xref: /openbmc/linux/arch/x86/kvm/cpuid.c (revision ac3cb6de)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Kernel-based Virtual Machine driver for Linux
4  * cpuid support routines
5  *
6  * derived from arch/x86/kvm/x86.c
7  *
8  * Copyright 2011 Red Hat, Inc. and/or its affiliates.
9  * Copyright IBM Corporation, 2008
10  */
11 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 
13 #include <linux/kvm_host.h>
14 #include <linux/export.h>
15 #include <linux/vmalloc.h>
16 #include <linux/uaccess.h>
17 #include <linux/sched/stat.h>
18 
19 #include <asm/processor.h>
20 #include <asm/user.h>
21 #include <asm/fpu/xstate.h>
22 #include <asm/sgx.h>
23 #include <asm/cpuid.h>
24 #include "cpuid.h"
25 #include "lapic.h"
26 #include "mmu.h"
27 #include "trace.h"
28 #include "pmu.h"
29 #include "xen.h"
30 
31 /*
32  * Unlike "struct cpuinfo_x86.x86_capability", kvm_cpu_caps doesn't need to be
33  * aligned to sizeof(unsigned long) because it's not accessed via bitops.
34  */
35 u32 kvm_cpu_caps[NR_KVM_CPU_CAPS] __read_mostly;
36 EXPORT_SYMBOL_GPL(kvm_cpu_caps);
37 
38 u32 xstate_required_size(u64 xstate_bv, bool compacted)
39 {
40 	int feature_bit = 0;
41 	u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
42 
43 	xstate_bv &= XFEATURE_MASK_EXTEND;
44 	while (xstate_bv) {
45 		if (xstate_bv & 0x1) {
46 		        u32 eax, ebx, ecx, edx, offset;
47 		        cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx);
48 			/* ECX[1]: 64B alignment in compacted form */
49 			if (compacted)
50 				offset = (ecx & 0x2) ? ALIGN(ret, 64) : ret;
51 			else
52 				offset = ebx;
53 			ret = max(ret, offset + eax);
54 		}
55 
56 		xstate_bv >>= 1;
57 		feature_bit++;
58 	}
59 
60 	return ret;
61 }
62 
63 #define F feature_bit
64 
65 /* Scattered Flag - For features that are scattered by cpufeatures.h. */
66 #define SF(name)						\
67 ({								\
68 	BUILD_BUG_ON(X86_FEATURE_##name >= MAX_CPU_FEATURES);	\
69 	(boot_cpu_has(X86_FEATURE_##name) ? F(name) : 0);	\
70 })
71 
72 /*
73  * Magic value used by KVM when querying userspace-provided CPUID entries and
74  * doesn't care about the CPIUD index because the index of the function in
75  * question is not significant.  Note, this magic value must have at least one
76  * bit set in bits[63:32] and must be consumed as a u64 by cpuid_entry2_find()
77  * to avoid false positives when processing guest CPUID input.
78  */
79 #define KVM_CPUID_INDEX_NOT_SIGNIFICANT -1ull
80 
81 static inline struct kvm_cpuid_entry2 *cpuid_entry2_find(
82 	struct kvm_cpuid_entry2 *entries, int nent, u32 function, u64 index)
83 {
84 	struct kvm_cpuid_entry2 *e;
85 	int i;
86 
87 	for (i = 0; i < nent; i++) {
88 		e = &entries[i];
89 
90 		if (e->function != function)
91 			continue;
92 
93 		/*
94 		 * If the index isn't significant, use the first entry with a
95 		 * matching function.  It's userspace's responsibilty to not
96 		 * provide "duplicate" entries in all cases.
97 		 */
98 		if (!(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) || e->index == index)
99 			return e;
100 
101 
102 		/*
103 		 * Similarly, use the first matching entry if KVM is doing a
104 		 * lookup (as opposed to emulating CPUID) for a function that's
105 		 * architecturally defined as not having a significant index.
106 		 */
107 		if (index == KVM_CPUID_INDEX_NOT_SIGNIFICANT) {
108 			/*
109 			 * Direct lookups from KVM should not diverge from what
110 			 * KVM defines internally (the architectural behavior).
111 			 */
112 			WARN_ON_ONCE(cpuid_function_is_indexed(function));
113 			return e;
114 		}
115 	}
116 
117 	return NULL;
118 }
119 
120 static int kvm_check_cpuid(struct kvm_vcpu *vcpu,
121 			   struct kvm_cpuid_entry2 *entries,
122 			   int nent)
123 {
124 	struct kvm_cpuid_entry2 *best;
125 	u64 xfeatures;
126 
127 	/*
128 	 * The existing code assumes virtual address is 48-bit or 57-bit in the
129 	 * canonical address checks; exit if it is ever changed.
130 	 */
131 	best = cpuid_entry2_find(entries, nent, 0x80000008,
132 				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
133 	if (best) {
134 		int vaddr_bits = (best->eax & 0xff00) >> 8;
135 
136 		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
137 			return -EINVAL;
138 	}
139 
140 	/*
141 	 * Exposing dynamic xfeatures to the guest requires additional
142 	 * enabling in the FPU, e.g. to expand the guest XSAVE state size.
143 	 */
144 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
145 	if (!best)
146 		return 0;
147 
148 	xfeatures = best->eax | ((u64)best->edx << 32);
149 	xfeatures &= XFEATURE_MASK_USER_DYNAMIC;
150 	if (!xfeatures)
151 		return 0;
152 
153 	return fpu_enable_guest_xfd_features(&vcpu->arch.guest_fpu, xfeatures);
154 }
155 
156 /* Check whether the supplied CPUID data is equal to what is already set for the vCPU. */
157 static int kvm_cpuid_check_equal(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
158 				 int nent)
159 {
160 	struct kvm_cpuid_entry2 *orig;
161 	int i;
162 
163 	if (nent != vcpu->arch.cpuid_nent)
164 		return -EINVAL;
165 
166 	for (i = 0; i < nent; i++) {
167 		orig = &vcpu->arch.cpuid_entries[i];
168 		if (e2[i].function != orig->function ||
169 		    e2[i].index != orig->index ||
170 		    e2[i].flags != orig->flags ||
171 		    e2[i].eax != orig->eax || e2[i].ebx != orig->ebx ||
172 		    e2[i].ecx != orig->ecx || e2[i].edx != orig->edx)
173 			return -EINVAL;
174 	}
175 
176 	return 0;
177 }
178 
179 static struct kvm_hypervisor_cpuid kvm_get_hypervisor_cpuid(struct kvm_vcpu *vcpu,
180 							    const char *sig)
181 {
182 	struct kvm_hypervisor_cpuid cpuid = {};
183 	struct kvm_cpuid_entry2 *entry;
184 	u32 base;
185 
186 	for_each_possible_hypervisor_cpuid_base(base) {
187 		entry = kvm_find_cpuid_entry(vcpu, base);
188 
189 		if (entry) {
190 			u32 signature[3];
191 
192 			signature[0] = entry->ebx;
193 			signature[1] = entry->ecx;
194 			signature[2] = entry->edx;
195 
196 			if (!memcmp(signature, sig, sizeof(signature))) {
197 				cpuid.base = base;
198 				cpuid.limit = entry->eax;
199 				break;
200 			}
201 		}
202 	}
203 
204 	return cpuid;
205 }
206 
207 static struct kvm_cpuid_entry2 *__kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu,
208 					      struct kvm_cpuid_entry2 *entries, int nent)
209 {
210 	u32 base = vcpu->arch.kvm_cpuid.base;
211 
212 	if (!base)
213 		return NULL;
214 
215 	return cpuid_entry2_find(entries, nent, base | KVM_CPUID_FEATURES,
216 				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
217 }
218 
219 static struct kvm_cpuid_entry2 *kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu)
220 {
221 	return __kvm_find_kvm_cpuid_features(vcpu, vcpu->arch.cpuid_entries,
222 					     vcpu->arch.cpuid_nent);
223 }
224 
225 void kvm_update_pv_runtime(struct kvm_vcpu *vcpu)
226 {
227 	struct kvm_cpuid_entry2 *best = kvm_find_kvm_cpuid_features(vcpu);
228 
229 	/*
230 	 * save the feature bitmap to avoid cpuid lookup for every PV
231 	 * operation
232 	 */
233 	if (best)
234 		vcpu->arch.pv_cpuid.features = best->eax;
235 }
236 
237 /*
238  * Calculate guest's supported XCR0 taking into account guest CPUID data and
239  * KVM's supported XCR0 (comprised of host's XCR0 and KVM_SUPPORTED_XCR0).
240  */
241 static u64 cpuid_get_supported_xcr0(struct kvm_cpuid_entry2 *entries, int nent)
242 {
243 	struct kvm_cpuid_entry2 *best;
244 
245 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
246 	if (!best)
247 		return 0;
248 
249 	return (best->eax | ((u64)best->edx << 32)) & kvm_caps.supported_xcr0;
250 }
251 
252 static void __kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *entries,
253 				       int nent)
254 {
255 	struct kvm_cpuid_entry2 *best;
256 
257 	best = cpuid_entry2_find(entries, nent, 1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
258 	if (best) {
259 		/* Update OSXSAVE bit */
260 		if (boot_cpu_has(X86_FEATURE_XSAVE))
261 			cpuid_entry_change(best, X86_FEATURE_OSXSAVE,
262 					   kvm_is_cr4_bit_set(vcpu, X86_CR4_OSXSAVE));
263 
264 		cpuid_entry_change(best, X86_FEATURE_APIC,
265 			   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
266 	}
267 
268 	best = cpuid_entry2_find(entries, nent, 7, 0);
269 	if (best && boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7)
270 		cpuid_entry_change(best, X86_FEATURE_OSPKE,
271 				   kvm_is_cr4_bit_set(vcpu, X86_CR4_PKE));
272 
273 	best = cpuid_entry2_find(entries, nent, 0xD, 0);
274 	if (best)
275 		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
276 
277 	best = cpuid_entry2_find(entries, nent, 0xD, 1);
278 	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
279 		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
280 		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
281 
282 	best = __kvm_find_kvm_cpuid_features(vcpu, entries, nent);
283 	if (kvm_hlt_in_guest(vcpu->kvm) && best &&
284 		(best->eax & (1 << KVM_FEATURE_PV_UNHALT)))
285 		best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
286 
287 	if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
288 		best = cpuid_entry2_find(entries, nent, 0x1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
289 		if (best)
290 			cpuid_entry_change(best, X86_FEATURE_MWAIT,
291 					   vcpu->arch.ia32_misc_enable_msr &
292 					   MSR_IA32_MISC_ENABLE_MWAIT);
293 	}
294 }
295 
296 void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu)
297 {
298 	__kvm_update_cpuid_runtime(vcpu, vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
299 }
300 EXPORT_SYMBOL_GPL(kvm_update_cpuid_runtime);
301 
302 static bool kvm_cpuid_has_hyperv(struct kvm_cpuid_entry2 *entries, int nent)
303 {
304 	struct kvm_cpuid_entry2 *entry;
305 
306 	entry = cpuid_entry2_find(entries, nent, HYPERV_CPUID_INTERFACE,
307 				  KVM_CPUID_INDEX_NOT_SIGNIFICANT);
308 	return entry && entry->eax == HYPERV_CPUID_SIGNATURE_EAX;
309 }
310 
311 static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
312 {
313 	struct kvm_lapic *apic = vcpu->arch.apic;
314 	struct kvm_cpuid_entry2 *best;
315 
316 	best = kvm_find_cpuid_entry(vcpu, 1);
317 	if (best && apic) {
318 		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
319 			apic->lapic_timer.timer_mode_mask = 3 << 17;
320 		else
321 			apic->lapic_timer.timer_mode_mask = 1 << 17;
322 
323 		kvm_apic_set_version(vcpu);
324 	}
325 
326 	vcpu->arch.guest_supported_xcr0 =
327 		cpuid_get_supported_xcr0(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
328 
329 	/*
330 	 * FP+SSE can always be saved/restored via KVM_{G,S}ET_XSAVE, even if
331 	 * XSAVE/XCRO are not exposed to the guest, and even if XSAVE isn't
332 	 * supported by the host.
333 	 */
334 	vcpu->arch.guest_fpu.fpstate->user_xfeatures = vcpu->arch.guest_supported_xcr0 |
335 						       XFEATURE_MASK_FPSSE;
336 
337 	kvm_update_pv_runtime(vcpu);
338 
339 	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
340 	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
341 
342 	kvm_pmu_refresh(vcpu);
343 	vcpu->arch.cr4_guest_rsvd_bits =
344 	    __cr4_reserved_bits(guest_cpuid_has, vcpu);
345 
346 	kvm_hv_set_cpuid(vcpu, kvm_cpuid_has_hyperv(vcpu->arch.cpuid_entries,
347 						    vcpu->arch.cpuid_nent));
348 
349 	/* Invoke the vendor callback only after the above state is updated. */
350 	static_call(kvm_x86_vcpu_after_set_cpuid)(vcpu);
351 
352 	/*
353 	 * Except for the MMU, which needs to do its thing any vendor specific
354 	 * adjustments to the reserved GPA bits.
355 	 */
356 	kvm_mmu_after_set_cpuid(vcpu);
357 }
358 
359 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
360 {
361 	struct kvm_cpuid_entry2 *best;
362 
363 	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
364 	if (!best || best->eax < 0x80000008)
365 		goto not_found;
366 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
367 	if (best)
368 		return best->eax & 0xff;
369 not_found:
370 	return 36;
371 }
372 
373 /*
374  * This "raw" version returns the reserved GPA bits without any adjustments for
375  * encryption technologies that usurp bits.  The raw mask should be used if and
376  * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
377  */
378 u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
379 {
380 	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
381 }
382 
383 static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
384                         int nent)
385 {
386 	int r;
387 
388 	__kvm_update_cpuid_runtime(vcpu, e2, nent);
389 
390 	/*
391 	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
392 	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
393 	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
394 	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
395 	 * the core vCPU model on the fly. It would've been better to forbid any
396 	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
397 	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
398 	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
399 	 * whether the supplied CPUID data is equal to what's already set.
400 	 */
401 	if (kvm_vcpu_has_run(vcpu)) {
402 		r = kvm_cpuid_check_equal(vcpu, e2, nent);
403 		if (r)
404 			return r;
405 
406 		kvfree(e2);
407 		return 0;
408 	}
409 
410 	if (kvm_cpuid_has_hyperv(e2, nent)) {
411 		r = kvm_hv_vcpu_init(vcpu);
412 		if (r)
413 			return r;
414 	}
415 
416 	r = kvm_check_cpuid(vcpu, e2, nent);
417 	if (r)
418 		return r;
419 
420 	kvfree(vcpu->arch.cpuid_entries);
421 	vcpu->arch.cpuid_entries = e2;
422 	vcpu->arch.cpuid_nent = nent;
423 
424 	vcpu->arch.kvm_cpuid = kvm_get_hypervisor_cpuid(vcpu, KVM_SIGNATURE);
425 	vcpu->arch.xen.cpuid = kvm_get_hypervisor_cpuid(vcpu, XEN_SIGNATURE);
426 	kvm_vcpu_after_set_cpuid(vcpu);
427 
428 	return 0;
429 }
430 
431 /* when an old userspace process fills a new kernel module */
432 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
433 			     struct kvm_cpuid *cpuid,
434 			     struct kvm_cpuid_entry __user *entries)
435 {
436 	int r, i;
437 	struct kvm_cpuid_entry *e = NULL;
438 	struct kvm_cpuid_entry2 *e2 = NULL;
439 
440 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
441 		return -E2BIG;
442 
443 	if (cpuid->nent) {
444 		e = vmemdup_user(entries, array_size(sizeof(*e), cpuid->nent));
445 		if (IS_ERR(e))
446 			return PTR_ERR(e);
447 
448 		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
449 		if (!e2) {
450 			r = -ENOMEM;
451 			goto out_free_cpuid;
452 		}
453 	}
454 	for (i = 0; i < cpuid->nent; i++) {
455 		e2[i].function = e[i].function;
456 		e2[i].eax = e[i].eax;
457 		e2[i].ebx = e[i].ebx;
458 		e2[i].ecx = e[i].ecx;
459 		e2[i].edx = e[i].edx;
460 		e2[i].index = 0;
461 		e2[i].flags = 0;
462 		e2[i].padding[0] = 0;
463 		e2[i].padding[1] = 0;
464 		e2[i].padding[2] = 0;
465 	}
466 
467 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
468 	if (r)
469 		kvfree(e2);
470 
471 out_free_cpuid:
472 	kvfree(e);
473 
474 	return r;
475 }
476 
477 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
478 			      struct kvm_cpuid2 *cpuid,
479 			      struct kvm_cpuid_entry2 __user *entries)
480 {
481 	struct kvm_cpuid_entry2 *e2 = NULL;
482 	int r;
483 
484 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
485 		return -E2BIG;
486 
487 	if (cpuid->nent) {
488 		e2 = vmemdup_user(entries, array_size(sizeof(*e2), cpuid->nent));
489 		if (IS_ERR(e2))
490 			return PTR_ERR(e2);
491 	}
492 
493 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
494 	if (r)
495 		kvfree(e2);
496 
497 	return r;
498 }
499 
500 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
501 			      struct kvm_cpuid2 *cpuid,
502 			      struct kvm_cpuid_entry2 __user *entries)
503 {
504 	if (cpuid->nent < vcpu->arch.cpuid_nent)
505 		return -E2BIG;
506 
507 	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
508 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
509 		return -EFAULT;
510 
511 	cpuid->nent = vcpu->arch.cpuid_nent;
512 	return 0;
513 }
514 
515 /* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */
516 static __always_inline void __kvm_cpu_cap_mask(unsigned int leaf)
517 {
518 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
519 	struct kvm_cpuid_entry2 entry;
520 
521 	reverse_cpuid_check(leaf);
522 
523 	cpuid_count(cpuid.function, cpuid.index,
524 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
525 
526 	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
527 }
528 
529 static __always_inline
530 void kvm_cpu_cap_init_kvm_defined(enum kvm_only_cpuid_leafs leaf, u32 mask)
531 {
532 	/* Use kvm_cpu_cap_mask for leafs that aren't KVM-only. */
533 	BUILD_BUG_ON(leaf < NCAPINTS);
534 
535 	kvm_cpu_caps[leaf] = mask;
536 
537 	__kvm_cpu_cap_mask(leaf);
538 }
539 
540 static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
541 {
542 	/* Use kvm_cpu_cap_init_kvm_defined for KVM-only leafs. */
543 	BUILD_BUG_ON(leaf >= NCAPINTS);
544 
545 	kvm_cpu_caps[leaf] &= mask;
546 
547 	__kvm_cpu_cap_mask(leaf);
548 }
549 
550 void kvm_set_cpu_caps(void)
551 {
552 #ifdef CONFIG_X86_64
553 	unsigned int f_gbpages = F(GBPAGES);
554 	unsigned int f_lm = F(LM);
555 	unsigned int f_xfd = F(XFD);
556 #else
557 	unsigned int f_gbpages = 0;
558 	unsigned int f_lm = 0;
559 	unsigned int f_xfd = 0;
560 #endif
561 	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
562 
563 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
564 		     sizeof(boot_cpu_data.x86_capability));
565 
566 	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
567 	       sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)));
568 
569 	kvm_cpu_cap_mask(CPUID_1_ECX,
570 		/*
571 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
572 		 * advertised to guests via CPUID!
573 		 */
574 		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
575 		0 /* DS-CPL, VMX, SMX, EST */ |
576 		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
577 		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
578 		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
579 		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
580 		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
581 		F(F16C) | F(RDRAND)
582 	);
583 	/* KVM emulates x2apic in software irrespective of host support. */
584 	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
585 
586 	kvm_cpu_cap_mask(CPUID_1_EDX,
587 		F(FPU) | F(VME) | F(DE) | F(PSE) |
588 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
589 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
590 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
591 		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
592 		0 /* Reserved, DS, ACPI */ | F(MMX) |
593 		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
594 		0 /* HTT, TM, Reserved, PBE */
595 	);
596 
597 	kvm_cpu_cap_mask(CPUID_7_0_EBX,
598 		F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) |
599 		F(FDP_EXCPTN_ONLY) | F(SMEP) | F(BMI2) | F(ERMS) | F(INVPCID) |
600 		F(RTM) | F(ZERO_FCS_FDS) | 0 /*MPX*/ | F(AVX512F) |
601 		F(AVX512DQ) | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) |
602 		F(CLFLUSHOPT) | F(CLWB) | 0 /*INTEL_PT*/ | F(AVX512PF) |
603 		F(AVX512ER) | F(AVX512CD) | F(SHA_NI) | F(AVX512BW) |
604 		F(AVX512VL));
605 
606 	kvm_cpu_cap_mask(CPUID_7_ECX,
607 		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
608 		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
609 		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
610 		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/ |
611 		F(SGX_LC) | F(BUS_LOCK_DETECT)
612 	);
613 	/* Set LA57 based on hardware capability. */
614 	if (cpuid_ecx(7) & F(LA57))
615 		kvm_cpu_cap_set(X86_FEATURE_LA57);
616 
617 	/*
618 	 * PKU not yet implemented for shadow paging and requires OSPKE
619 	 * to be set on the host. Clear it if that is not the case
620 	 */
621 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
622 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
623 
624 	kvm_cpu_cap_mask(CPUID_7_EDX,
625 		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
626 		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
627 		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
628 		F(SERIALIZE) | F(TSXLDTRK) | F(AVX512_FP16) |
629 		F(AMX_TILE) | F(AMX_INT8) | F(AMX_BF16) | F(FLUSH_L1D)
630 	);
631 
632 	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
633 	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
634 	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
635 
636 	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
637 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
638 	if (boot_cpu_has(X86_FEATURE_STIBP))
639 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
640 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
641 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
642 
643 	kvm_cpu_cap_mask(CPUID_7_1_EAX,
644 		F(AVX_VNNI) | F(AVX512_BF16) | F(CMPCCXADD) |
645 		F(FZRM) | F(FSRS) | F(FSRC) |
646 		F(AMX_FP16) | F(AVX_IFMA)
647 	);
648 
649 	kvm_cpu_cap_init_kvm_defined(CPUID_7_1_EDX,
650 		F(AVX_VNNI_INT8) | F(AVX_NE_CONVERT) | F(PREFETCHITI)
651 	);
652 
653 	kvm_cpu_cap_mask(CPUID_D_1_EAX,
654 		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES) | f_xfd
655 	);
656 
657 	kvm_cpu_cap_init_kvm_defined(CPUID_12_EAX,
658 		SF(SGX1) | SF(SGX2) | SF(SGX_EDECCSSA)
659 	);
660 
661 	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
662 		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
663 		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
664 		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
665 		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
666 		F(TOPOEXT) | 0 /* PERFCTR_CORE */
667 	);
668 
669 	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
670 		F(FPU) | F(VME) | F(DE) | F(PSE) |
671 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
672 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
673 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
674 		F(PAT) | F(PSE36) | 0 /* Reserved */ |
675 		F(NX) | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
676 		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
677 		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
678 	);
679 
680 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
681 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
682 
683 	kvm_cpu_cap_init_kvm_defined(CPUID_8000_0007_EDX,
684 		SF(CONSTANT_TSC)
685 	);
686 
687 	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
688 		F(CLZERO) | F(XSAVEERPTR) |
689 		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
690 		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON) |
691 		F(AMD_PSFD)
692 	);
693 
694 	/*
695 	 * AMD has separate bits for each SPEC_CTRL bit.
696 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
697 	 * record that in cpufeatures so use them.
698 	 */
699 	if (boot_cpu_has(X86_FEATURE_IBPB))
700 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
701 	if (boot_cpu_has(X86_FEATURE_IBRS))
702 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
703 	if (boot_cpu_has(X86_FEATURE_STIBP))
704 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
705 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
706 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
707 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
708 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
709 	/*
710 	 * The preference is to use SPEC CTRL MSR instead of the
711 	 * VIRT_SPEC MSR.
712 	 */
713 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
714 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
715 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
716 
717 	/*
718 	 * Hide all SVM features by default, SVM will set the cap bits for
719 	 * features it emulates and/or exposes for L1.
720 	 */
721 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
722 
723 	kvm_cpu_cap_mask(CPUID_8000_001F_EAX,
724 		0 /* SME */ | F(SEV) | 0 /* VM_PAGE_FLUSH */ | F(SEV_ES) |
725 		F(SME_COHERENT));
726 
727 	kvm_cpu_cap_mask(CPUID_8000_0021_EAX,
728 		F(NO_NESTED_DATA_BP) | F(LFENCE_RDTSC) | 0 /* SmmPgCfgLock */ |
729 		F(NULL_SEL_CLR_BASE) | F(AUTOIBRS) | 0 /* PrefetchCtlMsr */
730 	);
731 
732 	kvm_cpu_cap_init_kvm_defined(CPUID_8000_0022_EAX,
733 		F(PERFMON_V2)
734 	);
735 
736 	/*
737 	 * Synthesize "LFENCE is serializing" into the AMD-defined entry in
738 	 * KVM's supported CPUID if the feature is reported as supported by the
739 	 * kernel.  LFENCE_RDTSC was a Linux-defined synthetic feature long
740 	 * before AMD joined the bandwagon, e.g. LFENCE is serializing on most
741 	 * CPUs that support SSE2.  On CPUs that don't support AMD's leaf,
742 	 * kvm_cpu_cap_mask() will unfortunately drop the flag due to ANDing
743 	 * the mask with the raw host CPUID, and reporting support in AMD's
744 	 * leaf can make it easier for userspace to detect the feature.
745 	 */
746 	if (cpu_feature_enabled(X86_FEATURE_LFENCE_RDTSC))
747 		kvm_cpu_cap_set(X86_FEATURE_LFENCE_RDTSC);
748 	if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
749 		kvm_cpu_cap_set(X86_FEATURE_NULL_SEL_CLR_BASE);
750 	kvm_cpu_cap_set(X86_FEATURE_NO_SMM_CTL_MSR);
751 
752 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
753 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
754 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
755 		F(PMM) | F(PMM_EN)
756 	);
757 
758 	/*
759 	 * Hide RDTSCP and RDPID if either feature is reported as supported but
760 	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
761 	 * should never happen, but the guest will likely crash if RDTSCP or
762 	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
763 	 * the past.  For example, the sanity check may fire if this instance of
764 	 * KVM is running as L1 on top of an older, broken KVM.
765 	 */
766 	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
767 		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
768 		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
769 		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
770 		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
771 	}
772 }
773 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
774 
775 struct kvm_cpuid_array {
776 	struct kvm_cpuid_entry2 *entries;
777 	int maxnent;
778 	int nent;
779 };
780 
781 static struct kvm_cpuid_entry2 *get_next_cpuid(struct kvm_cpuid_array *array)
782 {
783 	if (array->nent >= array->maxnent)
784 		return NULL;
785 
786 	return &array->entries[array->nent++];
787 }
788 
789 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
790 					      u32 function, u32 index)
791 {
792 	struct kvm_cpuid_entry2 *entry = get_next_cpuid(array);
793 
794 	if (!entry)
795 		return NULL;
796 
797 	memset(entry, 0, sizeof(*entry));
798 	entry->function = function;
799 	entry->index = index;
800 	switch (function & 0xC0000000) {
801 	case 0x40000000:
802 		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
803 		return entry;
804 
805 	case 0x80000000:
806 		/*
807 		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
808 		 * would result in out-of-bounds calls to do_host_cpuid.
809 		 */
810 		{
811 			static int max_cpuid_80000000;
812 			if (!READ_ONCE(max_cpuid_80000000))
813 				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
814 			if (function > READ_ONCE(max_cpuid_80000000))
815 				return entry;
816 		}
817 		break;
818 
819 	default:
820 		break;
821 	}
822 
823 	cpuid_count(entry->function, entry->index,
824 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
825 
826 	if (cpuid_function_is_indexed(function))
827 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
828 
829 	return entry;
830 }
831 
832 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
833 {
834 	struct kvm_cpuid_entry2 *entry;
835 
836 	if (array->nent >= array->maxnent)
837 		return -E2BIG;
838 
839 	entry = &array->entries[array->nent];
840 	entry->function = func;
841 	entry->index = 0;
842 	entry->flags = 0;
843 
844 	switch (func) {
845 	case 0:
846 		entry->eax = 7;
847 		++array->nent;
848 		break;
849 	case 1:
850 		entry->ecx = F(MOVBE);
851 		++array->nent;
852 		break;
853 	case 7:
854 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
855 		entry->eax = 0;
856 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
857 			entry->ecx = F(RDPID);
858 		++array->nent;
859 		break;
860 	default:
861 		break;
862 	}
863 
864 	return 0;
865 }
866 
867 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
868 {
869 	struct kvm_cpuid_entry2 *entry;
870 	int r, i, max_idx;
871 
872 	/* all calls to cpuid_count() should be made on the same cpu */
873 	get_cpu();
874 
875 	r = -E2BIG;
876 
877 	entry = do_host_cpuid(array, function, 0);
878 	if (!entry)
879 		goto out;
880 
881 	switch (function) {
882 	case 0:
883 		/* Limited to the highest leaf implemented in KVM. */
884 		entry->eax = min(entry->eax, 0x1fU);
885 		break;
886 	case 1:
887 		cpuid_entry_override(entry, CPUID_1_EDX);
888 		cpuid_entry_override(entry, CPUID_1_ECX);
889 		break;
890 	case 2:
891 		/*
892 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
893 		 * CPUID(function=2, index=0) may return different results each
894 		 * time, with the least-significant byte in EAX enumerating the
895 		 * number of times software should do CPUID(2, 0).
896 		 *
897 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
898 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
899 		 * return 01H. Software should ignore this value and not
900 		 * interpret it as an informational descriptor", while AMD's
901 		 * APM states that CPUID(2) is reserved.
902 		 *
903 		 * WARN if a frankenstein CPU that supports virtualization and
904 		 * a stateful CPUID.0x2 is encountered.
905 		 */
906 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
907 		break;
908 	/* functions 4 and 0x8000001d have additional index. */
909 	case 4:
910 	case 0x8000001d:
911 		/*
912 		 * Read entries until the cache type in the previous entry is
913 		 * zero, i.e. indicates an invalid entry.
914 		 */
915 		for (i = 1; entry->eax & 0x1f; ++i) {
916 			entry = do_host_cpuid(array, function, i);
917 			if (!entry)
918 				goto out;
919 		}
920 		break;
921 	case 6: /* Thermal management */
922 		entry->eax = 0x4; /* allow ARAT */
923 		entry->ebx = 0;
924 		entry->ecx = 0;
925 		entry->edx = 0;
926 		break;
927 	/* function 7 has additional index. */
928 	case 7:
929 		entry->eax = min(entry->eax, 1u);
930 		cpuid_entry_override(entry, CPUID_7_0_EBX);
931 		cpuid_entry_override(entry, CPUID_7_ECX);
932 		cpuid_entry_override(entry, CPUID_7_EDX);
933 
934 		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
935 		if (entry->eax == 1) {
936 			entry = do_host_cpuid(array, function, 1);
937 			if (!entry)
938 				goto out;
939 
940 			cpuid_entry_override(entry, CPUID_7_1_EAX);
941 			cpuid_entry_override(entry, CPUID_7_1_EDX);
942 			entry->ebx = 0;
943 			entry->ecx = 0;
944 		}
945 		break;
946 	case 0xa: { /* Architectural Performance Monitoring */
947 		union cpuid10_eax eax;
948 		union cpuid10_edx edx;
949 
950 		if (!enable_pmu || !static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
951 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
952 			break;
953 		}
954 
955 		eax.split.version_id = kvm_pmu_cap.version;
956 		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
957 		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
958 		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
959 		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
960 		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
961 
962 		if (kvm_pmu_cap.version)
963 			edx.split.anythread_deprecated = 1;
964 		edx.split.reserved1 = 0;
965 		edx.split.reserved2 = 0;
966 
967 		entry->eax = eax.full;
968 		entry->ebx = kvm_pmu_cap.events_mask;
969 		entry->ecx = 0;
970 		entry->edx = edx.full;
971 		break;
972 	}
973 	case 0x1f:
974 	case 0xb:
975 		/*
976 		 * No topology; a valid topology is indicated by the presence
977 		 * of subleaf 1.
978 		 */
979 		entry->eax = entry->ebx = entry->ecx = 0;
980 		break;
981 	case 0xd: {
982 		u64 permitted_xcr0 = kvm_get_filtered_xcr0();
983 		u64 permitted_xss = kvm_caps.supported_xss;
984 
985 		entry->eax &= permitted_xcr0;
986 		entry->ebx = xstate_required_size(permitted_xcr0, false);
987 		entry->ecx = entry->ebx;
988 		entry->edx &= permitted_xcr0 >> 32;
989 		if (!permitted_xcr0)
990 			break;
991 
992 		entry = do_host_cpuid(array, function, 1);
993 		if (!entry)
994 			goto out;
995 
996 		cpuid_entry_override(entry, CPUID_D_1_EAX);
997 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
998 			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
999 							  true);
1000 		else {
1001 			WARN_ON_ONCE(permitted_xss != 0);
1002 			entry->ebx = 0;
1003 		}
1004 		entry->ecx &= permitted_xss;
1005 		entry->edx &= permitted_xss >> 32;
1006 
1007 		for (i = 2; i < 64; ++i) {
1008 			bool s_state;
1009 			if (permitted_xcr0 & BIT_ULL(i))
1010 				s_state = false;
1011 			else if (permitted_xss & BIT_ULL(i))
1012 				s_state = true;
1013 			else
1014 				continue;
1015 
1016 			entry = do_host_cpuid(array, function, i);
1017 			if (!entry)
1018 				goto out;
1019 
1020 			/*
1021 			 * The supported check above should have filtered out
1022 			 * invalid sub-leafs.  Only valid sub-leafs should
1023 			 * reach this point, and they should have a non-zero
1024 			 * save state size.  Furthermore, check whether the
1025 			 * processor agrees with permitted_xcr0/permitted_xss
1026 			 * on whether this is an XCR0- or IA32_XSS-managed area.
1027 			 */
1028 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
1029 				--array->nent;
1030 				continue;
1031 			}
1032 
1033 			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1034 				entry->ecx &= ~BIT_ULL(2);
1035 			entry->edx = 0;
1036 		}
1037 		break;
1038 	}
1039 	case 0x12:
1040 		/* Intel SGX */
1041 		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1042 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1043 			break;
1044 		}
1045 
1046 		/*
1047 		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1048 		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1049 		 * are restricted by kernel and KVM capabilities (like most
1050 		 * feature flags), while enclave size is unrestricted.
1051 		 */
1052 		cpuid_entry_override(entry, CPUID_12_EAX);
1053 		entry->ebx &= SGX_MISC_EXINFO;
1054 
1055 		entry = do_host_cpuid(array, function, 1);
1056 		if (!entry)
1057 			goto out;
1058 
1059 		/*
1060 		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1061 		 * feature flags.  Advertise all supported flags, including
1062 		 * privileged attributes that require explicit opt-in from
1063 		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1064 		 * expected to derive it from supported XCR0.
1065 		 */
1066 		entry->eax &= SGX_ATTR_PRIV_MASK | SGX_ATTR_UNPRIV_MASK;
1067 		entry->ebx &= 0;
1068 		break;
1069 	/* Intel PT */
1070 	case 0x14:
1071 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1072 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1073 			break;
1074 		}
1075 
1076 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1077 			if (!do_host_cpuid(array, function, i))
1078 				goto out;
1079 		}
1080 		break;
1081 	/* Intel AMX TILE */
1082 	case 0x1d:
1083 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1084 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1085 			break;
1086 		}
1087 
1088 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1089 			if (!do_host_cpuid(array, function, i))
1090 				goto out;
1091 		}
1092 		break;
1093 	case 0x1e: /* TMUL information */
1094 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1095 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1096 			break;
1097 		}
1098 		break;
1099 	case KVM_CPUID_SIGNATURE: {
1100 		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1101 		entry->eax = KVM_CPUID_FEATURES;
1102 		entry->ebx = sigptr[0];
1103 		entry->ecx = sigptr[1];
1104 		entry->edx = sigptr[2];
1105 		break;
1106 	}
1107 	case KVM_CPUID_FEATURES:
1108 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1109 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1110 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1111 			     (1 << KVM_FEATURE_ASYNC_PF) |
1112 			     (1 << KVM_FEATURE_PV_EOI) |
1113 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1114 			     (1 << KVM_FEATURE_PV_UNHALT) |
1115 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1116 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1117 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1118 			     (1 << KVM_FEATURE_POLL_CONTROL) |
1119 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1120 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1121 
1122 		if (sched_info_on())
1123 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1124 
1125 		entry->ebx = 0;
1126 		entry->ecx = 0;
1127 		entry->edx = 0;
1128 		break;
1129 	case 0x80000000:
1130 		entry->eax = min(entry->eax, 0x80000022);
1131 		/*
1132 		 * Serializing LFENCE is reported in a multitude of ways, and
1133 		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1134 		 * userspace by providing the CPUID leaf ourselves.
1135 		 *
1136 		 * However, only do it if the host has CPUID leaf 0x8000001d.
1137 		 * QEMU thinks that it can query the host blindly for that
1138 		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1139 		 * above.  The processor merrily returns values from the
1140 		 * highest Intel leaf which QEMU tries to use as the guest's
1141 		 * 0x8000001d.  Even worse, this can result in an infinite
1142 		 * loop if said highest leaf has no subleaves indexed by ECX.
1143 		 */
1144 		if (entry->eax >= 0x8000001d &&
1145 		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1146 		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1147 			entry->eax = max(entry->eax, 0x80000021);
1148 		break;
1149 	case 0x80000001:
1150 		entry->ebx &= ~GENMASK(27, 16);
1151 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1152 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1153 		break;
1154 	case 0x80000006:
1155 		/* Drop reserved bits, pass host L2 cache and TLB info. */
1156 		entry->edx &= ~GENMASK(17, 16);
1157 		break;
1158 	case 0x80000007: /* Advanced power management */
1159 		cpuid_entry_override(entry, CPUID_8000_0007_EDX);
1160 
1161 		/* mask against host */
1162 		entry->edx &= boot_cpu_data.x86_power;
1163 		entry->eax = entry->ebx = entry->ecx = 0;
1164 		break;
1165 	case 0x80000008: {
1166 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
1167 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
1168 		unsigned phys_as = entry->eax & 0xff;
1169 
1170 		/*
1171 		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1172 		 * the guest operates in the same PA space as the host, i.e.
1173 		 * reductions in MAXPHYADDR for memory encryption affect shadow
1174 		 * paging, too.
1175 		 *
1176 		 * If TDP is enabled but an explicit guest MAXPHYADDR is not
1177 		 * provided, use the raw bare metal MAXPHYADDR as reductions to
1178 		 * the HPAs do not affect GPAs.
1179 		 */
1180 		if (!tdp_enabled)
1181 			g_phys_as = boot_cpu_data.x86_phys_bits;
1182 		else if (!g_phys_as)
1183 			g_phys_as = phys_as;
1184 
1185 		entry->eax = g_phys_as | (virt_as << 8);
1186 		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
1187 		entry->edx = 0;
1188 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1189 		break;
1190 	}
1191 	case 0x8000000A:
1192 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1193 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1194 			break;
1195 		}
1196 		entry->eax = 1; /* SVM revision 1 */
1197 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1198 				   ASID emulation to nested SVM */
1199 		entry->ecx = 0; /* Reserved */
1200 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1201 		break;
1202 	case 0x80000019:
1203 		entry->ecx = entry->edx = 0;
1204 		break;
1205 	case 0x8000001a:
1206 		entry->eax &= GENMASK(2, 0);
1207 		entry->ebx = entry->ecx = entry->edx = 0;
1208 		break;
1209 	case 0x8000001e:
1210 		/* Do not return host topology information.  */
1211 		entry->eax = entry->ebx = entry->ecx = 0;
1212 		entry->edx = 0; /* reserved */
1213 		break;
1214 	case 0x8000001F:
1215 		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1216 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1217 		} else {
1218 			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1219 			/* Clear NumVMPL since KVM does not support VMPL.  */
1220 			entry->ebx &= ~GENMASK(31, 12);
1221 			/*
1222 			 * Enumerate '0' for "PA bits reduction", the adjusted
1223 			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1224 			 */
1225 			entry->ebx &= ~GENMASK(11, 6);
1226 		}
1227 		break;
1228 	case 0x80000020:
1229 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1230 		break;
1231 	case 0x80000021:
1232 		entry->ebx = entry->ecx = entry->edx = 0;
1233 		cpuid_entry_override(entry, CPUID_8000_0021_EAX);
1234 		break;
1235 	/* AMD Extended Performance Monitoring and Debug */
1236 	case 0x80000022: {
1237 		union cpuid_0x80000022_ebx ebx;
1238 
1239 		entry->ecx = entry->edx = 0;
1240 		if (!enable_pmu || !kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2)) {
1241 			entry->eax = entry->ebx;
1242 			break;
1243 		}
1244 
1245 		cpuid_entry_override(entry, CPUID_8000_0022_EAX);
1246 
1247 		if (kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2))
1248 			ebx.split.num_core_pmc = kvm_pmu_cap.num_counters_gp;
1249 		else if (kvm_cpu_cap_has(X86_FEATURE_PERFCTR_CORE))
1250 			ebx.split.num_core_pmc = AMD64_NUM_COUNTERS_CORE;
1251 		else
1252 			ebx.split.num_core_pmc = AMD64_NUM_COUNTERS;
1253 
1254 		entry->ebx = ebx.full;
1255 		break;
1256 	}
1257 	/*Add support for Centaur's CPUID instruction*/
1258 	case 0xC0000000:
1259 		/*Just support up to 0xC0000004 now*/
1260 		entry->eax = min(entry->eax, 0xC0000004);
1261 		break;
1262 	case 0xC0000001:
1263 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1264 		break;
1265 	case 3: /* Processor serial number */
1266 	case 5: /* MONITOR/MWAIT */
1267 	case 0xC0000002:
1268 	case 0xC0000003:
1269 	case 0xC0000004:
1270 	default:
1271 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1272 		break;
1273 	}
1274 
1275 	r = 0;
1276 
1277 out:
1278 	put_cpu();
1279 
1280 	return r;
1281 }
1282 
1283 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1284 			 unsigned int type)
1285 {
1286 	if (type == KVM_GET_EMULATED_CPUID)
1287 		return __do_cpuid_func_emulated(array, func);
1288 
1289 	return __do_cpuid_func(array, func);
1290 }
1291 
1292 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
1293 
1294 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1295 			  unsigned int type)
1296 {
1297 	u32 limit;
1298 	int r;
1299 
1300 	if (func == CENTAUR_CPUID_SIGNATURE &&
1301 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1302 		return 0;
1303 
1304 	r = do_cpuid_func(array, func, type);
1305 	if (r)
1306 		return r;
1307 
1308 	limit = array->entries[array->nent - 1].eax;
1309 	for (func = func + 1; func <= limit; ++func) {
1310 		r = do_cpuid_func(array, func, type);
1311 		if (r)
1312 			break;
1313 	}
1314 
1315 	return r;
1316 }
1317 
1318 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1319 				 __u32 num_entries, unsigned int ioctl_type)
1320 {
1321 	int i;
1322 	__u32 pad[3];
1323 
1324 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1325 		return false;
1326 
1327 	/*
1328 	 * We want to make sure that ->padding is being passed clean from
1329 	 * userspace in case we want to use it for something in the future.
1330 	 *
1331 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1332 	 * have to give ourselves satisfied only with the emulated side. /me
1333 	 * sheds a tear.
1334 	 */
1335 	for (i = 0; i < num_entries; i++) {
1336 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1337 			return true;
1338 
1339 		if (pad[0] || pad[1] || pad[2])
1340 			return true;
1341 	}
1342 	return false;
1343 }
1344 
1345 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1346 			    struct kvm_cpuid_entry2 __user *entries,
1347 			    unsigned int type)
1348 {
1349 	static const u32 funcs[] = {
1350 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1351 	};
1352 
1353 	struct kvm_cpuid_array array = {
1354 		.nent = 0,
1355 	};
1356 	int r, i;
1357 
1358 	if (cpuid->nent < 1)
1359 		return -E2BIG;
1360 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1361 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1362 
1363 	if (sanity_check_entries(entries, cpuid->nent, type))
1364 		return -EINVAL;
1365 
1366 	array.entries = kvcalloc(cpuid->nent, sizeof(struct kvm_cpuid_entry2), GFP_KERNEL);
1367 	if (!array.entries)
1368 		return -ENOMEM;
1369 
1370 	array.maxnent = cpuid->nent;
1371 
1372 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1373 		r = get_cpuid_func(&array, funcs[i], type);
1374 		if (r)
1375 			goto out_free;
1376 	}
1377 	cpuid->nent = array.nent;
1378 
1379 	if (copy_to_user(entries, array.entries,
1380 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1381 		r = -EFAULT;
1382 
1383 out_free:
1384 	kvfree(array.entries);
1385 	return r;
1386 }
1387 
1388 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
1389 						    u32 function, u32 index)
1390 {
1391 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1392 				 function, index);
1393 }
1394 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry_index);
1395 
1396 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1397 					      u32 function)
1398 {
1399 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1400 				 function, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
1401 }
1402 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1403 
1404 /*
1405  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1406  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1407  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1408  * range.  Centaur/VIA follows Intel semantics.
1409  *
1410  * A leaf is considered out-of-range if its function is higher than the maximum
1411  * supported leaf of its associated class or if its associated class does not
1412  * exist.
1413  *
1414  * There are three primary classes to be considered, with their respective
1415  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1416  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1417  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1418  *
1419  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1420  *  - Hypervisor: 0x40000000 - 0x4fffffff
1421  *  - Extended:   0x80000000 - 0xbfffffff
1422  *  - Centaur:    0xc0000000 - 0xcfffffff
1423  *
1424  * The Hypervisor class is further subdivided into sub-classes that each act as
1425  * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1426  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1427  * CPUID sub-classes are:
1428  *
1429  *  - HyperV:     0x40000000 - 0x400000ff
1430  *  - KVM:        0x40000100 - 0x400001ff
1431  */
1432 static struct kvm_cpuid_entry2 *
1433 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1434 {
1435 	struct kvm_cpuid_entry2 *basic, *class;
1436 	u32 function = *fn_ptr;
1437 
1438 	basic = kvm_find_cpuid_entry(vcpu, 0);
1439 	if (!basic)
1440 		return NULL;
1441 
1442 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1443 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1444 		return NULL;
1445 
1446 	if (function >= 0x40000000 && function <= 0x4fffffff)
1447 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1448 	else if (function >= 0xc0000000)
1449 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1450 	else
1451 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1452 
1453 	if (class && function <= class->eax)
1454 		return NULL;
1455 
1456 	/*
1457 	 * Leaf specific adjustments are also applied when redirecting to the
1458 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1459 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1460 	 * needs to be pulled from CPUID.0xb.1.
1461 	 */
1462 	*fn_ptr = basic->eax;
1463 
1464 	/*
1465 	 * The class does not exist or the requested function is out of range;
1466 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1467 	 * the original requested leaf is observed!
1468 	 */
1469 	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1470 }
1471 
1472 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1473 	       u32 *ecx, u32 *edx, bool exact_only)
1474 {
1475 	u32 orig_function = *eax, function = *eax, index = *ecx;
1476 	struct kvm_cpuid_entry2 *entry;
1477 	bool exact, used_max_basic = false;
1478 
1479 	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1480 	exact = !!entry;
1481 
1482 	if (!entry && !exact_only) {
1483 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1484 		used_max_basic = !!entry;
1485 	}
1486 
1487 	if (entry) {
1488 		*eax = entry->eax;
1489 		*ebx = entry->ebx;
1490 		*ecx = entry->ecx;
1491 		*edx = entry->edx;
1492 		if (function == 7 && index == 0) {
1493 			u64 data;
1494 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1495 			    (data & TSX_CTRL_CPUID_CLEAR))
1496 				*ebx &= ~(F(RTM) | F(HLE));
1497 		} else if (function == 0x80000007) {
1498 			if (kvm_hv_invtsc_suppressed(vcpu))
1499 				*edx &= ~SF(CONSTANT_TSC);
1500 		}
1501 	} else {
1502 		*eax = *ebx = *ecx = *edx = 0;
1503 		/*
1504 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1505 		 * and EDX is always the x2APIC ID, even for undefined
1506 		 * subleaves. Index 1 will exist iff the leaf is
1507 		 * implemented, so we pass through CL iff leaf 1
1508 		 * exists. EDX can be copied from any existing index.
1509 		 */
1510 		if (function == 0xb || function == 0x1f) {
1511 			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
1512 			if (entry) {
1513 				*ecx = index & 0xff;
1514 				*edx = entry->edx;
1515 			}
1516 		}
1517 	}
1518 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1519 			used_max_basic);
1520 	return exact;
1521 }
1522 EXPORT_SYMBOL_GPL(kvm_cpuid);
1523 
1524 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1525 {
1526 	u32 eax, ebx, ecx, edx;
1527 
1528 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1529 		return 1;
1530 
1531 	eax = kvm_rax_read(vcpu);
1532 	ecx = kvm_rcx_read(vcpu);
1533 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1534 	kvm_rax_write(vcpu, eax);
1535 	kvm_rbx_write(vcpu, ebx);
1536 	kvm_rcx_write(vcpu, ecx);
1537 	kvm_rdx_write(vcpu, edx);
1538 	return kvm_skip_emulated_instruction(vcpu);
1539 }
1540 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1541