xref: /linux/arch/x86/kvm/cpuid.c (revision 3349e275067f94ffb4141989aed9cbae7409429b)
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/lockdep.h"
15 #include <linux/export.h>
16 #include <linux/vmalloc.h>
17 #include <linux/uaccess.h>
18 #include <linux/sched/stat.h>
19 
20 #include <asm/processor.h>
21 #include <asm/user.h>
22 #include <asm/fpu/xstate.h>
23 #include <asm/sgx.h>
24 #include <asm/cpuid/api.h>
25 #include "cpuid.h"
26 #include "lapic.h"
27 #include "mmu.h"
28 #include "trace.h"
29 #include "pmu.h"
30 #include "xen.h"
31 
32 /*
33  * Unlike "struct cpuinfo_x86.x86_capability", kvm_cpu_caps doesn't need to be
34  * aligned to sizeof(unsigned long) because it's not accessed via bitops.
35  */
36 u32 kvm_cpu_caps[NR_KVM_CPU_CAPS] __read_mostly;
37 EXPORT_SYMBOL_GPL(kvm_cpu_caps);
38 
39 struct cpuid_xstate_sizes {
40 	u32 eax;
41 	u32 ebx;
42 	u32 ecx;
43 };
44 
45 static struct cpuid_xstate_sizes xstate_sizes[XFEATURE_MAX] __ro_after_init;
46 
kvm_init_xstate_sizes(void)47 void __init kvm_init_xstate_sizes(void)
48 {
49 	u32 ign;
50 	int i;
51 
52 	for (i = XFEATURE_YMM; i < ARRAY_SIZE(xstate_sizes); i++) {
53 		struct cpuid_xstate_sizes *xs = &xstate_sizes[i];
54 
55 		cpuid_count(0xD, i, &xs->eax, &xs->ebx, &xs->ecx, &ign);
56 	}
57 }
58 
xstate_required_size(u64 xstate_bv,bool compacted)59 u32 xstate_required_size(u64 xstate_bv, bool compacted)
60 {
61 	u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
62 	int i;
63 
64 	xstate_bv &= XFEATURE_MASK_EXTEND;
65 	for (i = XFEATURE_YMM; i < ARRAY_SIZE(xstate_sizes) && xstate_bv; i++) {
66 		struct cpuid_xstate_sizes *xs = &xstate_sizes[i];
67 		u32 offset;
68 
69 		if (!(xstate_bv & BIT_ULL(i)))
70 			continue;
71 
72 		/* ECX[1]: 64B alignment in compacted form */
73 		if (compacted)
74 			offset = (xs->ecx & 0x2) ? ALIGN(ret, 64) : ret;
75 		else
76 			offset = xs->ebx;
77 		ret = max(ret, offset + xs->eax);
78 		xstate_bv &= ~BIT_ULL(i);
79 	}
80 
81 	return ret;
82 }
83 
kvm_find_cpuid_entry2(struct kvm_cpuid_entry2 * entries,int nent,u32 function,u64 index)84 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry2(
85 	struct kvm_cpuid_entry2 *entries, int nent, u32 function, u64 index)
86 {
87 	struct kvm_cpuid_entry2 *e;
88 	int i;
89 
90 	/*
91 	 * KVM has a semi-arbitrary rule that querying the guest's CPUID model
92 	 * with IRQs disabled is disallowed.  The CPUID model can legitimately
93 	 * have over one hundred entries, i.e. the lookup is slow, and IRQs are
94 	 * typically disabled in KVM only when KVM is in a performance critical
95 	 * path, e.g. the core VM-Enter/VM-Exit run loop.  Nothing will break
96 	 * if this rule is violated, this assertion is purely to flag potential
97 	 * performance issues.  If this fires, consider moving the lookup out
98 	 * of the hotpath, e.g. by caching information during CPUID updates.
99 	 */
100 	lockdep_assert_irqs_enabled();
101 
102 	for (i = 0; i < nent; i++) {
103 		e = &entries[i];
104 
105 		if (e->function != function)
106 			continue;
107 
108 		/*
109 		 * If the index isn't significant, use the first entry with a
110 		 * matching function.  It's userspace's responsibility to not
111 		 * provide "duplicate" entries in all cases.
112 		 */
113 		if (!(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) || e->index == index)
114 			return e;
115 
116 
117 		/*
118 		 * Similarly, use the first matching entry if KVM is doing a
119 		 * lookup (as opposed to emulating CPUID) for a function that's
120 		 * architecturally defined as not having a significant index.
121 		 */
122 		if (index == KVM_CPUID_INDEX_NOT_SIGNIFICANT) {
123 			/*
124 			 * Direct lookups from KVM should not diverge from what
125 			 * KVM defines internally (the architectural behavior).
126 			 */
127 			WARN_ON_ONCE(cpuid_function_is_indexed(function));
128 			return e;
129 		}
130 	}
131 
132 	return NULL;
133 }
134 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry2);
135 
kvm_check_cpuid(struct kvm_vcpu * vcpu)136 static int kvm_check_cpuid(struct kvm_vcpu *vcpu)
137 {
138 	struct kvm_cpuid_entry2 *best;
139 	u64 xfeatures;
140 
141 	/*
142 	 * The existing code assumes virtual address is 48-bit or 57-bit in the
143 	 * canonical address checks; exit if it is ever changed.
144 	 */
145 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
146 	if (best) {
147 		int vaddr_bits = (best->eax & 0xff00) >> 8;
148 
149 		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
150 			return -EINVAL;
151 	}
152 
153 	/*
154 	 * Exposing dynamic xfeatures to the guest requires additional
155 	 * enabling in the FPU, e.g. to expand the guest XSAVE state size.
156 	 */
157 	best = kvm_find_cpuid_entry_index(vcpu, 0xd, 0);
158 	if (!best)
159 		return 0;
160 
161 	xfeatures = best->eax | ((u64)best->edx << 32);
162 	xfeatures &= XFEATURE_MASK_USER_DYNAMIC;
163 	if (!xfeatures)
164 		return 0;
165 
166 	return fpu_enable_guest_xfd_features(&vcpu->arch.guest_fpu, xfeatures);
167 }
168 
169 static u32 kvm_apply_cpuid_pv_features_quirk(struct kvm_vcpu *vcpu);
170 static void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu);
171 
172 /* Check whether the supplied CPUID data is equal to what is already set for the vCPU. */
kvm_cpuid_check_equal(struct kvm_vcpu * vcpu,struct kvm_cpuid_entry2 * e2,int nent)173 static int kvm_cpuid_check_equal(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
174 				 int nent)
175 {
176 	struct kvm_cpuid_entry2 *orig;
177 	int i;
178 
179 	/*
180 	 * Apply runtime CPUID updates to the incoming CPUID entries to avoid
181 	 * false positives due mismatches on KVM-owned feature flags.
182 	 *
183 	 * Note!  @e2 and @nent track the _old_ CPUID entries!
184 	 */
185 	kvm_update_cpuid_runtime(vcpu);
186 	kvm_apply_cpuid_pv_features_quirk(vcpu);
187 
188 	if (nent != vcpu->arch.cpuid_nent)
189 		return -EINVAL;
190 
191 	for (i = 0; i < nent; i++) {
192 		orig = &vcpu->arch.cpuid_entries[i];
193 		if (e2[i].function != orig->function ||
194 		    e2[i].index != orig->index ||
195 		    e2[i].flags != orig->flags ||
196 		    e2[i].eax != orig->eax || e2[i].ebx != orig->ebx ||
197 		    e2[i].ecx != orig->ecx || e2[i].edx != orig->edx)
198 			return -EINVAL;
199 	}
200 
201 	return 0;
202 }
203 
kvm_get_hypervisor_cpuid(struct kvm_vcpu * vcpu,const char * sig)204 static struct kvm_hypervisor_cpuid kvm_get_hypervisor_cpuid(struct kvm_vcpu *vcpu,
205 							    const char *sig)
206 {
207 	struct kvm_hypervisor_cpuid cpuid = {};
208 	struct kvm_cpuid_entry2 *entry;
209 	u32 base;
210 
211 	for_each_possible_cpuid_base_hypervisor(base) {
212 		entry = kvm_find_cpuid_entry(vcpu, base);
213 
214 		if (entry) {
215 			u32 signature[3];
216 
217 			signature[0] = entry->ebx;
218 			signature[1] = entry->ecx;
219 			signature[2] = entry->edx;
220 
221 			if (!memcmp(signature, sig, sizeof(signature))) {
222 				cpuid.base = base;
223 				cpuid.limit = entry->eax;
224 				break;
225 			}
226 		}
227 	}
228 
229 	return cpuid;
230 }
231 
kvm_apply_cpuid_pv_features_quirk(struct kvm_vcpu * vcpu)232 static u32 kvm_apply_cpuid_pv_features_quirk(struct kvm_vcpu *vcpu)
233 {
234 	struct kvm_hypervisor_cpuid kvm_cpuid;
235 	struct kvm_cpuid_entry2 *best;
236 
237 	kvm_cpuid = kvm_get_hypervisor_cpuid(vcpu, KVM_SIGNATURE);
238 	if (!kvm_cpuid.base)
239 		return 0;
240 
241 	best = kvm_find_cpuid_entry(vcpu, kvm_cpuid.base | KVM_CPUID_FEATURES);
242 	if (!best)
243 		return 0;
244 
245 	if (kvm_hlt_in_guest(vcpu->kvm))
246 		best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
247 
248 	return best->eax;
249 }
250 
251 /*
252  * Calculate guest's supported XCR0 taking into account guest CPUID data and
253  * KVM's supported XCR0 (comprised of host's XCR0 and KVM_SUPPORTED_XCR0).
254  */
cpuid_get_supported_xcr0(struct kvm_vcpu * vcpu)255 static u64 cpuid_get_supported_xcr0(struct kvm_vcpu *vcpu)
256 {
257 	struct kvm_cpuid_entry2 *best;
258 
259 	best = kvm_find_cpuid_entry_index(vcpu, 0xd, 0);
260 	if (!best)
261 		return 0;
262 
263 	return (best->eax | ((u64)best->edx << 32)) & kvm_caps.supported_xcr0;
264 }
265 
kvm_update_feature_runtime(struct kvm_vcpu * vcpu,struct kvm_cpuid_entry2 * entry,unsigned int x86_feature,bool has_feature)266 static __always_inline void kvm_update_feature_runtime(struct kvm_vcpu *vcpu,
267 						       struct kvm_cpuid_entry2 *entry,
268 						       unsigned int x86_feature,
269 						       bool has_feature)
270 {
271 	cpuid_entry_change(entry, x86_feature, has_feature);
272 	guest_cpu_cap_change(vcpu, x86_feature, has_feature);
273 }
274 
kvm_update_cpuid_runtime(struct kvm_vcpu * vcpu)275 static void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu)
276 {
277 	struct kvm_cpuid_entry2 *best;
278 
279 	vcpu->arch.cpuid_dynamic_bits_dirty = false;
280 
281 	best = kvm_find_cpuid_entry(vcpu, 1);
282 	if (best) {
283 		kvm_update_feature_runtime(vcpu, best, X86_FEATURE_OSXSAVE,
284 					   kvm_is_cr4_bit_set(vcpu, X86_CR4_OSXSAVE));
285 
286 		kvm_update_feature_runtime(vcpu, best, X86_FEATURE_APIC,
287 					   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
288 
289 		if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT))
290 			kvm_update_feature_runtime(vcpu, best, X86_FEATURE_MWAIT,
291 						   vcpu->arch.ia32_misc_enable_msr &
292 						   MSR_IA32_MISC_ENABLE_MWAIT);
293 	}
294 
295 	best = kvm_find_cpuid_entry_index(vcpu, 7, 0);
296 	if (best)
297 		kvm_update_feature_runtime(vcpu, best, X86_FEATURE_OSPKE,
298 					   kvm_is_cr4_bit_set(vcpu, X86_CR4_PKE));
299 
300 
301 	best = kvm_find_cpuid_entry_index(vcpu, 0xD, 0);
302 	if (best)
303 		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
304 
305 	best = kvm_find_cpuid_entry_index(vcpu, 0xD, 1);
306 	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
307 		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
308 		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
309 }
310 
kvm_cpuid_has_hyperv(struct kvm_vcpu * vcpu)311 static bool kvm_cpuid_has_hyperv(struct kvm_vcpu *vcpu)
312 {
313 #ifdef CONFIG_KVM_HYPERV
314 	struct kvm_cpuid_entry2 *entry;
315 
316 	entry = kvm_find_cpuid_entry(vcpu, HYPERV_CPUID_INTERFACE);
317 	return entry && entry->eax == HYPERV_CPUID_SIGNATURE_EAX;
318 #else
319 	return false;
320 #endif
321 }
322 
guest_cpuid_is_amd_or_hygon(struct kvm_vcpu * vcpu)323 static bool guest_cpuid_is_amd_or_hygon(struct kvm_vcpu *vcpu)
324 {
325 	struct kvm_cpuid_entry2 *entry;
326 
327 	entry = kvm_find_cpuid_entry(vcpu, 0);
328 	if (!entry)
329 		return false;
330 
331 	return is_guest_vendor_amd(entry->ebx, entry->ecx, entry->edx) ||
332 	       is_guest_vendor_hygon(entry->ebx, entry->ecx, entry->edx);
333 }
334 
335 /*
336  * This isn't truly "unsafe", but except for the cpu_caps initialization code,
337  * all register lookups should use __cpuid_entry_get_reg(), which provides
338  * compile-time validation of the input.
339  */
cpuid_get_reg_unsafe(struct kvm_cpuid_entry2 * entry,u32 reg)340 static u32 cpuid_get_reg_unsafe(struct kvm_cpuid_entry2 *entry, u32 reg)
341 {
342 	switch (reg) {
343 	case CPUID_EAX:
344 		return entry->eax;
345 	case CPUID_EBX:
346 		return entry->ebx;
347 	case CPUID_ECX:
348 		return entry->ecx;
349 	case CPUID_EDX:
350 		return entry->edx;
351 	default:
352 		WARN_ON_ONCE(1);
353 		return 0;
354 	}
355 }
356 
357 static int cpuid_func_emulated(struct kvm_cpuid_entry2 *entry, u32 func,
358 			       bool include_partially_emulated);
359 
kvm_vcpu_after_set_cpuid(struct kvm_vcpu * vcpu)360 void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
361 {
362 	struct kvm_lapic *apic = vcpu->arch.apic;
363 	struct kvm_cpuid_entry2 *best;
364 	struct kvm_cpuid_entry2 *entry;
365 	bool allow_gbpages;
366 	int i;
367 
368 	memset(vcpu->arch.cpu_caps, 0, sizeof(vcpu->arch.cpu_caps));
369 	BUILD_BUG_ON(ARRAY_SIZE(reverse_cpuid) != NR_KVM_CPU_CAPS);
370 
371 	/*
372 	 * Reset guest capabilities to userspace's guest CPUID definition, i.e.
373 	 * honor userspace's definition for features that don't require KVM or
374 	 * hardware management/support (or that KVM simply doesn't care about).
375 	 */
376 	for (i = 0; i < NR_KVM_CPU_CAPS; i++) {
377 		const struct cpuid_reg cpuid = reverse_cpuid[i];
378 		struct kvm_cpuid_entry2 emulated;
379 
380 		if (!cpuid.function)
381 			continue;
382 
383 		entry = kvm_find_cpuid_entry_index(vcpu, cpuid.function, cpuid.index);
384 		if (!entry)
385 			continue;
386 
387 		cpuid_func_emulated(&emulated, cpuid.function, true);
388 
389 		/*
390 		 * A vCPU has a feature if it's supported by KVM and is enabled
391 		 * in guest CPUID.  Note, this includes features that are
392 		 * supported by KVM but aren't advertised to userspace!
393 		 */
394 		vcpu->arch.cpu_caps[i] = kvm_cpu_caps[i] |
395 					 cpuid_get_reg_unsafe(&emulated, cpuid.reg);
396 		vcpu->arch.cpu_caps[i] &= cpuid_get_reg_unsafe(entry, cpuid.reg);
397 	}
398 
399 	kvm_update_cpuid_runtime(vcpu);
400 
401 	/*
402 	 * If TDP is enabled, let the guest use GBPAGES if they're supported in
403 	 * hardware.  The hardware page walker doesn't let KVM disable GBPAGES,
404 	 * i.e. won't treat them as reserved, and KVM doesn't redo the GVA->GPA
405 	 * walk for performance and complexity reasons.  Not to mention KVM
406 	 * _can't_ solve the problem because GVA->GPA walks aren't visible to
407 	 * KVM once a TDP translation is installed.  Mimic hardware behavior so
408 	 * that KVM's is at least consistent, i.e. doesn't randomly inject #PF.
409 	 * If TDP is disabled, honor *only* guest CPUID as KVM has full control
410 	 * and can install smaller shadow pages if the host lacks 1GiB support.
411 	 */
412 	allow_gbpages = tdp_enabled ? boot_cpu_has(X86_FEATURE_GBPAGES) :
413 				      guest_cpu_cap_has(vcpu, X86_FEATURE_GBPAGES);
414 	guest_cpu_cap_change(vcpu, X86_FEATURE_GBPAGES, allow_gbpages);
415 
416 	best = kvm_find_cpuid_entry(vcpu, 1);
417 	if (best && apic) {
418 		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
419 			apic->lapic_timer.timer_mode_mask = 3 << 17;
420 		else
421 			apic->lapic_timer.timer_mode_mask = 1 << 17;
422 
423 		kvm_apic_set_version(vcpu);
424 	}
425 
426 	vcpu->arch.guest_supported_xcr0 = cpuid_get_supported_xcr0(vcpu);
427 
428 	vcpu->arch.pv_cpuid.features = kvm_apply_cpuid_pv_features_quirk(vcpu);
429 
430 	vcpu->arch.is_amd_compatible = guest_cpuid_is_amd_or_hygon(vcpu);
431 	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
432 	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
433 
434 	kvm_pmu_refresh(vcpu);
435 
436 #define __kvm_cpu_cap_has(UNUSED_, f) kvm_cpu_cap_has(f)
437 	vcpu->arch.cr4_guest_rsvd_bits = __cr4_reserved_bits(__kvm_cpu_cap_has, UNUSED_) |
438 					 __cr4_reserved_bits(guest_cpu_cap_has, vcpu);
439 #undef __kvm_cpu_cap_has
440 
441 	kvm_hv_set_cpuid(vcpu, kvm_cpuid_has_hyperv(vcpu));
442 
443 	/* Invoke the vendor callback only after the above state is updated. */
444 	kvm_x86_call(vcpu_after_set_cpuid)(vcpu);
445 
446 	/*
447 	 * Except for the MMU, which needs to do its thing any vendor specific
448 	 * adjustments to the reserved GPA bits.
449 	 */
450 	kvm_mmu_after_set_cpuid(vcpu);
451 }
452 
cpuid_query_maxphyaddr(struct kvm_vcpu * vcpu)453 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
454 {
455 	struct kvm_cpuid_entry2 *best;
456 
457 	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
458 	if (!best || best->eax < 0x80000008)
459 		goto not_found;
460 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
461 	if (best)
462 		return best->eax & 0xff;
463 not_found:
464 	return 36;
465 }
466 
cpuid_query_maxguestphyaddr(struct kvm_vcpu * vcpu)467 int cpuid_query_maxguestphyaddr(struct kvm_vcpu *vcpu)
468 {
469 	struct kvm_cpuid_entry2 *best;
470 
471 	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
472 	if (!best || best->eax < 0x80000008)
473 		goto not_found;
474 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
475 	if (best)
476 		return (best->eax >> 16) & 0xff;
477 not_found:
478 	return 0;
479 }
480 
481 /*
482  * This "raw" version returns the reserved GPA bits without any adjustments for
483  * encryption technologies that usurp bits.  The raw mask should be used if and
484  * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
485  */
kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu * vcpu)486 u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
487 {
488 	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
489 }
490 
kvm_set_cpuid(struct kvm_vcpu * vcpu,struct kvm_cpuid_entry2 * e2,int nent)491 static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
492                         int nent)
493 {
494 	u32 vcpu_caps[NR_KVM_CPU_CAPS];
495 	int r;
496 
497 	/*
498 	 * Swap the existing (old) entries with the incoming (new) entries in
499 	 * order to massage the new entries, e.g. to account for dynamic bits
500 	 * that KVM controls, without clobbering the current guest CPUID, which
501 	 * KVM needs to preserve in order to unwind on failure.
502 	 *
503 	 * Similarly, save the vCPU's current cpu_caps so that the capabilities
504 	 * can be updated alongside the CPUID entries when performing runtime
505 	 * updates.  Full initialization is done if and only if the vCPU hasn't
506 	 * run, i.e. only if userspace is potentially changing CPUID features.
507 	 */
508 	swap(vcpu->arch.cpuid_entries, e2);
509 	swap(vcpu->arch.cpuid_nent, nent);
510 
511 	memcpy(vcpu_caps, vcpu->arch.cpu_caps, sizeof(vcpu_caps));
512 	BUILD_BUG_ON(sizeof(vcpu_caps) != sizeof(vcpu->arch.cpu_caps));
513 
514 	/*
515 	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
516 	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
517 	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
518 	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
519 	 * the core vCPU model on the fly. It would've been better to forbid any
520 	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
521 	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
522 	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
523 	 * whether the supplied CPUID data is equal to what's already set.
524 	 */
525 	if (kvm_vcpu_has_run(vcpu)) {
526 		r = kvm_cpuid_check_equal(vcpu, e2, nent);
527 		if (r)
528 			goto err;
529 		goto success;
530 	}
531 
532 #ifdef CONFIG_KVM_HYPERV
533 	if (kvm_cpuid_has_hyperv(vcpu)) {
534 		r = kvm_hv_vcpu_init(vcpu);
535 		if (r)
536 			goto err;
537 	}
538 #endif
539 
540 	r = kvm_check_cpuid(vcpu);
541 	if (r)
542 		goto err;
543 
544 #ifdef CONFIG_KVM_XEN
545 	vcpu->arch.xen.cpuid = kvm_get_hypervisor_cpuid(vcpu, XEN_SIGNATURE);
546 #endif
547 	kvm_vcpu_after_set_cpuid(vcpu);
548 
549 success:
550 	kvfree(e2);
551 	return 0;
552 
553 err:
554 	memcpy(vcpu->arch.cpu_caps, vcpu_caps, sizeof(vcpu_caps));
555 	swap(vcpu->arch.cpuid_entries, e2);
556 	swap(vcpu->arch.cpuid_nent, nent);
557 	return r;
558 }
559 
560 /* when an old userspace process fills a new kernel module */
kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu * vcpu,struct kvm_cpuid * cpuid,struct kvm_cpuid_entry __user * entries)561 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
562 			     struct kvm_cpuid *cpuid,
563 			     struct kvm_cpuid_entry __user *entries)
564 {
565 	int r, i;
566 	struct kvm_cpuid_entry *e = NULL;
567 	struct kvm_cpuid_entry2 *e2 = NULL;
568 
569 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
570 		return -E2BIG;
571 
572 	if (cpuid->nent) {
573 		e = vmemdup_array_user(entries, cpuid->nent, sizeof(*e));
574 		if (IS_ERR(e))
575 			return PTR_ERR(e);
576 
577 		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
578 		if (!e2) {
579 			r = -ENOMEM;
580 			goto out_free_cpuid;
581 		}
582 	}
583 	for (i = 0; i < cpuid->nent; i++) {
584 		e2[i].function = e[i].function;
585 		e2[i].eax = e[i].eax;
586 		e2[i].ebx = e[i].ebx;
587 		e2[i].ecx = e[i].ecx;
588 		e2[i].edx = e[i].edx;
589 		e2[i].index = 0;
590 		e2[i].flags = 0;
591 		e2[i].padding[0] = 0;
592 		e2[i].padding[1] = 0;
593 		e2[i].padding[2] = 0;
594 	}
595 
596 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
597 	if (r)
598 		kvfree(e2);
599 
600 out_free_cpuid:
601 	kvfree(e);
602 
603 	return r;
604 }
605 
kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu * vcpu,struct kvm_cpuid2 * cpuid,struct kvm_cpuid_entry2 __user * entries)606 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
607 			      struct kvm_cpuid2 *cpuid,
608 			      struct kvm_cpuid_entry2 __user *entries)
609 {
610 	struct kvm_cpuid_entry2 *e2 = NULL;
611 	int r;
612 
613 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
614 		return -E2BIG;
615 
616 	if (cpuid->nent) {
617 		e2 = vmemdup_array_user(entries, cpuid->nent, sizeof(*e2));
618 		if (IS_ERR(e2))
619 			return PTR_ERR(e2);
620 	}
621 
622 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
623 	if (r)
624 		kvfree(e2);
625 
626 	return r;
627 }
628 
kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu * vcpu,struct kvm_cpuid2 * cpuid,struct kvm_cpuid_entry2 __user * entries)629 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
630 			      struct kvm_cpuid2 *cpuid,
631 			      struct kvm_cpuid_entry2 __user *entries)
632 {
633 	if (cpuid->nent < vcpu->arch.cpuid_nent)
634 		return -E2BIG;
635 
636 	if (vcpu->arch.cpuid_dynamic_bits_dirty)
637 		kvm_update_cpuid_runtime(vcpu);
638 
639 	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
640 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
641 		return -EFAULT;
642 
643 	cpuid->nent = vcpu->arch.cpuid_nent;
644 	return 0;
645 }
646 
raw_cpuid_get(struct cpuid_reg cpuid)647 static __always_inline u32 raw_cpuid_get(struct cpuid_reg cpuid)
648 {
649 	struct kvm_cpuid_entry2 entry;
650 	u32 base;
651 
652 	/*
653 	 * KVM only supports features defined by Intel (0x0), AMD (0x80000000),
654 	 * and Centaur (0xc0000000).  WARN if a feature for new vendor base is
655 	 * defined, as this and other code would need to be updated.
656 	 */
657 	base = cpuid.function & 0xffff0000;
658 	if (WARN_ON_ONCE(base && base != 0x80000000 && base != 0xc0000000))
659 		return 0;
660 
661 	if (cpuid_eax(base) < cpuid.function)
662 		return 0;
663 
664 	cpuid_count(cpuid.function, cpuid.index,
665 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
666 
667 	return *__cpuid_entry_get_reg(&entry, cpuid.reg);
668 }
669 
670 /*
671  * For kernel-defined leafs, mask KVM's supported feature set with the kernel's
672  * capabilities as well as raw CPUID.  For KVM-defined leafs, consult only raw
673  * CPUID, as KVM is the one and only authority (in the kernel).
674  */
675 #define kvm_cpu_cap_init(leaf, feature_initializers...)			\
676 do {									\
677 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);	\
678 	const u32 __maybe_unused kvm_cpu_cap_init_in_progress = leaf;	\
679 	const u32 *kernel_cpu_caps = boot_cpu_data.x86_capability;	\
680 	u32 kvm_cpu_cap_passthrough = 0;				\
681 	u32 kvm_cpu_cap_synthesized = 0;				\
682 	u32 kvm_cpu_cap_emulated = 0;					\
683 	u32 kvm_cpu_cap_features = 0;					\
684 									\
685 	feature_initializers						\
686 									\
687 	kvm_cpu_caps[leaf] = kvm_cpu_cap_features;			\
688 									\
689 	if (leaf < NCAPINTS)						\
690 		kvm_cpu_caps[leaf] &= kernel_cpu_caps[leaf];		\
691 									\
692 	kvm_cpu_caps[leaf] |= kvm_cpu_cap_passthrough;			\
693 	kvm_cpu_caps[leaf] &= (raw_cpuid_get(cpuid) |			\
694 			       kvm_cpu_cap_synthesized);		\
695 	kvm_cpu_caps[leaf] |= kvm_cpu_cap_emulated;			\
696 } while (0)
697 
698 /*
699  * Assert that the feature bit being declared, e.g. via F(), is in the CPUID
700  * word that's being initialized.  Exempt 0x8000_0001.EDX usage of 0x1.EDX
701  * features, as AMD duplicated many 0x1.EDX features into 0x8000_0001.EDX.
702  */
703 #define KVM_VALIDATE_CPU_CAP_USAGE(name)				\
704 do {									\
705 	u32 __leaf = __feature_leaf(X86_FEATURE_##name);		\
706 									\
707 	BUILD_BUG_ON(__leaf != kvm_cpu_cap_init_in_progress);		\
708 } while (0)
709 
710 #define F(name)							\
711 ({								\
712 	KVM_VALIDATE_CPU_CAP_USAGE(name);			\
713 	kvm_cpu_cap_features |= feature_bit(name);		\
714 })
715 
716 /* Scattered Flag - For features that are scattered by cpufeatures.h. */
717 #define SCATTERED_F(name)					\
718 ({								\
719 	BUILD_BUG_ON(X86_FEATURE_##name >= MAX_CPU_FEATURES);	\
720 	KVM_VALIDATE_CPU_CAP_USAGE(name);			\
721 	if (boot_cpu_has(X86_FEATURE_##name))			\
722 		F(name);					\
723 })
724 
725 /* Features that KVM supports only on 64-bit kernels. */
726 #define X86_64_F(name)						\
727 ({								\
728 	KVM_VALIDATE_CPU_CAP_USAGE(name);			\
729 	if (IS_ENABLED(CONFIG_X86_64))				\
730 		F(name);					\
731 })
732 
733 /*
734  * Emulated Feature - For features that KVM emulates in software irrespective
735  * of host CPU/kernel support.
736  */
737 #define EMULATED_F(name)					\
738 ({								\
739 	kvm_cpu_cap_emulated |= feature_bit(name);		\
740 	F(name);						\
741 })
742 
743 /*
744  * Synthesized Feature - For features that are synthesized into boot_cpu_data,
745  * i.e. may not be present in the raw CPUID, but can still be advertised to
746  * userspace.  Primarily used for mitigation related feature flags.
747  */
748 #define SYNTHESIZED_F(name)					\
749 ({								\
750 	kvm_cpu_cap_synthesized |= feature_bit(name);		\
751 	F(name);						\
752 })
753 
754 /*
755  * Passthrough Feature - For features that KVM supports based purely on raw
756  * hardware CPUID, i.e. that KVM virtualizes even if the host kernel doesn't
757  * use the feature.  Simply force set the feature in KVM's capabilities, raw
758  * CPUID support will be factored in by kvm_cpu_cap_mask().
759  */
760 #define PASSTHROUGH_F(name)					\
761 ({								\
762 	kvm_cpu_cap_passthrough |= feature_bit(name);		\
763 	F(name);						\
764 })
765 
766 /*
767  * Aliased Features - For features in 0x8000_0001.EDX that are duplicates of
768  * identical 0x1.EDX features, and thus are aliased from 0x1 to 0x8000_0001.
769  */
770 #define ALIASED_1_EDX_F(name)							\
771 ({										\
772 	BUILD_BUG_ON(__feature_leaf(X86_FEATURE_##name) != CPUID_1_EDX);	\
773 	BUILD_BUG_ON(kvm_cpu_cap_init_in_progress != CPUID_8000_0001_EDX);	\
774 	kvm_cpu_cap_features |= feature_bit(name);				\
775 })
776 
777 /*
778  * Vendor Features - For features that KVM supports, but are added in later
779  * because they require additional vendor enabling.
780  */
781 #define VENDOR_F(name)						\
782 ({								\
783 	KVM_VALIDATE_CPU_CAP_USAGE(name);			\
784 })
785 
786 /*
787  * Runtime Features - For features that KVM dynamically sets/clears at runtime,
788  * e.g. when CR4 changes, but which are never advertised to userspace.
789  */
790 #define RUNTIME_F(name)						\
791 ({								\
792 	KVM_VALIDATE_CPU_CAP_USAGE(name);			\
793 })
794 
795 /*
796  * Undefine the MSR bit macro to avoid token concatenation issues when
797  * processing X86_FEATURE_SPEC_CTRL_SSBD.
798  */
799 #undef SPEC_CTRL_SSBD
800 
801 /* DS is defined by ptrace-abi.h on 32-bit builds. */
802 #undef DS
803 
kvm_set_cpu_caps(void)804 void kvm_set_cpu_caps(void)
805 {
806 	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
807 
808 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
809 		     sizeof(boot_cpu_data.x86_capability));
810 
811 	kvm_cpu_cap_init(CPUID_1_ECX,
812 		F(XMM3),
813 		F(PCLMULQDQ),
814 		VENDOR_F(DTES64),
815 		/*
816 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
817 		 * advertised to guests via CPUID!  MWAIT is also technically a
818 		 * runtime flag thanks to IA32_MISC_ENABLES; mark it as such so
819 		 * that KVM is aware that it's a known, unadvertised flag.
820 		 */
821 		RUNTIME_F(MWAIT),
822 		/* DS-CPL */
823 		VENDOR_F(VMX),
824 		/* SMX, EST */
825 		/* TM2 */
826 		F(SSSE3),
827 		/* CNXT-ID */
828 		/* Reserved */
829 		F(FMA),
830 		F(CX16),
831 		/* xTPR Update */
832 		F(PDCM),
833 		F(PCID),
834 		/* Reserved, DCA */
835 		F(XMM4_1),
836 		F(XMM4_2),
837 		EMULATED_F(X2APIC),
838 		F(MOVBE),
839 		F(POPCNT),
840 		EMULATED_F(TSC_DEADLINE_TIMER),
841 		F(AES),
842 		F(XSAVE),
843 		RUNTIME_F(OSXSAVE),
844 		F(AVX),
845 		F(F16C),
846 		F(RDRAND),
847 		EMULATED_F(HYPERVISOR),
848 	);
849 
850 	kvm_cpu_cap_init(CPUID_1_EDX,
851 		F(FPU),
852 		F(VME),
853 		F(DE),
854 		F(PSE),
855 		F(TSC),
856 		F(MSR),
857 		F(PAE),
858 		F(MCE),
859 		F(CX8),
860 		F(APIC),
861 		/* Reserved */
862 		F(SEP),
863 		F(MTRR),
864 		F(PGE),
865 		F(MCA),
866 		F(CMOV),
867 		F(PAT),
868 		F(PSE36),
869 		/* PSN */
870 		F(CLFLUSH),
871 		/* Reserved */
872 		VENDOR_F(DS),
873 		/* ACPI */
874 		F(MMX),
875 		F(FXSR),
876 		F(XMM),
877 		F(XMM2),
878 		F(SELFSNOOP),
879 		/* HTT, TM, Reserved, PBE */
880 	);
881 
882 	kvm_cpu_cap_init(CPUID_7_0_EBX,
883 		F(FSGSBASE),
884 		EMULATED_F(TSC_ADJUST),
885 		F(SGX),
886 		F(BMI1),
887 		F(HLE),
888 		F(AVX2),
889 		F(FDP_EXCPTN_ONLY),
890 		F(SMEP),
891 		F(BMI2),
892 		F(ERMS),
893 		F(INVPCID),
894 		F(RTM),
895 		F(ZERO_FCS_FDS),
896 		VENDOR_F(MPX),
897 		F(AVX512F),
898 		F(AVX512DQ),
899 		F(RDSEED),
900 		F(ADX),
901 		F(SMAP),
902 		F(AVX512IFMA),
903 		F(CLFLUSHOPT),
904 		F(CLWB),
905 		VENDOR_F(INTEL_PT),
906 		F(AVX512PF),
907 		F(AVX512ER),
908 		F(AVX512CD),
909 		F(SHA_NI),
910 		F(AVX512BW),
911 		F(AVX512VL),
912 	);
913 
914 	kvm_cpu_cap_init(CPUID_7_ECX,
915 		F(AVX512VBMI),
916 		PASSTHROUGH_F(LA57),
917 		F(PKU),
918 		RUNTIME_F(OSPKE),
919 		F(RDPID),
920 		F(AVX512_VPOPCNTDQ),
921 		F(UMIP),
922 		F(AVX512_VBMI2),
923 		F(GFNI),
924 		F(VAES),
925 		F(VPCLMULQDQ),
926 		F(AVX512_VNNI),
927 		F(AVX512_BITALG),
928 		F(CLDEMOTE),
929 		F(MOVDIRI),
930 		F(MOVDIR64B),
931 		VENDOR_F(WAITPKG),
932 		F(SGX_LC),
933 		F(BUS_LOCK_DETECT),
934 	);
935 
936 	/*
937 	 * PKU not yet implemented for shadow paging and requires OSPKE
938 	 * to be set on the host. Clear it if that is not the case
939 	 */
940 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
941 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
942 
943 	kvm_cpu_cap_init(CPUID_7_EDX,
944 		F(AVX512_4VNNIW),
945 		F(AVX512_4FMAPS),
946 		F(SPEC_CTRL),
947 		F(SPEC_CTRL_SSBD),
948 		EMULATED_F(ARCH_CAPABILITIES),
949 		F(INTEL_STIBP),
950 		F(MD_CLEAR),
951 		F(AVX512_VP2INTERSECT),
952 		F(FSRM),
953 		F(SERIALIZE),
954 		F(TSXLDTRK),
955 		F(AVX512_FP16),
956 		F(AMX_TILE),
957 		F(AMX_INT8),
958 		F(AMX_BF16),
959 		F(FLUSH_L1D),
960 	);
961 
962 	if (boot_cpu_has(X86_FEATURE_AMD_IBPB_RET) &&
963 	    boot_cpu_has(X86_FEATURE_AMD_IBPB) &&
964 	    boot_cpu_has(X86_FEATURE_AMD_IBRS))
965 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
966 	if (boot_cpu_has(X86_FEATURE_STIBP))
967 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
968 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
969 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
970 
971 	kvm_cpu_cap_init(CPUID_7_1_EAX,
972 		F(SHA512),
973 		F(SM3),
974 		F(SM4),
975 		F(AVX_VNNI),
976 		F(AVX512_BF16),
977 		F(CMPCCXADD),
978 		F(FZRM),
979 		F(FSRS),
980 		F(FSRC),
981 		F(WRMSRNS),
982 		F(AMX_FP16),
983 		F(AVX_IFMA),
984 		F(LAM),
985 	);
986 
987 	kvm_cpu_cap_init(CPUID_7_1_EDX,
988 		F(AVX_VNNI_INT8),
989 		F(AVX_NE_CONVERT),
990 		F(AMX_COMPLEX),
991 		F(AVX_VNNI_INT16),
992 		F(PREFETCHITI),
993 		F(AVX10),
994 	);
995 
996 	kvm_cpu_cap_init(CPUID_7_2_EDX,
997 		F(INTEL_PSFD),
998 		F(IPRED_CTRL),
999 		F(RRSBA_CTRL),
1000 		F(DDPD_U),
1001 		F(BHI_CTRL),
1002 		F(MCDT_NO),
1003 	);
1004 
1005 	kvm_cpu_cap_init(CPUID_D_1_EAX,
1006 		F(XSAVEOPT),
1007 		F(XSAVEC),
1008 		F(XGETBV1),
1009 		F(XSAVES),
1010 		X86_64_F(XFD),
1011 	);
1012 
1013 	kvm_cpu_cap_init(CPUID_12_EAX,
1014 		SCATTERED_F(SGX1),
1015 		SCATTERED_F(SGX2),
1016 		SCATTERED_F(SGX_EDECCSSA),
1017 	);
1018 
1019 	kvm_cpu_cap_init(CPUID_24_0_EBX,
1020 		F(AVX10_128),
1021 		F(AVX10_256),
1022 		F(AVX10_512),
1023 	);
1024 
1025 	kvm_cpu_cap_init(CPUID_8000_0001_ECX,
1026 		F(LAHF_LM),
1027 		F(CMP_LEGACY),
1028 		VENDOR_F(SVM),
1029 		/* ExtApicSpace */
1030 		F(CR8_LEGACY),
1031 		F(ABM),
1032 		F(SSE4A),
1033 		F(MISALIGNSSE),
1034 		F(3DNOWPREFETCH),
1035 		F(OSVW),
1036 		/* IBS */
1037 		F(XOP),
1038 		/* SKINIT, WDT, LWP */
1039 		F(FMA4),
1040 		F(TBM),
1041 		F(TOPOEXT),
1042 		VENDOR_F(PERFCTR_CORE),
1043 	);
1044 
1045 	kvm_cpu_cap_init(CPUID_8000_0001_EDX,
1046 		ALIASED_1_EDX_F(FPU),
1047 		ALIASED_1_EDX_F(VME),
1048 		ALIASED_1_EDX_F(DE),
1049 		ALIASED_1_EDX_F(PSE),
1050 		ALIASED_1_EDX_F(TSC),
1051 		ALIASED_1_EDX_F(MSR),
1052 		ALIASED_1_EDX_F(PAE),
1053 		ALIASED_1_EDX_F(MCE),
1054 		ALIASED_1_EDX_F(CX8),
1055 		ALIASED_1_EDX_F(APIC),
1056 		/* Reserved */
1057 		F(SYSCALL),
1058 		ALIASED_1_EDX_F(MTRR),
1059 		ALIASED_1_EDX_F(PGE),
1060 		ALIASED_1_EDX_F(MCA),
1061 		ALIASED_1_EDX_F(CMOV),
1062 		ALIASED_1_EDX_F(PAT),
1063 		ALIASED_1_EDX_F(PSE36),
1064 		/* Reserved */
1065 		F(NX),
1066 		/* Reserved */
1067 		F(MMXEXT),
1068 		ALIASED_1_EDX_F(MMX),
1069 		ALIASED_1_EDX_F(FXSR),
1070 		F(FXSR_OPT),
1071 		X86_64_F(GBPAGES),
1072 		F(RDTSCP),
1073 		/* Reserved */
1074 		X86_64_F(LM),
1075 		F(3DNOWEXT),
1076 		F(3DNOW),
1077 	);
1078 
1079 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
1080 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
1081 
1082 	kvm_cpu_cap_init(CPUID_8000_0007_EDX,
1083 		SCATTERED_F(CONSTANT_TSC),
1084 	);
1085 
1086 	kvm_cpu_cap_init(CPUID_8000_0008_EBX,
1087 		F(CLZERO),
1088 		F(XSAVEERPTR),
1089 		F(WBNOINVD),
1090 		F(AMD_IBPB),
1091 		F(AMD_IBRS),
1092 		F(AMD_SSBD),
1093 		F(VIRT_SSBD),
1094 		F(AMD_SSB_NO),
1095 		F(AMD_STIBP),
1096 		F(AMD_STIBP_ALWAYS_ON),
1097 		F(AMD_IBRS_SAME_MODE),
1098 		F(AMD_PSFD),
1099 		F(AMD_IBPB_RET),
1100 	);
1101 
1102 	/*
1103 	 * AMD has separate bits for each SPEC_CTRL bit.
1104 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
1105 	 * record that in cpufeatures so use them.
1106 	 */
1107 	if (boot_cpu_has(X86_FEATURE_IBPB)) {
1108 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
1109 		if (boot_cpu_has(X86_FEATURE_SPEC_CTRL) &&
1110 		    !boot_cpu_has_bug(X86_BUG_EIBRS_PBRSB))
1111 			kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB_RET);
1112 	}
1113 	if (boot_cpu_has(X86_FEATURE_IBRS))
1114 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
1115 	if (boot_cpu_has(X86_FEATURE_STIBP))
1116 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
1117 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
1118 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
1119 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
1120 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
1121 	/*
1122 	 * The preference is to use SPEC CTRL MSR instead of the
1123 	 * VIRT_SPEC MSR.
1124 	 */
1125 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
1126 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
1127 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
1128 
1129 	/* All SVM features required additional vendor module enabling. */
1130 	kvm_cpu_cap_init(CPUID_8000_000A_EDX,
1131 		VENDOR_F(NPT),
1132 		VENDOR_F(VMCBCLEAN),
1133 		VENDOR_F(FLUSHBYASID),
1134 		VENDOR_F(NRIPS),
1135 		VENDOR_F(TSCRATEMSR),
1136 		VENDOR_F(V_VMSAVE_VMLOAD),
1137 		VENDOR_F(LBRV),
1138 		VENDOR_F(PAUSEFILTER),
1139 		VENDOR_F(PFTHRESHOLD),
1140 		VENDOR_F(VGIF),
1141 		VENDOR_F(VNMI),
1142 		VENDOR_F(SVME_ADDR_CHK),
1143 	);
1144 
1145 	kvm_cpu_cap_init(CPUID_8000_001F_EAX,
1146 		VENDOR_F(SME),
1147 		VENDOR_F(SEV),
1148 		/* VM_PAGE_FLUSH */
1149 		VENDOR_F(SEV_ES),
1150 		F(SME_COHERENT),
1151 	);
1152 
1153 	kvm_cpu_cap_init(CPUID_8000_0021_EAX,
1154 		F(NO_NESTED_DATA_BP),
1155 		F(WRMSR_XX_BASE_NS),
1156 		/*
1157 		 * Synthesize "LFENCE is serializing" into the AMD-defined entry
1158 		 * in KVM's supported CPUID, i.e. if the feature is reported as
1159 		 * supported by the kernel.  LFENCE_RDTSC was a Linux-defined
1160 		 * synthetic feature long before AMD joined the bandwagon, e.g.
1161 		 * LFENCE is serializing on most CPUs that support SSE2.  On
1162 		 * CPUs that don't support AMD's leaf, ANDing with the raw host
1163 		 * CPUID will drop the flags, and reporting support in AMD's
1164 		 * leaf can make it easier for userspace to detect the feature.
1165 		 */
1166 		SYNTHESIZED_F(LFENCE_RDTSC),
1167 		/* SmmPgCfgLock */
1168 		F(NULL_SEL_CLR_BASE),
1169 		/* UpperAddressIgnore */
1170 		F(AUTOIBRS),
1171 		F(PREFETCHI),
1172 		EMULATED_F(NO_SMM_CTL_MSR),
1173 		/* PrefetchCtlMsr */
1174 		/* GpOnUserCpuid */
1175 		/* EPSF */
1176 		SYNTHESIZED_F(SBPB),
1177 		SYNTHESIZED_F(IBPB_BRTYPE),
1178 		SYNTHESIZED_F(SRSO_NO),
1179 		F(SRSO_USER_KERNEL_NO),
1180 	);
1181 
1182 	kvm_cpu_cap_init(CPUID_8000_0022_EAX,
1183 		F(PERFMON_V2),
1184 	);
1185 
1186 	if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
1187 		kvm_cpu_cap_set(X86_FEATURE_NULL_SEL_CLR_BASE);
1188 
1189 	kvm_cpu_cap_init(CPUID_C000_0001_EDX,
1190 		F(XSTORE),
1191 		F(XSTORE_EN),
1192 		F(XCRYPT),
1193 		F(XCRYPT_EN),
1194 		F(ACE2),
1195 		F(ACE2_EN),
1196 		F(PHE),
1197 		F(PHE_EN),
1198 		F(PMM),
1199 		F(PMM_EN),
1200 	);
1201 
1202 	/*
1203 	 * Hide RDTSCP and RDPID if either feature is reported as supported but
1204 	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
1205 	 * should never happen, but the guest will likely crash if RDTSCP or
1206 	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
1207 	 * the past.  For example, the sanity check may fire if this instance of
1208 	 * KVM is running as L1 on top of an older, broken KVM.
1209 	 */
1210 	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
1211 		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
1212 		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
1213 		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
1214 		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
1215 	}
1216 }
1217 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
1218 
1219 #undef F
1220 #undef SCATTERED_F
1221 #undef X86_64_F
1222 #undef EMULATED_F
1223 #undef SYNTHESIZED_F
1224 #undef PASSTHROUGH_F
1225 #undef ALIASED_1_EDX_F
1226 #undef VENDOR_F
1227 #undef RUNTIME_F
1228 
1229 struct kvm_cpuid_array {
1230 	struct kvm_cpuid_entry2 *entries;
1231 	int maxnent;
1232 	int nent;
1233 };
1234 
get_next_cpuid(struct kvm_cpuid_array * array)1235 static struct kvm_cpuid_entry2 *get_next_cpuid(struct kvm_cpuid_array *array)
1236 {
1237 	if (array->nent >= array->maxnent)
1238 		return NULL;
1239 
1240 	return &array->entries[array->nent++];
1241 }
1242 
do_host_cpuid(struct kvm_cpuid_array * array,u32 function,u32 index)1243 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
1244 					      u32 function, u32 index)
1245 {
1246 	struct kvm_cpuid_entry2 *entry = get_next_cpuid(array);
1247 
1248 	if (!entry)
1249 		return NULL;
1250 
1251 	memset(entry, 0, sizeof(*entry));
1252 	entry->function = function;
1253 	entry->index = index;
1254 	switch (function & 0xC0000000) {
1255 	case 0x40000000:
1256 		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
1257 		return entry;
1258 
1259 	case 0x80000000:
1260 		/*
1261 		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
1262 		 * would result in out-of-bounds calls to do_host_cpuid.
1263 		 */
1264 		{
1265 			static int max_cpuid_80000000;
1266 			if (!READ_ONCE(max_cpuid_80000000))
1267 				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
1268 			if (function > READ_ONCE(max_cpuid_80000000))
1269 				return entry;
1270 		}
1271 		break;
1272 
1273 	default:
1274 		break;
1275 	}
1276 
1277 	cpuid_count(entry->function, entry->index,
1278 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
1279 
1280 	if (cpuid_function_is_indexed(function))
1281 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
1282 
1283 	return entry;
1284 }
1285 
cpuid_func_emulated(struct kvm_cpuid_entry2 * entry,u32 func,bool include_partially_emulated)1286 static int cpuid_func_emulated(struct kvm_cpuid_entry2 *entry, u32 func,
1287 			       bool include_partially_emulated)
1288 {
1289 	memset(entry, 0, sizeof(*entry));
1290 
1291 	entry->function = func;
1292 	entry->index = 0;
1293 	entry->flags = 0;
1294 
1295 	switch (func) {
1296 	case 0:
1297 		entry->eax = 7;
1298 		return 1;
1299 	case 1:
1300 		entry->ecx = feature_bit(MOVBE);
1301 		/*
1302 		 * KVM allows userspace to enumerate MONITOR+MWAIT support to
1303 		 * the guest, but the MWAIT feature flag is never advertised
1304 		 * to userspace because MONITOR+MWAIT aren't virtualized by
1305 		 * hardware, can't be faithfully emulated in software (KVM
1306 		 * emulates them as NOPs), and allowing the guest to execute
1307 		 * them natively requires enabling a per-VM capability.
1308 		 */
1309 		if (include_partially_emulated)
1310 			entry->ecx |= feature_bit(MWAIT);
1311 		return 1;
1312 	case 7:
1313 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
1314 		entry->eax = 0;
1315 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
1316 			entry->ecx = feature_bit(RDPID);
1317 		return 1;
1318 	default:
1319 		return 0;
1320 	}
1321 }
1322 
__do_cpuid_func_emulated(struct kvm_cpuid_array * array,u32 func)1323 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
1324 {
1325 	if (array->nent >= array->maxnent)
1326 		return -E2BIG;
1327 
1328 	array->nent += cpuid_func_emulated(&array->entries[array->nent], func, false);
1329 	return 0;
1330 }
1331 
__do_cpuid_func(struct kvm_cpuid_array * array,u32 function)1332 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
1333 {
1334 	struct kvm_cpuid_entry2 *entry;
1335 	int r, i, max_idx;
1336 
1337 	/* all calls to cpuid_count() should be made on the same cpu */
1338 	get_cpu();
1339 
1340 	r = -E2BIG;
1341 
1342 	entry = do_host_cpuid(array, function, 0);
1343 	if (!entry)
1344 		goto out;
1345 
1346 	switch (function) {
1347 	case 0:
1348 		/* Limited to the highest leaf implemented in KVM. */
1349 		entry->eax = min(entry->eax, 0x24U);
1350 		break;
1351 	case 1:
1352 		cpuid_entry_override(entry, CPUID_1_EDX);
1353 		cpuid_entry_override(entry, CPUID_1_ECX);
1354 		break;
1355 	case 2:
1356 		/*
1357 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
1358 		 * CPUID(function=2, index=0) may return different results each
1359 		 * time, with the least-significant byte in EAX enumerating the
1360 		 * number of times software should do CPUID(2, 0).
1361 		 *
1362 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
1363 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
1364 		 * return 01H. Software should ignore this value and not
1365 		 * interpret it as an informational descriptor", while AMD's
1366 		 * APM states that CPUID(2) is reserved.
1367 		 *
1368 		 * WARN if a frankenstein CPU that supports virtualization and
1369 		 * a stateful CPUID.0x2 is encountered.
1370 		 */
1371 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
1372 		break;
1373 	/* functions 4 and 0x8000001d have additional index. */
1374 	case 4:
1375 	case 0x8000001d:
1376 		/*
1377 		 * Read entries until the cache type in the previous entry is
1378 		 * zero, i.e. indicates an invalid entry.
1379 		 */
1380 		for (i = 1; entry->eax & 0x1f; ++i) {
1381 			entry = do_host_cpuid(array, function, i);
1382 			if (!entry)
1383 				goto out;
1384 		}
1385 		break;
1386 	case 6: /* Thermal management */
1387 		entry->eax = 0x4; /* allow ARAT */
1388 		entry->ebx = 0;
1389 		entry->ecx = 0;
1390 		entry->edx = 0;
1391 		break;
1392 	/* function 7 has additional index. */
1393 	case 7:
1394 		max_idx = entry->eax = min(entry->eax, 2u);
1395 		cpuid_entry_override(entry, CPUID_7_0_EBX);
1396 		cpuid_entry_override(entry, CPUID_7_ECX);
1397 		cpuid_entry_override(entry, CPUID_7_EDX);
1398 
1399 		/* KVM only supports up to 0x7.2, capped above via min(). */
1400 		if (max_idx >= 1) {
1401 			entry = do_host_cpuid(array, function, 1);
1402 			if (!entry)
1403 				goto out;
1404 
1405 			cpuid_entry_override(entry, CPUID_7_1_EAX);
1406 			cpuid_entry_override(entry, CPUID_7_1_EDX);
1407 			entry->ebx = 0;
1408 			entry->ecx = 0;
1409 		}
1410 		if (max_idx >= 2) {
1411 			entry = do_host_cpuid(array, function, 2);
1412 			if (!entry)
1413 				goto out;
1414 
1415 			cpuid_entry_override(entry, CPUID_7_2_EDX);
1416 			entry->ecx = 0;
1417 			entry->ebx = 0;
1418 			entry->eax = 0;
1419 		}
1420 		break;
1421 	case 0xa: { /* Architectural Performance Monitoring */
1422 		union cpuid10_eax eax = { };
1423 		union cpuid10_edx edx = { };
1424 
1425 		if (!enable_pmu || !static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
1426 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1427 			break;
1428 		}
1429 
1430 		eax.split.version_id = kvm_pmu_cap.version;
1431 		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
1432 		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
1433 		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
1434 		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
1435 		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
1436 
1437 		if (kvm_pmu_cap.version)
1438 			edx.split.anythread_deprecated = 1;
1439 
1440 		entry->eax = eax.full;
1441 		entry->ebx = kvm_pmu_cap.events_mask;
1442 		entry->ecx = 0;
1443 		entry->edx = edx.full;
1444 		break;
1445 	}
1446 	case 0x1f:
1447 	case 0xb:
1448 		/*
1449 		 * No topology; a valid topology is indicated by the presence
1450 		 * of subleaf 1.
1451 		 */
1452 		entry->eax = entry->ebx = entry->ecx = 0;
1453 		break;
1454 	case 0xd: {
1455 		u64 permitted_xcr0 = kvm_get_filtered_xcr0();
1456 		u64 permitted_xss = kvm_caps.supported_xss;
1457 
1458 		entry->eax &= permitted_xcr0;
1459 		entry->ebx = xstate_required_size(permitted_xcr0, false);
1460 		entry->ecx = entry->ebx;
1461 		entry->edx &= permitted_xcr0 >> 32;
1462 		if (!permitted_xcr0)
1463 			break;
1464 
1465 		entry = do_host_cpuid(array, function, 1);
1466 		if (!entry)
1467 			goto out;
1468 
1469 		cpuid_entry_override(entry, CPUID_D_1_EAX);
1470 		if (entry->eax & (feature_bit(XSAVES) | feature_bit(XSAVEC)))
1471 			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
1472 							  true);
1473 		else {
1474 			WARN_ON_ONCE(permitted_xss != 0);
1475 			entry->ebx = 0;
1476 		}
1477 		entry->ecx &= permitted_xss;
1478 		entry->edx &= permitted_xss >> 32;
1479 
1480 		for (i = 2; i < 64; ++i) {
1481 			bool s_state;
1482 			if (permitted_xcr0 & BIT_ULL(i))
1483 				s_state = false;
1484 			else if (permitted_xss & BIT_ULL(i))
1485 				s_state = true;
1486 			else
1487 				continue;
1488 
1489 			entry = do_host_cpuid(array, function, i);
1490 			if (!entry)
1491 				goto out;
1492 
1493 			/*
1494 			 * The supported check above should have filtered out
1495 			 * invalid sub-leafs.  Only valid sub-leafs should
1496 			 * reach this point, and they should have a non-zero
1497 			 * save state size.  Furthermore, check whether the
1498 			 * processor agrees with permitted_xcr0/permitted_xss
1499 			 * on whether this is an XCR0- or IA32_XSS-managed area.
1500 			 */
1501 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
1502 				--array->nent;
1503 				continue;
1504 			}
1505 
1506 			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1507 				entry->ecx &= ~BIT_ULL(2);
1508 			entry->edx = 0;
1509 		}
1510 		break;
1511 	}
1512 	case 0x12:
1513 		/* Intel SGX */
1514 		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1515 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1516 			break;
1517 		}
1518 
1519 		/*
1520 		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1521 		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1522 		 * are restricted by kernel and KVM capabilities (like most
1523 		 * feature flags), while enclave size is unrestricted.
1524 		 */
1525 		cpuid_entry_override(entry, CPUID_12_EAX);
1526 		entry->ebx &= SGX_MISC_EXINFO;
1527 
1528 		entry = do_host_cpuid(array, function, 1);
1529 		if (!entry)
1530 			goto out;
1531 
1532 		/*
1533 		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1534 		 * feature flags.  Advertise all supported flags, including
1535 		 * privileged attributes that require explicit opt-in from
1536 		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1537 		 * expected to derive it from supported XCR0.
1538 		 */
1539 		entry->eax &= SGX_ATTR_PRIV_MASK | SGX_ATTR_UNPRIV_MASK;
1540 		entry->ebx &= 0;
1541 		break;
1542 	/* Intel PT */
1543 	case 0x14:
1544 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1545 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1546 			break;
1547 		}
1548 
1549 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1550 			if (!do_host_cpuid(array, function, i))
1551 				goto out;
1552 		}
1553 		break;
1554 	/* Intel AMX TILE */
1555 	case 0x1d:
1556 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1557 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1558 			break;
1559 		}
1560 
1561 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1562 			if (!do_host_cpuid(array, function, i))
1563 				goto out;
1564 		}
1565 		break;
1566 	case 0x1e: /* TMUL information */
1567 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1568 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1569 			break;
1570 		}
1571 		break;
1572 	case 0x24: {
1573 		u8 avx10_version;
1574 
1575 		if (!kvm_cpu_cap_has(X86_FEATURE_AVX10)) {
1576 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1577 			break;
1578 		}
1579 
1580 		/*
1581 		 * The AVX10 version is encoded in EBX[7:0].  Note, the version
1582 		 * is guaranteed to be >=1 if AVX10 is supported.  Note #2, the
1583 		 * version needs to be captured before overriding EBX features!
1584 		 */
1585 		avx10_version = min_t(u8, entry->ebx & 0xff, 1);
1586 		cpuid_entry_override(entry, CPUID_24_0_EBX);
1587 		entry->ebx |= avx10_version;
1588 
1589 		entry->eax = 0;
1590 		entry->ecx = 0;
1591 		entry->edx = 0;
1592 		break;
1593 	}
1594 	case KVM_CPUID_SIGNATURE: {
1595 		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1596 		entry->eax = KVM_CPUID_FEATURES;
1597 		entry->ebx = sigptr[0];
1598 		entry->ecx = sigptr[1];
1599 		entry->edx = sigptr[2];
1600 		break;
1601 	}
1602 	case KVM_CPUID_FEATURES:
1603 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1604 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1605 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1606 			     (1 << KVM_FEATURE_ASYNC_PF) |
1607 			     (1 << KVM_FEATURE_PV_EOI) |
1608 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1609 			     (1 << KVM_FEATURE_PV_UNHALT) |
1610 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1611 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1612 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1613 			     (1 << KVM_FEATURE_POLL_CONTROL) |
1614 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1615 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1616 
1617 		if (sched_info_on())
1618 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1619 
1620 		entry->ebx = 0;
1621 		entry->ecx = 0;
1622 		entry->edx = 0;
1623 		break;
1624 	case 0x80000000:
1625 		entry->eax = min(entry->eax, 0x80000022);
1626 		/*
1627 		 * Serializing LFENCE is reported in a multitude of ways, and
1628 		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1629 		 * userspace by providing the CPUID leaf ourselves.
1630 		 *
1631 		 * However, only do it if the host has CPUID leaf 0x8000001d.
1632 		 * QEMU thinks that it can query the host blindly for that
1633 		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1634 		 * above.  The processor merrily returns values from the
1635 		 * highest Intel leaf which QEMU tries to use as the guest's
1636 		 * 0x8000001d.  Even worse, this can result in an infinite
1637 		 * loop if said highest leaf has no subleaves indexed by ECX.
1638 		 */
1639 		if (entry->eax >= 0x8000001d &&
1640 		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1641 		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1642 			entry->eax = max(entry->eax, 0x80000021);
1643 		break;
1644 	case 0x80000001:
1645 		entry->ebx &= ~GENMASK(27, 16);
1646 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1647 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1648 		break;
1649 	case 0x80000005:
1650 		/*  Pass host L1 cache and TLB info. */
1651 		break;
1652 	case 0x80000006:
1653 		/* Drop reserved bits, pass host L2 cache and TLB info. */
1654 		entry->edx &= ~GENMASK(17, 16);
1655 		break;
1656 	case 0x80000007: /* Advanced power management */
1657 		cpuid_entry_override(entry, CPUID_8000_0007_EDX);
1658 
1659 		/* mask against host */
1660 		entry->edx &= boot_cpu_data.x86_power;
1661 		entry->eax = entry->ebx = entry->ecx = 0;
1662 		break;
1663 	case 0x80000008: {
1664 		/*
1665 		 * GuestPhysAddrSize (EAX[23:16]) is intended for software
1666 		 * use.
1667 		 *
1668 		 * KVM's ABI is to report the effective MAXPHYADDR for the
1669 		 * guest in PhysAddrSize (phys_as), and the maximum
1670 		 * *addressable* GPA in GuestPhysAddrSize (g_phys_as).
1671 		 *
1672 		 * GuestPhysAddrSize is valid if and only if TDP is enabled,
1673 		 * in which case the max GPA that can be addressed by KVM may
1674 		 * be less than the max GPA that can be legally generated by
1675 		 * the guest, e.g. if MAXPHYADDR>48 but the CPU doesn't
1676 		 * support 5-level TDP.
1677 		 */
1678 		unsigned int virt_as = max((entry->eax >> 8) & 0xff, 48U);
1679 		unsigned int phys_as, g_phys_as;
1680 
1681 		/*
1682 		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1683 		 * the guest operates in the same PA space as the host, i.e.
1684 		 * reductions in MAXPHYADDR for memory encryption affect shadow
1685 		 * paging, too.
1686 		 *
1687 		 * If TDP is enabled, use the raw bare metal MAXPHYADDR as
1688 		 * reductions to the HPAs do not affect GPAs.  The max
1689 		 * addressable GPA is the same as the max effective GPA, except
1690 		 * that it's capped at 48 bits if 5-level TDP isn't supported
1691 		 * (hardware processes bits 51:48 only when walking the fifth
1692 		 * level page table).
1693 		 */
1694 		if (!tdp_enabled) {
1695 			phys_as = boot_cpu_data.x86_phys_bits;
1696 			g_phys_as = 0;
1697 		} else {
1698 			phys_as = entry->eax & 0xff;
1699 			g_phys_as = phys_as;
1700 			if (kvm_mmu_get_max_tdp_level() < 5)
1701 				g_phys_as = min(g_phys_as, 48U);
1702 		}
1703 
1704 		entry->eax = phys_as | (virt_as << 8) | (g_phys_as << 16);
1705 		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
1706 		entry->edx = 0;
1707 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1708 		break;
1709 	}
1710 	case 0x8000000A:
1711 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1712 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1713 			break;
1714 		}
1715 		entry->eax = 1; /* SVM revision 1 */
1716 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1717 				   ASID emulation to nested SVM */
1718 		entry->ecx = 0; /* Reserved */
1719 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1720 		break;
1721 	case 0x80000019:
1722 		entry->ecx = entry->edx = 0;
1723 		break;
1724 	case 0x8000001a:
1725 		entry->eax &= GENMASK(2, 0);
1726 		entry->ebx = entry->ecx = entry->edx = 0;
1727 		break;
1728 	case 0x8000001e:
1729 		/* Do not return host topology information.  */
1730 		entry->eax = entry->ebx = entry->ecx = 0;
1731 		entry->edx = 0; /* reserved */
1732 		break;
1733 	case 0x8000001F:
1734 		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1735 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1736 		} else {
1737 			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1738 			/* Clear NumVMPL since KVM does not support VMPL.  */
1739 			entry->ebx &= ~GENMASK(31, 12);
1740 			/*
1741 			 * Enumerate '0' for "PA bits reduction", the adjusted
1742 			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1743 			 */
1744 			entry->ebx &= ~GENMASK(11, 6);
1745 		}
1746 		break;
1747 	case 0x80000020:
1748 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1749 		break;
1750 	case 0x80000021:
1751 		entry->ebx = entry->ecx = entry->edx = 0;
1752 		cpuid_entry_override(entry, CPUID_8000_0021_EAX);
1753 		break;
1754 	/* AMD Extended Performance Monitoring and Debug */
1755 	case 0x80000022: {
1756 		union cpuid_0x80000022_ebx ebx = { };
1757 
1758 		entry->ecx = entry->edx = 0;
1759 		if (!enable_pmu || !kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2)) {
1760 			entry->eax = entry->ebx = 0;
1761 			break;
1762 		}
1763 
1764 		cpuid_entry_override(entry, CPUID_8000_0022_EAX);
1765 
1766 		ebx.split.num_core_pmc = kvm_pmu_cap.num_counters_gp;
1767 		entry->ebx = ebx.full;
1768 		break;
1769 	}
1770 	/*Add support for Centaur's CPUID instruction*/
1771 	case 0xC0000000:
1772 		/*Just support up to 0xC0000004 now*/
1773 		entry->eax = min(entry->eax, 0xC0000004);
1774 		break;
1775 	case 0xC0000001:
1776 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1777 		break;
1778 	case 3: /* Processor serial number */
1779 	case 5: /* MONITOR/MWAIT */
1780 	case 0xC0000002:
1781 	case 0xC0000003:
1782 	case 0xC0000004:
1783 	default:
1784 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1785 		break;
1786 	}
1787 
1788 	r = 0;
1789 
1790 out:
1791 	put_cpu();
1792 
1793 	return r;
1794 }
1795 
do_cpuid_func(struct kvm_cpuid_array * array,u32 func,unsigned int type)1796 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1797 			 unsigned int type)
1798 {
1799 	if (type == KVM_GET_EMULATED_CPUID)
1800 		return __do_cpuid_func_emulated(array, func);
1801 
1802 	return __do_cpuid_func(array, func);
1803 }
1804 
1805 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
1806 
get_cpuid_func(struct kvm_cpuid_array * array,u32 func,unsigned int type)1807 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1808 			  unsigned int type)
1809 {
1810 	u32 limit;
1811 	int r;
1812 
1813 	if (func == CENTAUR_CPUID_SIGNATURE &&
1814 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1815 		return 0;
1816 
1817 	r = do_cpuid_func(array, func, type);
1818 	if (r)
1819 		return r;
1820 
1821 	limit = array->entries[array->nent - 1].eax;
1822 	for (func = func + 1; func <= limit; ++func) {
1823 		r = do_cpuid_func(array, func, type);
1824 		if (r)
1825 			break;
1826 	}
1827 
1828 	return r;
1829 }
1830 
sanity_check_entries(struct kvm_cpuid_entry2 __user * entries,__u32 num_entries,unsigned int ioctl_type)1831 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1832 				 __u32 num_entries, unsigned int ioctl_type)
1833 {
1834 	int i;
1835 	__u32 pad[3];
1836 
1837 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1838 		return false;
1839 
1840 	/*
1841 	 * We want to make sure that ->padding is being passed clean from
1842 	 * userspace in case we want to use it for something in the future.
1843 	 *
1844 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1845 	 * have to give ourselves satisfied only with the emulated side. /me
1846 	 * sheds a tear.
1847 	 */
1848 	for (i = 0; i < num_entries; i++) {
1849 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1850 			return true;
1851 
1852 		if (pad[0] || pad[1] || pad[2])
1853 			return true;
1854 	}
1855 	return false;
1856 }
1857 
kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 * cpuid,struct kvm_cpuid_entry2 __user * entries,unsigned int type)1858 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1859 			    struct kvm_cpuid_entry2 __user *entries,
1860 			    unsigned int type)
1861 {
1862 	static const u32 funcs[] = {
1863 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1864 	};
1865 
1866 	struct kvm_cpuid_array array = {
1867 		.nent = 0,
1868 	};
1869 	int r, i;
1870 
1871 	if (cpuid->nent < 1)
1872 		return -E2BIG;
1873 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1874 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1875 
1876 	if (sanity_check_entries(entries, cpuid->nent, type))
1877 		return -EINVAL;
1878 
1879 	array.entries = kvcalloc(cpuid->nent, sizeof(struct kvm_cpuid_entry2), GFP_KERNEL);
1880 	if (!array.entries)
1881 		return -ENOMEM;
1882 
1883 	array.maxnent = cpuid->nent;
1884 
1885 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1886 		r = get_cpuid_func(&array, funcs[i], type);
1887 		if (r)
1888 			goto out_free;
1889 	}
1890 	cpuid->nent = array.nent;
1891 
1892 	if (copy_to_user(entries, array.entries,
1893 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1894 		r = -EFAULT;
1895 
1896 out_free:
1897 	kvfree(array.entries);
1898 	return r;
1899 }
1900 
1901 /*
1902  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1903  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1904  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1905  * range.  Centaur/VIA follows Intel semantics.
1906  *
1907  * A leaf is considered out-of-range if its function is higher than the maximum
1908  * supported leaf of its associated class or if its associated class does not
1909  * exist.
1910  *
1911  * There are three primary classes to be considered, with their respective
1912  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1913  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1914  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1915  *
1916  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1917  *  - Hypervisor: 0x40000000 - 0x4fffffff
1918  *  - Extended:   0x80000000 - 0xbfffffff
1919  *  - Centaur:    0xc0000000 - 0xcfffffff
1920  *
1921  * The Hypervisor class is further subdivided into sub-classes that each act as
1922  * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1923  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1924  * CPUID sub-classes are:
1925  *
1926  *  - HyperV:     0x40000000 - 0x400000ff
1927  *  - KVM:        0x40000100 - 0x400001ff
1928  */
1929 static struct kvm_cpuid_entry2 *
get_out_of_range_cpuid_entry(struct kvm_vcpu * vcpu,u32 * fn_ptr,u32 index)1930 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1931 {
1932 	struct kvm_cpuid_entry2 *basic, *class;
1933 	u32 function = *fn_ptr;
1934 
1935 	basic = kvm_find_cpuid_entry(vcpu, 0);
1936 	if (!basic)
1937 		return NULL;
1938 
1939 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1940 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1941 		return NULL;
1942 
1943 	if (function >= 0x40000000 && function <= 0x4fffffff)
1944 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1945 	else if (function >= 0xc0000000)
1946 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1947 	else
1948 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1949 
1950 	if (class && function <= class->eax)
1951 		return NULL;
1952 
1953 	/*
1954 	 * Leaf specific adjustments are also applied when redirecting to the
1955 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1956 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1957 	 * needs to be pulled from CPUID.0xb.1.
1958 	 */
1959 	*fn_ptr = basic->eax;
1960 
1961 	/*
1962 	 * The class does not exist or the requested function is out of range;
1963 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1964 	 * the original requested leaf is observed!
1965 	 */
1966 	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1967 }
1968 
kvm_cpuid(struct kvm_vcpu * vcpu,u32 * eax,u32 * ebx,u32 * ecx,u32 * edx,bool exact_only)1969 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1970 	       u32 *ecx, u32 *edx, bool exact_only)
1971 {
1972 	u32 orig_function = *eax, function = *eax, index = *ecx;
1973 	struct kvm_cpuid_entry2 *entry;
1974 	bool exact, used_max_basic = false;
1975 
1976 	if (vcpu->arch.cpuid_dynamic_bits_dirty)
1977 		kvm_update_cpuid_runtime(vcpu);
1978 
1979 	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1980 	exact = !!entry;
1981 
1982 	if (!entry && !exact_only) {
1983 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1984 		used_max_basic = !!entry;
1985 	}
1986 
1987 	if (entry) {
1988 		*eax = entry->eax;
1989 		*ebx = entry->ebx;
1990 		*ecx = entry->ecx;
1991 		*edx = entry->edx;
1992 		if (function == 7 && index == 0) {
1993 			u64 data;
1994 			if ((*ebx & (feature_bit(RTM) | feature_bit(HLE))) &&
1995 			    !__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1996 			    (data & TSX_CTRL_CPUID_CLEAR))
1997 				*ebx &= ~(feature_bit(RTM) | feature_bit(HLE));
1998 		} else if (function == 0x80000007) {
1999 			if (kvm_hv_invtsc_suppressed(vcpu))
2000 				*edx &= ~feature_bit(CONSTANT_TSC);
2001 		} else if (IS_ENABLED(CONFIG_KVM_XEN) &&
2002 			   kvm_xen_is_tsc_leaf(vcpu, function)) {
2003 			/*
2004 			 * Update guest TSC frequency information if necessary.
2005 			 * Ignore failures, there is no sane value that can be
2006 			 * provided if KVM can't get the TSC frequency.
2007 			 */
2008 			if (kvm_check_request(KVM_REQ_CLOCK_UPDATE, vcpu))
2009 				kvm_guest_time_update(vcpu);
2010 
2011 			if (index == 1) {
2012 				*ecx = vcpu->arch.pvclock_tsc_mul;
2013 				*edx = vcpu->arch.pvclock_tsc_shift;
2014 			} else if (index == 2) {
2015 				*eax = vcpu->arch.hw_tsc_khz;
2016 			}
2017 		}
2018 	} else {
2019 		*eax = *ebx = *ecx = *edx = 0;
2020 		/*
2021 		 * When leaf 0BH or 1FH is defined, CL is pass-through
2022 		 * and EDX is always the x2APIC ID, even for undefined
2023 		 * subleaves. Index 1 will exist iff the leaf is
2024 		 * implemented, so we pass through CL iff leaf 1
2025 		 * exists. EDX can be copied from any existing index.
2026 		 */
2027 		if (function == 0xb || function == 0x1f) {
2028 			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
2029 			if (entry) {
2030 				*ecx = index & 0xff;
2031 				*edx = entry->edx;
2032 			}
2033 		}
2034 	}
2035 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
2036 			used_max_basic);
2037 	return exact;
2038 }
2039 EXPORT_SYMBOL_GPL(kvm_cpuid);
2040 
kvm_emulate_cpuid(struct kvm_vcpu * vcpu)2041 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
2042 {
2043 	u32 eax, ebx, ecx, edx;
2044 
2045 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
2046 		return 1;
2047 
2048 	eax = kvm_rax_read(vcpu);
2049 	ecx = kvm_rcx_read(vcpu);
2050 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
2051 	kvm_rax_write(vcpu, eax);
2052 	kvm_rbx_write(vcpu, ebx);
2053 	kvm_rcx_write(vcpu, ecx);
2054 	kvm_rdx_write(vcpu, edx);
2055 	return kvm_skip_emulated_instruction(vcpu);
2056 }
2057 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
2058