1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Kernel-based Virtual Machine driver for Linux
4 *
5 * Macros and functions to access KVM PTEs (also known as SPTEs)
6 *
7 * Copyright (C) 2006 Qumranet, Inc.
8 * Copyright 2020 Red Hat, Inc. and/or its affiliates.
9 */
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12 #include <linux/kvm_host.h>
13 #include "mmu.h"
14 #include "mmu_internal.h"
15 #include "x86.h"
16 #include "spte.h"
17
18 #include <asm/e820/api.h>
19 #include <asm/memtype.h>
20 #include <asm/vmx.h>
21
22 bool __read_mostly enable_mmio_caching = true;
23 static bool __ro_after_init allow_mmio_caching;
24 module_param_named(mmio_caching, enable_mmio_caching, bool, 0444);
25 EXPORT_SYMBOL_GPL(enable_mmio_caching);
26
27 u64 __read_mostly shadow_host_writable_mask;
28 u64 __read_mostly shadow_mmu_writable_mask;
29 u64 __read_mostly shadow_nx_mask;
30 u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
31 u64 __read_mostly shadow_user_mask;
32 u64 __read_mostly shadow_accessed_mask;
33 u64 __read_mostly shadow_dirty_mask;
34 u64 __read_mostly shadow_mmio_value;
35 u64 __read_mostly shadow_mmio_mask;
36 u64 __read_mostly shadow_mmio_access_mask;
37 u64 __read_mostly shadow_present_mask;
38 u64 __read_mostly shadow_memtype_mask;
39 u64 __read_mostly shadow_me_value;
40 u64 __read_mostly shadow_me_mask;
41 u64 __read_mostly shadow_acc_track_mask;
42
43 u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
44 u64 __read_mostly shadow_nonpresent_or_rsvd_lower_gfn_mask;
45
kvm_get_host_maxphyaddr(void)46 static u8 __init kvm_get_host_maxphyaddr(void)
47 {
48 /*
49 * boot_cpu_data.x86_phys_bits is reduced when MKTME or SME are detected
50 * in CPU detection code, but the processor treats those reduced bits as
51 * 'keyID' thus they are not reserved bits. Therefore KVM needs to look at
52 * the physical address bits reported by CPUID, i.e. the raw MAXPHYADDR,
53 * when reasoning about CPU behavior with respect to MAXPHYADDR.
54 */
55 if (likely(boot_cpu_data.extended_cpuid_level >= 0x80000008))
56 return cpuid_eax(0x80000008) & 0xff;
57
58 /*
59 * Quite weird to have VMX or SVM but not MAXPHYADDR; probably a VM with
60 * custom CPUID. Proceed with whatever the kernel found since these features
61 * aren't virtualizable (SME/SEV also require CPUIDs higher than 0x80000008).
62 */
63 return boot_cpu_data.x86_phys_bits;
64 }
65
kvm_mmu_spte_module_init(void)66 void __init kvm_mmu_spte_module_init(void)
67 {
68 /*
69 * Snapshot userspace's desire to allow MMIO caching. Whether or not
70 * KVM can actually enable MMIO caching depends on vendor-specific
71 * hardware capabilities and other module params that can't be resolved
72 * until the vendor module is loaded, i.e. enable_mmio_caching can and
73 * will change when the vendor module is (re)loaded.
74 */
75 allow_mmio_caching = enable_mmio_caching;
76
77 kvm_host.maxphyaddr = kvm_get_host_maxphyaddr();
78 }
79
generation_mmio_spte_mask(u64 gen)80 static u64 generation_mmio_spte_mask(u64 gen)
81 {
82 u64 mask;
83
84 WARN_ON_ONCE(gen & ~MMIO_SPTE_GEN_MASK);
85
86 mask = (gen << MMIO_SPTE_GEN_LOW_SHIFT) & MMIO_SPTE_GEN_LOW_MASK;
87 mask |= (gen << MMIO_SPTE_GEN_HIGH_SHIFT) & MMIO_SPTE_GEN_HIGH_MASK;
88 return mask;
89 }
90
make_mmio_spte(struct kvm_vcpu * vcpu,u64 gfn,unsigned int access)91 u64 make_mmio_spte(struct kvm_vcpu *vcpu, u64 gfn, unsigned int access)
92 {
93 u64 gen = kvm_vcpu_memslots(vcpu)->generation & MMIO_SPTE_GEN_MASK;
94 u64 spte = generation_mmio_spte_mask(gen);
95 u64 gpa = gfn << PAGE_SHIFT;
96
97 WARN_ON_ONCE(!vcpu->kvm->arch.shadow_mmio_value);
98
99 access &= shadow_mmio_access_mask;
100 spte |= vcpu->kvm->arch.shadow_mmio_value | access;
101 spte |= gpa | shadow_nonpresent_or_rsvd_mask;
102 spte |= (gpa & shadow_nonpresent_or_rsvd_mask)
103 << SHADOW_NONPRESENT_OR_RSVD_MASK_LEN;
104
105 return spte;
106 }
107
kvm_is_mmio_pfn(kvm_pfn_t pfn)108 static bool kvm_is_mmio_pfn(kvm_pfn_t pfn)
109 {
110 if (pfn_valid(pfn))
111 return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) &&
112 /*
113 * Some reserved pages, such as those from NVDIMM
114 * DAX devices, are not for MMIO, and can be mapped
115 * with cached memory type for better performance.
116 * However, the above check misconceives those pages
117 * as MMIO, and results in KVM mapping them with UC
118 * memory type, which would hurt the performance.
119 * Therefore, we check the host memory type in addition
120 * and only treat UC/UC-/WC pages as MMIO.
121 */
122 (!pat_enabled() || pat_pfn_immune_to_uc_mtrr(pfn));
123
124 return !e820__mapped_raw_any(pfn_to_hpa(pfn),
125 pfn_to_hpa(pfn + 1) - 1,
126 E820_TYPE_RAM);
127 }
128
129 /*
130 * Returns true if the SPTE has bits that may be set without holding mmu_lock.
131 * The caller is responsible for checking if the SPTE is shadow-present, and
132 * for determining whether or not the caller cares about non-leaf SPTEs.
133 */
spte_has_volatile_bits(u64 spte)134 bool spte_has_volatile_bits(u64 spte)
135 {
136 /*
137 * Always atomically update spte if it can be updated
138 * out of mmu-lock, it can ensure dirty bit is not lost,
139 * also, it can help us to get a stable is_writable_pte()
140 * to ensure tlb flush is not missed.
141 */
142 if (!is_writable_pte(spte) && is_mmu_writable_spte(spte))
143 return true;
144
145 if (is_access_track_spte(spte))
146 return true;
147
148 if (spte_ad_enabled(spte)) {
149 if (!(spte & shadow_accessed_mask) ||
150 (is_writable_pte(spte) && !(spte & shadow_dirty_mask)))
151 return true;
152 }
153
154 return false;
155 }
156
make_spte(struct kvm_vcpu * vcpu,struct kvm_mmu_page * sp,const struct kvm_memory_slot * slot,unsigned int pte_access,gfn_t gfn,kvm_pfn_t pfn,u64 old_spte,bool prefetch,bool can_unsync,bool host_writable,u64 * new_spte)157 bool make_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
158 const struct kvm_memory_slot *slot,
159 unsigned int pte_access, gfn_t gfn, kvm_pfn_t pfn,
160 u64 old_spte, bool prefetch, bool can_unsync,
161 bool host_writable, u64 *new_spte)
162 {
163 int level = sp->role.level;
164 u64 spte = SPTE_MMU_PRESENT_MASK;
165 bool wrprot = false;
166
167 /*
168 * For the EPT case, shadow_present_mask has no RWX bits set if
169 * exec-only page table entries are supported. In that case,
170 * ACC_USER_MASK and shadow_user_mask are used to represent
171 * read access. See FNAME(gpte_access) in paging_tmpl.h.
172 */
173 WARN_ON_ONCE((pte_access | shadow_present_mask) == SHADOW_NONPRESENT_VALUE);
174
175 if (sp->role.ad_disabled)
176 spte |= SPTE_TDP_AD_DISABLED;
177 else if (kvm_mmu_page_ad_need_write_protect(sp))
178 spte |= SPTE_TDP_AD_WRPROT_ONLY;
179
180 spte |= shadow_present_mask;
181 if (!prefetch)
182 spte |= spte_shadow_accessed_mask(spte);
183
184 /*
185 * For simplicity, enforce the NX huge page mitigation even if not
186 * strictly necessary. KVM could ignore the mitigation if paging is
187 * disabled in the guest, as the guest doesn't have any page tables to
188 * abuse. But to safely ignore the mitigation, KVM would have to
189 * ensure a new MMU is loaded (or all shadow pages zapped) when CR0.PG
190 * is toggled on, and that's a net negative for performance when TDP is
191 * enabled. When TDP is disabled, KVM will always switch to a new MMU
192 * when CR0.PG is toggled, but leveraging that to ignore the mitigation
193 * would tie make_spte() further to vCPU/MMU state, and add complexity
194 * just to optimize a mode that is anything but performance critical.
195 */
196 if (level > PG_LEVEL_4K && (pte_access & ACC_EXEC_MASK) &&
197 is_nx_huge_page_enabled(vcpu->kvm)) {
198 pte_access &= ~ACC_EXEC_MASK;
199 }
200
201 if (pte_access & ACC_EXEC_MASK)
202 spte |= shadow_x_mask;
203 else
204 spte |= shadow_nx_mask;
205
206 if (pte_access & ACC_USER_MASK)
207 spte |= shadow_user_mask;
208
209 if (level > PG_LEVEL_4K)
210 spte |= PT_PAGE_SIZE_MASK;
211
212 if (shadow_memtype_mask)
213 spte |= kvm_x86_call(get_mt_mask)(vcpu, gfn,
214 kvm_is_mmio_pfn(pfn));
215 if (host_writable)
216 spte |= shadow_host_writable_mask;
217 else
218 pte_access &= ~ACC_WRITE_MASK;
219
220 if (shadow_me_value && !kvm_is_mmio_pfn(pfn))
221 spte |= shadow_me_value;
222
223 spte |= (u64)pfn << PAGE_SHIFT;
224
225 if (pte_access & ACC_WRITE_MASK) {
226 spte |= PT_WRITABLE_MASK | shadow_mmu_writable_mask;
227
228 /*
229 * When overwriting an existing leaf SPTE, and the old SPTE was
230 * writable, skip trying to unsync shadow pages as any relevant
231 * shadow pages must already be unsync, i.e. the hash lookup is
232 * unnecessary (and expensive).
233 *
234 * The same reasoning applies to dirty page/folio accounting;
235 * KVM will mark the folio dirty using the old SPTE, thus
236 * there's no need to immediately mark the new SPTE as dirty.
237 *
238 * Note, both cases rely on KVM not changing PFNs without first
239 * zapping the old SPTE, which is guaranteed by both the shadow
240 * MMU and the TDP MMU.
241 */
242 if (is_last_spte(old_spte, level) && is_writable_pte(old_spte))
243 goto out;
244
245 /*
246 * Unsync shadow pages that are reachable by the new, writable
247 * SPTE. Write-protect the SPTE if the page can't be unsync'd,
248 * e.g. it's write-tracked (upper-level SPs) or has one or more
249 * shadow pages and unsync'ing pages is not allowed.
250 */
251 if (mmu_try_to_unsync_pages(vcpu->kvm, slot, gfn, can_unsync, prefetch)) {
252 wrprot = true;
253 pte_access &= ~ACC_WRITE_MASK;
254 spte &= ~(PT_WRITABLE_MASK | shadow_mmu_writable_mask);
255 }
256 }
257
258 if (pte_access & ACC_WRITE_MASK)
259 spte |= spte_shadow_dirty_mask(spte);
260
261 out:
262 if (prefetch)
263 spte = mark_spte_for_access_track(spte);
264
265 WARN_ONCE(is_rsvd_spte(&vcpu->arch.mmu->shadow_zero_check, spte, level),
266 "spte = 0x%llx, level = %d, rsvd bits = 0x%llx", spte, level,
267 get_rsvd_bits(&vcpu->arch.mmu->shadow_zero_check, spte, level));
268
269 if ((spte & PT_WRITABLE_MASK) && kvm_slot_dirty_track_enabled(slot)) {
270 /* Enforced by kvm_mmu_hugepage_adjust. */
271 WARN_ON_ONCE(level > PG_LEVEL_4K);
272 mark_page_dirty_in_slot(vcpu->kvm, slot, gfn);
273 }
274
275 *new_spte = spte;
276 return wrprot;
277 }
278
make_spte_executable(u64 spte)279 static u64 make_spte_executable(u64 spte)
280 {
281 bool is_access_track = is_access_track_spte(spte);
282
283 if (is_access_track)
284 spte = restore_acc_track_spte(spte);
285
286 spte &= ~shadow_nx_mask;
287 spte |= shadow_x_mask;
288
289 if (is_access_track)
290 spte = mark_spte_for_access_track(spte);
291
292 return spte;
293 }
294
295 /*
296 * Construct an SPTE that maps a sub-page of the given huge page SPTE where
297 * `index` identifies which sub-page.
298 *
299 * This is used during huge page splitting to build the SPTEs that make up the
300 * new page table.
301 */
make_huge_page_split_spte(struct kvm * kvm,u64 huge_spte,union kvm_mmu_page_role role,int index)302 u64 make_huge_page_split_spte(struct kvm *kvm, u64 huge_spte,
303 union kvm_mmu_page_role role, int index)
304 {
305 u64 child_spte = huge_spte;
306
307 KVM_BUG_ON(!is_shadow_present_pte(huge_spte) || !is_large_pte(huge_spte), kvm);
308
309 /*
310 * The child_spte already has the base address of the huge page being
311 * split. So we just have to OR in the offset to the page at the next
312 * lower level for the given index.
313 */
314 child_spte |= (index * KVM_PAGES_PER_HPAGE(role.level)) << PAGE_SHIFT;
315
316 if (role.level == PG_LEVEL_4K) {
317 child_spte &= ~PT_PAGE_SIZE_MASK;
318
319 /*
320 * When splitting to a 4K page where execution is allowed, mark
321 * the page executable as the NX hugepage mitigation no longer
322 * applies.
323 */
324 if ((role.access & ACC_EXEC_MASK) && is_nx_huge_page_enabled(kvm))
325 child_spte = make_spte_executable(child_spte);
326 }
327
328 return child_spte;
329 }
330
331
make_nonleaf_spte(u64 * child_pt,bool ad_disabled)332 u64 make_nonleaf_spte(u64 *child_pt, bool ad_disabled)
333 {
334 u64 spte = SPTE_MMU_PRESENT_MASK;
335
336 spte |= __pa(child_pt) | shadow_present_mask | PT_WRITABLE_MASK |
337 shadow_user_mask | shadow_x_mask | shadow_me_value;
338
339 if (ad_disabled)
340 spte |= SPTE_TDP_AD_DISABLED;
341 else
342 spte |= shadow_accessed_mask;
343
344 return spte;
345 }
346
mark_spte_for_access_track(u64 spte)347 u64 mark_spte_for_access_track(u64 spte)
348 {
349 if (spte_ad_enabled(spte))
350 return spte & ~shadow_accessed_mask;
351
352 if (is_access_track_spte(spte))
353 return spte;
354
355 check_spte_writable_invariants(spte);
356
357 WARN_ONCE(spte & (SHADOW_ACC_TRACK_SAVED_BITS_MASK <<
358 SHADOW_ACC_TRACK_SAVED_BITS_SHIFT),
359 "Access Tracking saved bit locations are not zero\n");
360
361 spte |= (spte & SHADOW_ACC_TRACK_SAVED_BITS_MASK) <<
362 SHADOW_ACC_TRACK_SAVED_BITS_SHIFT;
363 spte &= ~shadow_acc_track_mask;
364
365 return spte;
366 }
367
kvm_mmu_set_mmio_spte_mask(u64 mmio_value,u64 mmio_mask,u64 access_mask)368 void kvm_mmu_set_mmio_spte_mask(u64 mmio_value, u64 mmio_mask, u64 access_mask)
369 {
370 BUG_ON((u64)(unsigned)access_mask != access_mask);
371 WARN_ON(mmio_value & shadow_nonpresent_or_rsvd_lower_gfn_mask);
372
373 /*
374 * Reset to the original module param value to honor userspace's desire
375 * to (dis)allow MMIO caching. Update the param itself so that
376 * userspace can see whether or not KVM is actually using MMIO caching.
377 */
378 enable_mmio_caching = allow_mmio_caching;
379 if (!enable_mmio_caching)
380 mmio_value = 0;
381
382 /*
383 * The mask must contain only bits that are carved out specifically for
384 * the MMIO SPTE mask, e.g. to ensure there's no overlap with the MMIO
385 * generation.
386 */
387 if (WARN_ON(mmio_mask & ~SPTE_MMIO_ALLOWED_MASK))
388 mmio_value = 0;
389
390 /*
391 * Disable MMIO caching if the MMIO value collides with the bits that
392 * are used to hold the relocated GFN when the L1TF mitigation is
393 * enabled. This should never fire as there is no known hardware that
394 * can trigger this condition, e.g. SME/SEV CPUs that require a custom
395 * MMIO value are not susceptible to L1TF.
396 */
397 if (WARN_ON(mmio_value & (shadow_nonpresent_or_rsvd_mask <<
398 SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)))
399 mmio_value = 0;
400
401 /*
402 * The masked MMIO value must obviously match itself and a frozen SPTE
403 * must not get a false positive. Frozen SPTEs and MMIO SPTEs should
404 * never collide as MMIO must set some RWX bits, and frozen SPTEs must
405 * not set any RWX bits.
406 */
407 if (WARN_ON((mmio_value & mmio_mask) != mmio_value) ||
408 WARN_ON(mmio_value && (FROZEN_SPTE & mmio_mask) == mmio_value))
409 mmio_value = 0;
410
411 if (!mmio_value)
412 enable_mmio_caching = false;
413
414 shadow_mmio_value = mmio_value;
415 shadow_mmio_mask = mmio_mask;
416 shadow_mmio_access_mask = access_mask;
417 }
418 EXPORT_SYMBOL_GPL(kvm_mmu_set_mmio_spte_mask);
419
kvm_mmu_set_me_spte_mask(u64 me_value,u64 me_mask)420 void kvm_mmu_set_me_spte_mask(u64 me_value, u64 me_mask)
421 {
422 /* shadow_me_value must be a subset of shadow_me_mask */
423 if (WARN_ON(me_value & ~me_mask))
424 me_value = me_mask = 0;
425
426 shadow_me_value = me_value;
427 shadow_me_mask = me_mask;
428 }
429 EXPORT_SYMBOL_GPL(kvm_mmu_set_me_spte_mask);
430
kvm_mmu_set_ept_masks(bool has_ad_bits,bool has_exec_only)431 void kvm_mmu_set_ept_masks(bool has_ad_bits, bool has_exec_only)
432 {
433 shadow_user_mask = VMX_EPT_READABLE_MASK;
434 shadow_accessed_mask = has_ad_bits ? VMX_EPT_ACCESS_BIT : 0ull;
435 shadow_dirty_mask = has_ad_bits ? VMX_EPT_DIRTY_BIT : 0ull;
436 shadow_nx_mask = 0ull;
437 shadow_x_mask = VMX_EPT_EXECUTABLE_MASK;
438 /* VMX_EPT_SUPPRESS_VE_BIT is needed for W or X violation. */
439 shadow_present_mask =
440 (has_exec_only ? 0ull : VMX_EPT_READABLE_MASK) | VMX_EPT_SUPPRESS_VE_BIT;
441 /*
442 * EPT overrides the host MTRRs, and so KVM must program the desired
443 * memtype directly into the SPTEs. Note, this mask is just the mask
444 * of all bits that factor into the memtype, the actual memtype must be
445 * dynamically calculated, e.g. to ensure host MMIO is mapped UC.
446 */
447 shadow_memtype_mask = VMX_EPT_MT_MASK | VMX_EPT_IPAT_BIT;
448 shadow_acc_track_mask = VMX_EPT_RWX_MASK;
449 shadow_host_writable_mask = EPT_SPTE_HOST_WRITABLE;
450 shadow_mmu_writable_mask = EPT_SPTE_MMU_WRITABLE;
451
452 /*
453 * EPT Misconfigurations are generated if the value of bits 2:0
454 * of an EPT paging-structure entry is 110b (write/execute).
455 */
456 kvm_mmu_set_mmio_spte_mask(VMX_EPT_MISCONFIG_WX_VALUE,
457 VMX_EPT_RWX_MASK | VMX_EPT_SUPPRESS_VE_BIT, 0);
458 }
459 EXPORT_SYMBOL_GPL(kvm_mmu_set_ept_masks);
460
kvm_mmu_reset_all_pte_masks(void)461 void kvm_mmu_reset_all_pte_masks(void)
462 {
463 u8 low_phys_bits;
464 u64 mask;
465
466 /*
467 * If the CPU has 46 or less physical address bits, then set an
468 * appropriate mask to guard against L1TF attacks. Otherwise, it is
469 * assumed that the CPU is not vulnerable to L1TF.
470 *
471 * Some Intel CPUs address the L1 cache using more PA bits than are
472 * reported by CPUID. Use the PA width of the L1 cache when possible
473 * to achieve more effective mitigation, e.g. if system RAM overlaps
474 * the most significant bits of legal physical address space.
475 */
476 shadow_nonpresent_or_rsvd_mask = 0;
477 low_phys_bits = boot_cpu_data.x86_phys_bits;
478 if (boot_cpu_has_bug(X86_BUG_L1TF) &&
479 !WARN_ON_ONCE(boot_cpu_data.x86_cache_bits >=
480 52 - SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)) {
481 low_phys_bits = boot_cpu_data.x86_cache_bits
482 - SHADOW_NONPRESENT_OR_RSVD_MASK_LEN;
483 shadow_nonpresent_or_rsvd_mask =
484 rsvd_bits(low_phys_bits, boot_cpu_data.x86_cache_bits - 1);
485 }
486
487 shadow_nonpresent_or_rsvd_lower_gfn_mask =
488 GENMASK_ULL(low_phys_bits - 1, PAGE_SHIFT);
489
490 shadow_user_mask = PT_USER_MASK;
491 shadow_accessed_mask = PT_ACCESSED_MASK;
492 shadow_dirty_mask = PT_DIRTY_MASK;
493 shadow_nx_mask = PT64_NX_MASK;
494 shadow_x_mask = 0;
495 shadow_present_mask = PT_PRESENT_MASK;
496
497 /*
498 * For shadow paging and NPT, KVM uses PAT entry '0' to encode WB
499 * memtype in the SPTEs, i.e. relies on host MTRRs to provide the
500 * correct memtype (WB is the "weakest" memtype).
501 */
502 shadow_memtype_mask = 0;
503 shadow_acc_track_mask = 0;
504 shadow_me_mask = 0;
505 shadow_me_value = 0;
506
507 shadow_host_writable_mask = DEFAULT_SPTE_HOST_WRITABLE;
508 shadow_mmu_writable_mask = DEFAULT_SPTE_MMU_WRITABLE;
509
510 /*
511 * Set a reserved PA bit in MMIO SPTEs to generate page faults with
512 * PFEC.RSVD=1 on MMIO accesses. 64-bit PTEs (PAE, x86-64, and EPT
513 * paging) support a maximum of 52 bits of PA, i.e. if the CPU supports
514 * 52-bit physical addresses then there are no reserved PA bits in the
515 * PTEs and so the reserved PA approach must be disabled.
516 */
517 if (kvm_host.maxphyaddr < 52)
518 mask = BIT_ULL(51) | PT_PRESENT_MASK;
519 else
520 mask = 0;
521
522 kvm_mmu_set_mmio_spte_mask(mask, mask, ACC_WRITE_MASK | ACC_USER_MASK);
523 }
524