xref: /openbmc/linux/arch/x86/kvm/cpuid.c (revision 9a5d500c)
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 	int r;
505 
506 	r = -E2BIG;
507 	if (cpuid->nent < vcpu->arch.cpuid_nent)
508 		goto out;
509 	r = -EFAULT;
510 	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
511 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
512 		goto out;
513 	return 0;
514 
515 out:
516 	cpuid->nent = vcpu->arch.cpuid_nent;
517 	return r;
518 }
519 
520 /* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */
521 static __always_inline void __kvm_cpu_cap_mask(unsigned int leaf)
522 {
523 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
524 	struct kvm_cpuid_entry2 entry;
525 
526 	reverse_cpuid_check(leaf);
527 
528 	cpuid_count(cpuid.function, cpuid.index,
529 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
530 
531 	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
532 }
533 
534 static __always_inline
535 void kvm_cpu_cap_init_kvm_defined(enum kvm_only_cpuid_leafs leaf, u32 mask)
536 {
537 	/* Use kvm_cpu_cap_mask for leafs that aren't KVM-only. */
538 	BUILD_BUG_ON(leaf < NCAPINTS);
539 
540 	kvm_cpu_caps[leaf] = mask;
541 
542 	__kvm_cpu_cap_mask(leaf);
543 }
544 
545 static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
546 {
547 	/* Use kvm_cpu_cap_init_kvm_defined for KVM-only leafs. */
548 	BUILD_BUG_ON(leaf >= NCAPINTS);
549 
550 	kvm_cpu_caps[leaf] &= mask;
551 
552 	__kvm_cpu_cap_mask(leaf);
553 }
554 
555 void kvm_set_cpu_caps(void)
556 {
557 #ifdef CONFIG_X86_64
558 	unsigned int f_gbpages = F(GBPAGES);
559 	unsigned int f_lm = F(LM);
560 	unsigned int f_xfd = F(XFD);
561 #else
562 	unsigned int f_gbpages = 0;
563 	unsigned int f_lm = 0;
564 	unsigned int f_xfd = 0;
565 #endif
566 	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
567 
568 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
569 		     sizeof(boot_cpu_data.x86_capability));
570 
571 	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
572 	       sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)));
573 
574 	kvm_cpu_cap_mask(CPUID_1_ECX,
575 		/*
576 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
577 		 * advertised to guests via CPUID!
578 		 */
579 		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
580 		0 /* DS-CPL, VMX, SMX, EST */ |
581 		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
582 		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
583 		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
584 		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
585 		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
586 		F(F16C) | F(RDRAND)
587 	);
588 	/* KVM emulates x2apic in software irrespective of host support. */
589 	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
590 
591 	kvm_cpu_cap_mask(CPUID_1_EDX,
592 		F(FPU) | F(VME) | F(DE) | F(PSE) |
593 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
594 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
595 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
596 		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
597 		0 /* Reserved, DS, ACPI */ | F(MMX) |
598 		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
599 		0 /* HTT, TM, Reserved, PBE */
600 	);
601 
602 	kvm_cpu_cap_mask(CPUID_7_0_EBX,
603 		F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) |
604 		F(FDP_EXCPTN_ONLY) | F(SMEP) | F(BMI2) | F(ERMS) | F(INVPCID) |
605 		F(RTM) | F(ZERO_FCS_FDS) | 0 /*MPX*/ | F(AVX512F) |
606 		F(AVX512DQ) | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) |
607 		F(CLFLUSHOPT) | F(CLWB) | 0 /*INTEL_PT*/ | F(AVX512PF) |
608 		F(AVX512ER) | F(AVX512CD) | F(SHA_NI) | F(AVX512BW) |
609 		F(AVX512VL));
610 
611 	kvm_cpu_cap_mask(CPUID_7_ECX,
612 		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
613 		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
614 		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
615 		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/ |
616 		F(SGX_LC) | F(BUS_LOCK_DETECT)
617 	);
618 	/* Set LA57 based on hardware capability. */
619 	if (cpuid_ecx(7) & F(LA57))
620 		kvm_cpu_cap_set(X86_FEATURE_LA57);
621 
622 	/*
623 	 * PKU not yet implemented for shadow paging and requires OSPKE
624 	 * to be set on the host. Clear it if that is not the case
625 	 */
626 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
627 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
628 
629 	kvm_cpu_cap_mask(CPUID_7_EDX,
630 		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
631 		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
632 		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
633 		F(SERIALIZE) | F(TSXLDTRK) | F(AVX512_FP16) |
634 		F(AMX_TILE) | F(AMX_INT8) | F(AMX_BF16) | F(FLUSH_L1D)
635 	);
636 
637 	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
638 	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
639 	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
640 
641 	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
642 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
643 	if (boot_cpu_has(X86_FEATURE_STIBP))
644 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
645 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
646 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
647 
648 	kvm_cpu_cap_mask(CPUID_7_1_EAX,
649 		F(AVX_VNNI) | F(AVX512_BF16) | F(CMPCCXADD) |
650 		F(FZRM) | F(FSRS) | F(FSRC) |
651 		F(AMX_FP16) | F(AVX_IFMA)
652 	);
653 
654 	kvm_cpu_cap_init_kvm_defined(CPUID_7_1_EDX,
655 		F(AVX_VNNI_INT8) | F(AVX_NE_CONVERT) | F(PREFETCHITI)
656 	);
657 
658 	kvm_cpu_cap_mask(CPUID_D_1_EAX,
659 		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES) | f_xfd
660 	);
661 
662 	kvm_cpu_cap_init_kvm_defined(CPUID_12_EAX,
663 		SF(SGX1) | SF(SGX2) | SF(SGX_EDECCSSA)
664 	);
665 
666 	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
667 		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
668 		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
669 		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
670 		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
671 		F(TOPOEXT) | 0 /* PERFCTR_CORE */
672 	);
673 
674 	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
675 		F(FPU) | F(VME) | F(DE) | F(PSE) |
676 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
677 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
678 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
679 		F(PAT) | F(PSE36) | 0 /* Reserved */ |
680 		F(NX) | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
681 		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
682 		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
683 	);
684 
685 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
686 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
687 
688 	kvm_cpu_cap_init_kvm_defined(CPUID_8000_0007_EDX,
689 		SF(CONSTANT_TSC)
690 	);
691 
692 	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
693 		F(CLZERO) | F(XSAVEERPTR) |
694 		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
695 		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON) |
696 		F(AMD_PSFD)
697 	);
698 
699 	/*
700 	 * AMD has separate bits for each SPEC_CTRL bit.
701 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
702 	 * record that in cpufeatures so use them.
703 	 */
704 	if (boot_cpu_has(X86_FEATURE_IBPB))
705 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
706 	if (boot_cpu_has(X86_FEATURE_IBRS))
707 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
708 	if (boot_cpu_has(X86_FEATURE_STIBP))
709 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
710 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
711 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
712 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
713 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
714 	/*
715 	 * The preference is to use SPEC CTRL MSR instead of the
716 	 * VIRT_SPEC MSR.
717 	 */
718 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
719 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
720 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
721 
722 	/*
723 	 * Hide all SVM features by default, SVM will set the cap bits for
724 	 * features it emulates and/or exposes for L1.
725 	 */
726 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
727 
728 	kvm_cpu_cap_mask(CPUID_8000_001F_EAX,
729 		0 /* SME */ | F(SEV) | 0 /* VM_PAGE_FLUSH */ | F(SEV_ES) |
730 		F(SME_COHERENT));
731 
732 	kvm_cpu_cap_mask(CPUID_8000_0021_EAX,
733 		F(NO_NESTED_DATA_BP) | F(LFENCE_RDTSC) | 0 /* SmmPgCfgLock */ |
734 		F(NULL_SEL_CLR_BASE) | F(AUTOIBRS) | 0 /* PrefetchCtlMsr */
735 	);
736 
737 	/*
738 	 * Synthesize "LFENCE is serializing" into the AMD-defined entry in
739 	 * KVM's supported CPUID if the feature is reported as supported by the
740 	 * kernel.  LFENCE_RDTSC was a Linux-defined synthetic feature long
741 	 * before AMD joined the bandwagon, e.g. LFENCE is serializing on most
742 	 * CPUs that support SSE2.  On CPUs that don't support AMD's leaf,
743 	 * kvm_cpu_cap_mask() will unfortunately drop the flag due to ANDing
744 	 * the mask with the raw host CPUID, and reporting support in AMD's
745 	 * leaf can make it easier for userspace to detect the feature.
746 	 */
747 	if (cpu_feature_enabled(X86_FEATURE_LFENCE_RDTSC))
748 		kvm_cpu_cap_set(X86_FEATURE_LFENCE_RDTSC);
749 	if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
750 		kvm_cpu_cap_set(X86_FEATURE_NULL_SEL_CLR_BASE);
751 	kvm_cpu_cap_set(X86_FEATURE_NO_SMM_CTL_MSR);
752 
753 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
754 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
755 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
756 		F(PMM) | F(PMM_EN)
757 	);
758 
759 	/*
760 	 * Hide RDTSCP and RDPID if either feature is reported as supported but
761 	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
762 	 * should never happen, but the guest will likely crash if RDTSCP or
763 	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
764 	 * the past.  For example, the sanity check may fire if this instance of
765 	 * KVM is running as L1 on top of an older, broken KVM.
766 	 */
767 	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
768 		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
769 		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
770 		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
771 		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
772 	}
773 }
774 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
775 
776 struct kvm_cpuid_array {
777 	struct kvm_cpuid_entry2 *entries;
778 	int maxnent;
779 	int nent;
780 };
781 
782 static struct kvm_cpuid_entry2 *get_next_cpuid(struct kvm_cpuid_array *array)
783 {
784 	if (array->nent >= array->maxnent)
785 		return NULL;
786 
787 	return &array->entries[array->nent++];
788 }
789 
790 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
791 					      u32 function, u32 index)
792 {
793 	struct kvm_cpuid_entry2 *entry = get_next_cpuid(array);
794 
795 	if (!entry)
796 		return NULL;
797 
798 	memset(entry, 0, sizeof(*entry));
799 	entry->function = function;
800 	entry->index = index;
801 	switch (function & 0xC0000000) {
802 	case 0x40000000:
803 		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
804 		return entry;
805 
806 	case 0x80000000:
807 		/*
808 		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
809 		 * would result in out-of-bounds calls to do_host_cpuid.
810 		 */
811 		{
812 			static int max_cpuid_80000000;
813 			if (!READ_ONCE(max_cpuid_80000000))
814 				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
815 			if (function > READ_ONCE(max_cpuid_80000000))
816 				return entry;
817 		}
818 		break;
819 
820 	default:
821 		break;
822 	}
823 
824 	cpuid_count(entry->function, entry->index,
825 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
826 
827 	if (cpuid_function_is_indexed(function))
828 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
829 
830 	return entry;
831 }
832 
833 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
834 {
835 	struct kvm_cpuid_entry2 *entry;
836 
837 	if (array->nent >= array->maxnent)
838 		return -E2BIG;
839 
840 	entry = &array->entries[array->nent];
841 	entry->function = func;
842 	entry->index = 0;
843 	entry->flags = 0;
844 
845 	switch (func) {
846 	case 0:
847 		entry->eax = 7;
848 		++array->nent;
849 		break;
850 	case 1:
851 		entry->ecx = F(MOVBE);
852 		++array->nent;
853 		break;
854 	case 7:
855 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
856 		entry->eax = 0;
857 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
858 			entry->ecx = F(RDPID);
859 		++array->nent;
860 		break;
861 	default:
862 		break;
863 	}
864 
865 	return 0;
866 }
867 
868 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
869 {
870 	struct kvm_cpuid_entry2 *entry;
871 	int r, i, max_idx;
872 
873 	/* all calls to cpuid_count() should be made on the same cpu */
874 	get_cpu();
875 
876 	r = -E2BIG;
877 
878 	entry = do_host_cpuid(array, function, 0);
879 	if (!entry)
880 		goto out;
881 
882 	switch (function) {
883 	case 0:
884 		/* Limited to the highest leaf implemented in KVM. */
885 		entry->eax = min(entry->eax, 0x1fU);
886 		break;
887 	case 1:
888 		cpuid_entry_override(entry, CPUID_1_EDX);
889 		cpuid_entry_override(entry, CPUID_1_ECX);
890 		break;
891 	case 2:
892 		/*
893 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
894 		 * CPUID(function=2, index=0) may return different results each
895 		 * time, with the least-significant byte in EAX enumerating the
896 		 * number of times software should do CPUID(2, 0).
897 		 *
898 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
899 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
900 		 * return 01H. Software should ignore this value and not
901 		 * interpret it as an informational descriptor", while AMD's
902 		 * APM states that CPUID(2) is reserved.
903 		 *
904 		 * WARN if a frankenstein CPU that supports virtualization and
905 		 * a stateful CPUID.0x2 is encountered.
906 		 */
907 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
908 		break;
909 	/* functions 4 and 0x8000001d have additional index. */
910 	case 4:
911 	case 0x8000001d:
912 		/*
913 		 * Read entries until the cache type in the previous entry is
914 		 * zero, i.e. indicates an invalid entry.
915 		 */
916 		for (i = 1; entry->eax & 0x1f; ++i) {
917 			entry = do_host_cpuid(array, function, i);
918 			if (!entry)
919 				goto out;
920 		}
921 		break;
922 	case 6: /* Thermal management */
923 		entry->eax = 0x4; /* allow ARAT */
924 		entry->ebx = 0;
925 		entry->ecx = 0;
926 		entry->edx = 0;
927 		break;
928 	/* function 7 has additional index. */
929 	case 7:
930 		entry->eax = min(entry->eax, 1u);
931 		cpuid_entry_override(entry, CPUID_7_0_EBX);
932 		cpuid_entry_override(entry, CPUID_7_ECX);
933 		cpuid_entry_override(entry, CPUID_7_EDX);
934 
935 		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
936 		if (entry->eax == 1) {
937 			entry = do_host_cpuid(array, function, 1);
938 			if (!entry)
939 				goto out;
940 
941 			cpuid_entry_override(entry, CPUID_7_1_EAX);
942 			cpuid_entry_override(entry, CPUID_7_1_EDX);
943 			entry->ebx = 0;
944 			entry->ecx = 0;
945 		}
946 		break;
947 	case 0xa: { /* Architectural Performance Monitoring */
948 		union cpuid10_eax eax;
949 		union cpuid10_edx edx;
950 
951 		if (!static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
952 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
953 			break;
954 		}
955 
956 		eax.split.version_id = kvm_pmu_cap.version;
957 		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
958 		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
959 		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
960 		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
961 		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
962 
963 		if (kvm_pmu_cap.version)
964 			edx.split.anythread_deprecated = 1;
965 		edx.split.reserved1 = 0;
966 		edx.split.reserved2 = 0;
967 
968 		entry->eax = eax.full;
969 		entry->ebx = kvm_pmu_cap.events_mask;
970 		entry->ecx = 0;
971 		entry->edx = edx.full;
972 		break;
973 	}
974 	case 0x1f:
975 	case 0xb:
976 		/*
977 		 * No topology; a valid topology is indicated by the presence
978 		 * of subleaf 1.
979 		 */
980 		entry->eax = entry->ebx = entry->ecx = 0;
981 		break;
982 	case 0xd: {
983 		u64 permitted_xcr0 = kvm_get_filtered_xcr0();
984 		u64 permitted_xss = kvm_caps.supported_xss;
985 
986 		entry->eax &= permitted_xcr0;
987 		entry->ebx = xstate_required_size(permitted_xcr0, false);
988 		entry->ecx = entry->ebx;
989 		entry->edx &= permitted_xcr0 >> 32;
990 		if (!permitted_xcr0)
991 			break;
992 
993 		entry = do_host_cpuid(array, function, 1);
994 		if (!entry)
995 			goto out;
996 
997 		cpuid_entry_override(entry, CPUID_D_1_EAX);
998 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
999 			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
1000 							  true);
1001 		else {
1002 			WARN_ON_ONCE(permitted_xss != 0);
1003 			entry->ebx = 0;
1004 		}
1005 		entry->ecx &= permitted_xss;
1006 		entry->edx &= permitted_xss >> 32;
1007 
1008 		for (i = 2; i < 64; ++i) {
1009 			bool s_state;
1010 			if (permitted_xcr0 & BIT_ULL(i))
1011 				s_state = false;
1012 			else if (permitted_xss & BIT_ULL(i))
1013 				s_state = true;
1014 			else
1015 				continue;
1016 
1017 			entry = do_host_cpuid(array, function, i);
1018 			if (!entry)
1019 				goto out;
1020 
1021 			/*
1022 			 * The supported check above should have filtered out
1023 			 * invalid sub-leafs.  Only valid sub-leafs should
1024 			 * reach this point, and they should have a non-zero
1025 			 * save state size.  Furthermore, check whether the
1026 			 * processor agrees with permitted_xcr0/permitted_xss
1027 			 * on whether this is an XCR0- or IA32_XSS-managed area.
1028 			 */
1029 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
1030 				--array->nent;
1031 				continue;
1032 			}
1033 
1034 			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1035 				entry->ecx &= ~BIT_ULL(2);
1036 			entry->edx = 0;
1037 		}
1038 		break;
1039 	}
1040 	case 0x12:
1041 		/* Intel SGX */
1042 		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1043 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1044 			break;
1045 		}
1046 
1047 		/*
1048 		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1049 		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1050 		 * are restricted by kernel and KVM capabilities (like most
1051 		 * feature flags), while enclave size is unrestricted.
1052 		 */
1053 		cpuid_entry_override(entry, CPUID_12_EAX);
1054 		entry->ebx &= SGX_MISC_EXINFO;
1055 
1056 		entry = do_host_cpuid(array, function, 1);
1057 		if (!entry)
1058 			goto out;
1059 
1060 		/*
1061 		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1062 		 * feature flags.  Advertise all supported flags, including
1063 		 * privileged attributes that require explicit opt-in from
1064 		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1065 		 * expected to derive it from supported XCR0.
1066 		 */
1067 		entry->eax &= SGX_ATTR_PRIV_MASK | SGX_ATTR_UNPRIV_MASK;
1068 		entry->ebx &= 0;
1069 		break;
1070 	/* Intel PT */
1071 	case 0x14:
1072 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1073 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1074 			break;
1075 		}
1076 
1077 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1078 			if (!do_host_cpuid(array, function, i))
1079 				goto out;
1080 		}
1081 		break;
1082 	/* Intel AMX TILE */
1083 	case 0x1d:
1084 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1085 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1086 			break;
1087 		}
1088 
1089 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1090 			if (!do_host_cpuid(array, function, i))
1091 				goto out;
1092 		}
1093 		break;
1094 	case 0x1e: /* TMUL information */
1095 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1096 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1097 			break;
1098 		}
1099 		break;
1100 	case KVM_CPUID_SIGNATURE: {
1101 		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1102 		entry->eax = KVM_CPUID_FEATURES;
1103 		entry->ebx = sigptr[0];
1104 		entry->ecx = sigptr[1];
1105 		entry->edx = sigptr[2];
1106 		break;
1107 	}
1108 	case KVM_CPUID_FEATURES:
1109 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1110 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1111 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1112 			     (1 << KVM_FEATURE_ASYNC_PF) |
1113 			     (1 << KVM_FEATURE_PV_EOI) |
1114 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1115 			     (1 << KVM_FEATURE_PV_UNHALT) |
1116 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1117 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1118 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1119 			     (1 << KVM_FEATURE_POLL_CONTROL) |
1120 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1121 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1122 
1123 		if (sched_info_on())
1124 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1125 
1126 		entry->ebx = 0;
1127 		entry->ecx = 0;
1128 		entry->edx = 0;
1129 		break;
1130 	case 0x80000000:
1131 		entry->eax = min(entry->eax, 0x80000021);
1132 		/*
1133 		 * Serializing LFENCE is reported in a multitude of ways, and
1134 		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1135 		 * userspace by providing the CPUID leaf ourselves.
1136 		 *
1137 		 * However, only do it if the host has CPUID leaf 0x8000001d.
1138 		 * QEMU thinks that it can query the host blindly for that
1139 		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1140 		 * above.  The processor merrily returns values from the
1141 		 * highest Intel leaf which QEMU tries to use as the guest's
1142 		 * 0x8000001d.  Even worse, this can result in an infinite
1143 		 * loop if said highest leaf has no subleaves indexed by ECX.
1144 		 */
1145 		if (entry->eax >= 0x8000001d &&
1146 		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1147 		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1148 			entry->eax = max(entry->eax, 0x80000021);
1149 		break;
1150 	case 0x80000001:
1151 		entry->ebx &= ~GENMASK(27, 16);
1152 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1153 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1154 		break;
1155 	case 0x80000006:
1156 		/* Drop reserved bits, pass host L2 cache and TLB info. */
1157 		entry->edx &= ~GENMASK(17, 16);
1158 		break;
1159 	case 0x80000007: /* Advanced power management */
1160 		cpuid_entry_override(entry, CPUID_8000_0007_EDX);
1161 
1162 		/* mask against host */
1163 		entry->edx &= boot_cpu_data.x86_power;
1164 		entry->eax = entry->ebx = entry->ecx = 0;
1165 		break;
1166 	case 0x80000008: {
1167 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
1168 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
1169 		unsigned phys_as = entry->eax & 0xff;
1170 
1171 		/*
1172 		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1173 		 * the guest operates in the same PA space as the host, i.e.
1174 		 * reductions in MAXPHYADDR for memory encryption affect shadow
1175 		 * paging, too.
1176 		 *
1177 		 * If TDP is enabled but an explicit guest MAXPHYADDR is not
1178 		 * provided, use the raw bare metal MAXPHYADDR as reductions to
1179 		 * the HPAs do not affect GPAs.
1180 		 */
1181 		if (!tdp_enabled)
1182 			g_phys_as = boot_cpu_data.x86_phys_bits;
1183 		else if (!g_phys_as)
1184 			g_phys_as = phys_as;
1185 
1186 		entry->eax = g_phys_as | (virt_as << 8);
1187 		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
1188 		entry->edx = 0;
1189 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1190 		break;
1191 	}
1192 	case 0x8000000A:
1193 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1194 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1195 			break;
1196 		}
1197 		entry->eax = 1; /* SVM revision 1 */
1198 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1199 				   ASID emulation to nested SVM */
1200 		entry->ecx = 0; /* Reserved */
1201 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1202 		break;
1203 	case 0x80000019:
1204 		entry->ecx = entry->edx = 0;
1205 		break;
1206 	case 0x8000001a:
1207 		entry->eax &= GENMASK(2, 0);
1208 		entry->ebx = entry->ecx = entry->edx = 0;
1209 		break;
1210 	case 0x8000001e:
1211 		/* Do not return host topology information.  */
1212 		entry->eax = entry->ebx = entry->ecx = 0;
1213 		entry->edx = 0; /* reserved */
1214 		break;
1215 	case 0x8000001F:
1216 		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1217 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1218 		} else {
1219 			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1220 			/* Clear NumVMPL since KVM does not support VMPL.  */
1221 			entry->ebx &= ~GENMASK(31, 12);
1222 			/*
1223 			 * Enumerate '0' for "PA bits reduction", the adjusted
1224 			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1225 			 */
1226 			entry->ebx &= ~GENMASK(11, 6);
1227 		}
1228 		break;
1229 	case 0x80000020:
1230 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1231 		break;
1232 	case 0x80000021:
1233 		entry->ebx = entry->ecx = entry->edx = 0;
1234 		cpuid_entry_override(entry, CPUID_8000_0021_EAX);
1235 		break;
1236 	/*Add support for Centaur's CPUID instruction*/
1237 	case 0xC0000000:
1238 		/*Just support up to 0xC0000004 now*/
1239 		entry->eax = min(entry->eax, 0xC0000004);
1240 		break;
1241 	case 0xC0000001:
1242 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1243 		break;
1244 	case 3: /* Processor serial number */
1245 	case 5: /* MONITOR/MWAIT */
1246 	case 0xC0000002:
1247 	case 0xC0000003:
1248 	case 0xC0000004:
1249 	default:
1250 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1251 		break;
1252 	}
1253 
1254 	r = 0;
1255 
1256 out:
1257 	put_cpu();
1258 
1259 	return r;
1260 }
1261 
1262 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1263 			 unsigned int type)
1264 {
1265 	if (type == KVM_GET_EMULATED_CPUID)
1266 		return __do_cpuid_func_emulated(array, func);
1267 
1268 	return __do_cpuid_func(array, func);
1269 }
1270 
1271 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
1272 
1273 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1274 			  unsigned int type)
1275 {
1276 	u32 limit;
1277 	int r;
1278 
1279 	if (func == CENTAUR_CPUID_SIGNATURE &&
1280 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1281 		return 0;
1282 
1283 	r = do_cpuid_func(array, func, type);
1284 	if (r)
1285 		return r;
1286 
1287 	limit = array->entries[array->nent - 1].eax;
1288 	for (func = func + 1; func <= limit; ++func) {
1289 		r = do_cpuid_func(array, func, type);
1290 		if (r)
1291 			break;
1292 	}
1293 
1294 	return r;
1295 }
1296 
1297 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1298 				 __u32 num_entries, unsigned int ioctl_type)
1299 {
1300 	int i;
1301 	__u32 pad[3];
1302 
1303 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1304 		return false;
1305 
1306 	/*
1307 	 * We want to make sure that ->padding is being passed clean from
1308 	 * userspace in case we want to use it for something in the future.
1309 	 *
1310 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1311 	 * have to give ourselves satisfied only with the emulated side. /me
1312 	 * sheds a tear.
1313 	 */
1314 	for (i = 0; i < num_entries; i++) {
1315 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1316 			return true;
1317 
1318 		if (pad[0] || pad[1] || pad[2])
1319 			return true;
1320 	}
1321 	return false;
1322 }
1323 
1324 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1325 			    struct kvm_cpuid_entry2 __user *entries,
1326 			    unsigned int type)
1327 {
1328 	static const u32 funcs[] = {
1329 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1330 	};
1331 
1332 	struct kvm_cpuid_array array = {
1333 		.nent = 0,
1334 	};
1335 	int r, i;
1336 
1337 	if (cpuid->nent < 1)
1338 		return -E2BIG;
1339 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1340 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1341 
1342 	if (sanity_check_entries(entries, cpuid->nent, type))
1343 		return -EINVAL;
1344 
1345 	array.entries = kvcalloc(cpuid->nent, sizeof(struct kvm_cpuid_entry2), GFP_KERNEL);
1346 	if (!array.entries)
1347 		return -ENOMEM;
1348 
1349 	array.maxnent = cpuid->nent;
1350 
1351 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1352 		r = get_cpuid_func(&array, funcs[i], type);
1353 		if (r)
1354 			goto out_free;
1355 	}
1356 	cpuid->nent = array.nent;
1357 
1358 	if (copy_to_user(entries, array.entries,
1359 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1360 		r = -EFAULT;
1361 
1362 out_free:
1363 	kvfree(array.entries);
1364 	return r;
1365 }
1366 
1367 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
1368 						    u32 function, u32 index)
1369 {
1370 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1371 				 function, index);
1372 }
1373 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry_index);
1374 
1375 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1376 					      u32 function)
1377 {
1378 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1379 				 function, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
1380 }
1381 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1382 
1383 /*
1384  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1385  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1386  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1387  * range.  Centaur/VIA follows Intel semantics.
1388  *
1389  * A leaf is considered out-of-range if its function is higher than the maximum
1390  * supported leaf of its associated class or if its associated class does not
1391  * exist.
1392  *
1393  * There are three primary classes to be considered, with their respective
1394  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1395  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1396  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1397  *
1398  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1399  *  - Hypervisor: 0x40000000 - 0x4fffffff
1400  *  - Extended:   0x80000000 - 0xbfffffff
1401  *  - Centaur:    0xc0000000 - 0xcfffffff
1402  *
1403  * The Hypervisor class is further subdivided into sub-classes that each act as
1404  * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1405  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1406  * CPUID sub-classes are:
1407  *
1408  *  - HyperV:     0x40000000 - 0x400000ff
1409  *  - KVM:        0x40000100 - 0x400001ff
1410  */
1411 static struct kvm_cpuid_entry2 *
1412 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1413 {
1414 	struct kvm_cpuid_entry2 *basic, *class;
1415 	u32 function = *fn_ptr;
1416 
1417 	basic = kvm_find_cpuid_entry(vcpu, 0);
1418 	if (!basic)
1419 		return NULL;
1420 
1421 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1422 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1423 		return NULL;
1424 
1425 	if (function >= 0x40000000 && function <= 0x4fffffff)
1426 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1427 	else if (function >= 0xc0000000)
1428 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1429 	else
1430 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1431 
1432 	if (class && function <= class->eax)
1433 		return NULL;
1434 
1435 	/*
1436 	 * Leaf specific adjustments are also applied when redirecting to the
1437 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1438 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1439 	 * needs to be pulled from CPUID.0xb.1.
1440 	 */
1441 	*fn_ptr = basic->eax;
1442 
1443 	/*
1444 	 * The class does not exist or the requested function is out of range;
1445 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1446 	 * the original requested leaf is observed!
1447 	 */
1448 	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1449 }
1450 
1451 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1452 	       u32 *ecx, u32 *edx, bool exact_only)
1453 {
1454 	u32 orig_function = *eax, function = *eax, index = *ecx;
1455 	struct kvm_cpuid_entry2 *entry;
1456 	bool exact, used_max_basic = false;
1457 
1458 	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1459 	exact = !!entry;
1460 
1461 	if (!entry && !exact_only) {
1462 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1463 		used_max_basic = !!entry;
1464 	}
1465 
1466 	if (entry) {
1467 		*eax = entry->eax;
1468 		*ebx = entry->ebx;
1469 		*ecx = entry->ecx;
1470 		*edx = entry->edx;
1471 		if (function == 7 && index == 0) {
1472 			u64 data;
1473 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1474 			    (data & TSX_CTRL_CPUID_CLEAR))
1475 				*ebx &= ~(F(RTM) | F(HLE));
1476 		} else if (function == 0x80000007) {
1477 			if (kvm_hv_invtsc_suppressed(vcpu))
1478 				*edx &= ~SF(CONSTANT_TSC);
1479 		}
1480 	} else {
1481 		*eax = *ebx = *ecx = *edx = 0;
1482 		/*
1483 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1484 		 * and EDX is always the x2APIC ID, even for undefined
1485 		 * subleaves. Index 1 will exist iff the leaf is
1486 		 * implemented, so we pass through CL iff leaf 1
1487 		 * exists. EDX can be copied from any existing index.
1488 		 */
1489 		if (function == 0xb || function == 0x1f) {
1490 			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
1491 			if (entry) {
1492 				*ecx = index & 0xff;
1493 				*edx = entry->edx;
1494 			}
1495 		}
1496 	}
1497 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1498 			used_max_basic);
1499 	return exact;
1500 }
1501 EXPORT_SYMBOL_GPL(kvm_cpuid);
1502 
1503 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1504 {
1505 	u32 eax, ebx, ecx, edx;
1506 
1507 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1508 		return 1;
1509 
1510 	eax = kvm_rax_read(vcpu);
1511 	ecx = kvm_rcx_read(vcpu);
1512 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1513 	kvm_rax_write(vcpu, eax);
1514 	kvm_rbx_write(vcpu, ebx);
1515 	kvm_rcx_write(vcpu, ecx);
1516 	kvm_rdx_write(vcpu, edx);
1517 	return kvm_skip_emulated_instruction(vcpu);
1518 }
1519 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1520