xref: /openbmc/linux/arch/x86/kvm/cpuid.c (revision b8a94bfb)
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 
12 #include <linux/kvm_host.h>
13 #include <linux/export.h>
14 #include <linux/vmalloc.h>
15 #include <linux/uaccess.h>
16 #include <linux/sched/stat.h>
17 
18 #include <asm/processor.h>
19 #include <asm/user.h>
20 #include <asm/fpu/xstate.h>
21 #include <asm/sgx.h>
22 #include <asm/cpuid.h>
23 #include "cpuid.h"
24 #include "lapic.h"
25 #include "mmu.h"
26 #include "trace.h"
27 #include "pmu.h"
28 
29 /*
30  * Unlike "struct cpuinfo_x86.x86_capability", kvm_cpu_caps doesn't need to be
31  * aligned to sizeof(unsigned long) because it's not accessed via bitops.
32  */
33 u32 kvm_cpu_caps[NR_KVM_CPU_CAPS] __read_mostly;
34 EXPORT_SYMBOL_GPL(kvm_cpu_caps);
35 
36 u32 xstate_required_size(u64 xstate_bv, bool compacted)
37 {
38 	int feature_bit = 0;
39 	u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
40 
41 	xstate_bv &= XFEATURE_MASK_EXTEND;
42 	while (xstate_bv) {
43 		if (xstate_bv & 0x1) {
44 		        u32 eax, ebx, ecx, edx, offset;
45 		        cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx);
46 			/* ECX[1]: 64B alignment in compacted form */
47 			if (compacted)
48 				offset = (ecx & 0x2) ? ALIGN(ret, 64) : ret;
49 			else
50 				offset = ebx;
51 			ret = max(ret, offset + eax);
52 		}
53 
54 		xstate_bv >>= 1;
55 		feature_bit++;
56 	}
57 
58 	return ret;
59 }
60 
61 /*
62  * This one is tied to SSB in the user API, and not
63  * visible in /proc/cpuinfo.
64  */
65 #define KVM_X86_FEATURE_PSFD		(13*32+28) /* Predictive Store Forwarding Disable */
66 
67 #define F feature_bit
68 #define SF(name) (boot_cpu_has(X86_FEATURE_##name) ? F(name) : 0)
69 
70 /*
71  * Magic value used by KVM when querying userspace-provided CPUID entries and
72  * doesn't care about the CPIUD index because the index of the function in
73  * question is not significant.  Note, this magic value must have at least one
74  * bit set in bits[63:32] and must be consumed as a u64 by cpuid_entry2_find()
75  * to avoid false positives when processing guest CPUID input.
76  */
77 #define KVM_CPUID_INDEX_NOT_SIGNIFICANT -1ull
78 
79 static inline struct kvm_cpuid_entry2 *cpuid_entry2_find(
80 	struct kvm_cpuid_entry2 *entries, int nent, u32 function, u64 index)
81 {
82 	struct kvm_cpuid_entry2 *e;
83 	int i;
84 
85 	for (i = 0; i < nent; i++) {
86 		e = &entries[i];
87 
88 		if (e->function != function)
89 			continue;
90 
91 		/*
92 		 * If the index isn't significant, use the first entry with a
93 		 * matching function.  It's userspace's responsibilty to not
94 		 * provide "duplicate" entries in all cases.
95 		 */
96 		if (!(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) || e->index == index)
97 			return e;
98 
99 
100 		/*
101 		 * Similarly, use the first matching entry if KVM is doing a
102 		 * lookup (as opposed to emulating CPUID) for a function that's
103 		 * architecturally defined as not having a significant index.
104 		 */
105 		if (index == KVM_CPUID_INDEX_NOT_SIGNIFICANT) {
106 			/*
107 			 * Direct lookups from KVM should not diverge from what
108 			 * KVM defines internally (the architectural behavior).
109 			 */
110 			WARN_ON_ONCE(cpuid_function_is_indexed(function));
111 			return e;
112 		}
113 	}
114 
115 	return NULL;
116 }
117 
118 static int kvm_check_cpuid(struct kvm_vcpu *vcpu,
119 			   struct kvm_cpuid_entry2 *entries,
120 			   int nent)
121 {
122 	struct kvm_cpuid_entry2 *best;
123 	u64 xfeatures;
124 
125 	/*
126 	 * The existing code assumes virtual address is 48-bit or 57-bit in the
127 	 * canonical address checks; exit if it is ever changed.
128 	 */
129 	best = cpuid_entry2_find(entries, nent, 0x80000008,
130 				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
131 	if (best) {
132 		int vaddr_bits = (best->eax & 0xff00) >> 8;
133 
134 		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
135 			return -EINVAL;
136 	}
137 
138 	/*
139 	 * Exposing dynamic xfeatures to the guest requires additional
140 	 * enabling in the FPU, e.g. to expand the guest XSAVE state size.
141 	 */
142 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
143 	if (!best)
144 		return 0;
145 
146 	xfeatures = best->eax | ((u64)best->edx << 32);
147 	xfeatures &= XFEATURE_MASK_USER_DYNAMIC;
148 	if (!xfeatures)
149 		return 0;
150 
151 	return fpu_enable_guest_xfd_features(&vcpu->arch.guest_fpu, xfeatures);
152 }
153 
154 /* Check whether the supplied CPUID data is equal to what is already set for the vCPU. */
155 static int kvm_cpuid_check_equal(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
156 				 int nent)
157 {
158 	struct kvm_cpuid_entry2 *orig;
159 	int i;
160 
161 	if (nent != vcpu->arch.cpuid_nent)
162 		return -EINVAL;
163 
164 	for (i = 0; i < nent; i++) {
165 		orig = &vcpu->arch.cpuid_entries[i];
166 		if (e2[i].function != orig->function ||
167 		    e2[i].index != orig->index ||
168 		    e2[i].flags != orig->flags ||
169 		    e2[i].eax != orig->eax || e2[i].ebx != orig->ebx ||
170 		    e2[i].ecx != orig->ecx || e2[i].edx != orig->edx)
171 			return -EINVAL;
172 	}
173 
174 	return 0;
175 }
176 
177 static void kvm_update_kvm_cpuid_base(struct kvm_vcpu *vcpu)
178 {
179 	u32 function;
180 	struct kvm_cpuid_entry2 *entry;
181 
182 	vcpu->arch.kvm_cpuid_base = 0;
183 
184 	for_each_possible_hypervisor_cpuid_base(function) {
185 		entry = kvm_find_cpuid_entry(vcpu, function);
186 
187 		if (entry) {
188 			u32 signature[3];
189 
190 			signature[0] = entry->ebx;
191 			signature[1] = entry->ecx;
192 			signature[2] = entry->edx;
193 
194 			BUILD_BUG_ON(sizeof(signature) > sizeof(KVM_SIGNATURE));
195 			if (!memcmp(signature, KVM_SIGNATURE, sizeof(signature))) {
196 				vcpu->arch.kvm_cpuid_base = function;
197 				break;
198 			}
199 		}
200 	}
201 }
202 
203 static struct kvm_cpuid_entry2 *__kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu,
204 					      struct kvm_cpuid_entry2 *entries, int nent)
205 {
206 	u32 base = vcpu->arch.kvm_cpuid_base;
207 
208 	if (!base)
209 		return NULL;
210 
211 	return cpuid_entry2_find(entries, nent, base | KVM_CPUID_FEATURES,
212 				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
213 }
214 
215 static struct kvm_cpuid_entry2 *kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu)
216 {
217 	return __kvm_find_kvm_cpuid_features(vcpu, vcpu->arch.cpuid_entries,
218 					     vcpu->arch.cpuid_nent);
219 }
220 
221 void kvm_update_pv_runtime(struct kvm_vcpu *vcpu)
222 {
223 	struct kvm_cpuid_entry2 *best = kvm_find_kvm_cpuid_features(vcpu);
224 
225 	/*
226 	 * save the feature bitmap to avoid cpuid lookup for every PV
227 	 * operation
228 	 */
229 	if (best)
230 		vcpu->arch.pv_cpuid.features = best->eax;
231 }
232 
233 /*
234  * Calculate guest's supported XCR0 taking into account guest CPUID data and
235  * KVM's supported XCR0 (comprised of host's XCR0 and KVM_SUPPORTED_XCR0).
236  */
237 static u64 cpuid_get_supported_xcr0(struct kvm_cpuid_entry2 *entries, int nent)
238 {
239 	struct kvm_cpuid_entry2 *best;
240 
241 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
242 	if (!best)
243 		return 0;
244 
245 	return (best->eax | ((u64)best->edx << 32)) & kvm_caps.supported_xcr0;
246 }
247 
248 static void __kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *entries,
249 				       int nent)
250 {
251 	struct kvm_cpuid_entry2 *best;
252 	u64 guest_supported_xcr0 = cpuid_get_supported_xcr0(entries, nent);
253 
254 	best = cpuid_entry2_find(entries, nent, 1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
255 	if (best) {
256 		/* Update OSXSAVE bit */
257 		if (boot_cpu_has(X86_FEATURE_XSAVE))
258 			cpuid_entry_change(best, X86_FEATURE_OSXSAVE,
259 				   kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE));
260 
261 		cpuid_entry_change(best, X86_FEATURE_APIC,
262 			   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
263 	}
264 
265 	best = cpuid_entry2_find(entries, nent, 7, 0);
266 	if (best && boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7)
267 		cpuid_entry_change(best, X86_FEATURE_OSPKE,
268 				   kvm_read_cr4_bits(vcpu, X86_CR4_PKE));
269 
270 	best = cpuid_entry2_find(entries, nent, 0xD, 0);
271 	if (best)
272 		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
273 
274 	best = cpuid_entry2_find(entries, nent, 0xD, 1);
275 	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
276 		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
277 		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
278 
279 	best = __kvm_find_kvm_cpuid_features(vcpu, entries, nent);
280 	if (kvm_hlt_in_guest(vcpu->kvm) && best &&
281 		(best->eax & (1 << KVM_FEATURE_PV_UNHALT)))
282 		best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
283 
284 	if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
285 		best = cpuid_entry2_find(entries, nent, 0x1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
286 		if (best)
287 			cpuid_entry_change(best, X86_FEATURE_MWAIT,
288 					   vcpu->arch.ia32_misc_enable_msr &
289 					   MSR_IA32_MISC_ENABLE_MWAIT);
290 	}
291 
292 	/*
293 	 * Bits 127:0 of the allowed SECS.ATTRIBUTES (CPUID.0x12.0x1) enumerate
294 	 * the supported XSAVE Feature Request Mask (XFRM), i.e. the enclave's
295 	 * requested XCR0 value.  The enclave's XFRM must be a subset of XCRO
296 	 * at the time of EENTER, thus adjust the allowed XFRM by the guest's
297 	 * supported XCR0.  Similar to XCR0 handling, FP and SSE are forced to
298 	 * '1' even on CPUs that don't support XSAVE.
299 	 */
300 	best = cpuid_entry2_find(entries, nent, 0x12, 0x1);
301 	if (best) {
302 		best->ecx &= guest_supported_xcr0 & 0xffffffff;
303 		best->edx &= guest_supported_xcr0 >> 32;
304 		best->ecx |= XFEATURE_MASK_FPSSE;
305 	}
306 }
307 
308 void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu)
309 {
310 	__kvm_update_cpuid_runtime(vcpu, vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
311 }
312 EXPORT_SYMBOL_GPL(kvm_update_cpuid_runtime);
313 
314 static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
315 {
316 	struct kvm_lapic *apic = vcpu->arch.apic;
317 	struct kvm_cpuid_entry2 *best;
318 
319 	best = kvm_find_cpuid_entry(vcpu, 1);
320 	if (best && apic) {
321 		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
322 			apic->lapic_timer.timer_mode_mask = 3 << 17;
323 		else
324 			apic->lapic_timer.timer_mode_mask = 1 << 17;
325 
326 		kvm_apic_set_version(vcpu);
327 	}
328 
329 	vcpu->arch.guest_supported_xcr0 =
330 		cpuid_get_supported_xcr0(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
331 
332 	/*
333 	 * FP+SSE can always be saved/restored via KVM_{G,S}ET_XSAVE, even if
334 	 * XSAVE/XCRO are not exposed to the guest, and even if XSAVE isn't
335 	 * supported by the host.
336 	 */
337 	vcpu->arch.guest_fpu.fpstate->user_xfeatures = vcpu->arch.guest_supported_xcr0 |
338 						       XFEATURE_MASK_FPSSE;
339 
340 	kvm_update_pv_runtime(vcpu);
341 
342 	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
343 	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
344 
345 	kvm_pmu_refresh(vcpu);
346 	vcpu->arch.cr4_guest_rsvd_bits =
347 	    __cr4_reserved_bits(guest_cpuid_has, vcpu);
348 
349 	kvm_hv_set_cpuid(vcpu);
350 
351 	/* Invoke the vendor callback only after the above state is updated. */
352 	static_call(kvm_x86_vcpu_after_set_cpuid)(vcpu);
353 
354 	/*
355 	 * Except for the MMU, which needs to do its thing any vendor specific
356 	 * adjustments to the reserved GPA bits.
357 	 */
358 	kvm_mmu_after_set_cpuid(vcpu);
359 }
360 
361 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
362 {
363 	struct kvm_cpuid_entry2 *best;
364 
365 	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
366 	if (!best || best->eax < 0x80000008)
367 		goto not_found;
368 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
369 	if (best)
370 		return best->eax & 0xff;
371 not_found:
372 	return 36;
373 }
374 
375 /*
376  * This "raw" version returns the reserved GPA bits without any adjustments for
377  * encryption technologies that usurp bits.  The raw mask should be used if and
378  * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
379  */
380 u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
381 {
382 	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
383 }
384 
385 static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
386                         int nent)
387 {
388 	int r;
389 
390 	__kvm_update_cpuid_runtime(vcpu, e2, nent);
391 
392 	/*
393 	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
394 	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
395 	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
396 	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
397 	 * the core vCPU model on the fly. It would've been better to forbid any
398 	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
399 	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
400 	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
401 	 * whether the supplied CPUID data is equal to what's already set.
402 	 */
403 	if (vcpu->arch.last_vmentry_cpu != -1) {
404 		r = kvm_cpuid_check_equal(vcpu, e2, nent);
405 		if (r)
406 			return r;
407 
408 		kvfree(e2);
409 		return 0;
410 	}
411 
412 	r = kvm_check_cpuid(vcpu, e2, nent);
413 	if (r)
414 		return r;
415 
416 	kvfree(vcpu->arch.cpuid_entries);
417 	vcpu->arch.cpuid_entries = e2;
418 	vcpu->arch.cpuid_nent = nent;
419 
420 	kvm_update_kvm_cpuid_base(vcpu);
421 	kvm_vcpu_after_set_cpuid(vcpu);
422 
423 	return 0;
424 }
425 
426 /* when an old userspace process fills a new kernel module */
427 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
428 			     struct kvm_cpuid *cpuid,
429 			     struct kvm_cpuid_entry __user *entries)
430 {
431 	int r, i;
432 	struct kvm_cpuid_entry *e = NULL;
433 	struct kvm_cpuid_entry2 *e2 = NULL;
434 
435 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
436 		return -E2BIG;
437 
438 	if (cpuid->nent) {
439 		e = vmemdup_user(entries, array_size(sizeof(*e), cpuid->nent));
440 		if (IS_ERR(e))
441 			return PTR_ERR(e);
442 
443 		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
444 		if (!e2) {
445 			r = -ENOMEM;
446 			goto out_free_cpuid;
447 		}
448 	}
449 	for (i = 0; i < cpuid->nent; i++) {
450 		e2[i].function = e[i].function;
451 		e2[i].eax = e[i].eax;
452 		e2[i].ebx = e[i].ebx;
453 		e2[i].ecx = e[i].ecx;
454 		e2[i].edx = e[i].edx;
455 		e2[i].index = 0;
456 		e2[i].flags = 0;
457 		e2[i].padding[0] = 0;
458 		e2[i].padding[1] = 0;
459 		e2[i].padding[2] = 0;
460 	}
461 
462 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
463 	if (r)
464 		kvfree(e2);
465 
466 out_free_cpuid:
467 	kvfree(e);
468 
469 	return r;
470 }
471 
472 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
473 			      struct kvm_cpuid2 *cpuid,
474 			      struct kvm_cpuid_entry2 __user *entries)
475 {
476 	struct kvm_cpuid_entry2 *e2 = NULL;
477 	int r;
478 
479 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
480 		return -E2BIG;
481 
482 	if (cpuid->nent) {
483 		e2 = vmemdup_user(entries, array_size(sizeof(*e2), cpuid->nent));
484 		if (IS_ERR(e2))
485 			return PTR_ERR(e2);
486 	}
487 
488 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
489 	if (r)
490 		kvfree(e2);
491 
492 	return r;
493 }
494 
495 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
496 			      struct kvm_cpuid2 *cpuid,
497 			      struct kvm_cpuid_entry2 __user *entries)
498 {
499 	int r;
500 
501 	r = -E2BIG;
502 	if (cpuid->nent < vcpu->arch.cpuid_nent)
503 		goto out;
504 	r = -EFAULT;
505 	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
506 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
507 		goto out;
508 	return 0;
509 
510 out:
511 	cpuid->nent = vcpu->arch.cpuid_nent;
512 	return r;
513 }
514 
515 /* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */
516 static __always_inline void __kvm_cpu_cap_mask(unsigned int leaf)
517 {
518 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
519 	struct kvm_cpuid_entry2 entry;
520 
521 	reverse_cpuid_check(leaf);
522 
523 	cpuid_count(cpuid.function, cpuid.index,
524 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
525 
526 	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
527 }
528 
529 static __always_inline
530 void kvm_cpu_cap_init_scattered(enum kvm_only_cpuid_leafs leaf, u32 mask)
531 {
532 	/* Use kvm_cpu_cap_mask for non-scattered leafs. */
533 	BUILD_BUG_ON(leaf < NCAPINTS);
534 
535 	kvm_cpu_caps[leaf] = mask;
536 
537 	__kvm_cpu_cap_mask(leaf);
538 }
539 
540 static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
541 {
542 	/* Use kvm_cpu_cap_init_scattered for scattered leafs. */
543 	BUILD_BUG_ON(leaf >= NCAPINTS);
544 
545 	kvm_cpu_caps[leaf] &= mask;
546 
547 	__kvm_cpu_cap_mask(leaf);
548 }
549 
550 void kvm_set_cpu_caps(void)
551 {
552 #ifdef CONFIG_X86_64
553 	unsigned int f_gbpages = F(GBPAGES);
554 	unsigned int f_lm = F(LM);
555 	unsigned int f_xfd = F(XFD);
556 #else
557 	unsigned int f_gbpages = 0;
558 	unsigned int f_lm = 0;
559 	unsigned int f_xfd = 0;
560 #endif
561 	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
562 
563 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
564 		     sizeof(boot_cpu_data.x86_capability));
565 
566 	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
567 	       sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)));
568 
569 	kvm_cpu_cap_mask(CPUID_1_ECX,
570 		/*
571 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
572 		 * advertised to guests via CPUID!
573 		 */
574 		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
575 		0 /* DS-CPL, VMX, SMX, EST */ |
576 		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
577 		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
578 		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
579 		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
580 		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
581 		F(F16C) | F(RDRAND)
582 	);
583 	/* KVM emulates x2apic in software irrespective of host support. */
584 	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
585 
586 	kvm_cpu_cap_mask(CPUID_1_EDX,
587 		F(FPU) | F(VME) | F(DE) | F(PSE) |
588 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
589 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
590 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
591 		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
592 		0 /* Reserved, DS, ACPI */ | F(MMX) |
593 		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
594 		0 /* HTT, TM, Reserved, PBE */
595 	);
596 
597 	kvm_cpu_cap_mask(CPUID_7_0_EBX,
598 		F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) |
599 		F(FDP_EXCPTN_ONLY) | F(SMEP) | F(BMI2) | F(ERMS) | F(INVPCID) |
600 		F(RTM) | F(ZERO_FCS_FDS) | 0 /*MPX*/ | F(AVX512F) |
601 		F(AVX512DQ) | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) |
602 		F(CLFLUSHOPT) | F(CLWB) | 0 /*INTEL_PT*/ | F(AVX512PF) |
603 		F(AVX512ER) | F(AVX512CD) | F(SHA_NI) | F(AVX512BW) |
604 		F(AVX512VL));
605 
606 	kvm_cpu_cap_mask(CPUID_7_ECX,
607 		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
608 		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
609 		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
610 		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/ |
611 		F(SGX_LC) | F(BUS_LOCK_DETECT)
612 	);
613 	/* Set LA57 based on hardware capability. */
614 	if (cpuid_ecx(7) & F(LA57))
615 		kvm_cpu_cap_set(X86_FEATURE_LA57);
616 
617 	/*
618 	 * PKU not yet implemented for shadow paging and requires OSPKE
619 	 * to be set on the host. Clear it if that is not the case
620 	 */
621 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
622 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
623 
624 	kvm_cpu_cap_mask(CPUID_7_EDX,
625 		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
626 		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
627 		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
628 		F(SERIALIZE) | F(TSXLDTRK) | F(AVX512_FP16) |
629 		F(AMX_TILE) | F(AMX_INT8) | F(AMX_BF16)
630 	);
631 
632 	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
633 	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
634 	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
635 
636 	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
637 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
638 	if (boot_cpu_has(X86_FEATURE_STIBP))
639 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
640 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
641 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
642 
643 	kvm_cpu_cap_mask(CPUID_7_1_EAX,
644 		F(AVX_VNNI) | F(AVX512_BF16)
645 	);
646 
647 	kvm_cpu_cap_mask(CPUID_D_1_EAX,
648 		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES) | f_xfd
649 	);
650 
651 	kvm_cpu_cap_init_scattered(CPUID_12_EAX,
652 		SF(SGX1) | SF(SGX2)
653 	);
654 
655 	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
656 		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
657 		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
658 		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
659 		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
660 		F(TOPOEXT) | 0 /* PERFCTR_CORE */
661 	);
662 
663 	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
664 		F(FPU) | F(VME) | F(DE) | F(PSE) |
665 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
666 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
667 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
668 		F(PAT) | F(PSE36) | 0 /* Reserved */ |
669 		F(NX) | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
670 		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
671 		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
672 	);
673 
674 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
675 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
676 
677 	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
678 		F(CLZERO) | F(XSAVEERPTR) |
679 		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
680 		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON) |
681 		__feature_bit(KVM_X86_FEATURE_PSFD)
682 	);
683 
684 	/*
685 	 * AMD has separate bits for each SPEC_CTRL bit.
686 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
687 	 * record that in cpufeatures so use them.
688 	 */
689 	if (boot_cpu_has(X86_FEATURE_IBPB))
690 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
691 	if (boot_cpu_has(X86_FEATURE_IBRS))
692 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
693 	if (boot_cpu_has(X86_FEATURE_STIBP))
694 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
695 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
696 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
697 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
698 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
699 	/*
700 	 * The preference is to use SPEC CTRL MSR instead of the
701 	 * VIRT_SPEC MSR.
702 	 */
703 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
704 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
705 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
706 
707 	/*
708 	 * Hide all SVM features by default, SVM will set the cap bits for
709 	 * features it emulates and/or exposes for L1.
710 	 */
711 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
712 
713 	kvm_cpu_cap_mask(CPUID_8000_001F_EAX,
714 		0 /* SME */ | F(SEV) | 0 /* VM_PAGE_FLUSH */ | F(SEV_ES) |
715 		F(SME_COHERENT));
716 
717 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
718 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
719 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
720 		F(PMM) | F(PMM_EN)
721 	);
722 
723 	/*
724 	 * Hide RDTSCP and RDPID if either feature is reported as supported but
725 	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
726 	 * should never happen, but the guest will likely crash if RDTSCP or
727 	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
728 	 * the past.  For example, the sanity check may fire if this instance of
729 	 * KVM is running as L1 on top of an older, broken KVM.
730 	 */
731 	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
732 		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
733 		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
734 		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
735 		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
736 	}
737 }
738 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
739 
740 struct kvm_cpuid_array {
741 	struct kvm_cpuid_entry2 *entries;
742 	int maxnent;
743 	int nent;
744 };
745 
746 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
747 					      u32 function, u32 index)
748 {
749 	struct kvm_cpuid_entry2 *entry;
750 
751 	if (array->nent >= array->maxnent)
752 		return NULL;
753 
754 	entry = &array->entries[array->nent++];
755 
756 	memset(entry, 0, sizeof(*entry));
757 	entry->function = function;
758 	entry->index = index;
759 	switch (function & 0xC0000000) {
760 	case 0x40000000:
761 		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
762 		return entry;
763 
764 	case 0x80000000:
765 		/*
766 		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
767 		 * would result in out-of-bounds calls to do_host_cpuid.
768 		 */
769 		{
770 			static int max_cpuid_80000000;
771 			if (!READ_ONCE(max_cpuid_80000000))
772 				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
773 			if (function > READ_ONCE(max_cpuid_80000000))
774 				return entry;
775 		}
776 		break;
777 
778 	default:
779 		break;
780 	}
781 
782 	cpuid_count(entry->function, entry->index,
783 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
784 
785 	if (cpuid_function_is_indexed(function))
786 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
787 
788 	return entry;
789 }
790 
791 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
792 {
793 	struct kvm_cpuid_entry2 *entry;
794 
795 	if (array->nent >= array->maxnent)
796 		return -E2BIG;
797 
798 	entry = &array->entries[array->nent];
799 	entry->function = func;
800 	entry->index = 0;
801 	entry->flags = 0;
802 
803 	switch (func) {
804 	case 0:
805 		entry->eax = 7;
806 		++array->nent;
807 		break;
808 	case 1:
809 		entry->ecx = F(MOVBE);
810 		++array->nent;
811 		break;
812 	case 7:
813 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
814 		entry->eax = 0;
815 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
816 			entry->ecx = F(RDPID);
817 		++array->nent;
818 		break;
819 	default:
820 		break;
821 	}
822 
823 	return 0;
824 }
825 
826 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
827 {
828 	struct kvm_cpuid_entry2 *entry;
829 	int r, i, max_idx;
830 
831 	/* all calls to cpuid_count() should be made on the same cpu */
832 	get_cpu();
833 
834 	r = -E2BIG;
835 
836 	entry = do_host_cpuid(array, function, 0);
837 	if (!entry)
838 		goto out;
839 
840 	switch (function) {
841 	case 0:
842 		/* Limited to the highest leaf implemented in KVM. */
843 		entry->eax = min(entry->eax, 0x1fU);
844 		break;
845 	case 1:
846 		cpuid_entry_override(entry, CPUID_1_EDX);
847 		cpuid_entry_override(entry, CPUID_1_ECX);
848 		break;
849 	case 2:
850 		/*
851 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
852 		 * CPUID(function=2, index=0) may return different results each
853 		 * time, with the least-significant byte in EAX enumerating the
854 		 * number of times software should do CPUID(2, 0).
855 		 *
856 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
857 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
858 		 * return 01H. Software should ignore this value and not
859 		 * interpret it as an informational descriptor", while AMD's
860 		 * APM states that CPUID(2) is reserved.
861 		 *
862 		 * WARN if a frankenstein CPU that supports virtualization and
863 		 * a stateful CPUID.0x2 is encountered.
864 		 */
865 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
866 		break;
867 	/* functions 4 and 0x8000001d have additional index. */
868 	case 4:
869 	case 0x8000001d:
870 		/*
871 		 * Read entries until the cache type in the previous entry is
872 		 * zero, i.e. indicates an invalid entry.
873 		 */
874 		for (i = 1; entry->eax & 0x1f; ++i) {
875 			entry = do_host_cpuid(array, function, i);
876 			if (!entry)
877 				goto out;
878 		}
879 		break;
880 	case 6: /* Thermal management */
881 		entry->eax = 0x4; /* allow ARAT */
882 		entry->ebx = 0;
883 		entry->ecx = 0;
884 		entry->edx = 0;
885 		break;
886 	/* function 7 has additional index. */
887 	case 7:
888 		entry->eax = min(entry->eax, 1u);
889 		cpuid_entry_override(entry, CPUID_7_0_EBX);
890 		cpuid_entry_override(entry, CPUID_7_ECX);
891 		cpuid_entry_override(entry, CPUID_7_EDX);
892 
893 		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
894 		if (entry->eax == 1) {
895 			entry = do_host_cpuid(array, function, 1);
896 			if (!entry)
897 				goto out;
898 
899 			cpuid_entry_override(entry, CPUID_7_1_EAX);
900 			entry->ebx = 0;
901 			entry->ecx = 0;
902 			entry->edx = 0;
903 		}
904 		break;
905 	case 9:
906 		break;
907 	case 0xa: { /* Architectural Performance Monitoring */
908 		union cpuid10_eax eax;
909 		union cpuid10_edx edx;
910 
911 		if (!static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
912 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
913 			break;
914 		}
915 
916 		eax.split.version_id = kvm_pmu_cap.version;
917 		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
918 		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
919 		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
920 		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
921 		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
922 
923 		if (kvm_pmu_cap.version)
924 			edx.split.anythread_deprecated = 1;
925 		edx.split.reserved1 = 0;
926 		edx.split.reserved2 = 0;
927 
928 		entry->eax = eax.full;
929 		entry->ebx = kvm_pmu_cap.events_mask;
930 		entry->ecx = 0;
931 		entry->edx = edx.full;
932 		break;
933 	}
934 	/*
935 	 * Per Intel's SDM, the 0x1f is a superset of 0xb,
936 	 * thus they can be handled by common code.
937 	 */
938 	case 0x1f:
939 	case 0xb:
940 		/*
941 		 * Populate entries until the level type (ECX[15:8]) of the
942 		 * previous entry is zero.  Note, CPUID EAX.{0x1f,0xb}.0 is
943 		 * the starting entry, filled by the primary do_host_cpuid().
944 		 */
945 		for (i = 1; entry->ecx & 0xff00; ++i) {
946 			entry = do_host_cpuid(array, function, i);
947 			if (!entry)
948 				goto out;
949 		}
950 		break;
951 	case 0xd: {
952 		u64 permitted_xcr0 = kvm_caps.supported_xcr0 & xstate_get_guest_group_perm();
953 		u64 permitted_xss = kvm_caps.supported_xss;
954 
955 		entry->eax &= permitted_xcr0;
956 		entry->ebx = xstate_required_size(permitted_xcr0, false);
957 		entry->ecx = entry->ebx;
958 		entry->edx &= permitted_xcr0 >> 32;
959 		if (!permitted_xcr0)
960 			break;
961 
962 		entry = do_host_cpuid(array, function, 1);
963 		if (!entry)
964 			goto out;
965 
966 		cpuid_entry_override(entry, CPUID_D_1_EAX);
967 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
968 			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
969 							  true);
970 		else {
971 			WARN_ON_ONCE(permitted_xss != 0);
972 			entry->ebx = 0;
973 		}
974 		entry->ecx &= permitted_xss;
975 		entry->edx &= permitted_xss >> 32;
976 
977 		for (i = 2; i < 64; ++i) {
978 			bool s_state;
979 			if (permitted_xcr0 & BIT_ULL(i))
980 				s_state = false;
981 			else if (permitted_xss & BIT_ULL(i))
982 				s_state = true;
983 			else
984 				continue;
985 
986 			entry = do_host_cpuid(array, function, i);
987 			if (!entry)
988 				goto out;
989 
990 			/*
991 			 * The supported check above should have filtered out
992 			 * invalid sub-leafs.  Only valid sub-leafs should
993 			 * reach this point, and they should have a non-zero
994 			 * save state size.  Furthermore, check whether the
995 			 * processor agrees with permitted_xcr0/permitted_xss
996 			 * on whether this is an XCR0- or IA32_XSS-managed area.
997 			 */
998 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
999 				--array->nent;
1000 				continue;
1001 			}
1002 
1003 			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1004 				entry->ecx &= ~BIT_ULL(2);
1005 			entry->edx = 0;
1006 		}
1007 		break;
1008 	}
1009 	case 0x12:
1010 		/* Intel SGX */
1011 		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1012 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1013 			break;
1014 		}
1015 
1016 		/*
1017 		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1018 		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1019 		 * are restricted by kernel and KVM capabilities (like most
1020 		 * feature flags), while enclave size is unrestricted.
1021 		 */
1022 		cpuid_entry_override(entry, CPUID_12_EAX);
1023 		entry->ebx &= SGX_MISC_EXINFO;
1024 
1025 		entry = do_host_cpuid(array, function, 1);
1026 		if (!entry)
1027 			goto out;
1028 
1029 		/*
1030 		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1031 		 * feature flags.  Advertise all supported flags, including
1032 		 * privileged attributes that require explicit opt-in from
1033 		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1034 		 * expected to derive it from supported XCR0.
1035 		 */
1036 		entry->eax &= SGX_ATTR_DEBUG | SGX_ATTR_MODE64BIT |
1037 			      SGX_ATTR_PROVISIONKEY | SGX_ATTR_EINITTOKENKEY |
1038 			      SGX_ATTR_KSS;
1039 		entry->ebx &= 0;
1040 		break;
1041 	/* Intel PT */
1042 	case 0x14:
1043 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1044 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1045 			break;
1046 		}
1047 
1048 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1049 			if (!do_host_cpuid(array, function, i))
1050 				goto out;
1051 		}
1052 		break;
1053 	/* Intel AMX TILE */
1054 	case 0x1d:
1055 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1056 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1057 			break;
1058 		}
1059 
1060 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1061 			if (!do_host_cpuid(array, function, i))
1062 				goto out;
1063 		}
1064 		break;
1065 	case 0x1e: /* TMUL information */
1066 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1067 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1068 			break;
1069 		}
1070 		break;
1071 	case KVM_CPUID_SIGNATURE: {
1072 		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1073 		entry->eax = KVM_CPUID_FEATURES;
1074 		entry->ebx = sigptr[0];
1075 		entry->ecx = sigptr[1];
1076 		entry->edx = sigptr[2];
1077 		break;
1078 	}
1079 	case KVM_CPUID_FEATURES:
1080 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1081 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1082 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1083 			     (1 << KVM_FEATURE_ASYNC_PF) |
1084 			     (1 << KVM_FEATURE_PV_EOI) |
1085 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1086 			     (1 << KVM_FEATURE_PV_UNHALT) |
1087 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1088 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1089 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1090 			     (1 << KVM_FEATURE_POLL_CONTROL) |
1091 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1092 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1093 
1094 		if (sched_info_on())
1095 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1096 
1097 		entry->ebx = 0;
1098 		entry->ecx = 0;
1099 		entry->edx = 0;
1100 		break;
1101 	case 0x80000000:
1102 		entry->eax = min(entry->eax, 0x80000021);
1103 		/*
1104 		 * Serializing LFENCE is reported in a multitude of ways, and
1105 		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1106 		 * userspace by providing the CPUID leaf ourselves.
1107 		 *
1108 		 * However, only do it if the host has CPUID leaf 0x8000001d.
1109 		 * QEMU thinks that it can query the host blindly for that
1110 		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1111 		 * above.  The processor merrily returns values from the
1112 		 * highest Intel leaf which QEMU tries to use as the guest's
1113 		 * 0x8000001d.  Even worse, this can result in an infinite
1114 		 * loop if said highest leaf has no subleaves indexed by ECX.
1115 		 */
1116 		if (entry->eax >= 0x8000001d &&
1117 		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1118 		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1119 			entry->eax = max(entry->eax, 0x80000021);
1120 		break;
1121 	case 0x80000001:
1122 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1123 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1124 		break;
1125 	case 0x80000006:
1126 		/* L2 cache and TLB: pass through host info. */
1127 		break;
1128 	case 0x80000007: /* Advanced power management */
1129 		/* invariant TSC is CPUID.80000007H:EDX[8] */
1130 		entry->edx &= (1 << 8);
1131 		/* mask against host */
1132 		entry->edx &= boot_cpu_data.x86_power;
1133 		entry->eax = entry->ebx = entry->ecx = 0;
1134 		break;
1135 	case 0x80000008: {
1136 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
1137 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
1138 		unsigned phys_as = entry->eax & 0xff;
1139 
1140 		/*
1141 		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1142 		 * the guest operates in the same PA space as the host, i.e.
1143 		 * reductions in MAXPHYADDR for memory encryption affect shadow
1144 		 * paging, too.
1145 		 *
1146 		 * If TDP is enabled but an explicit guest MAXPHYADDR is not
1147 		 * provided, use the raw bare metal MAXPHYADDR as reductions to
1148 		 * the HPAs do not affect GPAs.
1149 		 */
1150 		if (!tdp_enabled)
1151 			g_phys_as = boot_cpu_data.x86_phys_bits;
1152 		else if (!g_phys_as)
1153 			g_phys_as = phys_as;
1154 
1155 		entry->eax = g_phys_as | (virt_as << 8);
1156 		entry->edx = 0;
1157 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1158 		break;
1159 	}
1160 	case 0x8000000A:
1161 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1162 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1163 			break;
1164 		}
1165 		entry->eax = 1; /* SVM revision 1 */
1166 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1167 				   ASID emulation to nested SVM */
1168 		entry->ecx = 0; /* Reserved */
1169 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1170 		break;
1171 	case 0x80000019:
1172 		entry->ecx = entry->edx = 0;
1173 		break;
1174 	case 0x8000001a:
1175 	case 0x8000001e:
1176 		break;
1177 	case 0x8000001F:
1178 		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1179 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1180 		} else {
1181 			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1182 
1183 			/*
1184 			 * Enumerate '0' for "PA bits reduction", the adjusted
1185 			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1186 			 */
1187 			entry->ebx &= ~GENMASK(11, 6);
1188 		}
1189 		break;
1190 	case 0x80000020:
1191 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1192 		break;
1193 	case 0x80000021:
1194 		entry->ebx = entry->ecx = entry->edx = 0;
1195 		/*
1196 		 * Pass down these bits:
1197 		 *    EAX      0      NNDBP, Processor ignores nested data breakpoints
1198 		 *    EAX      2      LAS, LFENCE always serializing
1199 		 *    EAX      6      NSCB, Null selector clear base
1200 		 *
1201 		 * Other defined bits are for MSRs that KVM does not expose:
1202 		 *   EAX      3      SPCL, SMM page configuration lock
1203 		 *   EAX      13     PCMSR, Prefetch control MSR
1204 		 */
1205 		entry->eax &= BIT(0) | BIT(2) | BIT(6);
1206 		if (static_cpu_has(X86_FEATURE_LFENCE_RDTSC))
1207 			entry->eax |= BIT(2);
1208 		if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
1209 			entry->eax |= BIT(6);
1210 		break;
1211 	/*Add support for Centaur's CPUID instruction*/
1212 	case 0xC0000000:
1213 		/*Just support up to 0xC0000004 now*/
1214 		entry->eax = min(entry->eax, 0xC0000004);
1215 		break;
1216 	case 0xC0000001:
1217 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1218 		break;
1219 	case 3: /* Processor serial number */
1220 	case 5: /* MONITOR/MWAIT */
1221 	case 0xC0000002:
1222 	case 0xC0000003:
1223 	case 0xC0000004:
1224 	default:
1225 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1226 		break;
1227 	}
1228 
1229 	r = 0;
1230 
1231 out:
1232 	put_cpu();
1233 
1234 	return r;
1235 }
1236 
1237 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1238 			 unsigned int type)
1239 {
1240 	if (type == KVM_GET_EMULATED_CPUID)
1241 		return __do_cpuid_func_emulated(array, func);
1242 
1243 	return __do_cpuid_func(array, func);
1244 }
1245 
1246 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
1247 
1248 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1249 			  unsigned int type)
1250 {
1251 	u32 limit;
1252 	int r;
1253 
1254 	if (func == CENTAUR_CPUID_SIGNATURE &&
1255 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1256 		return 0;
1257 
1258 	r = do_cpuid_func(array, func, type);
1259 	if (r)
1260 		return r;
1261 
1262 	limit = array->entries[array->nent - 1].eax;
1263 	for (func = func + 1; func <= limit; ++func) {
1264 		r = do_cpuid_func(array, func, type);
1265 		if (r)
1266 			break;
1267 	}
1268 
1269 	return r;
1270 }
1271 
1272 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1273 				 __u32 num_entries, unsigned int ioctl_type)
1274 {
1275 	int i;
1276 	__u32 pad[3];
1277 
1278 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1279 		return false;
1280 
1281 	/*
1282 	 * We want to make sure that ->padding is being passed clean from
1283 	 * userspace in case we want to use it for something in the future.
1284 	 *
1285 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1286 	 * have to give ourselves satisfied only with the emulated side. /me
1287 	 * sheds a tear.
1288 	 */
1289 	for (i = 0; i < num_entries; i++) {
1290 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1291 			return true;
1292 
1293 		if (pad[0] || pad[1] || pad[2])
1294 			return true;
1295 	}
1296 	return false;
1297 }
1298 
1299 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1300 			    struct kvm_cpuid_entry2 __user *entries,
1301 			    unsigned int type)
1302 {
1303 	static const u32 funcs[] = {
1304 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1305 	};
1306 
1307 	struct kvm_cpuid_array array = {
1308 		.nent = 0,
1309 	};
1310 	int r, i;
1311 
1312 	if (cpuid->nent < 1)
1313 		return -E2BIG;
1314 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1315 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1316 
1317 	if (sanity_check_entries(entries, cpuid->nent, type))
1318 		return -EINVAL;
1319 
1320 	array.entries = kvcalloc(sizeof(struct kvm_cpuid_entry2), cpuid->nent, GFP_KERNEL);
1321 	if (!array.entries)
1322 		return -ENOMEM;
1323 
1324 	array.maxnent = cpuid->nent;
1325 
1326 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1327 		r = get_cpuid_func(&array, funcs[i], type);
1328 		if (r)
1329 			goto out_free;
1330 	}
1331 	cpuid->nent = array.nent;
1332 
1333 	if (copy_to_user(entries, array.entries,
1334 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1335 		r = -EFAULT;
1336 
1337 out_free:
1338 	kvfree(array.entries);
1339 	return r;
1340 }
1341 
1342 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
1343 						    u32 function, u32 index)
1344 {
1345 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1346 				 function, index);
1347 }
1348 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry_index);
1349 
1350 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1351 					      u32 function)
1352 {
1353 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1354 				 function, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
1355 }
1356 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1357 
1358 /*
1359  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1360  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1361  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1362  * range.  Centaur/VIA follows Intel semantics.
1363  *
1364  * A leaf is considered out-of-range if its function is higher than the maximum
1365  * supported leaf of its associated class or if its associated class does not
1366  * exist.
1367  *
1368  * There are three primary classes to be considered, with their respective
1369  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1370  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1371  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1372  *
1373  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1374  *  - Hypervisor: 0x40000000 - 0x4fffffff
1375  *  - Extended:   0x80000000 - 0xbfffffff
1376  *  - Centaur:    0xc0000000 - 0xcfffffff
1377  *
1378  * The Hypervisor class is further subdivided into sub-classes that each act as
1379  * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1380  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1381  * CPUID sub-classes are:
1382  *
1383  *  - HyperV:     0x40000000 - 0x400000ff
1384  *  - KVM:        0x40000100 - 0x400001ff
1385  */
1386 static struct kvm_cpuid_entry2 *
1387 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1388 {
1389 	struct kvm_cpuid_entry2 *basic, *class;
1390 	u32 function = *fn_ptr;
1391 
1392 	basic = kvm_find_cpuid_entry(vcpu, 0);
1393 	if (!basic)
1394 		return NULL;
1395 
1396 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1397 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1398 		return NULL;
1399 
1400 	if (function >= 0x40000000 && function <= 0x4fffffff)
1401 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1402 	else if (function >= 0xc0000000)
1403 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1404 	else
1405 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1406 
1407 	if (class && function <= class->eax)
1408 		return NULL;
1409 
1410 	/*
1411 	 * Leaf specific adjustments are also applied when redirecting to the
1412 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1413 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1414 	 * needs to be pulled from CPUID.0xb.1.
1415 	 */
1416 	*fn_ptr = basic->eax;
1417 
1418 	/*
1419 	 * The class does not exist or the requested function is out of range;
1420 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1421 	 * the original requested leaf is observed!
1422 	 */
1423 	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1424 }
1425 
1426 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1427 	       u32 *ecx, u32 *edx, bool exact_only)
1428 {
1429 	u32 orig_function = *eax, function = *eax, index = *ecx;
1430 	struct kvm_cpuid_entry2 *entry;
1431 	bool exact, used_max_basic = false;
1432 
1433 	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1434 	exact = !!entry;
1435 
1436 	if (!entry && !exact_only) {
1437 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1438 		used_max_basic = !!entry;
1439 	}
1440 
1441 	if (entry) {
1442 		*eax = entry->eax;
1443 		*ebx = entry->ebx;
1444 		*ecx = entry->ecx;
1445 		*edx = entry->edx;
1446 		if (function == 7 && index == 0) {
1447 			u64 data;
1448 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1449 			    (data & TSX_CTRL_CPUID_CLEAR))
1450 				*ebx &= ~(F(RTM) | F(HLE));
1451 		}
1452 	} else {
1453 		*eax = *ebx = *ecx = *edx = 0;
1454 		/*
1455 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1456 		 * and EDX is always the x2APIC ID, even for undefined
1457 		 * subleaves. Index 1 will exist iff the leaf is
1458 		 * implemented, so we pass through CL iff leaf 1
1459 		 * exists. EDX can be copied from any existing index.
1460 		 */
1461 		if (function == 0xb || function == 0x1f) {
1462 			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
1463 			if (entry) {
1464 				*ecx = index & 0xff;
1465 				*edx = entry->edx;
1466 			}
1467 		}
1468 	}
1469 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1470 			used_max_basic);
1471 	return exact;
1472 }
1473 EXPORT_SYMBOL_GPL(kvm_cpuid);
1474 
1475 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1476 {
1477 	u32 eax, ebx, ecx, edx;
1478 
1479 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1480 		return 1;
1481 
1482 	eax = kvm_rax_read(vcpu);
1483 	ecx = kvm_rcx_read(vcpu);
1484 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1485 	kvm_rax_write(vcpu, eax);
1486 	kvm_rbx_write(vcpu, ebx);
1487 	kvm_rcx_write(vcpu, ecx);
1488 	kvm_rdx_write(vcpu, edx);
1489 	return kvm_skip_emulated_instruction(vcpu);
1490 }
1491 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1492