xref: /illumos-gate/usr/src/uts/intel/io/vmm/vmm_cpuid.c (revision b02947bf393d39f68e6c7fa8ccb98688f7f9c407)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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  * $FreeBSD$
29  */
30 /*
31  * This file and its contents are supplied under the terms of the
32  * Common Development and Distribution License ("CDDL"), version 1.0.
33  * You may only use this file in accordance with the terms of version
34  * 1.0 of the CDDL.
35  *
36  * A full copy of the text of the CDDL should have accompanied this
37  * source.  A copy of the CDDL is also available via the Internet at
38  * http://www.illumos.org/license/CDDL.
39  *
40  * Copyright 2014 Pluribus Networks Inc.
41  * Copyright 2018 Joyent, Inc.
42  * Copyright 2022 Oxide Computer Company
43  */
44 
45 #include <sys/types.h>
46 #include <sys/stdbool.h>
47 #include <sys/errno.h>
48 
49 #include <machine/md_var.h>
50 #include <machine/specialreg.h>
51 
52 #include <machine/vmm.h>
53 #include <sys/vmm_kernel.h>
54 
55 #include "vmm_host.h"
56 #include "vmm_util.h"
57 
58 /*
59  * CPUID Emulation
60  *
61  * All CPUID instruction exits are handled by the in-kernel emulation.
62  *
63  * ----------------
64  * Legacy Emulation
65  * ----------------
66  *
67  * Originally, the kernel vmm portion of bhyve relied on fixed logic to filter
68  * and/or generate CPUID results based on what was reported by the host CPU, as
69  * well as attributes of the VM (such as CPU topology, and enabled features).
70  * This is largely adequate to expose CPU capabilities to the guest in manner
71  * which allows it to operate properly.
72  *
73  * ------------------------------
74  * Userspace-Controlled Emulation
75  * ------------------------------
76  *
77  * In certain situations, more control over the CPUID emulation results present
78  * to the guest is desired.  Live migration between physical hosts is one such
79  * example, where the underlying CPUs, or at least their microcode, may differ
80  * between the source and destination.  In such cases, where changes to the
81  * CPUID results cannot be tolerated, the userspace portion of the VMM can be in
82  * complete control over the leaves which are presented to the guest.  It may
83  * still consult the "legacy" CPUID data for guidance about which CPU features
84  * are safe to expose (due to hypervisor limitations, etc).  This leaf
85  * information is configured on a per-vCPU basis.
86  *
87  * The emulation entries provided by userspace are expected to be in sorted
88  * order, running from lowest function and index to highest.
89  *
90  * For example:
91  * (func: 00h idx: 00h) ->
92  *     (flags: 0, eax: highest std leaf, ebx-edx: vendor id)
93  * (func: 0Dh idx: 00h) ->
94  *     (flags: VCE_FLAG_MATCH_INDEX, eax - edx: XCR0/XSAVE info)
95  * (func: 0Dh idx: 01h) ->
96  *     (flags: VCE_FLAG_MATCH_INDEX, eax - edx: XSAVE/XSAVEOPT details)
97  *     ...
98  * (func: 0Dh idx: 07H) ->
99  *     (flags: VCE_FLAG_MATCH_INDEX, eax - edx: AVX-512 details)
100  * (func: 8000000h idx: 0h) ->
101  *     (flags: 0, eax: highest extd leaf ...)
102  *     ...
103  */
104 
105 
106 #define	CPUID_TYPE_MASK	0xf0000000
107 #define	CPUID_TYPE_STD	0x00000000
108 #define	CPUID_TYPE_EXTD	0x80000000
109 
110 static const struct vcpu_cpuid_entry cpuid_empty_entry = { 0 };
111 
112 /*
113  * Given the CPUID configuration for a vCPU, locate the entry which matches the
114  * provided function/index tuple.  The entries list is walked in order, and the
115  * first valid match based on the function/index and flags will be emitted.
116  *
117  * If no match is found, but Intel-style fallback is configured, then the
118  * highest standard leaf encountered will be emitted.
119  */
120 static const struct vcpu_cpuid_entry *
121 cpuid_find_entry(const vcpu_cpuid_config_t *cfg, uint32_t func, uint32_t idx)
122 {
123 	const struct vcpu_cpuid_entry *last_std = NULL;
124 	const bool intel_fallback =
125 	    (cfg->vcc_flags & VCC_FLAG_INTEL_FALLBACK) != 0;
126 	bool matched_leaf = false;
127 
128 	ASSERT0(cfg->vcc_flags & VCC_FLAG_LEGACY_HANDLING);
129 
130 	for (uint_t i = 0; i < cfg->vcc_nent; i++) {
131 		const struct vcpu_cpuid_entry *ent = &cfg->vcc_entries[i];
132 		const bool ent_is_std =
133 		    (ent->vce_function & CPUID_TYPE_MASK) == CPUID_TYPE_STD;
134 		const bool ent_must_match_idx =
135 		    (ent->vce_flags & VCE_FLAG_MATCH_INDEX) != 0;
136 
137 		if (ent_is_std) {
138 			/*
139 			 * Keep track of the last "standard" leaf for
140 			 * Intel-style fallback behavior.
141 			 *
142 			 * This does currently not account for the sub-leaf
143 			 * index matching behavior for fallback described in the
144 			 * SDM.  It is not clear if any consumers rely on such
145 			 * matching when encountering fallback.
146 			 */
147 			last_std = ent;
148 		}
149 		if (ent->vce_function == func) {
150 			if (ent->vce_index == idx || !ent_must_match_idx) {
151 				return (ent);
152 			}
153 			/*
154 			 * Make note of when the top-level leaf matches, even
155 			 * when the index does not.
156 			 */
157 			matched_leaf = true;
158 		} else if (ent->vce_function > func) {
159 			if ((ent->vce_function & CPUID_TYPE_MASK) ==
160 			    (func & CPUID_TYPE_MASK)) {
161 				/*
162 				 * We are beyond a valid leaf to match, but have
163 				 * not exceeded the maximum leaf for this "type"
164 				 * (standard, extended, hvm, etc), so return an
165 				 * empty entry.
166 				 */
167 				return (&cpuid_empty_entry);
168 			} else {
169 				/*
170 				 * Otherwise, we can stop now, having gone
171 				 * beyond the last entry which could match the
172 				 * target function in a sorted list.
173 				 */
174 				break;
175 			}
176 		}
177 	}
178 
179 	if (matched_leaf || !intel_fallback) {
180 		return (&cpuid_empty_entry);
181 	} else {
182 		return (last_std);
183 	}
184 }
185 
186 void
187 vcpu_emulate_cpuid(struct vm *vm, int vcpuid, uint64_t *rax, uint64_t *rbx,
188     uint64_t *rcx, uint64_t *rdx)
189 {
190 	const vcpu_cpuid_config_t *cfg = vm_cpuid_config(vm, vcpuid);
191 
192 	ASSERT3P(rax, !=, NULL);
193 	ASSERT3P(rbx, !=, NULL);
194 	ASSERT3P(rcx, !=, NULL);
195 	ASSERT3P(rdx, !=, NULL);
196 
197 	/* Fall back to legacy handling if specified */
198 	if ((cfg->vcc_flags & VCC_FLAG_LEGACY_HANDLING) != 0) {
199 		uint32_t regs[4] = { *rax, 0, *rcx, 0 };
200 
201 		legacy_emulate_cpuid(vm, vcpuid, &regs[0], &regs[1], &regs[2],
202 		    &regs[3]);
203 		/* CPUID clears the upper 32-bits of the long-mode registers. */
204 		*rax = regs[0];
205 		*rbx = regs[1];
206 		*rcx = regs[2];
207 		*rdx = regs[3];
208 		return;
209 	}
210 
211 	const struct vcpu_cpuid_entry *ent = cpuid_find_entry(cfg, *rax, *rcx);
212 	ASSERT(ent != NULL);
213 	/* CPUID clears the upper 32-bits of the long-mode registers. */
214 	*rax = ent->vce_eax;
215 	*rbx = ent->vce_ebx;
216 	*rcx = ent->vce_ecx;
217 	*rdx = ent->vce_edx;
218 }
219 
220 /*
221  * Get the current CPUID emulation configuration for this vCPU.
222  *
223  * Only the existing flags will be emitted if the vCPU is configured for legacy
224  * operation via the VCC_FLAG_LEGACY_HANDLING flag.  If in userspace-controlled
225  * mode, then we will attempt to copy the existing entries into vcc_entries,
226  * its side specified by vcc_nent.
227  *
228  * Regardless of whether vcc_entries is adequately sized (or even present),
229  * vcc_nent will be set to the number of existing entries.
230  */
231 int
232 vm_get_cpuid(struct vm *vm, int vcpuid, vcpu_cpuid_config_t *res)
233 {
234 	if (vcpuid < 0 || vcpuid > VM_MAXCPU) {
235 		return (EINVAL);
236 	}
237 
238 	const vcpu_cpuid_config_t *src = vm_cpuid_config(vm, vcpuid);
239 	if (src->vcc_nent > res->vcc_nent) {
240 		res->vcc_nent = src->vcc_nent;
241 		return (E2BIG);
242 	} else if (src->vcc_nent != 0) {
243 		bcopy(src->vcc_entries, res->vcc_entries,
244 		    src->vcc_nent * sizeof (struct vcpu_cpuid_entry));
245 	}
246 	res->vcc_flags = src->vcc_flags;
247 	res->vcc_nent = src->vcc_nent;
248 	return (0);
249 }
250 
251 /*
252  * Set the CPUID emulation configuration for this vCPU.
253  *
254  * If VCC_FLAG_LEGACY_HANDLING is set in vcc_flags, then vcc_nent is expected to
255  * be set to 0, as configuring a list of entries would be useless when using the
256  * legacy handling.
257  *
258  * Any existing entries which are configured are freed, and the newly provided
259  * ones will be copied into their place.
260  */
261 int
262 vm_set_cpuid(struct vm *vm, int vcpuid, const vcpu_cpuid_config_t *src)
263 {
264 	if (vcpuid < 0 || vcpuid > VM_MAXCPU) {
265 		return (EINVAL);
266 	}
267 	if (src->vcc_nent > VMM_MAX_CPUID_ENTRIES) {
268 		return (EINVAL);
269 	}
270 	if ((src->vcc_flags & ~VCC_FLAGS_VALID) != 0) {
271 		return (EINVAL);
272 	}
273 	if ((src->vcc_flags & VCC_FLAG_LEGACY_HANDLING) != 0 &&
274 	    src->vcc_nent != 0) {
275 		/* No entries should be provided if using legacy handling */
276 		return (EINVAL);
277 	}
278 	for (uint_t i = 0; i < src->vcc_nent; i++) {
279 		/* Ensure all entries carry valid flags */
280 		if ((src->vcc_entries[i].vce_flags & ~VCE_FLAGS_VALID) != 0) {
281 			return (EINVAL);
282 		}
283 	}
284 
285 	vcpu_cpuid_config_t *cfg = vm_cpuid_config(vm, vcpuid);
286 
287 	/* Free any existing entries first */
288 	vcpu_cpuid_cleanup(cfg);
289 
290 	/* Copy supplied entries into freshly allocated space */
291 	if (src->vcc_nent != 0) {
292 		const size_t entries_sz =
293 		    src->vcc_nent * sizeof (struct vcpu_cpuid_entry);
294 
295 		cfg->vcc_nent = src->vcc_nent;
296 		cfg->vcc_entries = kmem_alloc(entries_sz, KM_SLEEP);
297 		bcopy(src->vcc_entries, cfg->vcc_entries, entries_sz);
298 	}
299 	cfg->vcc_flags = src->vcc_flags;
300 
301 	return (0);
302 }
303 
304 void
305 vcpu_cpuid_init(vcpu_cpuid_config_t *cfg)
306 {
307 	/* Default to legacy-style handling */
308 	cfg->vcc_flags = VCC_FLAG_LEGACY_HANDLING;
309 	cfg->vcc_nent = 0;
310 	cfg->vcc_entries = NULL;
311 }
312 
313 void
314 vcpu_cpuid_cleanup(vcpu_cpuid_config_t *cfg)
315 {
316 	if (cfg->vcc_nent != 0) {
317 		ASSERT3P(cfg->vcc_entries, !=, NULL);
318 
319 		kmem_free(cfg->vcc_entries,
320 		    cfg->vcc_nent * sizeof (struct vcpu_cpuid_entry));
321 
322 		cfg->vcc_nent = 0;
323 		cfg->vcc_entries = NULL;
324 	}
325 }
326 
327 static const char bhyve_id[12] = "bhyve bhyve ";
328 
329 /*
330  * Force exposition of the invariant TSC capability, regardless of whether the
331  * host CPU reports having it.
332  */
333 static int vmm_force_invariant_tsc = 0;
334 
335 #define	CPUID_0000_0000	(0x0)
336 #define	CPUID_0000_0001	(0x1)
337 #define	CPUID_0000_0002	(0x2)
338 #define	CPUID_0000_0003	(0x3)
339 #define	CPUID_0000_0004	(0x4)
340 #define	CPUID_0000_0006	(0x6)
341 #define	CPUID_0000_0007	(0x7)
342 #define	CPUID_0000_000A	(0xA)
343 #define	CPUID_0000_000B	(0xB)
344 #define	CPUID_0000_000D	(0xD)
345 #define	CPUID_0000_000F	(0xF)
346 #define	CPUID_0000_0010	(0x10)
347 #define	CPUID_0000_0015	(0x15)
348 #define	CPUID_8000_0000	(0x80000000)
349 #define	CPUID_8000_0001	(0x80000001)
350 #define	CPUID_8000_0002	(0x80000002)
351 #define	CPUID_8000_0003	(0x80000003)
352 #define	CPUID_8000_0004	(0x80000004)
353 #define	CPUID_8000_0006	(0x80000006)
354 #define	CPUID_8000_0007	(0x80000007)
355 #define	CPUID_8000_0008	(0x80000008)
356 #define	CPUID_8000_001D	(0x8000001D)
357 #define	CPUID_8000_001E	(0x8000001E)
358 
359 #define	CPUID_VM_HIGH	0x40000000
360 
361 /*
362  * CPUID instruction Fn0000_0001:
363  */
364 #define	CPUID_0000_0001_APICID_SHIFT	24
365 
366 
367 /*
368  * Round up to the next power of two, if necessary, and then take log2.
369  * Returns -1 if argument is zero.
370  */
371 static __inline int
372 log2(uint_t x)
373 {
374 	return (fls(x << (1 - powerof2(x))) - 1);
375 }
376 
377 /*
378  * The "legacy" bhyve cpuid emulation, which largly applies statically defined
379  * masks to the data provided by the host CPU.
380  */
381 void
382 legacy_emulate_cpuid(struct vm *vm, int vcpu_id, uint32_t *eax, uint32_t *ebx,
383     uint32_t *ecx, uint32_t *edx)
384 {
385 	const struct xsave_limits *limits;
386 	uint64_t cr4;
387 	int error, enable_invpcid, level, width = 0, x2apic_id = 0;
388 	unsigned int func, regs[4], logical_cpus = 0, param;
389 	enum x2apic_state x2apic_state;
390 	uint16_t cores, maxcpus, sockets, threads;
391 
392 	/*
393 	 * The function of CPUID is controlled through the provided value of
394 	 * %eax (and secondarily %ecx, for certain leaf data).
395 	 */
396 	func = (uint32_t)*eax;
397 	param = (uint32_t)*ecx;
398 
399 	/*
400 	 * Requests for invalid CPUID levels should map to the highest
401 	 * available level instead.
402 	 */
403 	if (cpu_exthigh != 0 && func >= 0x80000000) {
404 		if (func > cpu_exthigh)
405 			func = cpu_exthigh;
406 	} else if (func >= 0x40000000) {
407 		if (func > CPUID_VM_HIGH)
408 			func = CPUID_VM_HIGH;
409 	} else if (func > cpu_high) {
410 		func = cpu_high;
411 	}
412 
413 	/*
414 	 * In general the approach used for CPU topology is to
415 	 * advertise a flat topology where all CPUs are packages with
416 	 * no multi-core or SMT.
417 	 */
418 	switch (func) {
419 		/*
420 		 * Pass these through to the guest
421 		 */
422 		case CPUID_0000_0000:
423 		case CPUID_0000_0002:
424 		case CPUID_0000_0003:
425 		case CPUID_8000_0000:
426 		case CPUID_8000_0002:
427 		case CPUID_8000_0003:
428 		case CPUID_8000_0004:
429 		case CPUID_8000_0006:
430 			cpuid_count(func, param, regs);
431 			break;
432 		case CPUID_8000_0008:
433 			cpuid_count(func, param, regs);
434 			if (vmm_is_svm()) {
435 				/*
436 				 * As on Intel (0000_0007:0, EDX), mask out
437 				 * unsupported or unsafe AMD extended features
438 				 * (8000_0008 EBX).
439 				 */
440 				regs[1] &= (AMDFEID_CLZERO | AMDFEID_IRPERF |
441 				    AMDFEID_XSAVEERPTR);
442 
443 				vm_get_topology(vm, &sockets, &cores, &threads,
444 				    &maxcpus);
445 				/*
446 				 * Here, width is ApicIdCoreIdSize, present on
447 				 * at least Family 15h and newer.  It
448 				 * represents the "number of bits in the
449 				 * initial apicid that indicate thread id
450 				 * within a package."
451 				 *
452 				 * Our topo_probe_amd() uses it for
453 				 * pkg_id_shift and other OSes may rely on it.
454 				 */
455 				width = MIN(0xF, log2(threads * cores));
456 				if (width < 0x4)
457 					width = 0;
458 				logical_cpus = MIN(0xFF, threads * cores - 1);
459 				regs[2] = (width << AMDID_COREID_SIZE_SHIFT) |
460 				    logical_cpus;
461 			}
462 			break;
463 
464 		case CPUID_8000_0001:
465 			cpuid_count(func, param, regs);
466 
467 			/*
468 			 * Hide SVM from guest.
469 			 */
470 			regs[2] &= ~AMDID2_SVM;
471 
472 			/*
473 			 * Don't advertise extended performance counter MSRs
474 			 * to the guest.
475 			 */
476 			regs[2] &= ~AMDID2_PCXC;
477 			regs[2] &= ~AMDID2_PNXC;
478 			regs[2] &= ~AMDID2_PTSCEL2I;
479 
480 			/*
481 			 * Don't advertise Instruction Based Sampling feature.
482 			 */
483 			regs[2] &= ~AMDID2_IBS;
484 
485 			/* NodeID MSR not available */
486 			regs[2] &= ~AMDID2_NODE_ID;
487 
488 			/* Don't advertise the OS visible workaround feature */
489 			regs[2] &= ~AMDID2_OSVW;
490 
491 			/* Hide mwaitx/monitorx capability from the guest */
492 			regs[2] &= ~AMDID2_MWAITX;
493 
494 #ifndef __FreeBSD__
495 			/*
496 			 * Detection routines for TCE and FFXSR are missing
497 			 * from our vm_cpuid_capability() detection logic
498 			 * today.  Mask them out until that is remedied.
499 			 * They do not appear to be in common usage, so their
500 			 * absence should not cause undue trouble.
501 			 */
502 			regs[2] &= ~AMDID2_TCE;
503 			regs[3] &= ~AMDID_FFXSR;
504 #endif
505 
506 			/*
507 			 * Hide rdtscp/ia32_tsc_aux until we know how
508 			 * to deal with them.
509 			 */
510 			regs[3] &= ~AMDID_RDTSCP;
511 			break;
512 
513 		case CPUID_8000_0007:
514 			cpuid_count(func, param, regs);
515 			/*
516 			 * AMD uses this leaf to advertise the processor's
517 			 * power monitoring and RAS capabilities. These
518 			 * features are hardware-specific and exposing
519 			 * them to a guest doesn't make a lot of sense.
520 			 *
521 			 * Intel uses this leaf only to advertise the
522 			 * "Invariant TSC" feature with all other bits
523 			 * being reserved (set to zero).
524 			 */
525 			regs[0] = 0;
526 			regs[1] = 0;
527 			regs[2] = 0;
528 
529 			/*
530 			 * If the host system possesses an invariant TSC, then
531 			 * it is safe to expose to the guest.
532 			 *
533 			 * If there is measured skew between host TSCs, it will
534 			 * be properly offset so guests do not observe any
535 			 * change between CPU migrations.
536 			 */
537 			regs[3] &= AMDPM_TSC_INVARIANT;
538 
539 			/*
540 			 * Since illumos avoids deep C-states on CPUs which do
541 			 * not support an invariant TSC, it may be safe (and
542 			 * desired) to unconditionally expose that capability to
543 			 * the guest.
544 			 */
545 			if (vmm_force_invariant_tsc != 0) {
546 				regs[3] |= AMDPM_TSC_INVARIANT;
547 			}
548 			break;
549 
550 		case CPUID_8000_001D:
551 			/* AMD Cache topology, like 0000_0004 for Intel. */
552 			if (!vmm_is_svm())
553 				goto default_leaf;
554 
555 			/*
556 			 * Similar to Intel, generate a ficticious cache
557 			 * topology for the guest with L3 shared by the
558 			 * package, and L1 and L2 local to a core.
559 			 */
560 			vm_get_topology(vm, &sockets, &cores, &threads,
561 			    &maxcpus);
562 			switch (param) {
563 			case 0:
564 				logical_cpus = threads;
565 				level = 1;
566 				func = 1;	/* data cache */
567 				break;
568 			case 1:
569 				logical_cpus = threads;
570 				level = 2;
571 				func = 3;	/* unified cache */
572 				break;
573 			case 2:
574 				logical_cpus = threads * cores;
575 				level = 3;
576 				func = 3;	/* unified cache */
577 				break;
578 			default:
579 				logical_cpus = 0;
580 				level = 0;
581 				func = 0;
582 				break;
583 			}
584 
585 			logical_cpus = MIN(0xfff, logical_cpus - 1);
586 			regs[0] = (logical_cpus << 14) | (1 << 8) |
587 			    (level << 5) | func;
588 			regs[1] = (func > 0) ? (CACHE_LINE_SIZE - 1) : 0;
589 			regs[2] = 0;
590 			regs[3] = 0;
591 			break;
592 
593 		case CPUID_8000_001E:
594 			/*
595 			 * AMD Family 16h+ and Hygon Family 18h additional
596 			 * identifiers.
597 			 */
598 			if (!vmm_is_svm() || CPUID_TO_FAMILY(cpu_id) < 0x16)
599 				goto default_leaf;
600 
601 			vm_get_topology(vm, &sockets, &cores, &threads,
602 			    &maxcpus);
603 			regs[0] = vcpu_id;
604 			threads = MIN(0xFF, threads - 1);
605 			regs[1] = (threads << 8) |
606 			    (vcpu_id >> log2(threads + 1));
607 			/*
608 			 * XXX Bhyve topology cannot yet represent >1 node per
609 			 * processor.
610 			 */
611 			regs[2] = 0;
612 			regs[3] = 0;
613 			break;
614 
615 		case CPUID_0000_0001:
616 			do_cpuid(1, regs);
617 
618 			error = vm_get_x2apic_state(vm, vcpu_id, &x2apic_state);
619 			VERIFY0(error);
620 
621 			/*
622 			 * Override the APIC ID only in ebx
623 			 */
624 			regs[1] &= ~(CPUID_LOCAL_APIC_ID);
625 			regs[1] |= (vcpu_id << CPUID_0000_0001_APICID_SHIFT);
626 
627 			/*
628 			 * Don't expose VMX, SpeedStep, TME or SMX capability.
629 			 * Advertise x2APIC capability and Hypervisor guest.
630 			 */
631 			regs[2] &= ~(CPUID2_VMX | CPUID2_EST | CPUID2_TM2);
632 			regs[2] &= ~(CPUID2_SMX);
633 
634 			regs[2] |= CPUID2_HV;
635 
636 			if (x2apic_state != X2APIC_DISABLED)
637 				regs[2] |= CPUID2_X2APIC;
638 			else
639 				regs[2] &= ~CPUID2_X2APIC;
640 
641 			/*
642 			 * Only advertise CPUID2_XSAVE in the guest if
643 			 * the host is using XSAVE.
644 			 */
645 			if (!(regs[2] & CPUID2_OSXSAVE))
646 				regs[2] &= ~CPUID2_XSAVE;
647 
648 			/*
649 			 * If CPUID2_XSAVE is being advertised and the
650 			 * guest has set CR4_XSAVE, set
651 			 * CPUID2_OSXSAVE.
652 			 */
653 			regs[2] &= ~CPUID2_OSXSAVE;
654 			if (regs[2] & CPUID2_XSAVE) {
655 				error = vm_get_register(vm, vcpu_id,
656 				    VM_REG_GUEST_CR4, &cr4);
657 				VERIFY0(error);
658 				if (cr4 & CR4_XSAVE)
659 					regs[2] |= CPUID2_OSXSAVE;
660 			}
661 
662 			/*
663 			 * Hide monitor/mwait until we know how to deal with
664 			 * these instructions.
665 			 */
666 			regs[2] &= ~CPUID2_MON;
667 
668 			/*
669 			 * Hide the performance and debug features.
670 			 */
671 			regs[2] &= ~CPUID2_PDCM;
672 
673 			/*
674 			 * No TSC deadline support in the APIC yet
675 			 */
676 			regs[2] &= ~CPUID2_TSCDLT;
677 
678 			/*
679 			 * Hide thermal monitoring
680 			 */
681 			regs[3] &= ~(CPUID_ACPI | CPUID_TM);
682 
683 			/*
684 			 * Hide the debug store capability.
685 			 */
686 			regs[3] &= ~CPUID_DS;
687 
688 			/*
689 			 * Advertise the Machine Check and MTRR capability.
690 			 *
691 			 * Some guest OSes (e.g. Windows) will not boot if
692 			 * these features are absent.
693 			 */
694 			regs[3] |= (CPUID_MCA | CPUID_MCE | CPUID_MTRR);
695 
696 			vm_get_topology(vm, &sockets, &cores, &threads,
697 			    &maxcpus);
698 			logical_cpus = threads * cores;
699 			regs[1] &= ~CPUID_HTT_CORES;
700 			regs[1] |= (logical_cpus & 0xff) << 16;
701 			regs[3] |= CPUID_HTT;
702 			break;
703 
704 		case CPUID_0000_0004:
705 			cpuid_count(func, param, regs);
706 
707 			if (regs[0] || regs[1] || regs[2] || regs[3]) {
708 				vm_get_topology(vm, &sockets, &cores, &threads,
709 				    &maxcpus);
710 				regs[0] &= 0x3ff;
711 				regs[0] |= (cores - 1) << 26;
712 				/*
713 				 * Cache topology:
714 				 * - L1 and L2 are shared only by the logical
715 				 *   processors in a single core.
716 				 * - L3 and above are shared by all logical
717 				 *   processors in the package.
718 				 */
719 				logical_cpus = threads;
720 				level = (regs[0] >> 5) & 0x7;
721 				if (level >= 3)
722 					logical_cpus *= cores;
723 				regs[0] |= (logical_cpus - 1) << 14;
724 			}
725 			break;
726 
727 		case CPUID_0000_0007:
728 			regs[0] = 0;
729 			regs[1] = 0;
730 			regs[2] = 0;
731 			regs[3] = 0;
732 
733 			/* leaf 0 */
734 			if (param == 0) {
735 				cpuid_count(func, param, regs);
736 
737 				/* Only leaf 0 is supported */
738 				regs[0] = 0;
739 
740 				/*
741 				 * Expose known-safe features.
742 				 */
743 				regs[1] &= (CPUID_STDEXT_FSGSBASE |
744 				    CPUID_STDEXT_BMI1 | CPUID_STDEXT_HLE |
745 				    CPUID_STDEXT_AVX2 | CPUID_STDEXT_SMEP |
746 				    CPUID_STDEXT_BMI2 |
747 				    CPUID_STDEXT_ERMS | CPUID_STDEXT_RTM |
748 				    CPUID_STDEXT_AVX512F |
749 				    CPUID_STDEXT_RDSEED |
750 				    CPUID_STDEXT_SMAP |
751 				    CPUID_STDEXT_AVX512PF |
752 				    CPUID_STDEXT_AVX512ER |
753 				    CPUID_STDEXT_AVX512CD | CPUID_STDEXT_SHA);
754 				regs[2] = 0;
755 				regs[3] &= CPUID_STDEXT3_MD_CLEAR;
756 
757 				/* Advertise INVPCID if it is enabled. */
758 				error = vm_get_capability(vm, vcpu_id,
759 				    VM_CAP_ENABLE_INVPCID, &enable_invpcid);
760 				if (error == 0 && enable_invpcid)
761 					regs[1] |= CPUID_STDEXT_INVPCID;
762 			}
763 			break;
764 
765 		case CPUID_0000_0006:
766 			regs[0] = CPUTPM1_ARAT;
767 			regs[1] = 0;
768 			regs[2] = 0;
769 			regs[3] = 0;
770 			break;
771 
772 		case CPUID_0000_000A:
773 			/*
774 			 * Handle the access, but report 0 for
775 			 * all options
776 			 */
777 			regs[0] = 0;
778 			regs[1] = 0;
779 			regs[2] = 0;
780 			regs[3] = 0;
781 			break;
782 
783 		case CPUID_0000_000B:
784 			/*
785 			 * Intel processor topology enumeration
786 			 */
787 			if (vmm_is_intel()) {
788 				vm_get_topology(vm, &sockets, &cores, &threads,
789 				    &maxcpus);
790 				if (param == 0) {
791 					logical_cpus = threads;
792 					width = log2(logical_cpus);
793 					level = CPUID_TYPE_SMT;
794 					x2apic_id = vcpu_id;
795 				}
796 
797 				if (param == 1) {
798 					logical_cpus = threads * cores;
799 					width = log2(logical_cpus);
800 					level = CPUID_TYPE_CORE;
801 					x2apic_id = vcpu_id;
802 				}
803 
804 				if (param >= 2) {
805 					width = 0;
806 					logical_cpus = 0;
807 					level = 0;
808 					x2apic_id = 0;
809 				}
810 
811 				regs[0] = width & 0x1f;
812 				regs[1] = logical_cpus & 0xffff;
813 				regs[2] = (level << 8) | (param & 0xff);
814 				regs[3] = x2apic_id;
815 			} else {
816 				regs[0] = 0;
817 				regs[1] = 0;
818 				regs[2] = 0;
819 				regs[3] = 0;
820 			}
821 			break;
822 
823 		case CPUID_0000_000D:
824 			limits = vmm_get_xsave_limits();
825 			if (!limits->xsave_enabled) {
826 				regs[0] = 0;
827 				regs[1] = 0;
828 				regs[2] = 0;
829 				regs[3] = 0;
830 				break;
831 			}
832 
833 			cpuid_count(func, param, regs);
834 			switch (param) {
835 			case 0:
836 				/*
837 				 * Only permit the guest to use bits
838 				 * that are active in the host in
839 				 * %xcr0.  Also, claim that the
840 				 * maximum save area size is
841 				 * equivalent to the host's current
842 				 * save area size.  Since this runs
843 				 * "inside" of vmrun(), it runs with
844 				 * the guest's xcr0, so the current
845 				 * save area size is correct as-is.
846 				 */
847 				regs[0] &= limits->xcr0_allowed;
848 				regs[2] = limits->xsave_max_size;
849 				regs[3] &= (limits->xcr0_allowed >> 32);
850 				break;
851 			case 1:
852 				/* Only permit XSAVEOPT. */
853 				regs[0] &= CPUID_EXTSTATE_XSAVEOPT;
854 				regs[1] = 0;
855 				regs[2] = 0;
856 				regs[3] = 0;
857 				break;
858 			default:
859 				/*
860 				 * If the leaf is for a permitted feature,
861 				 * pass through as-is, otherwise return
862 				 * all zeroes.
863 				 */
864 				if (!(limits->xcr0_allowed & (1ul << param))) {
865 					regs[0] = 0;
866 					regs[1] = 0;
867 					regs[2] = 0;
868 					regs[3] = 0;
869 				}
870 				break;
871 			}
872 			break;
873 
874 		case CPUID_0000_000F:
875 		case CPUID_0000_0010:
876 			/*
877 			 * Do not report any Resource Director Technology
878 			 * capabilities.  Exposing control of cache or memory
879 			 * controller resource partitioning to the guest is not
880 			 * at all sensible.
881 			 *
882 			 * This is already hidden at a high level by masking of
883 			 * leaf 0x7.  Even still, a guest may look here for
884 			 * detailed capability information.
885 			 */
886 			regs[0] = 0;
887 			regs[1] = 0;
888 			regs[2] = 0;
889 			regs[3] = 0;
890 			break;
891 
892 		case CPUID_0000_0015:
893 			/*
894 			 * Don't report CPU TSC/Crystal ratio and clock
895 			 * values since guests may use these to derive the
896 			 * local APIC frequency..
897 			 */
898 			regs[0] = 0;
899 			regs[1] = 0;
900 			regs[2] = 0;
901 			regs[3] = 0;
902 			break;
903 
904 		case 0x40000000:
905 			regs[0] = CPUID_VM_HIGH;
906 			bcopy(bhyve_id, &regs[1], 4);
907 			bcopy(bhyve_id + 4, &regs[2], 4);
908 			bcopy(bhyve_id + 8, &regs[3], 4);
909 			break;
910 
911 		default:
912 default_leaf:
913 			/*
914 			 * The leaf value has already been clamped so
915 			 * simply pass this through.
916 			 */
917 			cpuid_count(func, param, regs);
918 			break;
919 	}
920 
921 	*eax = regs[0];
922 	*ebx = regs[1];
923 	*ecx = regs[2];
924 	*edx = regs[3];
925 }
926