xref: /linux/arch/x86/kvm/cpuid.c (revision a382b06d297e78ed7ac67afd0d8e8690406ac4ca)
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.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 	int feature_bit = 0;
62 	u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
63 
64 	xstate_bv &= XFEATURE_MASK_EXTEND;
65 	while (xstate_bv) {
66 		if (xstate_bv & 0x1) {
67 			struct cpuid_xstate_sizes *xs = &xstate_sizes[feature_bit];
68 			u32 offset;
69 
70 			/* ECX[1]: 64B alignment in compacted form */
71 			if (compacted)
72 				offset = (xs->ecx & 0x2) ? ALIGN(ret, 64) : ret;
73 			else
74 				offset = xs->ebx;
75 			ret = max(ret, offset + xs->eax);
76 		}
77 
78 		xstate_bv >>= 1;
79 		feature_bit++;
80 	}
81 
82 	return ret;
83 }
84 
85 /*
86  * Magic value used by KVM when querying userspace-provided CPUID entries and
87  * doesn't care about the CPIUD index because the index of the function in
88  * question is not significant.  Note, this magic value must have at least one
89  * bit set in bits[63:32] and must be consumed as a u64 by cpuid_entry2_find()
90  * to avoid false positives when processing guest CPUID input.
91  */
92 #define KVM_CPUID_INDEX_NOT_SIGNIFICANT -1ull
93 
cpuid_entry2_find(struct kvm_vcpu * vcpu,u32 function,u64 index)94 static struct kvm_cpuid_entry2 *cpuid_entry2_find(struct kvm_vcpu *vcpu,
95 						  u32 function, u64 index)
96 {
97 	struct kvm_cpuid_entry2 *e;
98 	int i;
99 
100 	/*
101 	 * KVM has a semi-arbitrary rule that querying the guest's CPUID model
102 	 * with IRQs disabled is disallowed.  The CPUID model can legitimately
103 	 * have over one hundred entries, i.e. the lookup is slow, and IRQs are
104 	 * typically disabled in KVM only when KVM is in a performance critical
105 	 * path, e.g. the core VM-Enter/VM-Exit run loop.  Nothing will break
106 	 * if this rule is violated, this assertion is purely to flag potential
107 	 * performance issues.  If this fires, consider moving the lookup out
108 	 * of the hotpath, e.g. by caching information during CPUID updates.
109 	 */
110 	lockdep_assert_irqs_enabled();
111 
112 	for (i = 0; i < vcpu->arch.cpuid_nent; i++) {
113 		e = &vcpu->arch.cpuid_entries[i];
114 
115 		if (e->function != function)
116 			continue;
117 
118 		/*
119 		 * If the index isn't significant, use the first entry with a
120 		 * matching function.  It's userspace's responsibility to not
121 		 * provide "duplicate" entries in all cases.
122 		 */
123 		if (!(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) || e->index == index)
124 			return e;
125 
126 
127 		/*
128 		 * Similarly, use the first matching entry if KVM is doing a
129 		 * lookup (as opposed to emulating CPUID) for a function that's
130 		 * architecturally defined as not having a significant index.
131 		 */
132 		if (index == KVM_CPUID_INDEX_NOT_SIGNIFICANT) {
133 			/*
134 			 * Direct lookups from KVM should not diverge from what
135 			 * KVM defines internally (the architectural behavior).
136 			 */
137 			WARN_ON_ONCE(cpuid_function_is_indexed(function));
138 			return e;
139 		}
140 	}
141 
142 	return NULL;
143 }
144 
kvm_find_cpuid_entry_index(struct kvm_vcpu * vcpu,u32 function,u32 index)145 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
146 						    u32 function, u32 index)
147 {
148 	return cpuid_entry2_find(vcpu, function, index);
149 }
150 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry_index);
151 
kvm_find_cpuid_entry(struct kvm_vcpu * vcpu,u32 function)152 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
153 					      u32 function)
154 {
155 	return cpuid_entry2_find(vcpu, function, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
156 }
157 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
158 
159 /*
160  * cpuid_entry2_find() and KVM_CPUID_INDEX_NOT_SIGNIFICANT should never be used
161  * directly outside of kvm_find_cpuid_entry() and kvm_find_cpuid_entry_index().
162  */
163 #undef KVM_CPUID_INDEX_NOT_SIGNIFICANT
164 
kvm_check_cpuid(struct kvm_vcpu * vcpu)165 static int kvm_check_cpuid(struct kvm_vcpu *vcpu)
166 {
167 	struct kvm_cpuid_entry2 *best;
168 	u64 xfeatures;
169 
170 	/*
171 	 * The existing code assumes virtual address is 48-bit or 57-bit in the
172 	 * canonical address checks; exit if it is ever changed.
173 	 */
174 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
175 	if (best) {
176 		int vaddr_bits = (best->eax & 0xff00) >> 8;
177 
178 		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
179 			return -EINVAL;
180 	}
181 
182 	/*
183 	 * Exposing dynamic xfeatures to the guest requires additional
184 	 * enabling in the FPU, e.g. to expand the guest XSAVE state size.
185 	 */
186 	best = kvm_find_cpuid_entry_index(vcpu, 0xd, 0);
187 	if (!best)
188 		return 0;
189 
190 	xfeatures = best->eax | ((u64)best->edx << 32);
191 	xfeatures &= XFEATURE_MASK_USER_DYNAMIC;
192 	if (!xfeatures)
193 		return 0;
194 
195 	return fpu_enable_guest_xfd_features(&vcpu->arch.guest_fpu, xfeatures);
196 }
197 
198 static u32 kvm_apply_cpuid_pv_features_quirk(struct kvm_vcpu *vcpu);
199 
200 /* 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)201 static int kvm_cpuid_check_equal(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
202 				 int nent)
203 {
204 	struct kvm_cpuid_entry2 *orig;
205 	int i;
206 
207 	/*
208 	 * Apply runtime CPUID updates to the incoming CPUID entries to avoid
209 	 * false positives due mismatches on KVM-owned feature flags.
210 	 *
211 	 * Note!  @e2 and @nent track the _old_ CPUID entries!
212 	 */
213 	kvm_update_cpuid_runtime(vcpu);
214 	kvm_apply_cpuid_pv_features_quirk(vcpu);
215 
216 	if (nent != vcpu->arch.cpuid_nent)
217 		return -EINVAL;
218 
219 	for (i = 0; i < nent; i++) {
220 		orig = &vcpu->arch.cpuid_entries[i];
221 		if (e2[i].function != orig->function ||
222 		    e2[i].index != orig->index ||
223 		    e2[i].flags != orig->flags ||
224 		    e2[i].eax != orig->eax || e2[i].ebx != orig->ebx ||
225 		    e2[i].ecx != orig->ecx || e2[i].edx != orig->edx)
226 			return -EINVAL;
227 	}
228 
229 	return 0;
230 }
231 
kvm_get_hypervisor_cpuid(struct kvm_vcpu * vcpu,const char * sig)232 static struct kvm_hypervisor_cpuid kvm_get_hypervisor_cpuid(struct kvm_vcpu *vcpu,
233 							    const char *sig)
234 {
235 	struct kvm_hypervisor_cpuid cpuid = {};
236 	struct kvm_cpuid_entry2 *entry;
237 	u32 base;
238 
239 	for_each_possible_hypervisor_cpuid_base(base) {
240 		entry = kvm_find_cpuid_entry(vcpu, base);
241 
242 		if (entry) {
243 			u32 signature[3];
244 
245 			signature[0] = entry->ebx;
246 			signature[1] = entry->ecx;
247 			signature[2] = entry->edx;
248 
249 			if (!memcmp(signature, sig, sizeof(signature))) {
250 				cpuid.base = base;
251 				cpuid.limit = entry->eax;
252 				break;
253 			}
254 		}
255 	}
256 
257 	return cpuid;
258 }
259 
kvm_apply_cpuid_pv_features_quirk(struct kvm_vcpu * vcpu)260 static u32 kvm_apply_cpuid_pv_features_quirk(struct kvm_vcpu *vcpu)
261 {
262 	struct kvm_hypervisor_cpuid kvm_cpuid;
263 	struct kvm_cpuid_entry2 *best;
264 
265 	kvm_cpuid = kvm_get_hypervisor_cpuid(vcpu, KVM_SIGNATURE);
266 	if (!kvm_cpuid.base)
267 		return 0;
268 
269 	best = kvm_find_cpuid_entry(vcpu, kvm_cpuid.base | KVM_CPUID_FEATURES);
270 	if (!best)
271 		return 0;
272 
273 	if (kvm_hlt_in_guest(vcpu->kvm))
274 		best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
275 
276 	return best->eax;
277 }
278 
279 /*
280  * Calculate guest's supported XCR0 taking into account guest CPUID data and
281  * KVM's supported XCR0 (comprised of host's XCR0 and KVM_SUPPORTED_XCR0).
282  */
cpuid_get_supported_xcr0(struct kvm_vcpu * vcpu)283 static u64 cpuid_get_supported_xcr0(struct kvm_vcpu *vcpu)
284 {
285 	struct kvm_cpuid_entry2 *best;
286 
287 	best = kvm_find_cpuid_entry_index(vcpu, 0xd, 0);
288 	if (!best)
289 		return 0;
290 
291 	return (best->eax | ((u64)best->edx << 32)) & kvm_caps.supported_xcr0;
292 }
293 
kvm_update_feature_runtime(struct kvm_vcpu * vcpu,struct kvm_cpuid_entry2 * entry,unsigned int x86_feature,bool has_feature)294 static __always_inline void kvm_update_feature_runtime(struct kvm_vcpu *vcpu,
295 						       struct kvm_cpuid_entry2 *entry,
296 						       unsigned int x86_feature,
297 						       bool has_feature)
298 {
299 	cpuid_entry_change(entry, x86_feature, has_feature);
300 	guest_cpu_cap_change(vcpu, x86_feature, has_feature);
301 }
302 
kvm_update_cpuid_runtime(struct kvm_vcpu * vcpu)303 void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu)
304 {
305 	struct kvm_cpuid_entry2 *best;
306 
307 	best = kvm_find_cpuid_entry(vcpu, 1);
308 	if (best) {
309 		kvm_update_feature_runtime(vcpu, best, X86_FEATURE_OSXSAVE,
310 					   kvm_is_cr4_bit_set(vcpu, X86_CR4_OSXSAVE));
311 
312 		kvm_update_feature_runtime(vcpu, best, X86_FEATURE_APIC,
313 					   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
314 
315 		if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT))
316 			kvm_update_feature_runtime(vcpu, best, X86_FEATURE_MWAIT,
317 						   vcpu->arch.ia32_misc_enable_msr &
318 						   MSR_IA32_MISC_ENABLE_MWAIT);
319 	}
320 
321 	best = kvm_find_cpuid_entry_index(vcpu, 7, 0);
322 	if (best)
323 		kvm_update_feature_runtime(vcpu, best, X86_FEATURE_OSPKE,
324 					   kvm_is_cr4_bit_set(vcpu, X86_CR4_PKE));
325 
326 
327 	best = kvm_find_cpuid_entry_index(vcpu, 0xD, 0);
328 	if (best)
329 		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
330 
331 	best = kvm_find_cpuid_entry_index(vcpu, 0xD, 1);
332 	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
333 		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
334 		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
335 }
336 EXPORT_SYMBOL_GPL(kvm_update_cpuid_runtime);
337 
kvm_cpuid_has_hyperv(struct kvm_vcpu * vcpu)338 static bool kvm_cpuid_has_hyperv(struct kvm_vcpu *vcpu)
339 {
340 #ifdef CONFIG_KVM_HYPERV
341 	struct kvm_cpuid_entry2 *entry;
342 
343 	entry = kvm_find_cpuid_entry(vcpu, HYPERV_CPUID_INTERFACE);
344 	return entry && entry->eax == HYPERV_CPUID_SIGNATURE_EAX;
345 #else
346 	return false;
347 #endif
348 }
349 
guest_cpuid_is_amd_or_hygon(struct kvm_vcpu * vcpu)350 static bool guest_cpuid_is_amd_or_hygon(struct kvm_vcpu *vcpu)
351 {
352 	struct kvm_cpuid_entry2 *entry;
353 
354 	entry = kvm_find_cpuid_entry(vcpu, 0);
355 	if (!entry)
356 		return false;
357 
358 	return is_guest_vendor_amd(entry->ebx, entry->ecx, entry->edx) ||
359 	       is_guest_vendor_hygon(entry->ebx, entry->ecx, entry->edx);
360 }
361 
362 /*
363  * This isn't truly "unsafe", but except for the cpu_caps initialization code,
364  * all register lookups should use __cpuid_entry_get_reg(), which provides
365  * compile-time validation of the input.
366  */
cpuid_get_reg_unsafe(struct kvm_cpuid_entry2 * entry,u32 reg)367 static u32 cpuid_get_reg_unsafe(struct kvm_cpuid_entry2 *entry, u32 reg)
368 {
369 	switch (reg) {
370 	case CPUID_EAX:
371 		return entry->eax;
372 	case CPUID_EBX:
373 		return entry->ebx;
374 	case CPUID_ECX:
375 		return entry->ecx;
376 	case CPUID_EDX:
377 		return entry->edx;
378 	default:
379 		WARN_ON_ONCE(1);
380 		return 0;
381 	}
382 }
383 
384 static int cpuid_func_emulated(struct kvm_cpuid_entry2 *entry, u32 func,
385 			       bool include_partially_emulated);
386 
kvm_vcpu_after_set_cpuid(struct kvm_vcpu * vcpu)387 void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
388 {
389 	struct kvm_lapic *apic = vcpu->arch.apic;
390 	struct kvm_cpuid_entry2 *best;
391 	struct kvm_cpuid_entry2 *entry;
392 	bool allow_gbpages;
393 	int i;
394 
395 	memset(vcpu->arch.cpu_caps, 0, sizeof(vcpu->arch.cpu_caps));
396 	BUILD_BUG_ON(ARRAY_SIZE(reverse_cpuid) != NR_KVM_CPU_CAPS);
397 
398 	/*
399 	 * Reset guest capabilities to userspace's guest CPUID definition, i.e.
400 	 * honor userspace's definition for features that don't require KVM or
401 	 * hardware management/support (or that KVM simply doesn't care about).
402 	 */
403 	for (i = 0; i < NR_KVM_CPU_CAPS; i++) {
404 		const struct cpuid_reg cpuid = reverse_cpuid[i];
405 		struct kvm_cpuid_entry2 emulated;
406 
407 		if (!cpuid.function)
408 			continue;
409 
410 		entry = kvm_find_cpuid_entry_index(vcpu, cpuid.function, cpuid.index);
411 		if (!entry)
412 			continue;
413 
414 		cpuid_func_emulated(&emulated, cpuid.function, true);
415 
416 		/*
417 		 * A vCPU has a feature if it's supported by KVM and is enabled
418 		 * in guest CPUID.  Note, this includes features that are
419 		 * supported by KVM but aren't advertised to userspace!
420 		 */
421 		vcpu->arch.cpu_caps[i] = kvm_cpu_caps[i] |
422 					 cpuid_get_reg_unsafe(&emulated, cpuid.reg);
423 		vcpu->arch.cpu_caps[i] &= cpuid_get_reg_unsafe(entry, cpuid.reg);
424 	}
425 
426 	kvm_update_cpuid_runtime(vcpu);
427 
428 	/*
429 	 * If TDP is enabled, let the guest use GBPAGES if they're supported in
430 	 * hardware.  The hardware page walker doesn't let KVM disable GBPAGES,
431 	 * i.e. won't treat them as reserved, and KVM doesn't redo the GVA->GPA
432 	 * walk for performance and complexity reasons.  Not to mention KVM
433 	 * _can't_ solve the problem because GVA->GPA walks aren't visible to
434 	 * KVM once a TDP translation is installed.  Mimic hardware behavior so
435 	 * that KVM's is at least consistent, i.e. doesn't randomly inject #PF.
436 	 * If TDP is disabled, honor *only* guest CPUID as KVM has full control
437 	 * and can install smaller shadow pages if the host lacks 1GiB support.
438 	 */
439 	allow_gbpages = tdp_enabled ? boot_cpu_has(X86_FEATURE_GBPAGES) :
440 				      guest_cpu_cap_has(vcpu, X86_FEATURE_GBPAGES);
441 	guest_cpu_cap_change(vcpu, X86_FEATURE_GBPAGES, allow_gbpages);
442 
443 	best = kvm_find_cpuid_entry(vcpu, 1);
444 	if (best && apic) {
445 		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
446 			apic->lapic_timer.timer_mode_mask = 3 << 17;
447 		else
448 			apic->lapic_timer.timer_mode_mask = 1 << 17;
449 
450 		kvm_apic_set_version(vcpu);
451 	}
452 
453 	vcpu->arch.guest_supported_xcr0 = cpuid_get_supported_xcr0(vcpu);
454 
455 	vcpu->arch.pv_cpuid.features = kvm_apply_cpuid_pv_features_quirk(vcpu);
456 
457 	vcpu->arch.is_amd_compatible = guest_cpuid_is_amd_or_hygon(vcpu);
458 	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
459 	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
460 
461 	kvm_pmu_refresh(vcpu);
462 
463 #define __kvm_cpu_cap_has(UNUSED_, f) kvm_cpu_cap_has(f)
464 	vcpu->arch.cr4_guest_rsvd_bits = __cr4_reserved_bits(__kvm_cpu_cap_has, UNUSED_) |
465 					 __cr4_reserved_bits(guest_cpu_cap_has, vcpu);
466 #undef __kvm_cpu_cap_has
467 
468 	kvm_hv_set_cpuid(vcpu, kvm_cpuid_has_hyperv(vcpu));
469 
470 	/* Invoke the vendor callback only after the above state is updated. */
471 	kvm_x86_call(vcpu_after_set_cpuid)(vcpu);
472 
473 	/*
474 	 * Except for the MMU, which needs to do its thing any vendor specific
475 	 * adjustments to the reserved GPA bits.
476 	 */
477 	kvm_mmu_after_set_cpuid(vcpu);
478 }
479 
cpuid_query_maxphyaddr(struct kvm_vcpu * vcpu)480 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
481 {
482 	struct kvm_cpuid_entry2 *best;
483 
484 	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
485 	if (!best || best->eax < 0x80000008)
486 		goto not_found;
487 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
488 	if (best)
489 		return best->eax & 0xff;
490 not_found:
491 	return 36;
492 }
493 
494 /*
495  * This "raw" version returns the reserved GPA bits without any adjustments for
496  * encryption technologies that usurp bits.  The raw mask should be used if and
497  * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
498  */
kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu * vcpu)499 u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
500 {
501 	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
502 }
503 
kvm_set_cpuid(struct kvm_vcpu * vcpu,struct kvm_cpuid_entry2 * e2,int nent)504 static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
505                         int nent)
506 {
507 	u32 vcpu_caps[NR_KVM_CPU_CAPS];
508 	int r;
509 
510 	/*
511 	 * Swap the existing (old) entries with the incoming (new) entries in
512 	 * order to massage the new entries, e.g. to account for dynamic bits
513 	 * that KVM controls, without clobbering the current guest CPUID, which
514 	 * KVM needs to preserve in order to unwind on failure.
515 	 *
516 	 * Similarly, save the vCPU's current cpu_caps so that the capabilities
517 	 * can be updated alongside the CPUID entries when performing runtime
518 	 * updates.  Full initialization is done if and only if the vCPU hasn't
519 	 * run, i.e. only if userspace is potentially changing CPUID features.
520 	 */
521 	swap(vcpu->arch.cpuid_entries, e2);
522 	swap(vcpu->arch.cpuid_nent, nent);
523 
524 	memcpy(vcpu_caps, vcpu->arch.cpu_caps, sizeof(vcpu_caps));
525 	BUILD_BUG_ON(sizeof(vcpu_caps) != sizeof(vcpu->arch.cpu_caps));
526 
527 	/*
528 	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
529 	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
530 	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
531 	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
532 	 * the core vCPU model on the fly. It would've been better to forbid any
533 	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
534 	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
535 	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
536 	 * whether the supplied CPUID data is equal to what's already set.
537 	 */
538 	if (kvm_vcpu_has_run(vcpu)) {
539 		r = kvm_cpuid_check_equal(vcpu, e2, nent);
540 		if (r)
541 			goto err;
542 		goto success;
543 	}
544 
545 #ifdef CONFIG_KVM_HYPERV
546 	if (kvm_cpuid_has_hyperv(vcpu)) {
547 		r = kvm_hv_vcpu_init(vcpu);
548 		if (r)
549 			goto err;
550 	}
551 #endif
552 
553 	r = kvm_check_cpuid(vcpu);
554 	if (r)
555 		goto err;
556 
557 #ifdef CONFIG_KVM_XEN
558 	vcpu->arch.xen.cpuid = kvm_get_hypervisor_cpuid(vcpu, XEN_SIGNATURE);
559 #endif
560 	kvm_vcpu_after_set_cpuid(vcpu);
561 
562 success:
563 	kvfree(e2);
564 	return 0;
565 
566 err:
567 	memcpy(vcpu->arch.cpu_caps, vcpu_caps, sizeof(vcpu_caps));
568 	swap(vcpu->arch.cpuid_entries, e2);
569 	swap(vcpu->arch.cpuid_nent, nent);
570 	return r;
571 }
572 
573 /* 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)574 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
575 			     struct kvm_cpuid *cpuid,
576 			     struct kvm_cpuid_entry __user *entries)
577 {
578 	int r, i;
579 	struct kvm_cpuid_entry *e = NULL;
580 	struct kvm_cpuid_entry2 *e2 = NULL;
581 
582 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
583 		return -E2BIG;
584 
585 	if (cpuid->nent) {
586 		e = vmemdup_array_user(entries, cpuid->nent, sizeof(*e));
587 		if (IS_ERR(e))
588 			return PTR_ERR(e);
589 
590 		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
591 		if (!e2) {
592 			r = -ENOMEM;
593 			goto out_free_cpuid;
594 		}
595 	}
596 	for (i = 0; i < cpuid->nent; i++) {
597 		e2[i].function = e[i].function;
598 		e2[i].eax = e[i].eax;
599 		e2[i].ebx = e[i].ebx;
600 		e2[i].ecx = e[i].ecx;
601 		e2[i].edx = e[i].edx;
602 		e2[i].index = 0;
603 		e2[i].flags = 0;
604 		e2[i].padding[0] = 0;
605 		e2[i].padding[1] = 0;
606 		e2[i].padding[2] = 0;
607 	}
608 
609 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
610 	if (r)
611 		kvfree(e2);
612 
613 out_free_cpuid:
614 	kvfree(e);
615 
616 	return r;
617 }
618 
kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu * vcpu,struct kvm_cpuid2 * cpuid,struct kvm_cpuid_entry2 __user * entries)619 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
620 			      struct kvm_cpuid2 *cpuid,
621 			      struct kvm_cpuid_entry2 __user *entries)
622 {
623 	struct kvm_cpuid_entry2 *e2 = NULL;
624 	int r;
625 
626 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
627 		return -E2BIG;
628 
629 	if (cpuid->nent) {
630 		e2 = vmemdup_array_user(entries, cpuid->nent, sizeof(*e2));
631 		if (IS_ERR(e2))
632 			return PTR_ERR(e2);
633 	}
634 
635 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
636 	if (r)
637 		kvfree(e2);
638 
639 	return r;
640 }
641 
kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu * vcpu,struct kvm_cpuid2 * cpuid,struct kvm_cpuid_entry2 __user * entries)642 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
643 			      struct kvm_cpuid2 *cpuid,
644 			      struct kvm_cpuid_entry2 __user *entries)
645 {
646 	if (cpuid->nent < vcpu->arch.cpuid_nent)
647 		return -E2BIG;
648 
649 	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
650 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
651 		return -EFAULT;
652 
653 	cpuid->nent = vcpu->arch.cpuid_nent;
654 	return 0;
655 }
656 
raw_cpuid_get(struct cpuid_reg cpuid)657 static __always_inline u32 raw_cpuid_get(struct cpuid_reg cpuid)
658 {
659 	struct kvm_cpuid_entry2 entry;
660 	u32 base;
661 
662 	/*
663 	 * KVM only supports features defined by Intel (0x0), AMD (0x80000000),
664 	 * and Centaur (0xc0000000).  WARN if a feature for new vendor base is
665 	 * defined, as this and other code would need to be updated.
666 	 */
667 	base = cpuid.function & 0xffff0000;
668 	if (WARN_ON_ONCE(base && base != 0x80000000 && base != 0xc0000000))
669 		return 0;
670 
671 	if (cpuid_eax(base) < cpuid.function)
672 		return 0;
673 
674 	cpuid_count(cpuid.function, cpuid.index,
675 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
676 
677 	return *__cpuid_entry_get_reg(&entry, cpuid.reg);
678 }
679 
680 /*
681  * For kernel-defined leafs, mask KVM's supported feature set with the kernel's
682  * capabilities as well as raw CPUID.  For KVM-defined leafs, consult only raw
683  * CPUID, as KVM is the one and only authority (in the kernel).
684  */
685 #define kvm_cpu_cap_init(leaf, feature_initializers...)			\
686 do {									\
687 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);	\
688 	const u32 __maybe_unused kvm_cpu_cap_init_in_progress = leaf;	\
689 	const u32 *kernel_cpu_caps = boot_cpu_data.x86_capability;	\
690 	u32 kvm_cpu_cap_passthrough = 0;				\
691 	u32 kvm_cpu_cap_synthesized = 0;				\
692 	u32 kvm_cpu_cap_emulated = 0;					\
693 	u32 kvm_cpu_cap_features = 0;					\
694 									\
695 	feature_initializers						\
696 									\
697 	kvm_cpu_caps[leaf] = kvm_cpu_cap_features;			\
698 									\
699 	if (leaf < NCAPINTS)						\
700 		kvm_cpu_caps[leaf] &= kernel_cpu_caps[leaf];		\
701 									\
702 	kvm_cpu_caps[leaf] |= kvm_cpu_cap_passthrough;			\
703 	kvm_cpu_caps[leaf] &= (raw_cpuid_get(cpuid) |			\
704 			       kvm_cpu_cap_synthesized);		\
705 	kvm_cpu_caps[leaf] |= kvm_cpu_cap_emulated;			\
706 } while (0)
707 
708 /*
709  * Assert that the feature bit being declared, e.g. via F(), is in the CPUID
710  * word that's being initialized.  Exempt 0x8000_0001.EDX usage of 0x1.EDX
711  * features, as AMD duplicated many 0x1.EDX features into 0x8000_0001.EDX.
712  */
713 #define KVM_VALIDATE_CPU_CAP_USAGE(name)				\
714 do {									\
715 	u32 __leaf = __feature_leaf(X86_FEATURE_##name);		\
716 									\
717 	BUILD_BUG_ON(__leaf != kvm_cpu_cap_init_in_progress);		\
718 } while (0)
719 
720 #define F(name)							\
721 ({								\
722 	KVM_VALIDATE_CPU_CAP_USAGE(name);			\
723 	kvm_cpu_cap_features |= feature_bit(name);		\
724 })
725 
726 /* Scattered Flag - For features that are scattered by cpufeatures.h. */
727 #define SCATTERED_F(name)					\
728 ({								\
729 	BUILD_BUG_ON(X86_FEATURE_##name >= MAX_CPU_FEATURES);	\
730 	KVM_VALIDATE_CPU_CAP_USAGE(name);			\
731 	if (boot_cpu_has(X86_FEATURE_##name))			\
732 		F(name);					\
733 })
734 
735 /* Features that KVM supports only on 64-bit kernels. */
736 #define X86_64_F(name)						\
737 ({								\
738 	KVM_VALIDATE_CPU_CAP_USAGE(name);			\
739 	if (IS_ENABLED(CONFIG_X86_64))				\
740 		F(name);					\
741 })
742 
743 /*
744  * Emulated Feature - For features that KVM emulates in software irrespective
745  * of host CPU/kernel support.
746  */
747 #define EMULATED_F(name)					\
748 ({								\
749 	kvm_cpu_cap_emulated |= feature_bit(name);		\
750 	F(name);						\
751 })
752 
753 /*
754  * Synthesized Feature - For features that are synthesized into boot_cpu_data,
755  * i.e. may not be present in the raw CPUID, but can still be advertised to
756  * userspace.  Primarily used for mitigation related feature flags.
757  */
758 #define SYNTHESIZED_F(name)					\
759 ({								\
760 	kvm_cpu_cap_synthesized |= feature_bit(name);		\
761 	F(name);						\
762 })
763 
764 /*
765  * Passthrough Feature - For features that KVM supports based purely on raw
766  * hardware CPUID, i.e. that KVM virtualizes even if the host kernel doesn't
767  * use the feature.  Simply force set the feature in KVM's capabilities, raw
768  * CPUID support will be factored in by kvm_cpu_cap_mask().
769  */
770 #define PASSTHROUGH_F(name)					\
771 ({								\
772 	kvm_cpu_cap_passthrough |= feature_bit(name);		\
773 	F(name);						\
774 })
775 
776 /*
777  * Aliased Features - For features in 0x8000_0001.EDX that are duplicates of
778  * identical 0x1.EDX features, and thus are aliased from 0x1 to 0x8000_0001.
779  */
780 #define ALIASED_1_EDX_F(name)							\
781 ({										\
782 	BUILD_BUG_ON(__feature_leaf(X86_FEATURE_##name) != CPUID_1_EDX);	\
783 	BUILD_BUG_ON(kvm_cpu_cap_init_in_progress != CPUID_8000_0001_EDX);	\
784 	kvm_cpu_cap_features |= feature_bit(name);				\
785 })
786 
787 /*
788  * Vendor Features - For features that KVM supports, but are added in later
789  * because they require additional vendor enabling.
790  */
791 #define VENDOR_F(name)						\
792 ({								\
793 	KVM_VALIDATE_CPU_CAP_USAGE(name);			\
794 })
795 
796 /*
797  * Runtime Features - For features that KVM dynamically sets/clears at runtime,
798  * e.g. when CR4 changes, but which are never advertised to userspace.
799  */
800 #define RUNTIME_F(name)						\
801 ({								\
802 	KVM_VALIDATE_CPU_CAP_USAGE(name);			\
803 })
804 
805 /*
806  * Undefine the MSR bit macro to avoid token concatenation issues when
807  * processing X86_FEATURE_SPEC_CTRL_SSBD.
808  */
809 #undef SPEC_CTRL_SSBD
810 
811 /* DS is defined by ptrace-abi.h on 32-bit builds. */
812 #undef DS
813 
kvm_set_cpu_caps(void)814 void kvm_set_cpu_caps(void)
815 {
816 	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
817 
818 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
819 		     sizeof(boot_cpu_data.x86_capability));
820 
821 	kvm_cpu_cap_init(CPUID_1_ECX,
822 		F(XMM3),
823 		F(PCLMULQDQ),
824 		VENDOR_F(DTES64),
825 		/*
826 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
827 		 * advertised to guests via CPUID!  MWAIT is also technically a
828 		 * runtime flag thanks to IA32_MISC_ENABLES; mark it as such so
829 		 * that KVM is aware that it's a known, unadvertised flag.
830 		 */
831 		RUNTIME_F(MWAIT),
832 		/* DS-CPL */
833 		VENDOR_F(VMX),
834 		/* SMX, EST */
835 		/* TM2 */
836 		F(SSSE3),
837 		/* CNXT-ID */
838 		/* Reserved */
839 		F(FMA),
840 		F(CX16),
841 		/* xTPR Update */
842 		F(PDCM),
843 		F(PCID),
844 		/* Reserved, DCA */
845 		F(XMM4_1),
846 		F(XMM4_2),
847 		EMULATED_F(X2APIC),
848 		F(MOVBE),
849 		F(POPCNT),
850 		EMULATED_F(TSC_DEADLINE_TIMER),
851 		F(AES),
852 		F(XSAVE),
853 		RUNTIME_F(OSXSAVE),
854 		F(AVX),
855 		F(F16C),
856 		F(RDRAND),
857 		EMULATED_F(HYPERVISOR),
858 	);
859 
860 	kvm_cpu_cap_init(CPUID_1_EDX,
861 		F(FPU),
862 		F(VME),
863 		F(DE),
864 		F(PSE),
865 		F(TSC),
866 		F(MSR),
867 		F(PAE),
868 		F(MCE),
869 		F(CX8),
870 		F(APIC),
871 		/* Reserved */
872 		F(SEP),
873 		F(MTRR),
874 		F(PGE),
875 		F(MCA),
876 		F(CMOV),
877 		F(PAT),
878 		F(PSE36),
879 		/* PSN */
880 		F(CLFLUSH),
881 		/* Reserved */
882 		VENDOR_F(DS),
883 		/* ACPI */
884 		F(MMX),
885 		F(FXSR),
886 		F(XMM),
887 		F(XMM2),
888 		F(SELFSNOOP),
889 		/* HTT, TM, Reserved, PBE */
890 	);
891 
892 	kvm_cpu_cap_init(CPUID_7_0_EBX,
893 		F(FSGSBASE),
894 		EMULATED_F(TSC_ADJUST),
895 		F(SGX),
896 		F(BMI1),
897 		F(HLE),
898 		F(AVX2),
899 		F(FDP_EXCPTN_ONLY),
900 		F(SMEP),
901 		F(BMI2),
902 		F(ERMS),
903 		F(INVPCID),
904 		F(RTM),
905 		F(ZERO_FCS_FDS),
906 		VENDOR_F(MPX),
907 		F(AVX512F),
908 		F(AVX512DQ),
909 		F(RDSEED),
910 		F(ADX),
911 		F(SMAP),
912 		F(AVX512IFMA),
913 		F(CLFLUSHOPT),
914 		F(CLWB),
915 		VENDOR_F(INTEL_PT),
916 		F(AVX512PF),
917 		F(AVX512ER),
918 		F(AVX512CD),
919 		F(SHA_NI),
920 		F(AVX512BW),
921 		F(AVX512VL),
922 	);
923 
924 	kvm_cpu_cap_init(CPUID_7_ECX,
925 		F(AVX512VBMI),
926 		PASSTHROUGH_F(LA57),
927 		F(PKU),
928 		RUNTIME_F(OSPKE),
929 		F(RDPID),
930 		F(AVX512_VPOPCNTDQ),
931 		F(UMIP),
932 		F(AVX512_VBMI2),
933 		F(GFNI),
934 		F(VAES),
935 		F(VPCLMULQDQ),
936 		F(AVX512_VNNI),
937 		F(AVX512_BITALG),
938 		F(CLDEMOTE),
939 		F(MOVDIRI),
940 		F(MOVDIR64B),
941 		VENDOR_F(WAITPKG),
942 		F(SGX_LC),
943 		F(BUS_LOCK_DETECT),
944 	);
945 
946 	/*
947 	 * PKU not yet implemented for shadow paging and requires OSPKE
948 	 * to be set on the host. Clear it if that is not the case
949 	 */
950 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
951 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
952 
953 	kvm_cpu_cap_init(CPUID_7_EDX,
954 		F(AVX512_4VNNIW),
955 		F(AVX512_4FMAPS),
956 		F(SPEC_CTRL),
957 		F(SPEC_CTRL_SSBD),
958 		EMULATED_F(ARCH_CAPABILITIES),
959 		F(INTEL_STIBP),
960 		F(MD_CLEAR),
961 		F(AVX512_VP2INTERSECT),
962 		F(FSRM),
963 		F(SERIALIZE),
964 		F(TSXLDTRK),
965 		F(AVX512_FP16),
966 		F(AMX_TILE),
967 		F(AMX_INT8),
968 		F(AMX_BF16),
969 		F(FLUSH_L1D),
970 	);
971 
972 	if (boot_cpu_has(X86_FEATURE_AMD_IBPB_RET) &&
973 	    boot_cpu_has(X86_FEATURE_AMD_IBPB) &&
974 	    boot_cpu_has(X86_FEATURE_AMD_IBRS))
975 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
976 	if (boot_cpu_has(X86_FEATURE_STIBP))
977 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
978 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
979 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
980 
981 	kvm_cpu_cap_init(CPUID_7_1_EAX,
982 		F(SHA512),
983 		F(SM3),
984 		F(SM4),
985 		F(AVX_VNNI),
986 		F(AVX512_BF16),
987 		F(CMPCCXADD),
988 		F(FZRM),
989 		F(FSRS),
990 		F(FSRC),
991 		F(AMX_FP16),
992 		F(AVX_IFMA),
993 		F(LAM),
994 	);
995 
996 	kvm_cpu_cap_init(CPUID_7_1_EDX,
997 		F(AVX_VNNI_INT8),
998 		F(AVX_NE_CONVERT),
999 		F(AMX_COMPLEX),
1000 		F(AVX_VNNI_INT16),
1001 		F(PREFETCHITI),
1002 		F(AVX10),
1003 	);
1004 
1005 	kvm_cpu_cap_init(CPUID_7_2_EDX,
1006 		F(INTEL_PSFD),
1007 		F(IPRED_CTRL),
1008 		F(RRSBA_CTRL),
1009 		F(DDPD_U),
1010 		F(BHI_CTRL),
1011 		F(MCDT_NO),
1012 	);
1013 
1014 	kvm_cpu_cap_init(CPUID_D_1_EAX,
1015 		F(XSAVEOPT),
1016 		F(XSAVEC),
1017 		F(XGETBV1),
1018 		F(XSAVES),
1019 		X86_64_F(XFD),
1020 	);
1021 
1022 	kvm_cpu_cap_init(CPUID_12_EAX,
1023 		SCATTERED_F(SGX1),
1024 		SCATTERED_F(SGX2),
1025 		SCATTERED_F(SGX_EDECCSSA),
1026 	);
1027 
1028 	kvm_cpu_cap_init(CPUID_24_0_EBX,
1029 		F(AVX10_128),
1030 		F(AVX10_256),
1031 		F(AVX10_512),
1032 	);
1033 
1034 	kvm_cpu_cap_init(CPUID_8000_0001_ECX,
1035 		F(LAHF_LM),
1036 		F(CMP_LEGACY),
1037 		VENDOR_F(SVM),
1038 		/* ExtApicSpace */
1039 		F(CR8_LEGACY),
1040 		F(ABM),
1041 		F(SSE4A),
1042 		F(MISALIGNSSE),
1043 		F(3DNOWPREFETCH),
1044 		F(OSVW),
1045 		/* IBS */
1046 		F(XOP),
1047 		/* SKINIT, WDT, LWP */
1048 		F(FMA4),
1049 		F(TBM),
1050 		F(TOPOEXT),
1051 		VENDOR_F(PERFCTR_CORE),
1052 	);
1053 
1054 	kvm_cpu_cap_init(CPUID_8000_0001_EDX,
1055 		ALIASED_1_EDX_F(FPU),
1056 		ALIASED_1_EDX_F(VME),
1057 		ALIASED_1_EDX_F(DE),
1058 		ALIASED_1_EDX_F(PSE),
1059 		ALIASED_1_EDX_F(TSC),
1060 		ALIASED_1_EDX_F(MSR),
1061 		ALIASED_1_EDX_F(PAE),
1062 		ALIASED_1_EDX_F(MCE),
1063 		ALIASED_1_EDX_F(CX8),
1064 		ALIASED_1_EDX_F(APIC),
1065 		/* Reserved */
1066 		F(SYSCALL),
1067 		ALIASED_1_EDX_F(MTRR),
1068 		ALIASED_1_EDX_F(PGE),
1069 		ALIASED_1_EDX_F(MCA),
1070 		ALIASED_1_EDX_F(CMOV),
1071 		ALIASED_1_EDX_F(PAT),
1072 		ALIASED_1_EDX_F(PSE36),
1073 		/* Reserved */
1074 		F(NX),
1075 		/* Reserved */
1076 		F(MMXEXT),
1077 		ALIASED_1_EDX_F(MMX),
1078 		ALIASED_1_EDX_F(FXSR),
1079 		F(FXSR_OPT),
1080 		X86_64_F(GBPAGES),
1081 		F(RDTSCP),
1082 		/* Reserved */
1083 		X86_64_F(LM),
1084 		F(3DNOWEXT),
1085 		F(3DNOW),
1086 	);
1087 
1088 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
1089 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
1090 
1091 	kvm_cpu_cap_init(CPUID_8000_0007_EDX,
1092 		SCATTERED_F(CONSTANT_TSC),
1093 	);
1094 
1095 	kvm_cpu_cap_init(CPUID_8000_0008_EBX,
1096 		F(CLZERO),
1097 		F(XSAVEERPTR),
1098 		F(WBNOINVD),
1099 		F(AMD_IBPB),
1100 		F(AMD_IBRS),
1101 		F(AMD_SSBD),
1102 		F(VIRT_SSBD),
1103 		F(AMD_SSB_NO),
1104 		F(AMD_STIBP),
1105 		F(AMD_STIBP_ALWAYS_ON),
1106 		F(AMD_PSFD),
1107 		F(AMD_IBPB_RET),
1108 	);
1109 
1110 	/*
1111 	 * AMD has separate bits for each SPEC_CTRL bit.
1112 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
1113 	 * record that in cpufeatures so use them.
1114 	 */
1115 	if (boot_cpu_has(X86_FEATURE_IBPB)) {
1116 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
1117 		if (boot_cpu_has(X86_FEATURE_SPEC_CTRL) &&
1118 		    !boot_cpu_has_bug(X86_BUG_EIBRS_PBRSB))
1119 			kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB_RET);
1120 	}
1121 	if (boot_cpu_has(X86_FEATURE_IBRS))
1122 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
1123 	if (boot_cpu_has(X86_FEATURE_STIBP))
1124 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
1125 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
1126 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
1127 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
1128 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
1129 	/*
1130 	 * The preference is to use SPEC CTRL MSR instead of the
1131 	 * VIRT_SPEC MSR.
1132 	 */
1133 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
1134 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
1135 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
1136 
1137 	/* All SVM features required additional vendor module enabling. */
1138 	kvm_cpu_cap_init(CPUID_8000_000A_EDX,
1139 		VENDOR_F(NPT),
1140 		VENDOR_F(VMCBCLEAN),
1141 		VENDOR_F(FLUSHBYASID),
1142 		VENDOR_F(NRIPS),
1143 		VENDOR_F(TSCRATEMSR),
1144 		VENDOR_F(V_VMSAVE_VMLOAD),
1145 		VENDOR_F(LBRV),
1146 		VENDOR_F(PAUSEFILTER),
1147 		VENDOR_F(PFTHRESHOLD),
1148 		VENDOR_F(VGIF),
1149 		VENDOR_F(VNMI),
1150 		VENDOR_F(SVME_ADDR_CHK),
1151 	);
1152 
1153 	kvm_cpu_cap_init(CPUID_8000_001F_EAX,
1154 		VENDOR_F(SME),
1155 		VENDOR_F(SEV),
1156 		/* VM_PAGE_FLUSH */
1157 		VENDOR_F(SEV_ES),
1158 		F(SME_COHERENT),
1159 	);
1160 
1161 	kvm_cpu_cap_init(CPUID_8000_0021_EAX,
1162 		F(NO_NESTED_DATA_BP),
1163 		/*
1164 		 * Synthesize "LFENCE is serializing" into the AMD-defined entry
1165 		 * in KVM's supported CPUID, i.e. if the feature is reported as
1166 		 * supported by the kernel.  LFENCE_RDTSC was a Linux-defined
1167 		 * synthetic feature long before AMD joined the bandwagon, e.g.
1168 		 * LFENCE is serializing on most CPUs that support SSE2.  On
1169 		 * CPUs that don't support AMD's leaf, ANDing with the raw host
1170 		 * CPUID will drop the flags, and reporting support in AMD's
1171 		 * leaf can make it easier for userspace to detect the feature.
1172 		 */
1173 		SYNTHESIZED_F(LFENCE_RDTSC),
1174 		/* SmmPgCfgLock */
1175 		F(NULL_SEL_CLR_BASE),
1176 		F(AUTOIBRS),
1177 		EMULATED_F(NO_SMM_CTL_MSR),
1178 		/* PrefetchCtlMsr */
1179 		F(WRMSR_XX_BASE_NS),
1180 		SYNTHESIZED_F(SBPB),
1181 		SYNTHESIZED_F(IBPB_BRTYPE),
1182 		SYNTHESIZED_F(SRSO_NO),
1183 		F(SRSO_USER_KERNEL_NO),
1184 	);
1185 
1186 	kvm_cpu_cap_init(CPUID_8000_0022_EAX,
1187 		F(PERFMON_V2),
1188 	);
1189 
1190 	if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
1191 		kvm_cpu_cap_set(X86_FEATURE_NULL_SEL_CLR_BASE);
1192 
1193 	kvm_cpu_cap_init(CPUID_C000_0001_EDX,
1194 		F(XSTORE),
1195 		F(XSTORE_EN),
1196 		F(XCRYPT),
1197 		F(XCRYPT_EN),
1198 		F(ACE2),
1199 		F(ACE2_EN),
1200 		F(PHE),
1201 		F(PHE_EN),
1202 		F(PMM),
1203 		F(PMM_EN),
1204 	);
1205 
1206 	/*
1207 	 * Hide RDTSCP and RDPID if either feature is reported as supported but
1208 	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
1209 	 * should never happen, but the guest will likely crash if RDTSCP or
1210 	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
1211 	 * the past.  For example, the sanity check may fire if this instance of
1212 	 * KVM is running as L1 on top of an older, broken KVM.
1213 	 */
1214 	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
1215 		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
1216 		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
1217 		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
1218 		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
1219 	}
1220 }
1221 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
1222 
1223 #undef F
1224 #undef SCATTERED_F
1225 #undef X86_64_F
1226 #undef EMULATED_F
1227 #undef SYNTHESIZED_F
1228 #undef PASSTHROUGH_F
1229 #undef ALIASED_1_EDX_F
1230 #undef VENDOR_F
1231 #undef RUNTIME_F
1232 
1233 struct kvm_cpuid_array {
1234 	struct kvm_cpuid_entry2 *entries;
1235 	int maxnent;
1236 	int nent;
1237 };
1238 
get_next_cpuid(struct kvm_cpuid_array * array)1239 static struct kvm_cpuid_entry2 *get_next_cpuid(struct kvm_cpuid_array *array)
1240 {
1241 	if (array->nent >= array->maxnent)
1242 		return NULL;
1243 
1244 	return &array->entries[array->nent++];
1245 }
1246 
do_host_cpuid(struct kvm_cpuid_array * array,u32 function,u32 index)1247 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
1248 					      u32 function, u32 index)
1249 {
1250 	struct kvm_cpuid_entry2 *entry = get_next_cpuid(array);
1251 
1252 	if (!entry)
1253 		return NULL;
1254 
1255 	memset(entry, 0, sizeof(*entry));
1256 	entry->function = function;
1257 	entry->index = index;
1258 	switch (function & 0xC0000000) {
1259 	case 0x40000000:
1260 		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
1261 		return entry;
1262 
1263 	case 0x80000000:
1264 		/*
1265 		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
1266 		 * would result in out-of-bounds calls to do_host_cpuid.
1267 		 */
1268 		{
1269 			static int max_cpuid_80000000;
1270 			if (!READ_ONCE(max_cpuid_80000000))
1271 				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
1272 			if (function > READ_ONCE(max_cpuid_80000000))
1273 				return entry;
1274 		}
1275 		break;
1276 
1277 	default:
1278 		break;
1279 	}
1280 
1281 	cpuid_count(entry->function, entry->index,
1282 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
1283 
1284 	if (cpuid_function_is_indexed(function))
1285 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
1286 
1287 	return entry;
1288 }
1289 
cpuid_func_emulated(struct kvm_cpuid_entry2 * entry,u32 func,bool include_partially_emulated)1290 static int cpuid_func_emulated(struct kvm_cpuid_entry2 *entry, u32 func,
1291 			       bool include_partially_emulated)
1292 {
1293 	memset(entry, 0, sizeof(*entry));
1294 
1295 	entry->function = func;
1296 	entry->index = 0;
1297 	entry->flags = 0;
1298 
1299 	switch (func) {
1300 	case 0:
1301 		entry->eax = 7;
1302 		return 1;
1303 	case 1:
1304 		entry->ecx = feature_bit(MOVBE);
1305 		/*
1306 		 * KVM allows userspace to enumerate MONITOR+MWAIT support to
1307 		 * the guest, but the MWAIT feature flag is never advertised
1308 		 * to userspace because MONITOR+MWAIT aren't virtualized by
1309 		 * hardware, can't be faithfully emulated in software (KVM
1310 		 * emulates them as NOPs), and allowing the guest to execute
1311 		 * them natively requires enabling a per-VM capability.
1312 		 */
1313 		if (include_partially_emulated)
1314 			entry->ecx |= feature_bit(MWAIT);
1315 		return 1;
1316 	case 7:
1317 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
1318 		entry->eax = 0;
1319 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
1320 			entry->ecx = feature_bit(RDPID);
1321 		return 1;
1322 	default:
1323 		return 0;
1324 	}
1325 }
1326 
__do_cpuid_func_emulated(struct kvm_cpuid_array * array,u32 func)1327 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
1328 {
1329 	if (array->nent >= array->maxnent)
1330 		return -E2BIG;
1331 
1332 	array->nent += cpuid_func_emulated(&array->entries[array->nent], func, false);
1333 	return 0;
1334 }
1335 
__do_cpuid_func(struct kvm_cpuid_array * array,u32 function)1336 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
1337 {
1338 	struct kvm_cpuid_entry2 *entry;
1339 	int r, i, max_idx;
1340 
1341 	/* all calls to cpuid_count() should be made on the same cpu */
1342 	get_cpu();
1343 
1344 	r = -E2BIG;
1345 
1346 	entry = do_host_cpuid(array, function, 0);
1347 	if (!entry)
1348 		goto out;
1349 
1350 	switch (function) {
1351 	case 0:
1352 		/* Limited to the highest leaf implemented in KVM. */
1353 		entry->eax = min(entry->eax, 0x24U);
1354 		break;
1355 	case 1:
1356 		cpuid_entry_override(entry, CPUID_1_EDX);
1357 		cpuid_entry_override(entry, CPUID_1_ECX);
1358 		break;
1359 	case 2:
1360 		/*
1361 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
1362 		 * CPUID(function=2, index=0) may return different results each
1363 		 * time, with the least-significant byte in EAX enumerating the
1364 		 * number of times software should do CPUID(2, 0).
1365 		 *
1366 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
1367 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
1368 		 * return 01H. Software should ignore this value and not
1369 		 * interpret it as an informational descriptor", while AMD's
1370 		 * APM states that CPUID(2) is reserved.
1371 		 *
1372 		 * WARN if a frankenstein CPU that supports virtualization and
1373 		 * a stateful CPUID.0x2 is encountered.
1374 		 */
1375 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
1376 		break;
1377 	/* functions 4 and 0x8000001d have additional index. */
1378 	case 4:
1379 	case 0x8000001d:
1380 		/*
1381 		 * Read entries until the cache type in the previous entry is
1382 		 * zero, i.e. indicates an invalid entry.
1383 		 */
1384 		for (i = 1; entry->eax & 0x1f; ++i) {
1385 			entry = do_host_cpuid(array, function, i);
1386 			if (!entry)
1387 				goto out;
1388 		}
1389 		break;
1390 	case 6: /* Thermal management */
1391 		entry->eax = 0x4; /* allow ARAT */
1392 		entry->ebx = 0;
1393 		entry->ecx = 0;
1394 		entry->edx = 0;
1395 		break;
1396 	/* function 7 has additional index. */
1397 	case 7:
1398 		max_idx = entry->eax = min(entry->eax, 2u);
1399 		cpuid_entry_override(entry, CPUID_7_0_EBX);
1400 		cpuid_entry_override(entry, CPUID_7_ECX);
1401 		cpuid_entry_override(entry, CPUID_7_EDX);
1402 
1403 		/* KVM only supports up to 0x7.2, capped above via min(). */
1404 		if (max_idx >= 1) {
1405 			entry = do_host_cpuid(array, function, 1);
1406 			if (!entry)
1407 				goto out;
1408 
1409 			cpuid_entry_override(entry, CPUID_7_1_EAX);
1410 			cpuid_entry_override(entry, CPUID_7_1_EDX);
1411 			entry->ebx = 0;
1412 			entry->ecx = 0;
1413 		}
1414 		if (max_idx >= 2) {
1415 			entry = do_host_cpuid(array, function, 2);
1416 			if (!entry)
1417 				goto out;
1418 
1419 			cpuid_entry_override(entry, CPUID_7_2_EDX);
1420 			entry->ecx = 0;
1421 			entry->ebx = 0;
1422 			entry->eax = 0;
1423 		}
1424 		break;
1425 	case 0xa: { /* Architectural Performance Monitoring */
1426 		union cpuid10_eax eax;
1427 		union cpuid10_edx edx;
1428 
1429 		if (!enable_pmu || !static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
1430 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1431 			break;
1432 		}
1433 
1434 		eax.split.version_id = kvm_pmu_cap.version;
1435 		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
1436 		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
1437 		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
1438 		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
1439 		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
1440 
1441 		if (kvm_pmu_cap.version)
1442 			edx.split.anythread_deprecated = 1;
1443 		edx.split.reserved1 = 0;
1444 		edx.split.reserved2 = 0;
1445 
1446 		entry->eax = eax.full;
1447 		entry->ebx = kvm_pmu_cap.events_mask;
1448 		entry->ecx = 0;
1449 		entry->edx = edx.full;
1450 		break;
1451 	}
1452 	case 0x1f:
1453 	case 0xb:
1454 		/*
1455 		 * No topology; a valid topology is indicated by the presence
1456 		 * of subleaf 1.
1457 		 */
1458 		entry->eax = entry->ebx = entry->ecx = 0;
1459 		break;
1460 	case 0xd: {
1461 		u64 permitted_xcr0 = kvm_get_filtered_xcr0();
1462 		u64 permitted_xss = kvm_caps.supported_xss;
1463 
1464 		entry->eax &= permitted_xcr0;
1465 		entry->ebx = xstate_required_size(permitted_xcr0, false);
1466 		entry->ecx = entry->ebx;
1467 		entry->edx &= permitted_xcr0 >> 32;
1468 		if (!permitted_xcr0)
1469 			break;
1470 
1471 		entry = do_host_cpuid(array, function, 1);
1472 		if (!entry)
1473 			goto out;
1474 
1475 		cpuid_entry_override(entry, CPUID_D_1_EAX);
1476 		if (entry->eax & (feature_bit(XSAVES) | feature_bit(XSAVEC)))
1477 			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
1478 							  true);
1479 		else {
1480 			WARN_ON_ONCE(permitted_xss != 0);
1481 			entry->ebx = 0;
1482 		}
1483 		entry->ecx &= permitted_xss;
1484 		entry->edx &= permitted_xss >> 32;
1485 
1486 		for (i = 2; i < 64; ++i) {
1487 			bool s_state;
1488 			if (permitted_xcr0 & BIT_ULL(i))
1489 				s_state = false;
1490 			else if (permitted_xss & BIT_ULL(i))
1491 				s_state = true;
1492 			else
1493 				continue;
1494 
1495 			entry = do_host_cpuid(array, function, i);
1496 			if (!entry)
1497 				goto out;
1498 
1499 			/*
1500 			 * The supported check above should have filtered out
1501 			 * invalid sub-leafs.  Only valid sub-leafs should
1502 			 * reach this point, and they should have a non-zero
1503 			 * save state size.  Furthermore, check whether the
1504 			 * processor agrees with permitted_xcr0/permitted_xss
1505 			 * on whether this is an XCR0- or IA32_XSS-managed area.
1506 			 */
1507 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
1508 				--array->nent;
1509 				continue;
1510 			}
1511 
1512 			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1513 				entry->ecx &= ~BIT_ULL(2);
1514 			entry->edx = 0;
1515 		}
1516 		break;
1517 	}
1518 	case 0x12:
1519 		/* Intel SGX */
1520 		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1521 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1522 			break;
1523 		}
1524 
1525 		/*
1526 		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1527 		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1528 		 * are restricted by kernel and KVM capabilities (like most
1529 		 * feature flags), while enclave size is unrestricted.
1530 		 */
1531 		cpuid_entry_override(entry, CPUID_12_EAX);
1532 		entry->ebx &= SGX_MISC_EXINFO;
1533 
1534 		entry = do_host_cpuid(array, function, 1);
1535 		if (!entry)
1536 			goto out;
1537 
1538 		/*
1539 		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1540 		 * feature flags.  Advertise all supported flags, including
1541 		 * privileged attributes that require explicit opt-in from
1542 		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1543 		 * expected to derive it from supported XCR0.
1544 		 */
1545 		entry->eax &= SGX_ATTR_PRIV_MASK | SGX_ATTR_UNPRIV_MASK;
1546 		entry->ebx &= 0;
1547 		break;
1548 	/* Intel PT */
1549 	case 0x14:
1550 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1551 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1552 			break;
1553 		}
1554 
1555 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1556 			if (!do_host_cpuid(array, function, i))
1557 				goto out;
1558 		}
1559 		break;
1560 	/* Intel AMX TILE */
1561 	case 0x1d:
1562 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1563 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1564 			break;
1565 		}
1566 
1567 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1568 			if (!do_host_cpuid(array, function, i))
1569 				goto out;
1570 		}
1571 		break;
1572 	case 0x1e: /* TMUL information */
1573 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1574 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1575 			break;
1576 		}
1577 		break;
1578 	case 0x24: {
1579 		u8 avx10_version;
1580 
1581 		if (!kvm_cpu_cap_has(X86_FEATURE_AVX10)) {
1582 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1583 			break;
1584 		}
1585 
1586 		/*
1587 		 * The AVX10 version is encoded in EBX[7:0].  Note, the version
1588 		 * is guaranteed to be >=1 if AVX10 is supported.  Note #2, the
1589 		 * version needs to be captured before overriding EBX features!
1590 		 */
1591 		avx10_version = min_t(u8, entry->ebx & 0xff, 1);
1592 		cpuid_entry_override(entry, CPUID_24_0_EBX);
1593 		entry->ebx |= avx10_version;
1594 
1595 		entry->eax = 0;
1596 		entry->ecx = 0;
1597 		entry->edx = 0;
1598 		break;
1599 	}
1600 	case KVM_CPUID_SIGNATURE: {
1601 		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1602 		entry->eax = KVM_CPUID_FEATURES;
1603 		entry->ebx = sigptr[0];
1604 		entry->ecx = sigptr[1];
1605 		entry->edx = sigptr[2];
1606 		break;
1607 	}
1608 	case KVM_CPUID_FEATURES:
1609 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1610 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1611 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1612 			     (1 << KVM_FEATURE_ASYNC_PF) |
1613 			     (1 << KVM_FEATURE_PV_EOI) |
1614 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1615 			     (1 << KVM_FEATURE_PV_UNHALT) |
1616 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1617 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1618 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1619 			     (1 << KVM_FEATURE_POLL_CONTROL) |
1620 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1621 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1622 
1623 		if (sched_info_on())
1624 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1625 
1626 		entry->ebx = 0;
1627 		entry->ecx = 0;
1628 		entry->edx = 0;
1629 		break;
1630 	case 0x80000000:
1631 		entry->eax = min(entry->eax, 0x80000022);
1632 		/*
1633 		 * Serializing LFENCE is reported in a multitude of ways, and
1634 		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1635 		 * userspace by providing the CPUID leaf ourselves.
1636 		 *
1637 		 * However, only do it if the host has CPUID leaf 0x8000001d.
1638 		 * QEMU thinks that it can query the host blindly for that
1639 		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1640 		 * above.  The processor merrily returns values from the
1641 		 * highest Intel leaf which QEMU tries to use as the guest's
1642 		 * 0x8000001d.  Even worse, this can result in an infinite
1643 		 * loop if said highest leaf has no subleaves indexed by ECX.
1644 		 */
1645 		if (entry->eax >= 0x8000001d &&
1646 		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1647 		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1648 			entry->eax = max(entry->eax, 0x80000021);
1649 		break;
1650 	case 0x80000001:
1651 		entry->ebx &= ~GENMASK(27, 16);
1652 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1653 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1654 		break;
1655 	case 0x80000005:
1656 		/*  Pass host L1 cache and TLB info. */
1657 		break;
1658 	case 0x80000006:
1659 		/* Drop reserved bits, pass host L2 cache and TLB info. */
1660 		entry->edx &= ~GENMASK(17, 16);
1661 		break;
1662 	case 0x80000007: /* Advanced power management */
1663 		cpuid_entry_override(entry, CPUID_8000_0007_EDX);
1664 
1665 		/* mask against host */
1666 		entry->edx &= boot_cpu_data.x86_power;
1667 		entry->eax = entry->ebx = entry->ecx = 0;
1668 		break;
1669 	case 0x80000008: {
1670 		/*
1671 		 * GuestPhysAddrSize (EAX[23:16]) is intended for software
1672 		 * use.
1673 		 *
1674 		 * KVM's ABI is to report the effective MAXPHYADDR for the
1675 		 * guest in PhysAddrSize (phys_as), and the maximum
1676 		 * *addressable* GPA in GuestPhysAddrSize (g_phys_as).
1677 		 *
1678 		 * GuestPhysAddrSize is valid if and only if TDP is enabled,
1679 		 * in which case the max GPA that can be addressed by KVM may
1680 		 * be less than the max GPA that can be legally generated by
1681 		 * the guest, e.g. if MAXPHYADDR>48 but the CPU doesn't
1682 		 * support 5-level TDP.
1683 		 */
1684 		unsigned int virt_as = max((entry->eax >> 8) & 0xff, 48U);
1685 		unsigned int phys_as, g_phys_as;
1686 
1687 		/*
1688 		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1689 		 * the guest operates in the same PA space as the host, i.e.
1690 		 * reductions in MAXPHYADDR for memory encryption affect shadow
1691 		 * paging, too.
1692 		 *
1693 		 * If TDP is enabled, use the raw bare metal MAXPHYADDR as
1694 		 * reductions to the HPAs do not affect GPAs.  The max
1695 		 * addressable GPA is the same as the max effective GPA, except
1696 		 * that it's capped at 48 bits if 5-level TDP isn't supported
1697 		 * (hardware processes bits 51:48 only when walking the fifth
1698 		 * level page table).
1699 		 */
1700 		if (!tdp_enabled) {
1701 			phys_as = boot_cpu_data.x86_phys_bits;
1702 			g_phys_as = 0;
1703 		} else {
1704 			phys_as = entry->eax & 0xff;
1705 			g_phys_as = phys_as;
1706 			if (kvm_mmu_get_max_tdp_level() < 5)
1707 				g_phys_as = min(g_phys_as, 48);
1708 		}
1709 
1710 		entry->eax = phys_as | (virt_as << 8) | (g_phys_as << 16);
1711 		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
1712 		entry->edx = 0;
1713 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1714 		break;
1715 	}
1716 	case 0x8000000A:
1717 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1718 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1719 			break;
1720 		}
1721 		entry->eax = 1; /* SVM revision 1 */
1722 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1723 				   ASID emulation to nested SVM */
1724 		entry->ecx = 0; /* Reserved */
1725 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1726 		break;
1727 	case 0x80000019:
1728 		entry->ecx = entry->edx = 0;
1729 		break;
1730 	case 0x8000001a:
1731 		entry->eax &= GENMASK(2, 0);
1732 		entry->ebx = entry->ecx = entry->edx = 0;
1733 		break;
1734 	case 0x8000001e:
1735 		/* Do not return host topology information.  */
1736 		entry->eax = entry->ebx = entry->ecx = 0;
1737 		entry->edx = 0; /* reserved */
1738 		break;
1739 	case 0x8000001F:
1740 		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1741 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1742 		} else {
1743 			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1744 			/* Clear NumVMPL since KVM does not support VMPL.  */
1745 			entry->ebx &= ~GENMASK(31, 12);
1746 			/*
1747 			 * Enumerate '0' for "PA bits reduction", the adjusted
1748 			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1749 			 */
1750 			entry->ebx &= ~GENMASK(11, 6);
1751 		}
1752 		break;
1753 	case 0x80000020:
1754 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1755 		break;
1756 	case 0x80000021:
1757 		entry->ebx = entry->ecx = entry->edx = 0;
1758 		cpuid_entry_override(entry, CPUID_8000_0021_EAX);
1759 		break;
1760 	/* AMD Extended Performance Monitoring and Debug */
1761 	case 0x80000022: {
1762 		union cpuid_0x80000022_ebx ebx;
1763 
1764 		entry->ecx = entry->edx = 0;
1765 		if (!enable_pmu || !kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2)) {
1766 			entry->eax = entry->ebx = 0;
1767 			break;
1768 		}
1769 
1770 		cpuid_entry_override(entry, CPUID_8000_0022_EAX);
1771 
1772 		if (kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2))
1773 			ebx.split.num_core_pmc = kvm_pmu_cap.num_counters_gp;
1774 		else if (kvm_cpu_cap_has(X86_FEATURE_PERFCTR_CORE))
1775 			ebx.split.num_core_pmc = AMD64_NUM_COUNTERS_CORE;
1776 		else
1777 			ebx.split.num_core_pmc = AMD64_NUM_COUNTERS;
1778 
1779 		entry->ebx = ebx.full;
1780 		break;
1781 	}
1782 	/*Add support for Centaur's CPUID instruction*/
1783 	case 0xC0000000:
1784 		/*Just support up to 0xC0000004 now*/
1785 		entry->eax = min(entry->eax, 0xC0000004);
1786 		break;
1787 	case 0xC0000001:
1788 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1789 		break;
1790 	case 3: /* Processor serial number */
1791 	case 5: /* MONITOR/MWAIT */
1792 	case 0xC0000002:
1793 	case 0xC0000003:
1794 	case 0xC0000004:
1795 	default:
1796 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1797 		break;
1798 	}
1799 
1800 	r = 0;
1801 
1802 out:
1803 	put_cpu();
1804 
1805 	return r;
1806 }
1807 
do_cpuid_func(struct kvm_cpuid_array * array,u32 func,unsigned int type)1808 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1809 			 unsigned int type)
1810 {
1811 	if (type == KVM_GET_EMULATED_CPUID)
1812 		return __do_cpuid_func_emulated(array, func);
1813 
1814 	return __do_cpuid_func(array, func);
1815 }
1816 
1817 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
1818 
get_cpuid_func(struct kvm_cpuid_array * array,u32 func,unsigned int type)1819 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1820 			  unsigned int type)
1821 {
1822 	u32 limit;
1823 	int r;
1824 
1825 	if (func == CENTAUR_CPUID_SIGNATURE &&
1826 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1827 		return 0;
1828 
1829 	r = do_cpuid_func(array, func, type);
1830 	if (r)
1831 		return r;
1832 
1833 	limit = array->entries[array->nent - 1].eax;
1834 	for (func = func + 1; func <= limit; ++func) {
1835 		r = do_cpuid_func(array, func, type);
1836 		if (r)
1837 			break;
1838 	}
1839 
1840 	return r;
1841 }
1842 
sanity_check_entries(struct kvm_cpuid_entry2 __user * entries,__u32 num_entries,unsigned int ioctl_type)1843 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1844 				 __u32 num_entries, unsigned int ioctl_type)
1845 {
1846 	int i;
1847 	__u32 pad[3];
1848 
1849 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1850 		return false;
1851 
1852 	/*
1853 	 * We want to make sure that ->padding is being passed clean from
1854 	 * userspace in case we want to use it for something in the future.
1855 	 *
1856 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1857 	 * have to give ourselves satisfied only with the emulated side. /me
1858 	 * sheds a tear.
1859 	 */
1860 	for (i = 0; i < num_entries; i++) {
1861 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1862 			return true;
1863 
1864 		if (pad[0] || pad[1] || pad[2])
1865 			return true;
1866 	}
1867 	return false;
1868 }
1869 
kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 * cpuid,struct kvm_cpuid_entry2 __user * entries,unsigned int type)1870 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1871 			    struct kvm_cpuid_entry2 __user *entries,
1872 			    unsigned int type)
1873 {
1874 	static const u32 funcs[] = {
1875 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1876 	};
1877 
1878 	struct kvm_cpuid_array array = {
1879 		.nent = 0,
1880 	};
1881 	int r, i;
1882 
1883 	if (cpuid->nent < 1)
1884 		return -E2BIG;
1885 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1886 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1887 
1888 	if (sanity_check_entries(entries, cpuid->nent, type))
1889 		return -EINVAL;
1890 
1891 	array.entries = kvcalloc(cpuid->nent, sizeof(struct kvm_cpuid_entry2), GFP_KERNEL);
1892 	if (!array.entries)
1893 		return -ENOMEM;
1894 
1895 	array.maxnent = cpuid->nent;
1896 
1897 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1898 		r = get_cpuid_func(&array, funcs[i], type);
1899 		if (r)
1900 			goto out_free;
1901 	}
1902 	cpuid->nent = array.nent;
1903 
1904 	if (copy_to_user(entries, array.entries,
1905 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1906 		r = -EFAULT;
1907 
1908 out_free:
1909 	kvfree(array.entries);
1910 	return r;
1911 }
1912 
1913 /*
1914  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1915  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1916  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1917  * range.  Centaur/VIA follows Intel semantics.
1918  *
1919  * A leaf is considered out-of-range if its function is higher than the maximum
1920  * supported leaf of its associated class or if its associated class does not
1921  * exist.
1922  *
1923  * There are three primary classes to be considered, with their respective
1924  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1925  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1926  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1927  *
1928  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1929  *  - Hypervisor: 0x40000000 - 0x4fffffff
1930  *  - Extended:   0x80000000 - 0xbfffffff
1931  *  - Centaur:    0xc0000000 - 0xcfffffff
1932  *
1933  * The Hypervisor class is further subdivided into sub-classes that each act as
1934  * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1935  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1936  * CPUID sub-classes are:
1937  *
1938  *  - HyperV:     0x40000000 - 0x400000ff
1939  *  - KVM:        0x40000100 - 0x400001ff
1940  */
1941 static struct kvm_cpuid_entry2 *
get_out_of_range_cpuid_entry(struct kvm_vcpu * vcpu,u32 * fn_ptr,u32 index)1942 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1943 {
1944 	struct kvm_cpuid_entry2 *basic, *class;
1945 	u32 function = *fn_ptr;
1946 
1947 	basic = kvm_find_cpuid_entry(vcpu, 0);
1948 	if (!basic)
1949 		return NULL;
1950 
1951 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1952 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1953 		return NULL;
1954 
1955 	if (function >= 0x40000000 && function <= 0x4fffffff)
1956 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1957 	else if (function >= 0xc0000000)
1958 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1959 	else
1960 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1961 
1962 	if (class && function <= class->eax)
1963 		return NULL;
1964 
1965 	/*
1966 	 * Leaf specific adjustments are also applied when redirecting to the
1967 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1968 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1969 	 * needs to be pulled from CPUID.0xb.1.
1970 	 */
1971 	*fn_ptr = basic->eax;
1972 
1973 	/*
1974 	 * The class does not exist or the requested function is out of range;
1975 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1976 	 * the original requested leaf is observed!
1977 	 */
1978 	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1979 }
1980 
kvm_cpuid(struct kvm_vcpu * vcpu,u32 * eax,u32 * ebx,u32 * ecx,u32 * edx,bool exact_only)1981 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1982 	       u32 *ecx, u32 *edx, bool exact_only)
1983 {
1984 	u32 orig_function = *eax, function = *eax, index = *ecx;
1985 	struct kvm_cpuid_entry2 *entry;
1986 	bool exact, used_max_basic = false;
1987 
1988 	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1989 	exact = !!entry;
1990 
1991 	if (!entry && !exact_only) {
1992 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1993 		used_max_basic = !!entry;
1994 	}
1995 
1996 	if (entry) {
1997 		*eax = entry->eax;
1998 		*ebx = entry->ebx;
1999 		*ecx = entry->ecx;
2000 		*edx = entry->edx;
2001 		if (function == 7 && index == 0) {
2002 			u64 data;
2003 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
2004 			    (data & TSX_CTRL_CPUID_CLEAR))
2005 				*ebx &= ~(feature_bit(RTM) | feature_bit(HLE));
2006 		} else if (function == 0x80000007) {
2007 			if (kvm_hv_invtsc_suppressed(vcpu))
2008 				*edx &= ~feature_bit(CONSTANT_TSC);
2009 		}
2010 	} else {
2011 		*eax = *ebx = *ecx = *edx = 0;
2012 		/*
2013 		 * When leaf 0BH or 1FH is defined, CL is pass-through
2014 		 * and EDX is always the x2APIC ID, even for undefined
2015 		 * subleaves. Index 1 will exist iff the leaf is
2016 		 * implemented, so we pass through CL iff leaf 1
2017 		 * exists. EDX can be copied from any existing index.
2018 		 */
2019 		if (function == 0xb || function == 0x1f) {
2020 			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
2021 			if (entry) {
2022 				*ecx = index & 0xff;
2023 				*edx = entry->edx;
2024 			}
2025 		}
2026 	}
2027 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
2028 			used_max_basic);
2029 	return exact;
2030 }
2031 EXPORT_SYMBOL_GPL(kvm_cpuid);
2032 
kvm_emulate_cpuid(struct kvm_vcpu * vcpu)2033 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
2034 {
2035 	u32 eax, ebx, ecx, edx;
2036 
2037 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
2038 		return 1;
2039 
2040 	eax = kvm_rax_read(vcpu);
2041 	ecx = kvm_rcx_read(vcpu);
2042 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
2043 	kvm_rax_write(vcpu, eax);
2044 	kvm_rbx_write(vcpu, ebx);
2045 	kvm_rcx_write(vcpu, ecx);
2046 	kvm_rdx_write(vcpu, edx);
2047 	return kvm_skip_emulated_instruction(vcpu);
2048 }
2049 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
2050