xref: /freebsd/sys/amd64/vmm/x86.c (revision fe75646a0234a261c0013bf1840fdac4acaf0cec)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2011 NetApp, Inc.
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  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/param.h>
30 #include <sys/pcpu.h>
31 #include <sys/systm.h>
32 #include <sys/sysctl.h>
33 
34 #include <machine/clock.h>
35 #include <machine/cpufunc.h>
36 #include <machine/md_var.h>
37 #include <machine/segments.h>
38 #include <machine/specialreg.h>
39 
40 #include <machine/vmm.h>
41 
42 #include "vmm_host.h"
43 #include "vmm_ktr.h"
44 #include "vmm_util.h"
45 #include "x86.h"
46 
47 SYSCTL_DECL(_hw_vmm);
48 static SYSCTL_NODE(_hw_vmm, OID_AUTO, topology, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
49     NULL);
50 
51 #define	CPUID_VM_HIGH		0x40000000
52 
53 static const char bhyve_id[12] = "bhyve bhyve ";
54 
55 static uint64_t bhyve_xcpuids;
56 SYSCTL_ULONG(_hw_vmm, OID_AUTO, bhyve_xcpuids, CTLFLAG_RW, &bhyve_xcpuids, 0,
57     "Number of times an unknown cpuid leaf was accessed");
58 
59 static int cpuid_leaf_b = 1;
60 SYSCTL_INT(_hw_vmm_topology, OID_AUTO, cpuid_leaf_b, CTLFLAG_RDTUN,
61     &cpuid_leaf_b, 0, NULL);
62 
63 /*
64  * Round up to the next power of two, if necessary, and then take log2.
65  * Returns -1 if argument is zero.
66  */
67 static __inline int
68 log2(u_int x)
69 {
70 
71 	return (fls(x << (1 - powerof2(x))) - 1);
72 }
73 
74 int
75 x86_emulate_cpuid(struct vcpu *vcpu, uint64_t *rax, uint64_t *rbx,
76     uint64_t *rcx, uint64_t *rdx)
77 {
78 	struct vm *vm = vcpu_vm(vcpu);
79 	int vcpu_id = vcpu_vcpuid(vcpu);
80 	const struct xsave_limits *limits;
81 	uint64_t cr4;
82 	int error, enable_invpcid, enable_rdpid, enable_rdtscp, level,
83 	    width, x2apic_id;
84 	unsigned int func, regs[4], logical_cpus, param;
85 	enum x2apic_state x2apic_state;
86 	uint16_t cores, maxcpus, sockets, threads;
87 
88 	/*
89 	 * The function of CPUID is controlled through the provided value of
90 	 * %eax (and secondarily %ecx, for certain leaf data).
91 	 */
92 	func = (uint32_t)*rax;
93 	param = (uint32_t)*rcx;
94 
95 	VCPU_CTR2(vm, vcpu_id, "cpuid %#x,%#x", func, param);
96 
97 	/*
98 	 * Requests for invalid CPUID levels should map to the highest
99 	 * available level instead.
100 	 */
101 	if (cpu_exthigh != 0 && func >= 0x80000000) {
102 		if (func > cpu_exthigh)
103 			func = cpu_exthigh;
104 	} else if (func >= 0x40000000) {
105 		if (func > CPUID_VM_HIGH)
106 			func = CPUID_VM_HIGH;
107 	} else if (func > cpu_high) {
108 		func = cpu_high;
109 	}
110 
111 	/*
112 	 * In general the approach used for CPU topology is to
113 	 * advertise a flat topology where all CPUs are packages with
114 	 * no multi-core or SMT.
115 	 */
116 	switch (func) {
117 		/*
118 		 * Pass these through to the guest
119 		 */
120 		case CPUID_0000_0000:
121 		case CPUID_0000_0002:
122 		case CPUID_0000_0003:
123 		case CPUID_8000_0000:
124 		case CPUID_8000_0002:
125 		case CPUID_8000_0003:
126 		case CPUID_8000_0004:
127 		case CPUID_8000_0006:
128 			cpuid_count(func, param, regs);
129 			break;
130 		case CPUID_8000_0008:
131 			cpuid_count(func, param, regs);
132 			if (vmm_is_svm()) {
133 				/*
134 				 * As on Intel (0000_0007:0, EDX), mask out
135 				 * unsupported or unsafe AMD extended features
136 				 * (8000_0008 EBX).
137 				 */
138 				regs[1] &= (AMDFEID_CLZERO | AMDFEID_IRPERF |
139 				    AMDFEID_XSAVEERPTR);
140 
141 				vm_get_topology(vm, &sockets, &cores, &threads,
142 				    &maxcpus);
143 				/*
144 				 * Here, width is ApicIdCoreIdSize, present on
145 				 * at least Family 15h and newer.  It
146 				 * represents the "number of bits in the
147 				 * initial apicid that indicate thread id
148 				 * within a package."
149 				 *
150 				 * Our topo_probe_amd() uses it for
151 				 * pkg_id_shift and other OSes may rely on it.
152 				 */
153 				width = MIN(0xF, log2(threads * cores));
154 				if (width < 0x4)
155 					width = 0;
156 				logical_cpus = MIN(0xFF, threads * cores - 1);
157 				regs[2] = (width << AMDID_COREID_SIZE_SHIFT) | logical_cpus;
158 			}
159 			break;
160 
161 		case CPUID_8000_0001:
162 			cpuid_count(func, param, regs);
163 
164 			/*
165 			 * Hide SVM from guest.
166 			 */
167 			regs[2] &= ~AMDID2_SVM;
168 
169 			/*
170 			 * Don't advertise extended performance counter MSRs
171 			 * to the guest.
172 			 */
173 			regs[2] &= ~AMDID2_PCXC;
174 			regs[2] &= ~AMDID2_PNXC;
175 			regs[2] &= ~AMDID2_PTSCEL2I;
176 
177 			/*
178 			 * Don't advertise Instruction Based Sampling feature.
179 			 */
180 			regs[2] &= ~AMDID2_IBS;
181 
182 			/* NodeID MSR not available */
183 			regs[2] &= ~AMDID2_NODE_ID;
184 
185 			/* Don't advertise the OS visible workaround feature */
186 			regs[2] &= ~AMDID2_OSVW;
187 
188 			/* Hide mwaitx/monitorx capability from the guest */
189 			regs[2] &= ~AMDID2_MWAITX;
190 
191 			/* Advertise RDTSCP if it is enabled. */
192 			error = vm_get_capability(vcpu,
193 			    VM_CAP_RDTSCP, &enable_rdtscp);
194 			if (error == 0 && enable_rdtscp)
195 				regs[3] |= AMDID_RDTSCP;
196 			else
197 				regs[3] &= ~AMDID_RDTSCP;
198 			break;
199 
200 		case CPUID_8000_0007:
201 			/*
202 			 * AMD uses this leaf to advertise the processor's
203 			 * power monitoring and RAS capabilities. These
204 			 * features are hardware-specific and exposing
205 			 * them to a guest doesn't make a lot of sense.
206 			 *
207 			 * Intel uses this leaf only to advertise the
208 			 * "Invariant TSC" feature with all other bits
209 			 * being reserved (set to zero).
210 			 */
211 			regs[0] = 0;
212 			regs[1] = 0;
213 			regs[2] = 0;
214 			regs[3] = 0;
215 
216 			/*
217 			 * "Invariant TSC" can be advertised to the guest if:
218 			 * - host TSC frequency is invariant
219 			 * - host TSCs are synchronized across physical cpus
220 			 *
221 			 * XXX This still falls short because the vcpu
222 			 * can observe the TSC moving backwards as it
223 			 * migrates across physical cpus. But at least
224 			 * it should discourage the guest from using the
225 			 * TSC to keep track of time.
226 			 */
227 			if (tsc_is_invariant && smp_tsc)
228 				regs[3] |= AMDPM_TSC_INVARIANT;
229 			break;
230 
231 		case CPUID_8000_001D:
232 			/* AMD Cache topology, like 0000_0004 for Intel. */
233 			if (!vmm_is_svm())
234 				goto default_leaf;
235 
236 			/*
237 			 * Similar to Intel, generate a ficticious cache
238 			 * topology for the guest with L3 shared by the
239 			 * package, and L1 and L2 local to a core.
240 			 */
241 			vm_get_topology(vm, &sockets, &cores, &threads,
242 			    &maxcpus);
243 			switch (param) {
244 			case 0:
245 				logical_cpus = threads;
246 				level = 1;
247 				func = 1;	/* data cache */
248 				break;
249 			case 1:
250 				logical_cpus = threads;
251 				level = 2;
252 				func = 3;	/* unified cache */
253 				break;
254 			case 2:
255 				logical_cpus = threads * cores;
256 				level = 3;
257 				func = 3;	/* unified cache */
258 				break;
259 			default:
260 				logical_cpus = 0;
261 				level = 0;
262 				func = 0;
263 				break;
264 			}
265 
266 			logical_cpus = MIN(0xfff, logical_cpus - 1);
267 			regs[0] = (logical_cpus << 14) | (1 << 8) |
268 			    (level << 5) | func;
269 			regs[1] = (func > 0) ? (CACHE_LINE_SIZE - 1) : 0;
270 			regs[2] = 0;
271 			regs[3] = 0;
272 			break;
273 
274 		case CPUID_8000_001E:
275 			/*
276 			 * AMD Family 16h+ and Hygon Family 18h additional
277 			 * identifiers.
278 			 */
279 			if (!vmm_is_svm() || CPUID_TO_FAMILY(cpu_id) < 0x16)
280 				goto default_leaf;
281 
282 			vm_get_topology(vm, &sockets, &cores, &threads,
283 			    &maxcpus);
284 			regs[0] = vcpu_id;
285 			threads = MIN(0xFF, threads - 1);
286 			regs[1] = (threads << 8) |
287 			    (vcpu_id >> log2(threads + 1));
288 			/*
289 			 * XXX Bhyve topology cannot yet represent >1 node per
290 			 * processor.
291 			 */
292 			regs[2] = 0;
293 			regs[3] = 0;
294 			break;
295 
296 		case CPUID_0000_0001:
297 			do_cpuid(1, regs);
298 
299 			error = vm_get_x2apic_state(vcpu, &x2apic_state);
300 			if (error) {
301 				panic("x86_emulate_cpuid: error %d "
302 				      "fetching x2apic state", error);
303 			}
304 
305 			/*
306 			 * Override the APIC ID only in ebx
307 			 */
308 			regs[1] &= ~(CPUID_LOCAL_APIC_ID);
309 			regs[1] |= (vcpu_id << CPUID_0000_0001_APICID_SHIFT);
310 
311 			/*
312 			 * Don't expose VMX, SpeedStep, TME or SMX capability.
313 			 * Advertise x2APIC capability and Hypervisor guest.
314 			 */
315 			regs[2] &= ~(CPUID2_VMX | CPUID2_EST | CPUID2_TM2);
316 			regs[2] &= ~(CPUID2_SMX);
317 
318 			regs[2] |= CPUID2_HV;
319 
320 			if (x2apic_state != X2APIC_DISABLED)
321 				regs[2] |= CPUID2_X2APIC;
322 			else
323 				regs[2] &= ~CPUID2_X2APIC;
324 
325 			/*
326 			 * Only advertise CPUID2_XSAVE in the guest if
327 			 * the host is using XSAVE.
328 			 */
329 			if (!(regs[2] & CPUID2_OSXSAVE))
330 				regs[2] &= ~CPUID2_XSAVE;
331 
332 			/*
333 			 * If CPUID2_XSAVE is being advertised and the
334 			 * guest has set CR4_XSAVE, set
335 			 * CPUID2_OSXSAVE.
336 			 */
337 			regs[2] &= ~CPUID2_OSXSAVE;
338 			if (regs[2] & CPUID2_XSAVE) {
339 				error = vm_get_register(vcpu,
340 				    VM_REG_GUEST_CR4, &cr4);
341 				if (error)
342 					panic("x86_emulate_cpuid: error %d "
343 					      "fetching %%cr4", error);
344 				if (cr4 & CR4_XSAVE)
345 					regs[2] |= CPUID2_OSXSAVE;
346 			}
347 
348 			/*
349 			 * Hide monitor/mwait until we know how to deal with
350 			 * these instructions.
351 			 */
352 			regs[2] &= ~CPUID2_MON;
353 
354                         /*
355 			 * Hide the performance and debug features.
356 			 */
357 			regs[2] &= ~CPUID2_PDCM;
358 
359 			/*
360 			 * No TSC deadline support in the APIC yet
361 			 */
362 			regs[2] &= ~CPUID2_TSCDLT;
363 
364 			/*
365 			 * Hide thermal monitoring
366 			 */
367 			regs[3] &= ~(CPUID_ACPI | CPUID_TM);
368 
369 			/*
370 			 * Hide the debug store capability.
371 			 */
372 			regs[3] &= ~CPUID_DS;
373 
374 			/*
375 			 * Advertise the Machine Check and MTRR capability.
376 			 *
377 			 * Some guest OSes (e.g. Windows) will not boot if
378 			 * these features are absent.
379 			 */
380 			regs[3] |= (CPUID_MCA | CPUID_MCE | CPUID_MTRR);
381 
382 			vm_get_topology(vm, &sockets, &cores, &threads,
383 			    &maxcpus);
384 			logical_cpus = threads * cores;
385 			regs[1] &= ~CPUID_HTT_CORES;
386 			regs[1] |= (logical_cpus & 0xff) << 16;
387 			regs[3] |= CPUID_HTT;
388 			break;
389 
390 		case CPUID_0000_0004:
391 			cpuid_count(func, param, regs);
392 
393 			if (regs[0] || regs[1] || regs[2] || regs[3]) {
394 				vm_get_topology(vm, &sockets, &cores, &threads,
395 				    &maxcpus);
396 				regs[0] &= 0x3ff;
397 				regs[0] |= (cores - 1) << 26;
398 				/*
399 				 * Cache topology:
400 				 * - L1 and L2 are shared only by the logical
401 				 *   processors in a single core.
402 				 * - L3 and above are shared by all logical
403 				 *   processors in the package.
404 				 */
405 				logical_cpus = threads;
406 				level = (regs[0] >> 5) & 0x7;
407 				if (level >= 3)
408 					logical_cpus *= cores;
409 				regs[0] |= (logical_cpus - 1) << 14;
410 			}
411 			break;
412 
413 		case CPUID_0000_0007:
414 			regs[0] = 0;
415 			regs[1] = 0;
416 			regs[2] = 0;
417 			regs[3] = 0;
418 
419 			/* leaf 0 */
420 			if (param == 0) {
421 				cpuid_count(func, param, regs);
422 
423 				/* Only leaf 0 is supported */
424 				regs[0] = 0;
425 
426 				/*
427 				 * Expose known-safe features.
428 				 */
429 				regs[1] &= CPUID_STDEXT_FSGSBASE |
430 				    CPUID_STDEXT_BMI1 | CPUID_STDEXT_HLE |
431 				    CPUID_STDEXT_AVX2 | CPUID_STDEXT_SMEP |
432 				    CPUID_STDEXT_BMI2 |
433 				    CPUID_STDEXT_ERMS | CPUID_STDEXT_RTM |
434 				    CPUID_STDEXT_AVX512F |
435 				    CPUID_STDEXT_AVX512DQ |
436 				    CPUID_STDEXT_RDSEED |
437 				    CPUID_STDEXT_SMAP |
438 				    CPUID_STDEXT_AVX512PF |
439 				    CPUID_STDEXT_AVX512ER |
440 				    CPUID_STDEXT_AVX512CD | CPUID_STDEXT_SHA |
441 				    CPUID_STDEXT_AVX512BW |
442 				    CPUID_STDEXT_AVX512VL;
443 				regs[2] &= CPUID_STDEXT2_VAES |
444 				    CPUID_STDEXT2_VPCLMULQDQ;
445 				regs[3] &= CPUID_STDEXT3_MD_CLEAR;
446 
447 				/* Advertise RDPID if it is enabled. */
448 				error = vm_get_capability(vcpu, VM_CAP_RDPID,
449 				    &enable_rdpid);
450 				if (error == 0 && enable_rdpid)
451 					regs[2] |= CPUID_STDEXT2_RDPID;
452 
453 				/* Advertise INVPCID if it is enabled. */
454 				error = vm_get_capability(vcpu,
455 				    VM_CAP_ENABLE_INVPCID, &enable_invpcid);
456 				if (error == 0 && enable_invpcid)
457 					regs[1] |= CPUID_STDEXT_INVPCID;
458 			}
459 			break;
460 
461 		case CPUID_0000_0006:
462 			regs[0] = CPUTPM1_ARAT;
463 			regs[1] = 0;
464 			regs[2] = 0;
465 			regs[3] = 0;
466 			break;
467 
468 		case CPUID_0000_000A:
469 			/*
470 			 * Handle the access, but report 0 for
471 			 * all options
472 			 */
473 			regs[0] = 0;
474 			regs[1] = 0;
475 			regs[2] = 0;
476 			regs[3] = 0;
477 			break;
478 
479 		case CPUID_0000_000B:
480 			/*
481 			 * Intel processor topology enumeration
482 			 */
483 			if (vmm_is_intel()) {
484 				vm_get_topology(vm, &sockets, &cores, &threads,
485 				    &maxcpus);
486 				if (param == 0) {
487 					logical_cpus = threads;
488 					width = log2(logical_cpus);
489 					level = CPUID_TYPE_SMT;
490 					x2apic_id = vcpu_id;
491 				}
492 
493 				if (param == 1) {
494 					logical_cpus = threads * cores;
495 					width = log2(logical_cpus);
496 					level = CPUID_TYPE_CORE;
497 					x2apic_id = vcpu_id;
498 				}
499 
500 				if (!cpuid_leaf_b || param >= 2) {
501 					width = 0;
502 					logical_cpus = 0;
503 					level = 0;
504 					x2apic_id = 0;
505 				}
506 
507 				regs[0] = width & 0x1f;
508 				regs[1] = logical_cpus & 0xffff;
509 				regs[2] = (level << 8) | (param & 0xff);
510 				regs[3] = x2apic_id;
511 			} else {
512 				regs[0] = 0;
513 				regs[1] = 0;
514 				regs[2] = 0;
515 				regs[3] = 0;
516 			}
517 			break;
518 
519 		case CPUID_0000_000D:
520 			limits = vmm_get_xsave_limits();
521 			if (!limits->xsave_enabled) {
522 				regs[0] = 0;
523 				regs[1] = 0;
524 				regs[2] = 0;
525 				regs[3] = 0;
526 				break;
527 			}
528 
529 			cpuid_count(func, param, regs);
530 			switch (param) {
531 			case 0:
532 				/*
533 				 * Only permit the guest to use bits
534 				 * that are active in the host in
535 				 * %xcr0.  Also, claim that the
536 				 * maximum save area size is
537 				 * equivalent to the host's current
538 				 * save area size.  Since this runs
539 				 * "inside" of vmrun(), it runs with
540 				 * the guest's xcr0, so the current
541 				 * save area size is correct as-is.
542 				 */
543 				regs[0] &= limits->xcr0_allowed;
544 				regs[2] = limits->xsave_max_size;
545 				regs[3] &= (limits->xcr0_allowed >> 32);
546 				break;
547 			case 1:
548 				/* Only permit XSAVEOPT. */
549 				regs[0] &= CPUID_EXTSTATE_XSAVEOPT;
550 				regs[1] = 0;
551 				regs[2] = 0;
552 				regs[3] = 0;
553 				break;
554 			default:
555 				/*
556 				 * If the leaf is for a permitted feature,
557 				 * pass through as-is, otherwise return
558 				 * all zeroes.
559 				 */
560 				if (!(limits->xcr0_allowed & (1ul << param))) {
561 					regs[0] = 0;
562 					regs[1] = 0;
563 					regs[2] = 0;
564 					regs[3] = 0;
565 				}
566 				break;
567 			}
568 			break;
569 
570 		case CPUID_0000_000F:
571 		case CPUID_0000_0010:
572 			/*
573 			 * Do not report any Resource Director Technology
574 			 * capabilities.  Exposing control of cache or memory
575 			 * controller resource partitioning to the guest is not
576 			 * at all sensible.
577 			 *
578 			 * This is already hidden at a high level by masking of
579 			 * leaf 0x7.  Even still, a guest may look here for
580 			 * detailed capability information.
581 			 */
582 			regs[0] = 0;
583 			regs[1] = 0;
584 			regs[2] = 0;
585 			regs[3] = 0;
586 			break;
587 
588 		case CPUID_0000_0015:
589 			/*
590 			 * Don't report CPU TSC/Crystal ratio and clock
591 			 * values since guests may use these to derive the
592 			 * local APIC frequency..
593 			 */
594 			regs[0] = 0;
595 			regs[1] = 0;
596 			regs[2] = 0;
597 			regs[3] = 0;
598 			break;
599 
600 		case 0x40000000:
601 			regs[0] = CPUID_VM_HIGH;
602 			bcopy(bhyve_id, &regs[1], 4);
603 			bcopy(bhyve_id + 4, &regs[2], 4);
604 			bcopy(bhyve_id + 8, &regs[3], 4);
605 			break;
606 
607 		default:
608 default_leaf:
609 			/*
610 			 * The leaf value has already been clamped so
611 			 * simply pass this through, keeping count of
612 			 * how many unhandled leaf values have been seen.
613 			 */
614 			atomic_add_long(&bhyve_xcpuids, 1);
615 			cpuid_count(func, param, regs);
616 			break;
617 	}
618 
619 	/*
620 	 * CPUID clears the upper 32-bits of the long-mode registers.
621 	 */
622 	*rax = regs[0];
623 	*rbx = regs[1];
624 	*rcx = regs[2];
625 	*rdx = regs[3];
626 
627 	return (1);
628 }
629 
630 bool
631 vm_cpuid_capability(struct vcpu *vcpu, enum vm_cpuid_capability cap)
632 {
633 	bool rv;
634 
635 	KASSERT(cap > 0 && cap < VCC_LAST, ("%s: invalid vm_cpu_capability %d",
636 	    __func__, cap));
637 
638 	/*
639 	 * Simply passthrough the capabilities of the host cpu for now.
640 	 */
641 	rv = false;
642 	switch (cap) {
643 	case VCC_NO_EXECUTE:
644 		if (amd_feature & AMDID_NX)
645 			rv = true;
646 		break;
647 	case VCC_FFXSR:
648 		if (amd_feature & AMDID_FFXSR)
649 			rv = true;
650 		break;
651 	case VCC_TCE:
652 		if (amd_feature2 & AMDID2_TCE)
653 			rv = true;
654 		break;
655 	default:
656 		panic("%s: unknown vm_cpu_capability %d", __func__, cap);
657 	}
658 	return (rv);
659 }
660 
661 int
662 vm_rdmtrr(struct vm_mtrr *mtrr, u_int num, uint64_t *val)
663 {
664 	switch (num) {
665 	case MSR_MTRRcap:
666 		*val = MTRR_CAP_WC | MTRR_CAP_FIXED | VMM_MTRR_VAR_MAX;
667 		break;
668 	case MSR_MTRRdefType:
669 		*val = mtrr->def_type;
670 		break;
671 	case MSR_MTRR4kBase ... MSR_MTRR4kBase + 7:
672 		*val = mtrr->fixed4k[num - MSR_MTRR4kBase];
673 		break;
674 	case MSR_MTRR16kBase ... MSR_MTRR16kBase + 1:
675 		*val = mtrr->fixed16k[num - MSR_MTRR16kBase];
676 		break;
677 	case MSR_MTRR64kBase:
678 		*val = mtrr->fixed64k;
679 		break;
680 	case MSR_MTRRVarBase ... MSR_MTRRVarBase + (VMM_MTRR_VAR_MAX * 2) - 1: {
681 		u_int offset = num - MSR_MTRRVarBase;
682 		if (offset % 2 == 0) {
683 			*val = mtrr->var[offset / 2].base;
684 		} else {
685 			*val = mtrr->var[offset / 2].mask;
686 		}
687 		break;
688 	}
689 	default:
690 		return (-1);
691 	}
692 
693 	return (0);
694 }
695 
696 int
697 vm_wrmtrr(struct vm_mtrr *mtrr, u_int num, uint64_t val)
698 {
699 	switch (num) {
700 	case MSR_MTRRcap:
701 		/* MTRRCAP is read only */
702 		return (-1);
703 	case MSR_MTRRdefType:
704 		if (val & ~VMM_MTRR_DEF_MASK) {
705 			/* generate #GP on writes to reserved fields */
706 			return (-1);
707 		}
708 		mtrr->def_type = val;
709 		break;
710 	case MSR_MTRR4kBase ... MSR_MTRR4kBase + 7:
711 		mtrr->fixed4k[num - MSR_MTRR4kBase] = val;
712 		break;
713 	case MSR_MTRR16kBase ... MSR_MTRR16kBase + 1:
714 		mtrr->fixed16k[num - MSR_MTRR16kBase] = val;
715 		break;
716 	case MSR_MTRR64kBase:
717 		mtrr->fixed64k = val;
718 		break;
719 	case MSR_MTRRVarBase ... MSR_MTRRVarBase + (VMM_MTRR_VAR_MAX * 2) - 1: {
720 		u_int offset = num - MSR_MTRRVarBase;
721 		if (offset % 2 == 0) {
722 			if (val & ~VMM_MTRR_PHYSBASE_MASK) {
723 				/* generate #GP on writes to reserved fields */
724 				return (-1);
725 			}
726 			mtrr->var[offset / 2].base = val;
727 		} else {
728 			if (val & ~VMM_MTRR_PHYSMASK_MASK) {
729 				/* generate #GP on writes to reserved fields */
730 				return (-1);
731 			}
732 			mtrr->var[offset / 2].mask = val;
733 		}
734 		break;
735 	}
736 	default:
737 		return (-1);
738 	}
739 
740 	return (0);
741 }
742