1 // SPDX-License-Identifier: GPL-2.0-only
3 * Kernel-based Virtual Machine driver for Linux
5 * Macros and functions to access KVM PTEs (also known as SPTEs)
7 * Copyright (C) 2006 Qumranet, Inc.
8 * Copyright 2020 Red Hat, Inc. and/or its affiliates.
12 #include <linux/kvm_host.h>
14 #include "mmu_internal.h"
18 #include <asm/e820/api.h>
19 #include <asm/memtype.h>
22 static bool __read_mostly enable_mmio_caching = true;
23 module_param_named(mmio_caching, enable_mmio_caching, bool, 0444);
25 u64 __read_mostly shadow_host_writable_mask;
26 u64 __read_mostly shadow_mmu_writable_mask;
27 u64 __read_mostly shadow_nx_mask;
28 u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
29 u64 __read_mostly shadow_user_mask;
30 u64 __read_mostly shadow_accessed_mask;
31 u64 __read_mostly shadow_dirty_mask;
32 u64 __read_mostly shadow_mmio_value;
33 u64 __read_mostly shadow_mmio_mask;
34 u64 __read_mostly shadow_mmio_access_mask;
35 u64 __read_mostly shadow_present_mask;
36 u64 __read_mostly shadow_me_mask;
37 u64 __read_mostly shadow_acc_track_mask;
39 u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
40 u64 __read_mostly shadow_nonpresent_or_rsvd_lower_gfn_mask;
42 u8 __read_mostly shadow_phys_bits;
44 static u64 generation_mmio_spte_mask(u64 gen)
48 WARN_ON(gen & ~MMIO_SPTE_GEN_MASK);
50 mask = (gen << MMIO_SPTE_GEN_LOW_SHIFT) & MMIO_SPTE_GEN_LOW_MASK;
51 mask |= (gen << MMIO_SPTE_GEN_HIGH_SHIFT) & MMIO_SPTE_GEN_HIGH_MASK;
55 u64 make_mmio_spte(struct kvm_vcpu *vcpu, u64 gfn, unsigned int access)
57 u64 gen = kvm_vcpu_memslots(vcpu)->generation & MMIO_SPTE_GEN_MASK;
58 u64 spte = generation_mmio_spte_mask(gen);
59 u64 gpa = gfn << PAGE_SHIFT;
61 WARN_ON_ONCE(!shadow_mmio_value);
63 access &= shadow_mmio_access_mask;
64 spte |= shadow_mmio_value | access;
65 spte |= gpa | shadow_nonpresent_or_rsvd_mask;
66 spte |= (gpa & shadow_nonpresent_or_rsvd_mask)
67 << SHADOW_NONPRESENT_OR_RSVD_MASK_LEN;
72 static bool kvm_is_mmio_pfn(kvm_pfn_t pfn)
75 return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) &&
77 * Some reserved pages, such as those from NVDIMM
78 * DAX devices, are not for MMIO, and can be mapped
79 * with cached memory type for better performance.
80 * However, the above check misconceives those pages
81 * as MMIO, and results in KVM mapping them with UC
82 * memory type, which would hurt the performance.
83 * Therefore, we check the host memory type in addition
84 * and only treat UC/UC-/WC pages as MMIO.
86 (!pat_enabled() || pat_pfn_immune_to_uc_mtrr(pfn));
88 return !e820__mapped_raw_any(pfn_to_hpa(pfn),
89 pfn_to_hpa(pfn + 1) - 1,
93 bool make_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
94 const struct kvm_memory_slot *slot,
95 unsigned int pte_access, gfn_t gfn, kvm_pfn_t pfn,
96 u64 old_spte, bool prefetch, bool can_unsync,
97 bool host_writable, u64 *new_spte)
99 int level = sp->role.level;
100 u64 spte = SPTE_MMU_PRESENT_MASK;
103 if (sp->role.ad_disabled)
104 spte |= SPTE_TDP_AD_DISABLED_MASK;
105 else if (kvm_mmu_page_ad_need_write_protect(sp))
106 spte |= SPTE_TDP_AD_WRPROT_ONLY_MASK;
109 * For the EPT case, shadow_present_mask is 0 if hardware
110 * supports exec-only page table entries. In that case,
111 * ACC_USER_MASK and shadow_user_mask are used to represent
112 * read access. See FNAME(gpte_access) in paging_tmpl.h.
114 spte |= shadow_present_mask;
116 spte |= spte_shadow_accessed_mask(spte);
118 if (level > PG_LEVEL_4K && (pte_access & ACC_EXEC_MASK) &&
119 is_nx_huge_page_enabled()) {
120 pte_access &= ~ACC_EXEC_MASK;
123 if (pte_access & ACC_EXEC_MASK)
124 spte |= shadow_x_mask;
126 spte |= shadow_nx_mask;
128 if (pte_access & ACC_USER_MASK)
129 spte |= shadow_user_mask;
131 if (level > PG_LEVEL_4K)
132 spte |= PT_PAGE_SIZE_MASK;
134 spte |= static_call(kvm_x86_get_mt_mask)(vcpu, gfn,
135 kvm_is_mmio_pfn(pfn));
138 spte |= shadow_host_writable_mask;
140 pte_access &= ~ACC_WRITE_MASK;
142 if (!kvm_is_mmio_pfn(pfn))
143 spte |= shadow_me_mask;
145 spte |= (u64)pfn << PAGE_SHIFT;
147 if (pte_access & ACC_WRITE_MASK) {
148 spte |= PT_WRITABLE_MASK | shadow_mmu_writable_mask;
151 * Optimization: for pte sync, if spte was writable the hash
152 * lookup is unnecessary (and expensive). Write protection
153 * is responsibility of kvm_mmu_get_page / kvm_mmu_sync_roots.
154 * Same reasoning can be applied to dirty page accounting.
156 if (is_writable_pte(old_spte))
160 * Unsync shadow pages that are reachable by the new, writable
161 * SPTE. Write-protect the SPTE if the page can't be unsync'd,
162 * e.g. it's write-tracked (upper-level SPs) or has one or more
163 * shadow pages and unsync'ing pages is not allowed.
165 if (mmu_try_to_unsync_pages(vcpu->kvm, slot, gfn, can_unsync, prefetch)) {
166 pgprintk("%s: found shadow page for %llx, marking ro\n",
169 pte_access &= ~ACC_WRITE_MASK;
170 spte &= ~(PT_WRITABLE_MASK | shadow_mmu_writable_mask);
174 if (pte_access & ACC_WRITE_MASK)
175 spte |= spte_shadow_dirty_mask(spte);
179 spte = mark_spte_for_access_track(spte);
181 WARN_ONCE(is_rsvd_spte(&vcpu->arch.mmu->shadow_zero_check, spte, level),
182 "spte = 0x%llx, level = %d, rsvd bits = 0x%llx", spte, level,
183 get_rsvd_bits(&vcpu->arch.mmu->shadow_zero_check, spte, level));
185 if ((spte & PT_WRITABLE_MASK) && kvm_slot_dirty_track_enabled(slot)) {
186 /* Enforced by kvm_mmu_hugepage_adjust. */
187 WARN_ON(level > PG_LEVEL_4K);
188 mark_page_dirty_in_slot(vcpu->kvm, slot, gfn);
195 static u64 make_spte_executable(u64 spte)
197 bool is_access_track = is_access_track_spte(spte);
200 spte = restore_acc_track_spte(spte);
202 spte &= ~shadow_nx_mask;
203 spte |= shadow_x_mask;
206 spte = mark_spte_for_access_track(spte);
212 * Construct an SPTE that maps a sub-page of the given huge page SPTE where
213 * `index` identifies which sub-page.
215 * This is used during huge page splitting to build the SPTEs that make up the
218 u64 make_huge_page_split_spte(u64 huge_spte, int huge_level, int index)
223 if (WARN_ON_ONCE(!is_shadow_present_pte(huge_spte)))
226 if (WARN_ON_ONCE(!is_large_pte(huge_spte)))
229 child_spte = huge_spte;
230 child_level = huge_level - 1;
233 * The child_spte already has the base address of the huge page being
234 * split. So we just have to OR in the offset to the page at the next
235 * lower level for the given index.
237 child_spte |= (index * KVM_PAGES_PER_HPAGE(child_level)) << PAGE_SHIFT;
239 if (child_level == PG_LEVEL_4K) {
240 child_spte &= ~PT_PAGE_SIZE_MASK;
243 * When splitting to a 4K page, mark the page executable as the
244 * NX hugepage mitigation no longer applies.
246 if (is_nx_huge_page_enabled())
247 child_spte = make_spte_executable(child_spte);
254 u64 make_nonleaf_spte(u64 *child_pt, bool ad_disabled)
256 u64 spte = SPTE_MMU_PRESENT_MASK;
258 spte |= __pa(child_pt) | shadow_present_mask | PT_WRITABLE_MASK |
259 shadow_user_mask | shadow_x_mask | shadow_me_mask;
262 spte |= SPTE_TDP_AD_DISABLED_MASK;
264 spte |= shadow_accessed_mask;
269 u64 kvm_mmu_changed_pte_notifier_make_spte(u64 old_spte, kvm_pfn_t new_pfn)
273 new_spte = old_spte & ~PT64_BASE_ADDR_MASK;
274 new_spte |= (u64)new_pfn << PAGE_SHIFT;
276 new_spte &= ~PT_WRITABLE_MASK;
277 new_spte &= ~shadow_host_writable_mask;
278 new_spte &= ~shadow_mmu_writable_mask;
280 new_spte = mark_spte_for_access_track(new_spte);
285 static u8 kvm_get_shadow_phys_bits(void)
288 * boot_cpu_data.x86_phys_bits is reduced when MKTME or SME are detected
289 * in CPU detection code, but the processor treats those reduced bits as
290 * 'keyID' thus they are not reserved bits. Therefore KVM needs to look at
291 * the physical address bits reported by CPUID.
293 if (likely(boot_cpu_data.extended_cpuid_level >= 0x80000008))
294 return cpuid_eax(0x80000008) & 0xff;
297 * Quite weird to have VMX or SVM but not MAXPHYADDR; probably a VM with
298 * custom CPUID. Proceed with whatever the kernel found since these features
299 * aren't virtualizable (SME/SEV also require CPUIDs higher than 0x80000008).
301 return boot_cpu_data.x86_phys_bits;
304 u64 mark_spte_for_access_track(u64 spte)
306 if (spte_ad_enabled(spte))
307 return spte & ~shadow_accessed_mask;
309 if (is_access_track_spte(spte))
312 check_spte_writable_invariants(spte);
314 WARN_ONCE(spte & (SHADOW_ACC_TRACK_SAVED_BITS_MASK <<
315 SHADOW_ACC_TRACK_SAVED_BITS_SHIFT),
316 "kvm: Access Tracking saved bit locations are not zero\n");
318 spte |= (spte & SHADOW_ACC_TRACK_SAVED_BITS_MASK) <<
319 SHADOW_ACC_TRACK_SAVED_BITS_SHIFT;
320 spte &= ~shadow_acc_track_mask;
325 void kvm_mmu_set_mmio_spte_mask(u64 mmio_value, u64 mmio_mask, u64 access_mask)
327 BUG_ON((u64)(unsigned)access_mask != access_mask);
328 WARN_ON(mmio_value & shadow_nonpresent_or_rsvd_lower_gfn_mask);
330 if (!enable_mmio_caching)
334 * Disable MMIO caching if the MMIO value collides with the bits that
335 * are used to hold the relocated GFN when the L1TF mitigation is
336 * enabled. This should never fire as there is no known hardware that
337 * can trigger this condition, e.g. SME/SEV CPUs that require a custom
338 * MMIO value are not susceptible to L1TF.
340 if (WARN_ON(mmio_value & (shadow_nonpresent_or_rsvd_mask <<
341 SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)))
345 * The masked MMIO value must obviously match itself and a removed SPTE
346 * must not get a false positive. Removed SPTEs and MMIO SPTEs should
347 * never collide as MMIO must set some RWX bits, and removed SPTEs must
348 * not set any RWX bits.
350 if (WARN_ON((mmio_value & mmio_mask) != mmio_value) ||
351 WARN_ON(mmio_value && (REMOVED_SPTE & mmio_mask) == mmio_value))
354 shadow_mmio_value = mmio_value;
355 shadow_mmio_mask = mmio_mask;
356 shadow_mmio_access_mask = access_mask;
358 EXPORT_SYMBOL_GPL(kvm_mmu_set_mmio_spte_mask);
360 void kvm_mmu_set_ept_masks(bool has_ad_bits, bool has_exec_only)
362 shadow_user_mask = VMX_EPT_READABLE_MASK;
363 shadow_accessed_mask = has_ad_bits ? VMX_EPT_ACCESS_BIT : 0ull;
364 shadow_dirty_mask = has_ad_bits ? VMX_EPT_DIRTY_BIT : 0ull;
365 shadow_nx_mask = 0ull;
366 shadow_x_mask = VMX_EPT_EXECUTABLE_MASK;
367 shadow_present_mask = has_exec_only ? 0ull : VMX_EPT_READABLE_MASK;
368 shadow_acc_track_mask = VMX_EPT_RWX_MASK;
369 shadow_me_mask = 0ull;
371 shadow_host_writable_mask = EPT_SPTE_HOST_WRITABLE;
372 shadow_mmu_writable_mask = EPT_SPTE_MMU_WRITABLE;
375 * EPT Misconfigurations are generated if the value of bits 2:0
376 * of an EPT paging-structure entry is 110b (write/execute).
378 kvm_mmu_set_mmio_spte_mask(VMX_EPT_MISCONFIG_WX_VALUE,
379 VMX_EPT_RWX_MASK, 0);
381 EXPORT_SYMBOL_GPL(kvm_mmu_set_ept_masks);
383 void kvm_mmu_reset_all_pte_masks(void)
388 shadow_phys_bits = kvm_get_shadow_phys_bits();
391 * If the CPU has 46 or less physical address bits, then set an
392 * appropriate mask to guard against L1TF attacks. Otherwise, it is
393 * assumed that the CPU is not vulnerable to L1TF.
395 * Some Intel CPUs address the L1 cache using more PA bits than are
396 * reported by CPUID. Use the PA width of the L1 cache when possible
397 * to achieve more effective mitigation, e.g. if system RAM overlaps
398 * the most significant bits of legal physical address space.
400 shadow_nonpresent_or_rsvd_mask = 0;
401 low_phys_bits = boot_cpu_data.x86_phys_bits;
402 if (boot_cpu_has_bug(X86_BUG_L1TF) &&
403 !WARN_ON_ONCE(boot_cpu_data.x86_cache_bits >=
404 52 - SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)) {
405 low_phys_bits = boot_cpu_data.x86_cache_bits
406 - SHADOW_NONPRESENT_OR_RSVD_MASK_LEN;
407 shadow_nonpresent_or_rsvd_mask =
408 rsvd_bits(low_phys_bits, boot_cpu_data.x86_cache_bits - 1);
411 shadow_nonpresent_or_rsvd_lower_gfn_mask =
412 GENMASK_ULL(low_phys_bits - 1, PAGE_SHIFT);
414 shadow_user_mask = PT_USER_MASK;
415 shadow_accessed_mask = PT_ACCESSED_MASK;
416 shadow_dirty_mask = PT_DIRTY_MASK;
417 shadow_nx_mask = PT64_NX_MASK;
419 shadow_present_mask = PT_PRESENT_MASK;
420 shadow_acc_track_mask = 0;
421 shadow_me_mask = sme_me_mask;
423 shadow_host_writable_mask = DEFAULT_SPTE_HOST_WRITABLE;
424 shadow_mmu_writable_mask = DEFAULT_SPTE_MMU_WRITABLE;
427 * Set a reserved PA bit in MMIO SPTEs to generate page faults with
428 * PFEC.RSVD=1 on MMIO accesses. 64-bit PTEs (PAE, x86-64, and EPT
429 * paging) support a maximum of 52 bits of PA, i.e. if the CPU supports
430 * 52-bit physical addresses then there are no reserved PA bits in the
431 * PTEs and so the reserved PA approach must be disabled.
433 if (shadow_phys_bits < 52)
434 mask = BIT_ULL(51) | PT_PRESENT_MASK;
438 kvm_mmu_set_mmio_spte_mask(mask, mask, ACC_WRITE_MASK | ACC_USER_MASK);