
967 		}
968 	}
969 
970 	return EIO;
971 }
972 
973 int
974 acpicpu_md_tstate_set(struct acpicpu_tstate *ts)
975 {
976 	uint64_t val;
977 	uint8_t i;
978 
979 	val = ts->ts_control;
980 	val = val & __BITS(0, 4);
981 
982 	wrmsr(MSR_THERM_CONTROL, val);
983 
984 	if (ts->ts_status == 0) {
985 		DELAY(ts->ts_latency);
986 		return 0;
987 	}
988 
989 	for (i = val = 0; i < ACPICPU_T_STATE_RETRY; i++) {
990 
991 		val = rdmsr(MSR_THERM_CONTROL);
992 
993 		if (val == ts->ts_status)
994 			return 0;
995 
996 		DELAY(ts->ts_latency);
997 	}
998 
999 	return EAGAIN;
1000 }
1001 
1002 /*
1003  * A kludge for backwards compatibility.
1004  */
1005 static int
1006 acpicpu_md_pstate_sysctl_init(void)
1007 {
1008 	const struct sysctlnode	*fnode, *mnode, *rnode;
1009 	const char *str;
1010 	int rv;
1011 
1012 	switch (cpu_vendor) {
1013 
1014 	case CPUVENDOR_IDT:
1015 	case CPUVENDOR_INTEL:
1016 		str = "est";
1017 		break;
1018 
1019 	case CPUVENDOR_AMD:
1020 		str = "powernow";
1021 		break;
1022 
1023 	default:
1024 		return ENODEV;
1025 	}
1026 
1027 
1028 	rv = sysctl_createv(&acpicpu_log, 0, NULL, &rnode,
1029 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "machdep", NULL,
1030 	    NULL, 0, NULL, 0, CTL_MACHDEP, CTL_EOL);
1031 
1032 	if (rv != 0)
1033 		goto fail;
1034 
1035 	rv = sysctl_createv(&acpicpu_log, 0, &rnode, &mnode,
1036 	    0, CTLTYPE_NODE, str, NULL,
1037 	    NULL, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1038 
1039 	if (rv != 0)
1040 		goto fail;
1041 
1042 	rv = sysctl_createv(&acpicpu_log, 0, &mnode, &fnode,
1043 	    0, CTLTYPE_NODE, "frequency", NULL,
1044 	    NULL, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1045 
1046 	if (rv != 0)
1047 		goto fail;
1048 
1049 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1050 	    CTLFLAG_READWRITE, CTLTYPE_INT, "target", NULL,
1051 	    acpicpu_md_pstate_sysctl_set, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1052 
1053 	if (rv != 0)
1054 		goto fail;
1055 
1056 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1057 	    CTLFLAG_READONLY, CTLTYPE_INT, "current", NULL,
1058 	    acpicpu_md_pstate_sysctl_get, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1059 
1060 	if (rv != 0)
1061 		goto fail;
1062 
1063 	rv = sysctl_createv(&acpicpu_log, 0, &fnode, &rnode,
1064 	    CTLFLAG_READONLY, CTLTYPE_STRING, "available", NULL,
1065 	    acpicpu_md_pstate_sysctl_all, 0, NULL, 0, CTL_CREATE, CTL_EOL);
1066 
1067 	if (rv != 0)
1068 		goto fail;
1069 
1070 	return 0;
1071 
1072 fail:
1073 	if (acpicpu_log != NULL) {
1074 		sysctl_teardown(&acpicpu_log);
1075 		acpicpu_log = NULL;
1076 	}
1077 
1078 	return rv;
1079 }
1080 
1081 static int
1082 acpicpu_md_pstate_sysctl_get(SYSCTLFN_ARGS)
1083 {
1084 	struct sysctlnode node;
1085 	uint32_t freq;
1086 	int err;
1087 
1088 	freq = cpufreq_get(curcpu());
1089 
1090 	if (freq == 0)
1091 		return ENXIO;
1092 
1093 	node = *rnode;
1094 	node.sysctl_data = &freq;
1095 
1096 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1097 
1098 	if (err != 0 || newp == NULL)
1099 		return err;
1100 
1101 	return 0;
1102 }
1103 
1104 static int
1105 acpicpu_md_pstate_sysctl_set(SYSCTLFN_ARGS)
1106 {
1107 	struct sysctlnode node;
1108 	uint32_t freq;
1109 	int err;
1110 
1111 	freq = cpufreq_get(curcpu());
1112 
1113 	if (freq == 0)
1114 		return ENXIO;
1115 
1116 	node = *rnode;
1117 	node.sysctl_data = &freq;
1118 
1119 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1120 
1121 	if (err != 0 || newp == NULL)
1122 		return err;
1123 
1124 	cpufreq_set_all(freq);
1125 
1126 	return 0;
1127 }
1128 
1129 static int
1130 acpicpu_md_pstate_sysctl_all(SYSCTLFN_ARGS)
1131 {
1132 	struct cpu_info *ci = curcpu();
1133 	struct acpicpu_softc *sc;
1134 	struct sysctlnode node;
1135 	char buf[1024];
1136 	size_t len;
1137 	uint32_t i;
1138 	int err;
1139 
1140 	sc = acpicpu_sc[ci->ci_acpiid];
1141 
1142 	if (sc == NULL)
1143 		return ENXIO;
1144 
1145 	(void)memset(&buf, 0, sizeof(buf));
1146 
1147 	mutex_enter(&sc->sc_mtx);
1148 
1149 	for (len = 0, i = sc->sc_pstate_max; i < sc->sc_pstate_count; i++) {
1150 
1151 		if (sc->sc_pstate[i].ps_freq == 0)
1152 			continue;
1153 
1154 		if (len >= sizeof(buf))
1155 			break;
1156 		len += snprintf(buf + len, sizeof(buf) - len, "%u%s",
1157 		    sc->sc_pstate[i].ps_freq,
1158 		    i < (sc->sc_pstate_count - 1) ? " " : "");
1159 	}
1160 
1161 	mutex_exit(&sc->sc_mtx);
1162 
1163 	node = *rnode;
1164 	node.sysctl_data = buf;
1165 
1166 	err = sysctl_lookup(SYSCTLFN_CALL(&node));
1167 
1168 	if (err != 0 || newp == NULL)
1169 		return err;
1170 
1171 	return 0;
1172 }
1173 
1174