1c50d8ae3SPaolo Bonzini // SPDX-License-Identifier: GPL-2.0-only 2c50d8ae3SPaolo Bonzini /* 3c50d8ae3SPaolo Bonzini * Kernel-based Virtual Machine driver for Linux 4c50d8ae3SPaolo Bonzini * 5c50d8ae3SPaolo Bonzini * This module enables machines with Intel VT-x extensions to run virtual 6c50d8ae3SPaolo Bonzini * machines without emulation or binary translation. 7c50d8ae3SPaolo Bonzini * 8c50d8ae3SPaolo Bonzini * MMU support 9c50d8ae3SPaolo Bonzini * 10c50d8ae3SPaolo Bonzini * Copyright (C) 2006 Qumranet, Inc. 11c50d8ae3SPaolo Bonzini * Copyright 2010 Red Hat, Inc. and/or its affiliates. 12c50d8ae3SPaolo Bonzini * 13c50d8ae3SPaolo Bonzini * Authors: 14c50d8ae3SPaolo Bonzini * Yaniv Kamay <yaniv@qumranet.com> 15c50d8ae3SPaolo Bonzini * Avi Kivity <avi@qumranet.com> 16c50d8ae3SPaolo Bonzini */ 178d20bd63SSean Christopherson #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 18c50d8ae3SPaolo Bonzini 19c50d8ae3SPaolo Bonzini #include "irq.h" 2088197e6aS彭浩(Richard) #include "ioapic.h" 21c50d8ae3SPaolo Bonzini #include "mmu.h" 226ca9a6f3SSean Christopherson #include "mmu_internal.h" 23fe5db27dSBen Gardon #include "tdp_mmu.h" 24c50d8ae3SPaolo Bonzini #include "x86.h" 25c50d8ae3SPaolo Bonzini #include "kvm_cache_regs.h" 26b0b42197SPaolo Bonzini #include "smm.h" 272f728d66SSean Christopherson #include "kvm_emulate.h" 2858ea7cf7SSean Christopherson #include "page_track.h" 29c50d8ae3SPaolo Bonzini #include "cpuid.h" 305a9624afSPaolo Bonzini #include "spte.h" 31c50d8ae3SPaolo Bonzini 32c50d8ae3SPaolo Bonzini #include <linux/kvm_host.h> 33c50d8ae3SPaolo Bonzini #include <linux/types.h> 34c50d8ae3SPaolo Bonzini #include <linux/string.h> 35c50d8ae3SPaolo Bonzini #include <linux/mm.h> 36c50d8ae3SPaolo Bonzini #include <linux/highmem.h> 37c50d8ae3SPaolo Bonzini #include <linux/moduleparam.h> 38c50d8ae3SPaolo Bonzini #include <linux/export.h> 39c50d8ae3SPaolo Bonzini #include <linux/swap.h> 40c50d8ae3SPaolo Bonzini #include <linux/hugetlb.h> 41c50d8ae3SPaolo Bonzini #include <linux/compiler.h> 42c50d8ae3SPaolo Bonzini #include <linux/srcu.h> 43c50d8ae3SPaolo Bonzini #include <linux/slab.h> 44c50d8ae3SPaolo Bonzini #include <linux/sched/signal.h> 45c50d8ae3SPaolo Bonzini #include <linux/uaccess.h> 46c50d8ae3SPaolo Bonzini #include <linux/hash.h> 47c50d8ae3SPaolo Bonzini #include <linux/kern_levels.h> 4811b36fe7SChristophe JAILLET #include <linux/kstrtox.h> 49c50d8ae3SPaolo Bonzini #include <linux/kthread.h> 5066a5c40fSTanzir Hasan #include <linux/wordpart.h> 51c50d8ae3SPaolo Bonzini 52c50d8ae3SPaolo Bonzini #include <asm/page.h> 53eb243d1dSIngo Molnar #include <asm/memtype.h> 54c50d8ae3SPaolo Bonzini #include <asm/cmpxchg.h> 55c50d8ae3SPaolo Bonzini #include <asm/io.h> 564a98623dSSean Christopherson #include <asm/set_memory.h> 5765efc4dcSThomas Gleixner #include <asm/spec-ctrl.h> 58c50d8ae3SPaolo Bonzini #include <asm/vmx.h> 5958ea7cf7SSean Christopherson 60c50d8ae3SPaolo Bonzini #include "trace.h" 61c50d8ae3SPaolo Bonzini 620b210fafSSean Christopherson static bool nx_hugepage_mitigation_hard_disabled; 630b210fafSSean Christopherson 64a9d6496dSShaokun Zhang int __read_mostly nx_huge_pages = -1; 654dfe4f40SJunaid Shahid static uint __read_mostly nx_huge_pages_recovery_period_ms; 66c50d8ae3SPaolo Bonzini #ifdef CONFIG_PREEMPT_RT 67c50d8ae3SPaolo Bonzini /* Recovery can cause latency spikes, disable it for PREEMPT_RT. */ 68c50d8ae3SPaolo Bonzini static uint __read_mostly nx_huge_pages_recovery_ratio = 0; 69c50d8ae3SPaolo Bonzini #else 70c50d8ae3SPaolo Bonzini static uint __read_mostly nx_huge_pages_recovery_ratio = 60; 71c50d8ae3SPaolo Bonzini #endif 72c50d8ae3SPaolo Bonzini 730b210fafSSean Christopherson static int get_nx_huge_pages(char *buffer, const struct kernel_param *kp); 74c50d8ae3SPaolo Bonzini static int set_nx_huge_pages(const char *val, const struct kernel_param *kp); 754dfe4f40SJunaid Shahid static int set_nx_huge_pages_recovery_param(const char *val, const struct kernel_param *kp); 76c50d8ae3SPaolo Bonzini 77d5d6c18dSJoe Perches static const struct kernel_param_ops nx_huge_pages_ops = { 78c50d8ae3SPaolo Bonzini .set = set_nx_huge_pages, 790b210fafSSean Christopherson .get = get_nx_huge_pages, 80c50d8ae3SPaolo Bonzini }; 81c50d8ae3SPaolo Bonzini 824dfe4f40SJunaid Shahid static const struct kernel_param_ops nx_huge_pages_recovery_param_ops = { 834dfe4f40SJunaid Shahid .set = set_nx_huge_pages_recovery_param, 84c50d8ae3SPaolo Bonzini .get = param_get_uint, 85c50d8ae3SPaolo Bonzini }; 86c50d8ae3SPaolo Bonzini 87c50d8ae3SPaolo Bonzini module_param_cb(nx_huge_pages, &nx_huge_pages_ops, &nx_huge_pages, 0644); 88c50d8ae3SPaolo Bonzini __MODULE_PARM_TYPE(nx_huge_pages, "bool"); 894dfe4f40SJunaid Shahid module_param_cb(nx_huge_pages_recovery_ratio, &nx_huge_pages_recovery_param_ops, 90c50d8ae3SPaolo Bonzini &nx_huge_pages_recovery_ratio, 0644); 91c50d8ae3SPaolo Bonzini __MODULE_PARM_TYPE(nx_huge_pages_recovery_ratio, "uint"); 924dfe4f40SJunaid Shahid module_param_cb(nx_huge_pages_recovery_period_ms, &nx_huge_pages_recovery_param_ops, 934dfe4f40SJunaid Shahid &nx_huge_pages_recovery_period_ms, 0644); 944dfe4f40SJunaid Shahid __MODULE_PARM_TYPE(nx_huge_pages_recovery_period_ms, "uint"); 95c50d8ae3SPaolo Bonzini 9671fe7013SSean Christopherson static bool __read_mostly force_flush_and_sync_on_reuse; 9771fe7013SSean Christopherson module_param_named(flush_on_reuse, force_flush_and_sync_on_reuse, bool, 0644); 9871fe7013SSean Christopherson 99c50d8ae3SPaolo Bonzini /* 100c50d8ae3SPaolo Bonzini * When setting this variable to true it enables Two-Dimensional-Paging 101c50d8ae3SPaolo Bonzini * where the hardware walks 2 page tables: 102c50d8ae3SPaolo Bonzini * 1. the guest-virtual to guest-physical 103c50d8ae3SPaolo Bonzini * 2. while doing 1. it walks guest-physical to host-physical 104c50d8ae3SPaolo Bonzini * If the hardware supports that we don't need to do shadow paging. 105c50d8ae3SPaolo Bonzini */ 106c50d8ae3SPaolo Bonzini bool tdp_enabled = false; 107c50d8ae3SPaolo Bonzini 1087f604e92SDavid Matlack static bool __ro_after_init tdp_mmu_allowed; 1091f98f2bdSDavid Matlack 1101f98f2bdSDavid Matlack #ifdef CONFIG_X86_64 1111f98f2bdSDavid Matlack bool __read_mostly tdp_mmu_enabled = true; 1121f98f2bdSDavid Matlack module_param_named(tdp_mmu, tdp_mmu_enabled, bool, 0444); 1131f98f2bdSDavid Matlack #endif 1141f98f2bdSDavid Matlack 1151d92d2e8SSean Christopherson static int max_huge_page_level __read_mostly; 116746700d2SWei Huang static int tdp_root_level __read_mostly; 11783013059SSean Christopherson static int max_tdp_level __read_mostly; 118703c335dSSean Christopherson 119c50d8ae3SPaolo Bonzini #define PTE_PREFETCH_NUM 8 120c50d8ae3SPaolo Bonzini 121c50d8ae3SPaolo Bonzini #include <trace/events/kvm.h> 122c50d8ae3SPaolo Bonzini 123dc1cff96SPeter Xu /* make pte_list_desc fit well in cache lines */ 12413236e25SPeter Xu #define PTE_LIST_EXT 14 125c50d8ae3SPaolo Bonzini 12613236e25SPeter Xu /* 127141705b7SLai Jiangshan * struct pte_list_desc is the core data structure used to implement a custom 128141705b7SLai Jiangshan * list for tracking a set of related SPTEs, e.g. all the SPTEs that map a 129141705b7SLai Jiangshan * given GFN when used in the context of rmaps. Using a custom list allows KVM 130141705b7SLai Jiangshan * to optimize for the common case where many GFNs will have at most a handful 131141705b7SLai Jiangshan * of SPTEs pointing at them, i.e. allows packing multiple SPTEs into a small 132141705b7SLai Jiangshan * memory footprint, which in turn improves runtime performance by exploiting 133141705b7SLai Jiangshan * cache locality. 134141705b7SLai Jiangshan * 135141705b7SLai Jiangshan * A list is comprised of one or more pte_list_desc objects (descriptors). 136141705b7SLai Jiangshan * Each individual descriptor stores up to PTE_LIST_EXT SPTEs. If a descriptor 137141705b7SLai Jiangshan * is full and a new SPTEs needs to be added, a new descriptor is allocated and 138141705b7SLai Jiangshan * becomes the head of the list. This means that by definitions, all tail 139141705b7SLai Jiangshan * descriptors are full. 140141705b7SLai Jiangshan * 141141705b7SLai Jiangshan * Note, the meta data fields are deliberately placed at the start of the 142141705b7SLai Jiangshan * structure to optimize the cacheline layout; accessing the descriptor will 143141705b7SLai Jiangshan * touch only a single cacheline so long as @spte_count<=6 (or if only the 144141705b7SLai Jiangshan * descriptors metadata is accessed). 14513236e25SPeter Xu */ 146c50d8ae3SPaolo Bonzini struct pte_list_desc { 147c50d8ae3SPaolo Bonzini struct pte_list_desc *more; 148141705b7SLai Jiangshan /* The number of PTEs stored in _this_ descriptor. */ 149141705b7SLai Jiangshan u32 spte_count; 150141705b7SLai Jiangshan /* The number of PTEs stored in all tails of this descriptor. */ 151141705b7SLai Jiangshan u32 tail_count; 15213236e25SPeter Xu u64 *sptes[PTE_LIST_EXT]; 153c50d8ae3SPaolo Bonzini }; 154c50d8ae3SPaolo Bonzini 155c50d8ae3SPaolo Bonzini struct kvm_shadow_walk_iterator { 156c50d8ae3SPaolo Bonzini u64 addr; 157c50d8ae3SPaolo Bonzini hpa_t shadow_addr; 158c50d8ae3SPaolo Bonzini u64 *sptep; 159c50d8ae3SPaolo Bonzini int level; 160c50d8ae3SPaolo Bonzini unsigned index; 161c50d8ae3SPaolo Bonzini }; 162c50d8ae3SPaolo Bonzini 163c50d8ae3SPaolo Bonzini #define for_each_shadow_entry_using_root(_vcpu, _root, _addr, _walker) \ 164c50d8ae3SPaolo Bonzini for (shadow_walk_init_using_root(&(_walker), (_vcpu), \ 165c50d8ae3SPaolo Bonzini (_root), (_addr)); \ 166c50d8ae3SPaolo Bonzini shadow_walk_okay(&(_walker)); \ 167c50d8ae3SPaolo Bonzini shadow_walk_next(&(_walker))) 168c50d8ae3SPaolo Bonzini 169c50d8ae3SPaolo Bonzini #define for_each_shadow_entry(_vcpu, _addr, _walker) \ 170c50d8ae3SPaolo Bonzini for (shadow_walk_init(&(_walker), _vcpu, _addr); \ 171c50d8ae3SPaolo Bonzini shadow_walk_okay(&(_walker)); \ 172c50d8ae3SPaolo Bonzini shadow_walk_next(&(_walker))) 173c50d8ae3SPaolo Bonzini 174c50d8ae3SPaolo Bonzini #define for_each_shadow_entry_lockless(_vcpu, _addr, _walker, spte) \ 175c50d8ae3SPaolo Bonzini for (shadow_walk_init(&(_walker), _vcpu, _addr); \ 176c50d8ae3SPaolo Bonzini shadow_walk_okay(&(_walker)) && \ 177c50d8ae3SPaolo Bonzini ({ spte = mmu_spte_get_lockless(_walker.sptep); 1; }); \ 178c50d8ae3SPaolo Bonzini __shadow_walk_next(&(_walker), spte)) 179c50d8ae3SPaolo Bonzini 180c50d8ae3SPaolo Bonzini static struct kmem_cache *pte_list_desc_cache; 18102c00b3aSBen Gardon struct kmem_cache *mmu_page_header_cache; 182c50d8ae3SPaolo Bonzini static struct percpu_counter kvm_total_used_mmu_pages; 183c50d8ae3SPaolo Bonzini 184c50d8ae3SPaolo Bonzini static void mmu_spte_set(u64 *sptep, u64 spte); 185c50d8ae3SPaolo Bonzini 186594e91a1SSean Christopherson struct kvm_mmu_role_regs { 187594e91a1SSean Christopherson const unsigned long cr0; 188594e91a1SSean Christopherson const unsigned long cr4; 189594e91a1SSean Christopherson const u64 efer; 190594e91a1SSean Christopherson }; 191594e91a1SSean Christopherson 192c50d8ae3SPaolo Bonzini #define CREATE_TRACE_POINTS 193c50d8ae3SPaolo Bonzini #include "mmutrace.h" 194c50d8ae3SPaolo Bonzini 195594e91a1SSean Christopherson /* 196594e91a1SSean Christopherson * Yes, lot's of underscores. They're a hint that you probably shouldn't be 1977a458f0eSPaolo Bonzini * reading from the role_regs. Once the root_role is constructed, it becomes 198594e91a1SSean Christopherson * the single source of truth for the MMU's state. 199594e91a1SSean Christopherson */ 200594e91a1SSean Christopherson #define BUILD_MMU_ROLE_REGS_ACCESSOR(reg, name, flag) \ 20182ffa13fSPaolo Bonzini static inline bool __maybe_unused \ 20282ffa13fSPaolo Bonzini ____is_##reg##_##name(const struct kvm_mmu_role_regs *regs) \ 203594e91a1SSean Christopherson { \ 204594e91a1SSean Christopherson return !!(regs->reg & flag); \ 205594e91a1SSean Christopherson } 206594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr0, pg, X86_CR0_PG); 207594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr0, wp, X86_CR0_WP); 208594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pse, X86_CR4_PSE); 209594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pae, X86_CR4_PAE); 210594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, smep, X86_CR4_SMEP); 211594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, smap, X86_CR4_SMAP); 212594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pke, X86_CR4_PKE); 213594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, la57, X86_CR4_LA57); 214594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(efer, nx, EFER_NX); 215594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(efer, lma, EFER_LMA); 216594e91a1SSean Christopherson 21760667724SSean Christopherson /* 21860667724SSean Christopherson * The MMU itself (with a valid role) is the single source of truth for the 21960667724SSean Christopherson * MMU. Do not use the regs used to build the MMU/role, nor the vCPU. The 22060667724SSean Christopherson * regs don't account for dependencies, e.g. clearing CR4 bits if CR0.PG=1, 22160667724SSean Christopherson * and the vCPU may be incorrect/irrelevant. 22260667724SSean Christopherson */ 22360667724SSean Christopherson #define BUILD_MMU_ROLE_ACCESSOR(base_or_ext, reg, name) \ 2244ac21457SPaolo Bonzini static inline bool __maybe_unused is_##reg##_##name(struct kvm_mmu *mmu) \ 22560667724SSean Christopherson { \ 226e5ed0fb0SPaolo Bonzini return !!(mmu->cpu_role. base_or_ext . reg##_##name); \ 22760667724SSean Christopherson } 22860667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(base, cr0, wp); 22960667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(ext, cr4, pse); 23060667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(ext, cr4, smep); 23160667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(ext, cr4, smap); 23260667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(ext, cr4, pke); 23360667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(ext, cr4, la57); 23460667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(base, efer, nx); 23556b321f9SPaolo Bonzini BUILD_MMU_ROLE_ACCESSOR(ext, efer, lma); 23660667724SSean Christopherson 237faf72962SPaolo Bonzini static inline bool is_cr0_pg(struct kvm_mmu *mmu) 238faf72962SPaolo Bonzini { 239faf72962SPaolo Bonzini return mmu->cpu_role.base.level > 0; 240faf72962SPaolo Bonzini } 241faf72962SPaolo Bonzini 242faf72962SPaolo Bonzini static inline bool is_cr4_pae(struct kvm_mmu *mmu) 243faf72962SPaolo Bonzini { 244faf72962SPaolo Bonzini return !mmu->cpu_role.base.has_4_byte_gpte; 245faf72962SPaolo Bonzini } 246faf72962SPaolo Bonzini 247594e91a1SSean Christopherson static struct kvm_mmu_role_regs vcpu_to_role_regs(struct kvm_vcpu *vcpu) 248594e91a1SSean Christopherson { 249594e91a1SSean Christopherson struct kvm_mmu_role_regs regs = { 250594e91a1SSean Christopherson .cr0 = kvm_read_cr0_bits(vcpu, KVM_MMU_CR0_ROLE_BITS), 251594e91a1SSean Christopherson .cr4 = kvm_read_cr4_bits(vcpu, KVM_MMU_CR4_ROLE_BITS), 252594e91a1SSean Christopherson .efer = vcpu->arch.efer, 253594e91a1SSean Christopherson }; 254594e91a1SSean Christopherson 255594e91a1SSean Christopherson return regs; 256594e91a1SSean Christopherson } 257c50d8ae3SPaolo Bonzini 2582fdcc1b3SPaolo Bonzini static unsigned long get_guest_cr3(struct kvm_vcpu *vcpu) 259c50d8ae3SPaolo Bonzini { 2602fdcc1b3SPaolo Bonzini return kvm_read_cr3(vcpu); 261c50d8ae3SPaolo Bonzini } 262c50d8ae3SPaolo Bonzini 2632fdcc1b3SPaolo Bonzini static inline unsigned long kvm_mmu_get_guest_pgd(struct kvm_vcpu *vcpu, 2642fdcc1b3SPaolo Bonzini struct kvm_mmu *mmu) 265c50d8ae3SPaolo Bonzini { 266aefb2f2eSBreno Leitao if (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) && mmu->get_guest_pgd == get_guest_cr3) 2672fdcc1b3SPaolo Bonzini return kvm_read_cr3(vcpu); 268c50d8ae3SPaolo Bonzini 2692fdcc1b3SPaolo Bonzini return mmu->get_guest_pgd(vcpu); 2702fdcc1b3SPaolo Bonzini } 271c50d8ae3SPaolo Bonzini 2728a1300ffSSean Christopherson static inline bool kvm_available_flush_remote_tlbs_range(void) 273c50d8ae3SPaolo Bonzini { 2740277022aSSean Christopherson #if IS_ENABLED(CONFIG_HYPERV) 2758a1300ffSSean Christopherson return kvm_x86_ops.flush_remote_tlbs_range; 2760277022aSSean Christopherson #else 2770277022aSSean Christopherson return false; 2780277022aSSean Christopherson #endif 279c50d8ae3SPaolo Bonzini } 280c50d8ae3SPaolo Bonzini 2811b2dc736SHou Wenlong static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index); 2821b2dc736SHou Wenlong 2831b2dc736SHou Wenlong /* Flush the range of guest memory mapped by the given SPTE. */ 2841b2dc736SHou Wenlong static void kvm_flush_remote_tlbs_sptep(struct kvm *kvm, u64 *sptep) 2851b2dc736SHou Wenlong { 2861b2dc736SHou Wenlong struct kvm_mmu_page *sp = sptep_to_sp(sptep); 2871b2dc736SHou Wenlong gfn_t gfn = kvm_mmu_page_get_gfn(sp, spte_index(sptep)); 2881b2dc736SHou Wenlong 2891b2dc736SHou Wenlong kvm_flush_remote_tlbs_gfn(kvm, gfn, sp->role.level); 2901b2dc736SHou Wenlong } 2911b2dc736SHou Wenlong 2928f79b064SBen Gardon static void mark_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, u64 gfn, 2938f79b064SBen Gardon unsigned int access) 2948f79b064SBen Gardon { 295c236d962SSean Christopherson u64 spte = make_mmio_spte(vcpu, gfn, access); 2968f79b064SBen Gardon 297c236d962SSean Christopherson trace_mark_mmio_spte(sptep, gfn, spte); 298c236d962SSean Christopherson mmu_spte_set(sptep, spte); 299c50d8ae3SPaolo Bonzini } 300c50d8ae3SPaolo Bonzini 301c50d8ae3SPaolo Bonzini static gfn_t get_mmio_spte_gfn(u64 spte) 302c50d8ae3SPaolo Bonzini { 303c50d8ae3SPaolo Bonzini u64 gpa = spte & shadow_nonpresent_or_rsvd_lower_gfn_mask; 304c50d8ae3SPaolo Bonzini 3058a967d65SPaolo Bonzini gpa |= (spte >> SHADOW_NONPRESENT_OR_RSVD_MASK_LEN) 306c50d8ae3SPaolo Bonzini & shadow_nonpresent_or_rsvd_mask; 307c50d8ae3SPaolo Bonzini 308c50d8ae3SPaolo Bonzini return gpa >> PAGE_SHIFT; 309c50d8ae3SPaolo Bonzini } 310c50d8ae3SPaolo Bonzini 311c50d8ae3SPaolo Bonzini static unsigned get_mmio_spte_access(u64 spte) 312c50d8ae3SPaolo Bonzini { 313c50d8ae3SPaolo Bonzini return spte & shadow_mmio_access_mask; 314c50d8ae3SPaolo Bonzini } 315c50d8ae3SPaolo Bonzini 316c50d8ae3SPaolo Bonzini static bool check_mmio_spte(struct kvm_vcpu *vcpu, u64 spte) 317c50d8ae3SPaolo Bonzini { 318c50d8ae3SPaolo Bonzini u64 kvm_gen, spte_gen, gen; 319c50d8ae3SPaolo Bonzini 320c50d8ae3SPaolo Bonzini gen = kvm_vcpu_memslots(vcpu)->generation; 321c50d8ae3SPaolo Bonzini if (unlikely(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS)) 322c50d8ae3SPaolo Bonzini return false; 323c50d8ae3SPaolo Bonzini 324c50d8ae3SPaolo Bonzini kvm_gen = gen & MMIO_SPTE_GEN_MASK; 325c50d8ae3SPaolo Bonzini spte_gen = get_mmio_spte_generation(spte); 326c50d8ae3SPaolo Bonzini 327c50d8ae3SPaolo Bonzini trace_check_mmio_spte(spte, kvm_gen, spte_gen); 328c50d8ae3SPaolo Bonzini return likely(kvm_gen == spte_gen); 329c50d8ae3SPaolo Bonzini } 330c50d8ae3SPaolo Bonzini 331c50d8ae3SPaolo Bonzini static int is_cpuid_PSE36(void) 332c50d8ae3SPaolo Bonzini { 333c50d8ae3SPaolo Bonzini return 1; 334c50d8ae3SPaolo Bonzini } 335c50d8ae3SPaolo Bonzini 336c50d8ae3SPaolo Bonzini #ifdef CONFIG_X86_64 337c50d8ae3SPaolo Bonzini static void __set_spte(u64 *sptep, u64 spte) 338c50d8ae3SPaolo Bonzini { 339837d557aSSean Christopherson KVM_MMU_WARN_ON(is_ept_ve_possible(spte)); 340c50d8ae3SPaolo Bonzini WRITE_ONCE(*sptep, spte); 341c50d8ae3SPaolo Bonzini } 342c50d8ae3SPaolo Bonzini 343c50d8ae3SPaolo Bonzini static void __update_clear_spte_fast(u64 *sptep, u64 spte) 344c50d8ae3SPaolo Bonzini { 345837d557aSSean Christopherson KVM_MMU_WARN_ON(is_ept_ve_possible(spte)); 346c50d8ae3SPaolo Bonzini WRITE_ONCE(*sptep, spte); 347c50d8ae3SPaolo Bonzini } 348c50d8ae3SPaolo Bonzini 349c50d8ae3SPaolo Bonzini static u64 __update_clear_spte_slow(u64 *sptep, u64 spte) 350c50d8ae3SPaolo Bonzini { 351837d557aSSean Christopherson KVM_MMU_WARN_ON(is_ept_ve_possible(spte)); 352c50d8ae3SPaolo Bonzini return xchg(sptep, spte); 353c50d8ae3SPaolo Bonzini } 354c50d8ae3SPaolo Bonzini 355c50d8ae3SPaolo Bonzini static u64 __get_spte_lockless(u64 *sptep) 356c50d8ae3SPaolo Bonzini { 357c50d8ae3SPaolo Bonzini return READ_ONCE(*sptep); 358c50d8ae3SPaolo Bonzini } 359c50d8ae3SPaolo Bonzini #else 360c50d8ae3SPaolo Bonzini union split_spte { 361c50d8ae3SPaolo Bonzini struct { 362c50d8ae3SPaolo Bonzini u32 spte_low; 363c50d8ae3SPaolo Bonzini u32 spte_high; 364c50d8ae3SPaolo Bonzini }; 365c50d8ae3SPaolo Bonzini u64 spte; 366c50d8ae3SPaolo Bonzini }; 367c50d8ae3SPaolo Bonzini 368c50d8ae3SPaolo Bonzini static void count_spte_clear(u64 *sptep, u64 spte) 369c50d8ae3SPaolo Bonzini { 37057354682SSean Christopherson struct kvm_mmu_page *sp = sptep_to_sp(sptep); 371c50d8ae3SPaolo Bonzini 372c50d8ae3SPaolo Bonzini if (is_shadow_present_pte(spte)) 373c50d8ae3SPaolo Bonzini return; 374c50d8ae3SPaolo Bonzini 375c50d8ae3SPaolo Bonzini /* Ensure the spte is completely set before we increase the count */ 376c50d8ae3SPaolo Bonzini smp_wmb(); 377c50d8ae3SPaolo Bonzini sp->clear_spte_count++; 378c50d8ae3SPaolo Bonzini } 379c50d8ae3SPaolo Bonzini 380c50d8ae3SPaolo Bonzini static void __set_spte(u64 *sptep, u64 spte) 381c50d8ae3SPaolo Bonzini { 382c50d8ae3SPaolo Bonzini union split_spte *ssptep, sspte; 383c50d8ae3SPaolo Bonzini 384c50d8ae3SPaolo Bonzini ssptep = (union split_spte *)sptep; 385c50d8ae3SPaolo Bonzini sspte = (union split_spte)spte; 386c50d8ae3SPaolo Bonzini 387c50d8ae3SPaolo Bonzini ssptep->spte_high = sspte.spte_high; 388c50d8ae3SPaolo Bonzini 389c50d8ae3SPaolo Bonzini /* 390c50d8ae3SPaolo Bonzini * If we map the spte from nonpresent to present, We should store 391c50d8ae3SPaolo Bonzini * the high bits firstly, then set present bit, so cpu can not 392c50d8ae3SPaolo Bonzini * fetch this spte while we are setting the spte. 393c50d8ae3SPaolo Bonzini */ 394c50d8ae3SPaolo Bonzini smp_wmb(); 395c50d8ae3SPaolo Bonzini 396c50d8ae3SPaolo Bonzini WRITE_ONCE(ssptep->spte_low, sspte.spte_low); 397c50d8ae3SPaolo Bonzini } 398c50d8ae3SPaolo Bonzini 399c50d8ae3SPaolo Bonzini static void __update_clear_spte_fast(u64 *sptep, u64 spte) 400c50d8ae3SPaolo Bonzini { 401c50d8ae3SPaolo Bonzini union split_spte *ssptep, sspte; 402c50d8ae3SPaolo Bonzini 403c50d8ae3SPaolo Bonzini ssptep = (union split_spte *)sptep; 404c50d8ae3SPaolo Bonzini sspte = (union split_spte)spte; 405c50d8ae3SPaolo Bonzini 406c50d8ae3SPaolo Bonzini WRITE_ONCE(ssptep->spte_low, sspte.spte_low); 407c50d8ae3SPaolo Bonzini 408c50d8ae3SPaolo Bonzini /* 409c50d8ae3SPaolo Bonzini * If we map the spte from present to nonpresent, we should clear 410c50d8ae3SPaolo Bonzini * present bit firstly to avoid vcpu fetch the old high bits. 411c50d8ae3SPaolo Bonzini */ 412c50d8ae3SPaolo Bonzini smp_wmb(); 413c50d8ae3SPaolo Bonzini 414c50d8ae3SPaolo Bonzini ssptep->spte_high = sspte.spte_high; 415c50d8ae3SPaolo Bonzini count_spte_clear(sptep, spte); 416c50d8ae3SPaolo Bonzini } 417c50d8ae3SPaolo Bonzini 418c50d8ae3SPaolo Bonzini static u64 __update_clear_spte_slow(u64 *sptep, u64 spte) 419c50d8ae3SPaolo Bonzini { 420c50d8ae3SPaolo Bonzini union split_spte *ssptep, sspte, orig; 421c50d8ae3SPaolo Bonzini 422c50d8ae3SPaolo Bonzini ssptep = (union split_spte *)sptep; 423c50d8ae3SPaolo Bonzini sspte = (union split_spte)spte; 424c50d8ae3SPaolo Bonzini 425c50d8ae3SPaolo Bonzini /* xchg acts as a barrier before the setting of the high bits */ 426c50d8ae3SPaolo Bonzini orig.spte_low = xchg(&ssptep->spte_low, sspte.spte_low); 427c50d8ae3SPaolo Bonzini orig.spte_high = ssptep->spte_high; 428c50d8ae3SPaolo Bonzini ssptep->spte_high = sspte.spte_high; 429c50d8ae3SPaolo Bonzini count_spte_clear(sptep, spte); 430c50d8ae3SPaolo Bonzini 431c50d8ae3SPaolo Bonzini return orig.spte; 432c50d8ae3SPaolo Bonzini } 433c50d8ae3SPaolo Bonzini 434c50d8ae3SPaolo Bonzini /* 435c50d8ae3SPaolo Bonzini * The idea using the light way get the spte on x86_32 guest is from 436c50d8ae3SPaolo Bonzini * gup_get_pte (mm/gup.c). 437c50d8ae3SPaolo Bonzini * 438f3b65bbaSPaolo Bonzini * An spte tlb flush may be pending, because they are coalesced and 439f3b65bbaSPaolo Bonzini * we are running out of the MMU lock. Therefore 440c50d8ae3SPaolo Bonzini * we need to protect against in-progress updates of the spte. 441c50d8ae3SPaolo Bonzini * 442c50d8ae3SPaolo Bonzini * Reading the spte while an update is in progress may get the old value 443c50d8ae3SPaolo Bonzini * for the high part of the spte. The race is fine for a present->non-present 444c50d8ae3SPaolo Bonzini * change (because the high part of the spte is ignored for non-present spte), 445c50d8ae3SPaolo Bonzini * but for a present->present change we must reread the spte. 446c50d8ae3SPaolo Bonzini * 447c50d8ae3SPaolo Bonzini * All such changes are done in two steps (present->non-present and 448c50d8ae3SPaolo Bonzini * non-present->present), hence it is enough to count the number of 449c50d8ae3SPaolo Bonzini * present->non-present updates: if it changed while reading the spte, 450c50d8ae3SPaolo Bonzini * we might have hit the race. This is done using clear_spte_count. 451c50d8ae3SPaolo Bonzini */ 452c50d8ae3SPaolo Bonzini static u64 __get_spte_lockless(u64 *sptep) 453c50d8ae3SPaolo Bonzini { 45457354682SSean Christopherson struct kvm_mmu_page *sp = sptep_to_sp(sptep); 455c50d8ae3SPaolo Bonzini union split_spte spte, *orig = (union split_spte *)sptep; 456c50d8ae3SPaolo Bonzini int count; 457c50d8ae3SPaolo Bonzini 458c50d8ae3SPaolo Bonzini retry: 459c50d8ae3SPaolo Bonzini count = sp->clear_spte_count; 460c50d8ae3SPaolo Bonzini smp_rmb(); 461c50d8ae3SPaolo Bonzini 462c50d8ae3SPaolo Bonzini spte.spte_low = orig->spte_low; 463c50d8ae3SPaolo Bonzini smp_rmb(); 464c50d8ae3SPaolo Bonzini 465c50d8ae3SPaolo Bonzini spte.spte_high = orig->spte_high; 466c50d8ae3SPaolo Bonzini smp_rmb(); 467c50d8ae3SPaolo Bonzini 468c50d8ae3SPaolo Bonzini if (unlikely(spte.spte_low != orig->spte_low || 469c50d8ae3SPaolo Bonzini count != sp->clear_spte_count)) 470c50d8ae3SPaolo Bonzini goto retry; 471c50d8ae3SPaolo Bonzini 472c50d8ae3SPaolo Bonzini return spte.spte; 473c50d8ae3SPaolo Bonzini } 474c50d8ae3SPaolo Bonzini #endif 475c50d8ae3SPaolo Bonzini 476c50d8ae3SPaolo Bonzini /* Rules for using mmu_spte_set: 477c50d8ae3SPaolo Bonzini * Set the sptep from nonpresent to present. 478c50d8ae3SPaolo Bonzini * Note: the sptep being assigned *must* be either not present 479c50d8ae3SPaolo Bonzini * or in a state where the hardware will not attempt to update 480c50d8ae3SPaolo Bonzini * the spte. 481c50d8ae3SPaolo Bonzini */ 482c50d8ae3SPaolo Bonzini static void mmu_spte_set(u64 *sptep, u64 new_spte) 483c50d8ae3SPaolo Bonzini { 48420ba462dSSean Christopherson WARN_ON_ONCE(is_shadow_present_pte(*sptep)); 485c50d8ae3SPaolo Bonzini __set_spte(sptep, new_spte); 486c50d8ae3SPaolo Bonzini } 487c50d8ae3SPaolo Bonzini 488c50d8ae3SPaolo Bonzini /* 489c50d8ae3SPaolo Bonzini * Update the SPTE (excluding the PFN), but do not track changes in its 490c50d8ae3SPaolo Bonzini * accessed/dirty status. 491c50d8ae3SPaolo Bonzini */ 492c50d8ae3SPaolo Bonzini static u64 mmu_spte_update_no_track(u64 *sptep, u64 new_spte) 493c50d8ae3SPaolo Bonzini { 494c50d8ae3SPaolo Bonzini u64 old_spte = *sptep; 495c50d8ae3SPaolo Bonzini 49620ba462dSSean Christopherson WARN_ON_ONCE(!is_shadow_present_pte(new_spte)); 497115111efSDavid Matlack check_spte_writable_invariants(new_spte); 498c50d8ae3SPaolo Bonzini 499c50d8ae3SPaolo Bonzini if (!is_shadow_present_pte(old_spte)) { 500c50d8ae3SPaolo Bonzini mmu_spte_set(sptep, new_spte); 501c50d8ae3SPaolo Bonzini return old_spte; 502c50d8ae3SPaolo Bonzini } 503c50d8ae3SPaolo Bonzini 504c50d8ae3SPaolo Bonzini if (!spte_has_volatile_bits(old_spte)) 505c50d8ae3SPaolo Bonzini __update_clear_spte_fast(sptep, new_spte); 506c50d8ae3SPaolo Bonzini else 507c50d8ae3SPaolo Bonzini old_spte = __update_clear_spte_slow(sptep, new_spte); 508c50d8ae3SPaolo Bonzini 50920ba462dSSean Christopherson WARN_ON_ONCE(spte_to_pfn(old_spte) != spte_to_pfn(new_spte)); 510c50d8ae3SPaolo Bonzini 511c50d8ae3SPaolo Bonzini return old_spte; 512c50d8ae3SPaolo Bonzini } 513c50d8ae3SPaolo Bonzini 514c50d8ae3SPaolo Bonzini /* Rules for using mmu_spte_update: 515c50d8ae3SPaolo Bonzini * Update the state bits, it means the mapped pfn is not changed. 516c50d8ae3SPaolo Bonzini * 51702844ac1SDavid Matlack * Whenever an MMU-writable SPTE is overwritten with a read-only SPTE, remote 51802844ac1SDavid Matlack * TLBs must be flushed. Otherwise rmap_write_protect will find a read-only 51902844ac1SDavid Matlack * spte, even though the writable spte might be cached on a CPU's TLB. 520c50d8ae3SPaolo Bonzini * 521c50d8ae3SPaolo Bonzini * Returns true if the TLB needs to be flushed 522c50d8ae3SPaolo Bonzini */ 523c50d8ae3SPaolo Bonzini static bool mmu_spte_update(u64 *sptep, u64 new_spte) 524c50d8ae3SPaolo Bonzini { 525c50d8ae3SPaolo Bonzini bool flush = false; 526c50d8ae3SPaolo Bonzini u64 old_spte = mmu_spte_update_no_track(sptep, new_spte); 527c50d8ae3SPaolo Bonzini 528c50d8ae3SPaolo Bonzini if (!is_shadow_present_pte(old_spte)) 529c50d8ae3SPaolo Bonzini return false; 530c50d8ae3SPaolo Bonzini 531c50d8ae3SPaolo Bonzini /* 532c50d8ae3SPaolo Bonzini * For the spte updated out of mmu-lock is safe, since 533c50d8ae3SPaolo Bonzini * we always atomically update it, see the comments in 534c50d8ae3SPaolo Bonzini * spte_has_volatile_bits(). 535c50d8ae3SPaolo Bonzini */ 536706c9c55SSean Christopherson if (is_mmu_writable_spte(old_spte) && 537c50d8ae3SPaolo Bonzini !is_writable_pte(new_spte)) 538c50d8ae3SPaolo Bonzini flush = true; 539c50d8ae3SPaolo Bonzini 540c50d8ae3SPaolo Bonzini /* 541c50d8ae3SPaolo Bonzini * Flush TLB when accessed/dirty states are changed in the page tables, 542c50d8ae3SPaolo Bonzini * to guarantee consistency between TLB and page tables. 543c50d8ae3SPaolo Bonzini */ 544c50d8ae3SPaolo Bonzini 545c50d8ae3SPaolo Bonzini if (is_accessed_spte(old_spte) && !is_accessed_spte(new_spte)) { 546c50d8ae3SPaolo Bonzini flush = true; 547c50d8ae3SPaolo Bonzini kvm_set_pfn_accessed(spte_to_pfn(old_spte)); 548c50d8ae3SPaolo Bonzini } 549c50d8ae3SPaolo Bonzini 550c50d8ae3SPaolo Bonzini if (is_dirty_spte(old_spte) && !is_dirty_spte(new_spte)) { 551c50d8ae3SPaolo Bonzini flush = true; 552c50d8ae3SPaolo Bonzini kvm_set_pfn_dirty(spte_to_pfn(old_spte)); 553c50d8ae3SPaolo Bonzini } 554c50d8ae3SPaolo Bonzini 555c50d8ae3SPaolo Bonzini return flush; 556c50d8ae3SPaolo Bonzini } 557c50d8ae3SPaolo Bonzini 558c50d8ae3SPaolo Bonzini /* 559c50d8ae3SPaolo Bonzini * Rules for using mmu_spte_clear_track_bits: 560c50d8ae3SPaolo Bonzini * It sets the sptep from present to nonpresent, and track the 561c50d8ae3SPaolo Bonzini * state bits, it is used to clear the last level sptep. 5627fa2a347SSean Christopherson * Returns the old PTE. 563c50d8ae3SPaolo Bonzini */ 56435d539c3SSean Christopherson static u64 mmu_spte_clear_track_bits(struct kvm *kvm, u64 *sptep) 565c50d8ae3SPaolo Bonzini { 566c50d8ae3SPaolo Bonzini kvm_pfn_t pfn; 567c50d8ae3SPaolo Bonzini u64 old_spte = *sptep; 56871f51d2cSMingwei Zhang int level = sptep_to_sp(sptep)->role.level; 569b14b2690SSean Christopherson struct page *page; 570c50d8ae3SPaolo Bonzini 57154eb3ef5SSean Christopherson if (!is_shadow_present_pte(old_spte) || 57254eb3ef5SSean Christopherson !spte_has_volatile_bits(old_spte)) 573d8fa2031SSean Christopherson __update_clear_spte_fast(sptep, SHADOW_NONPRESENT_VALUE); 574c50d8ae3SPaolo Bonzini else 575d8fa2031SSean Christopherson old_spte = __update_clear_spte_slow(sptep, SHADOW_NONPRESENT_VALUE); 576c50d8ae3SPaolo Bonzini 577c50d8ae3SPaolo Bonzini if (!is_shadow_present_pte(old_spte)) 5787fa2a347SSean Christopherson return old_spte; 579c50d8ae3SPaolo Bonzini 58071f51d2cSMingwei Zhang kvm_update_page_stats(kvm, level, -1); 58171f51d2cSMingwei Zhang 582c50d8ae3SPaolo Bonzini pfn = spte_to_pfn(old_spte); 583c50d8ae3SPaolo Bonzini 584c50d8ae3SPaolo Bonzini /* 585b14b2690SSean Christopherson * KVM doesn't hold a reference to any pages mapped into the guest, and 586b14b2690SSean Christopherson * instead uses the mmu_notifier to ensure that KVM unmaps any pages 587b14b2690SSean Christopherson * before they are reclaimed. Sanity check that, if the pfn is backed 588b14b2690SSean Christopherson * by a refcounted page, the refcount is elevated. 589c50d8ae3SPaolo Bonzini */ 590b14b2690SSean Christopherson page = kvm_pfn_to_refcounted_page(pfn); 59120ba462dSSean Christopherson WARN_ON_ONCE(page && !page_count(page)); 592c50d8ae3SPaolo Bonzini 593c50d8ae3SPaolo Bonzini if (is_accessed_spte(old_spte)) 594c50d8ae3SPaolo Bonzini kvm_set_pfn_accessed(pfn); 595c50d8ae3SPaolo Bonzini 596c50d8ae3SPaolo Bonzini if (is_dirty_spte(old_spte)) 597c50d8ae3SPaolo Bonzini kvm_set_pfn_dirty(pfn); 598c50d8ae3SPaolo Bonzini 5997fa2a347SSean Christopherson return old_spte; 600c50d8ae3SPaolo Bonzini } 601c50d8ae3SPaolo Bonzini 602c50d8ae3SPaolo Bonzini /* 603c50d8ae3SPaolo Bonzini * Rules for using mmu_spte_clear_no_track: 604c50d8ae3SPaolo Bonzini * Directly clear spte without caring the state bits of sptep, 605c50d8ae3SPaolo Bonzini * it is used to set the upper level spte. 606c50d8ae3SPaolo Bonzini */ 607c50d8ae3SPaolo Bonzini static void mmu_spte_clear_no_track(u64 *sptep) 608c50d8ae3SPaolo Bonzini { 609d8fa2031SSean Christopherson __update_clear_spte_fast(sptep, SHADOW_NONPRESENT_VALUE); 610c50d8ae3SPaolo Bonzini } 611c50d8ae3SPaolo Bonzini 612c50d8ae3SPaolo Bonzini static u64 mmu_spte_get_lockless(u64 *sptep) 613c50d8ae3SPaolo Bonzini { 614c50d8ae3SPaolo Bonzini return __get_spte_lockless(sptep); 615c50d8ae3SPaolo Bonzini } 616c50d8ae3SPaolo Bonzini 617c50d8ae3SPaolo Bonzini /* Returns the Accessed status of the PTE and resets it at the same time. */ 618c50d8ae3SPaolo Bonzini static bool mmu_spte_age(u64 *sptep) 619c50d8ae3SPaolo Bonzini { 620c50d8ae3SPaolo Bonzini u64 spte = mmu_spte_get_lockless(sptep); 621c50d8ae3SPaolo Bonzini 622c50d8ae3SPaolo Bonzini if (!is_accessed_spte(spte)) 623c50d8ae3SPaolo Bonzini return false; 624c50d8ae3SPaolo Bonzini 625c50d8ae3SPaolo Bonzini if (spte_ad_enabled(spte)) { 626c50d8ae3SPaolo Bonzini clear_bit((ffs(shadow_accessed_mask) - 1), 627c50d8ae3SPaolo Bonzini (unsigned long *)sptep); 628c50d8ae3SPaolo Bonzini } else { 629c50d8ae3SPaolo Bonzini /* 630c50d8ae3SPaolo Bonzini * Capture the dirty status of the page, so that it doesn't get 631c50d8ae3SPaolo Bonzini * lost when the SPTE is marked for access tracking. 632c50d8ae3SPaolo Bonzini */ 633c50d8ae3SPaolo Bonzini if (is_writable_pte(spte)) 634c50d8ae3SPaolo Bonzini kvm_set_pfn_dirty(spte_to_pfn(spte)); 635c50d8ae3SPaolo Bonzini 636c50d8ae3SPaolo Bonzini spte = mark_spte_for_access_track(spte); 637c50d8ae3SPaolo Bonzini mmu_spte_update_no_track(sptep, spte); 638c50d8ae3SPaolo Bonzini } 639c50d8ae3SPaolo Bonzini 640c50d8ae3SPaolo Bonzini return true; 641c50d8ae3SPaolo Bonzini } 642c50d8ae3SPaolo Bonzini 64378fdd2f0SSean Christopherson static inline bool is_tdp_mmu_active(struct kvm_vcpu *vcpu) 64478fdd2f0SSean Christopherson { 64578fdd2f0SSean Christopherson return tdp_mmu_enabled && vcpu->arch.mmu->root_role.direct; 64678fdd2f0SSean Christopherson } 64778fdd2f0SSean Christopherson 648c50d8ae3SPaolo Bonzini static void walk_shadow_page_lockless_begin(struct kvm_vcpu *vcpu) 649c50d8ae3SPaolo Bonzini { 65078fdd2f0SSean Christopherson if (is_tdp_mmu_active(vcpu)) { 651c5c8c7c5SDavid Matlack kvm_tdp_mmu_walk_lockless_begin(); 652c5c8c7c5SDavid Matlack } else { 653c50d8ae3SPaolo Bonzini /* 654c50d8ae3SPaolo Bonzini * Prevent page table teardown by making any free-er wait during 655c50d8ae3SPaolo Bonzini * kvm_flush_remote_tlbs() IPI to all active vcpus. 656c50d8ae3SPaolo Bonzini */ 657c50d8ae3SPaolo Bonzini local_irq_disable(); 658c50d8ae3SPaolo Bonzini 659c50d8ae3SPaolo Bonzini /* 660c50d8ae3SPaolo Bonzini * Make sure a following spte read is not reordered ahead of the write 661c50d8ae3SPaolo Bonzini * to vcpu->mode. 662c50d8ae3SPaolo Bonzini */ 663c50d8ae3SPaolo Bonzini smp_store_mb(vcpu->mode, READING_SHADOW_PAGE_TABLES); 664c50d8ae3SPaolo Bonzini } 665c5c8c7c5SDavid Matlack } 666c50d8ae3SPaolo Bonzini 667c50d8ae3SPaolo Bonzini static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu) 668c50d8ae3SPaolo Bonzini { 66978fdd2f0SSean Christopherson if (is_tdp_mmu_active(vcpu)) { 670c5c8c7c5SDavid Matlack kvm_tdp_mmu_walk_lockless_end(); 671c5c8c7c5SDavid Matlack } else { 672c50d8ae3SPaolo Bonzini /* 673c50d8ae3SPaolo Bonzini * Make sure the write to vcpu->mode is not reordered in front of 674c50d8ae3SPaolo Bonzini * reads to sptes. If it does, kvm_mmu_commit_zap_page() can see us 675c50d8ae3SPaolo Bonzini * OUTSIDE_GUEST_MODE and proceed to free the shadow page table. 676c50d8ae3SPaolo Bonzini */ 677c50d8ae3SPaolo Bonzini smp_store_release(&vcpu->mode, OUTSIDE_GUEST_MODE); 678c50d8ae3SPaolo Bonzini local_irq_enable(); 679c50d8ae3SPaolo Bonzini } 680c5c8c7c5SDavid Matlack } 681c50d8ae3SPaolo Bonzini 682378f5cd6SSean Christopherson static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu, bool maybe_indirect) 683c50d8ae3SPaolo Bonzini { 684c50d8ae3SPaolo Bonzini int r; 685c50d8ae3SPaolo Bonzini 686531281adSSean Christopherson /* 1 rmap, 1 parent PTE per level, and the prefetched rmaps. */ 68794ce87efSSean Christopherson r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache, 688531281adSSean Christopherson 1 + PT64_ROOT_MAX_LEVEL + PTE_PREFETCH_NUM); 689c50d8ae3SPaolo Bonzini if (r) 690c50d8ae3SPaolo Bonzini return r; 69194ce87efSSean Christopherson r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_shadow_page_cache, 692171a90d7SSean Christopherson PT64_ROOT_MAX_LEVEL); 693171a90d7SSean Christopherson if (r) 694171a90d7SSean Christopherson return r; 695378f5cd6SSean Christopherson if (maybe_indirect) { 6966a97575dSDavid Matlack r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_shadowed_info_cache, 697171a90d7SSean Christopherson PT64_ROOT_MAX_LEVEL); 698c50d8ae3SPaolo Bonzini if (r) 699c50d8ae3SPaolo Bonzini return r; 700378f5cd6SSean Christopherson } 70194ce87efSSean Christopherson return kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache, 702531281adSSean Christopherson PT64_ROOT_MAX_LEVEL); 703c50d8ae3SPaolo Bonzini } 704c50d8ae3SPaolo Bonzini 705c50d8ae3SPaolo Bonzini static void mmu_free_memory_caches(struct kvm_vcpu *vcpu) 706c50d8ae3SPaolo Bonzini { 70794ce87efSSean Christopherson kvm_mmu_free_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache); 70894ce87efSSean Christopherson kvm_mmu_free_memory_cache(&vcpu->arch.mmu_shadow_page_cache); 7096a97575dSDavid Matlack kvm_mmu_free_memory_cache(&vcpu->arch.mmu_shadowed_info_cache); 71094ce87efSSean Christopherson kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache); 711c50d8ae3SPaolo Bonzini } 712c50d8ae3SPaolo Bonzini 713c50d8ae3SPaolo Bonzini static void mmu_free_pte_list_desc(struct pte_list_desc *pte_list_desc) 714c50d8ae3SPaolo Bonzini { 715c50d8ae3SPaolo Bonzini kmem_cache_free(pte_list_desc_cache, pte_list_desc); 716c50d8ae3SPaolo Bonzini } 717c50d8ae3SPaolo Bonzini 7186a97575dSDavid Matlack static bool sp_has_gptes(struct kvm_mmu_page *sp); 7196a97575dSDavid Matlack 720c50d8ae3SPaolo Bonzini static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index) 721c50d8ae3SPaolo Bonzini { 72284e5ffd0SLai Jiangshan if (sp->role.passthrough) 72384e5ffd0SLai Jiangshan return sp->gfn; 72484e5ffd0SLai Jiangshan 725c50d8ae3SPaolo Bonzini if (!sp->role.direct) 7266a97575dSDavid Matlack return sp->shadowed_translation[index] >> PAGE_SHIFT; 727c50d8ae3SPaolo Bonzini 7282ca3129eSSean Christopherson return sp->gfn + (index << ((sp->role.level - 1) * SPTE_LEVEL_BITS)); 729c50d8ae3SPaolo Bonzini } 730c50d8ae3SPaolo Bonzini 7316a97575dSDavid Matlack /* 7326a97575dSDavid Matlack * For leaf SPTEs, fetch the *guest* access permissions being shadowed. Note 7336a97575dSDavid Matlack * that the SPTE itself may have a more constrained access permissions that 7346a97575dSDavid Matlack * what the guest enforces. For example, a guest may create an executable 7356a97575dSDavid Matlack * huge PTE but KVM may disallow execution to mitigate iTLB multihit. 7366a97575dSDavid Matlack */ 7376a97575dSDavid Matlack static u32 kvm_mmu_page_get_access(struct kvm_mmu_page *sp, int index) 738c50d8ae3SPaolo Bonzini { 7396a97575dSDavid Matlack if (sp_has_gptes(sp)) 7406a97575dSDavid Matlack return sp->shadowed_translation[index] & ACC_ALL; 7416a97575dSDavid Matlack 7426a97575dSDavid Matlack /* 7436a97575dSDavid Matlack * For direct MMUs (e.g. TDP or non-paging guests) or passthrough SPs, 7446a97575dSDavid Matlack * KVM is not shadowing any guest page tables, so the "guest access 7456a97575dSDavid Matlack * permissions" are just ACC_ALL. 7466a97575dSDavid Matlack * 7476a97575dSDavid Matlack * For direct SPs in indirect MMUs (shadow paging), i.e. when KVM 7486a97575dSDavid Matlack * is shadowing a guest huge page with small pages, the guest access 7496a97575dSDavid Matlack * permissions being shadowed are the access permissions of the huge 7506a97575dSDavid Matlack * page. 7516a97575dSDavid Matlack * 7526a97575dSDavid Matlack * In both cases, sp->role.access contains the correct access bits. 7536a97575dSDavid Matlack */ 7546a97575dSDavid Matlack return sp->role.access; 7556a97575dSDavid Matlack } 7566a97575dSDavid Matlack 75772ae5822SSean Christopherson static void kvm_mmu_page_set_translation(struct kvm_mmu_page *sp, int index, 75872ae5822SSean Christopherson gfn_t gfn, unsigned int access) 7596a97575dSDavid Matlack { 7606a97575dSDavid Matlack if (sp_has_gptes(sp)) { 7616a97575dSDavid Matlack sp->shadowed_translation[index] = (gfn << PAGE_SHIFT) | access; 76284e5ffd0SLai Jiangshan return; 76384e5ffd0SLai Jiangshan } 76484e5ffd0SLai Jiangshan 7656a97575dSDavid Matlack WARN_ONCE(access != kvm_mmu_page_get_access(sp, index), 7666a97575dSDavid Matlack "access mismatch under %s page %llx (expected %u, got %u)\n", 7676a97575dSDavid Matlack sp->role.passthrough ? "passthrough" : "direct", 7686a97575dSDavid Matlack sp->gfn, kvm_mmu_page_get_access(sp, index), access); 7696a97575dSDavid Matlack 7706a97575dSDavid Matlack WARN_ONCE(gfn != kvm_mmu_page_get_gfn(sp, index), 7716a97575dSDavid Matlack "gfn mismatch under %s page %llx (expected %llx, got %llx)\n", 7726a97575dSDavid Matlack sp->role.passthrough ? "passthrough" : "direct", 7736a97575dSDavid Matlack sp->gfn, kvm_mmu_page_get_gfn(sp, index), gfn); 774c50d8ae3SPaolo Bonzini } 775c50d8ae3SPaolo Bonzini 77672ae5822SSean Christopherson static void kvm_mmu_page_set_access(struct kvm_mmu_page *sp, int index, 77772ae5822SSean Christopherson unsigned int access) 7786a97575dSDavid Matlack { 7796a97575dSDavid Matlack gfn_t gfn = kvm_mmu_page_get_gfn(sp, index); 7806a97575dSDavid Matlack 7816a97575dSDavid Matlack kvm_mmu_page_set_translation(sp, index, gfn, access); 782c50d8ae3SPaolo Bonzini } 783c50d8ae3SPaolo Bonzini 784c50d8ae3SPaolo Bonzini /* 785c50d8ae3SPaolo Bonzini * Return the pointer to the large page information for a given gfn, 786c50d8ae3SPaolo Bonzini * handling slots that are not large page aligned. 787c50d8ae3SPaolo Bonzini */ 788c50d8ae3SPaolo Bonzini static struct kvm_lpage_info *lpage_info_slot(gfn_t gfn, 7898ca6f063SBen Gardon const struct kvm_memory_slot *slot, int level) 790c50d8ae3SPaolo Bonzini { 791c50d8ae3SPaolo Bonzini unsigned long idx; 792c50d8ae3SPaolo Bonzini 793c50d8ae3SPaolo Bonzini idx = gfn_to_index(gfn, slot->base_gfn, level); 794c50d8ae3SPaolo Bonzini return &slot->arch.lpage_info[level - 2][idx]; 795c50d8ae3SPaolo Bonzini } 796c50d8ae3SPaolo Bonzini 79790b4fe17SChao Peng /* 79890b4fe17SChao Peng * The most significant bit in disallow_lpage tracks whether or not memory 79990b4fe17SChao Peng * attributes are mixed, i.e. not identical for all gfns at the current level. 80090b4fe17SChao Peng * The lower order bits are used to refcount other cases where a hugepage is 80190b4fe17SChao Peng * disallowed, e.g. if KVM has shadow a page table at the gfn. 80290b4fe17SChao Peng */ 80390b4fe17SChao Peng #define KVM_LPAGE_MIXED_FLAG BIT(31) 80490b4fe17SChao Peng 805269e9552SHamza Mahfooz static void update_gfn_disallow_lpage_count(const struct kvm_memory_slot *slot, 806c50d8ae3SPaolo Bonzini gfn_t gfn, int count) 807c50d8ae3SPaolo Bonzini { 808c50d8ae3SPaolo Bonzini struct kvm_lpage_info *linfo; 80990b4fe17SChao Peng int old, i; 810c50d8ae3SPaolo Bonzini 8113bae0459SSean Christopherson for (i = PG_LEVEL_2M; i <= KVM_MAX_HUGEPAGE_LEVEL; ++i) { 812c50d8ae3SPaolo Bonzini linfo = lpage_info_slot(gfn, slot, i); 81390b4fe17SChao Peng 81490b4fe17SChao Peng old = linfo->disallow_lpage; 815c50d8ae3SPaolo Bonzini linfo->disallow_lpage += count; 81690b4fe17SChao Peng WARN_ON_ONCE((old ^ linfo->disallow_lpage) & KVM_LPAGE_MIXED_FLAG); 817c50d8ae3SPaolo Bonzini } 818c50d8ae3SPaolo Bonzini } 819c50d8ae3SPaolo Bonzini 820269e9552SHamza Mahfooz void kvm_mmu_gfn_disallow_lpage(const struct kvm_memory_slot *slot, gfn_t gfn) 821c50d8ae3SPaolo Bonzini { 822c50d8ae3SPaolo Bonzini update_gfn_disallow_lpage_count(slot, gfn, 1); 823c50d8ae3SPaolo Bonzini } 824c50d8ae3SPaolo Bonzini 825269e9552SHamza Mahfooz void kvm_mmu_gfn_allow_lpage(const struct kvm_memory_slot *slot, gfn_t gfn) 826c50d8ae3SPaolo Bonzini { 827c50d8ae3SPaolo Bonzini update_gfn_disallow_lpage_count(slot, gfn, -1); 828c50d8ae3SPaolo Bonzini } 829c50d8ae3SPaolo Bonzini 830c50d8ae3SPaolo Bonzini static void account_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp) 831c50d8ae3SPaolo Bonzini { 832c50d8ae3SPaolo Bonzini struct kvm_memslots *slots; 833c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot; 834c50d8ae3SPaolo Bonzini gfn_t gfn; 835c50d8ae3SPaolo Bonzini 836c50d8ae3SPaolo Bonzini kvm->arch.indirect_shadow_pages++; 837226d9b8fSSean Christopherson /* 838226d9b8fSSean Christopherson * Ensure indirect_shadow_pages is elevated prior to re-reading guest 839226d9b8fSSean Christopherson * child PTEs in FNAME(gpte_changed), i.e. guarantee either in-flight 840226d9b8fSSean Christopherson * emulated writes are visible before re-reading guest PTEs, or that 841226d9b8fSSean Christopherson * an emulated write will see the elevated count and acquire mmu_lock 842226d9b8fSSean Christopherson * to update SPTEs. Pairs with the smp_mb() in kvm_mmu_track_write(). 843226d9b8fSSean Christopherson */ 844226d9b8fSSean Christopherson smp_mb(); 845226d9b8fSSean Christopherson 846c50d8ae3SPaolo Bonzini gfn = sp->gfn; 847c50d8ae3SPaolo Bonzini slots = kvm_memslots_for_spte_role(kvm, sp->role); 848c50d8ae3SPaolo Bonzini slot = __gfn_to_memslot(slots, gfn); 849c50d8ae3SPaolo Bonzini 850c50d8ae3SPaolo Bonzini /* the non-leaf shadow pages are keeping readonly. */ 8513bae0459SSean Christopherson if (sp->role.level > PG_LEVEL_4K) 85296316a06SSean Christopherson return __kvm_write_track_add_gfn(kvm, slot, gfn); 853c50d8ae3SPaolo Bonzini 854c50d8ae3SPaolo Bonzini kvm_mmu_gfn_disallow_lpage(slot, gfn); 855be911771SDavid Matlack 856be911771SDavid Matlack if (kvm_mmu_slot_gfn_write_protect(kvm, slot, gfn, PG_LEVEL_4K)) 8574ad980aeSHou Wenlong kvm_flush_remote_tlbs_gfn(kvm, gfn, PG_LEVEL_4K); 858c50d8ae3SPaolo Bonzini } 859c50d8ae3SPaolo Bonzini 86061f94478SSean Christopherson void track_possible_nx_huge_page(struct kvm *kvm, struct kvm_mmu_page *sp) 861c50d8ae3SPaolo Bonzini { 862428e9216SSean Christopherson /* 863428e9216SSean Christopherson * If it's possible to replace the shadow page with an NX huge page, 864428e9216SSean Christopherson * i.e. if the shadow page is the only thing currently preventing KVM 865428e9216SSean Christopherson * from using a huge page, add the shadow page to the list of "to be 866428e9216SSean Christopherson * zapped for NX recovery" pages. Note, the shadow page can already be 867428e9216SSean Christopherson * on the list if KVM is reusing an existing shadow page, i.e. if KVM 868428e9216SSean Christopherson * links a shadow page at multiple points. 869428e9216SSean Christopherson */ 87061f94478SSean Christopherson if (!list_empty(&sp->possible_nx_huge_page_link)) 871c50d8ae3SPaolo Bonzini return; 872c50d8ae3SPaolo Bonzini 873c50d8ae3SPaolo Bonzini ++kvm->stat.nx_lpage_splits; 87455c510e2SSean Christopherson list_add_tail(&sp->possible_nx_huge_page_link, 87555c510e2SSean Christopherson &kvm->arch.possible_nx_huge_pages); 876c50d8ae3SPaolo Bonzini } 877c50d8ae3SPaolo Bonzini 87861f94478SSean Christopherson static void account_nx_huge_page(struct kvm *kvm, struct kvm_mmu_page *sp, 87961f94478SSean Christopherson bool nx_huge_page_possible) 88061f94478SSean Christopherson { 88161f94478SSean Christopherson sp->nx_huge_page_disallowed = true; 88261f94478SSean Christopherson 88361f94478SSean Christopherson if (nx_huge_page_possible) 88461f94478SSean Christopherson track_possible_nx_huge_page(kvm, sp); 885c50d8ae3SPaolo Bonzini } 886c50d8ae3SPaolo Bonzini 887c50d8ae3SPaolo Bonzini static void unaccount_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp) 888c50d8ae3SPaolo Bonzini { 889c50d8ae3SPaolo Bonzini struct kvm_memslots *slots; 890c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot; 891c50d8ae3SPaolo Bonzini gfn_t gfn; 892c50d8ae3SPaolo Bonzini 893c50d8ae3SPaolo Bonzini kvm->arch.indirect_shadow_pages--; 894c50d8ae3SPaolo Bonzini gfn = sp->gfn; 895c50d8ae3SPaolo Bonzini slots = kvm_memslots_for_spte_role(kvm, sp->role); 896c50d8ae3SPaolo Bonzini slot = __gfn_to_memslot(slots, gfn); 8973bae0459SSean Christopherson if (sp->role.level > PG_LEVEL_4K) 89896316a06SSean Christopherson return __kvm_write_track_remove_gfn(kvm, slot, gfn); 899c50d8ae3SPaolo Bonzini 900c50d8ae3SPaolo Bonzini kvm_mmu_gfn_allow_lpage(slot, gfn); 901c50d8ae3SPaolo Bonzini } 902c50d8ae3SPaolo Bonzini 90361f94478SSean Christopherson void untrack_possible_nx_huge_page(struct kvm *kvm, struct kvm_mmu_page *sp) 904c50d8ae3SPaolo Bonzini { 90555c510e2SSean Christopherson if (list_empty(&sp->possible_nx_huge_page_link)) 906428e9216SSean Christopherson return; 907428e9216SSean Christopherson 908c50d8ae3SPaolo Bonzini --kvm->stat.nx_lpage_splits; 90955c510e2SSean Christopherson list_del_init(&sp->possible_nx_huge_page_link); 910c50d8ae3SPaolo Bonzini } 911c50d8ae3SPaolo Bonzini 91261f94478SSean Christopherson static void unaccount_nx_huge_page(struct kvm *kvm, struct kvm_mmu_page *sp) 91361f94478SSean Christopherson { 91461f94478SSean Christopherson sp->nx_huge_page_disallowed = false; 91561f94478SSean Christopherson 91661f94478SSean Christopherson untrack_possible_nx_huge_page(kvm, sp); 917c50d8ae3SPaolo Bonzini } 918c50d8ae3SPaolo Bonzini 919f3d90f90SSean Christopherson static struct kvm_memory_slot *gfn_to_memslot_dirty_bitmap(struct kvm_vcpu *vcpu, 920f3d90f90SSean Christopherson gfn_t gfn, 921c50d8ae3SPaolo Bonzini bool no_dirty_log) 922c50d8ae3SPaolo Bonzini { 923c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot; 924c50d8ae3SPaolo Bonzini 925c50d8ae3SPaolo Bonzini slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn); 92691b0d268SPaolo Bonzini if (!slot || slot->flags & KVM_MEMSLOT_INVALID) 92791b0d268SPaolo Bonzini return NULL; 928044c59c4SPeter Xu if (no_dirty_log && kvm_slot_dirty_track_enabled(slot)) 92991b0d268SPaolo Bonzini return NULL; 930c50d8ae3SPaolo Bonzini 931c50d8ae3SPaolo Bonzini return slot; 932c50d8ae3SPaolo Bonzini } 933c50d8ae3SPaolo Bonzini 934c50d8ae3SPaolo Bonzini /* 935c50d8ae3SPaolo Bonzini * About rmap_head encoding: 936c50d8ae3SPaolo Bonzini * 937c50d8ae3SPaolo Bonzini * If the bit zero of rmap_head->val is clear, then it points to the only spte 938c50d8ae3SPaolo Bonzini * in this rmap chain. Otherwise, (rmap_head->val & ~1) points to a struct 939c50d8ae3SPaolo Bonzini * pte_list_desc containing more mappings. 940c50d8ae3SPaolo Bonzini */ 941c50d8ae3SPaolo Bonzini 942c50d8ae3SPaolo Bonzini /* 943c50d8ae3SPaolo Bonzini * Returns the number of pointers in the rmap chain, not counting the new one. 944c50d8ae3SPaolo Bonzini */ 9452ff9039aSDavid Matlack static int pte_list_add(struct kvm_mmu_memory_cache *cache, u64 *spte, 946c50d8ae3SPaolo Bonzini struct kvm_rmap_head *rmap_head) 947c50d8ae3SPaolo Bonzini { 948c50d8ae3SPaolo Bonzini struct pte_list_desc *desc; 94913236e25SPeter Xu int count = 0; 950c50d8ae3SPaolo Bonzini 951c50d8ae3SPaolo Bonzini if (!rmap_head->val) { 952c50d8ae3SPaolo Bonzini rmap_head->val = (unsigned long)spte; 953c50d8ae3SPaolo Bonzini } else if (!(rmap_head->val & 1)) { 9542ff9039aSDavid Matlack desc = kvm_mmu_memory_cache_alloc(cache); 955c50d8ae3SPaolo Bonzini desc->sptes[0] = (u64 *)rmap_head->val; 956c50d8ae3SPaolo Bonzini desc->sptes[1] = spte; 95713236e25SPeter Xu desc->spte_count = 2; 958141705b7SLai Jiangshan desc->tail_count = 0; 959c50d8ae3SPaolo Bonzini rmap_head->val = (unsigned long)desc | 1; 960c50d8ae3SPaolo Bonzini ++count; 961c50d8ae3SPaolo Bonzini } else { 962c50d8ae3SPaolo Bonzini desc = (struct pte_list_desc *)(rmap_head->val & ~1ul); 963141705b7SLai Jiangshan count = desc->tail_count + desc->spte_count; 964141705b7SLai Jiangshan 965141705b7SLai Jiangshan /* 966141705b7SLai Jiangshan * If the previous head is full, allocate a new head descriptor 967141705b7SLai Jiangshan * as tail descriptors are always kept full. 968141705b7SLai Jiangshan */ 969141705b7SLai Jiangshan if (desc->spte_count == PTE_LIST_EXT) { 970141705b7SLai Jiangshan desc = kvm_mmu_memory_cache_alloc(cache); 971141705b7SLai Jiangshan desc->more = (struct pte_list_desc *)(rmap_head->val & ~1ul); 97213236e25SPeter Xu desc->spte_count = 0; 973141705b7SLai Jiangshan desc->tail_count = count; 974141705b7SLai Jiangshan rmap_head->val = (unsigned long)desc | 1; 975c6c4f961SLi RongQing } 97613236e25SPeter Xu desc->sptes[desc->spte_count++] = spte; 977c50d8ae3SPaolo Bonzini } 978c50d8ae3SPaolo Bonzini return count; 979c50d8ae3SPaolo Bonzini } 980c50d8ae3SPaolo Bonzini 981069f30c6SMingwei Zhang static void pte_list_desc_remove_entry(struct kvm *kvm, 982069f30c6SMingwei Zhang struct kvm_rmap_head *rmap_head, 983141705b7SLai Jiangshan struct pte_list_desc *desc, int i) 984c50d8ae3SPaolo Bonzini { 985141705b7SLai Jiangshan struct pte_list_desc *head_desc = (struct pte_list_desc *)(rmap_head->val & ~1ul); 986141705b7SLai Jiangshan int j = head_desc->spte_count - 1; 987c50d8ae3SPaolo Bonzini 988141705b7SLai Jiangshan /* 989141705b7SLai Jiangshan * The head descriptor should never be empty. A new head is added only 990141705b7SLai Jiangshan * when adding an entry and the previous head is full, and heads are 991141705b7SLai Jiangshan * removed (this flow) when they become empty. 992141705b7SLai Jiangshan */ 99352e322edSSean Christopherson KVM_BUG_ON_DATA_CORRUPTION(j < 0, kvm); 994141705b7SLai Jiangshan 995141705b7SLai Jiangshan /* 996141705b7SLai Jiangshan * Replace the to-be-freed SPTE with the last valid entry from the head 997141705b7SLai Jiangshan * descriptor to ensure that tail descriptors are full at all times. 998141705b7SLai Jiangshan * Note, this also means that tail_count is stable for each descriptor. 999141705b7SLai Jiangshan */ 1000141705b7SLai Jiangshan desc->sptes[i] = head_desc->sptes[j]; 1001141705b7SLai Jiangshan head_desc->sptes[j] = NULL; 1002141705b7SLai Jiangshan head_desc->spte_count--; 1003141705b7SLai Jiangshan if (head_desc->spte_count) 1004c50d8ae3SPaolo Bonzini return; 1005141705b7SLai Jiangshan 1006141705b7SLai Jiangshan /* 1007141705b7SLai Jiangshan * The head descriptor is empty. If there are no tail descriptors, 100854aa699eSBjorn Helgaas * nullify the rmap head to mark the list as empty, else point the rmap 1009141705b7SLai Jiangshan * head at the next descriptor, i.e. the new head. 1010141705b7SLai Jiangshan */ 1011141705b7SLai Jiangshan if (!head_desc->more) 1012fe3c2b4cSMiaohe Lin rmap_head->val = 0; 1013c50d8ae3SPaolo Bonzini else 1014141705b7SLai Jiangshan rmap_head->val = (unsigned long)head_desc->more | 1; 1015141705b7SLai Jiangshan mmu_free_pte_list_desc(head_desc); 1016c50d8ae3SPaolo Bonzini } 1017c50d8ae3SPaolo Bonzini 1018069f30c6SMingwei Zhang static void pte_list_remove(struct kvm *kvm, u64 *spte, 1019069f30c6SMingwei Zhang struct kvm_rmap_head *rmap_head) 1020c50d8ae3SPaolo Bonzini { 1021c50d8ae3SPaolo Bonzini struct pte_list_desc *desc; 1022c50d8ae3SPaolo Bonzini int i; 1023c50d8ae3SPaolo Bonzini 102452e322edSSean Christopherson if (KVM_BUG_ON_DATA_CORRUPTION(!rmap_head->val, kvm)) 102552e322edSSean Christopherson return; 102652e322edSSean Christopherson 102752e322edSSean Christopherson if (!(rmap_head->val & 1)) { 102852e322edSSean Christopherson if (KVM_BUG_ON_DATA_CORRUPTION((u64 *)rmap_head->val != spte, kvm)) 102952e322edSSean Christopherson return; 103052e322edSSean Christopherson 1031c50d8ae3SPaolo Bonzini rmap_head->val = 0; 1032c50d8ae3SPaolo Bonzini } else { 1033c50d8ae3SPaolo Bonzini desc = (struct pte_list_desc *)(rmap_head->val & ~1ul); 1034c50d8ae3SPaolo Bonzini while (desc) { 103513236e25SPeter Xu for (i = 0; i < desc->spte_count; ++i) { 1036c50d8ae3SPaolo Bonzini if (desc->sptes[i] == spte) { 1037069f30c6SMingwei Zhang pte_list_desc_remove_entry(kvm, rmap_head, 1038069f30c6SMingwei Zhang desc, i); 1039c50d8ae3SPaolo Bonzini return; 1040c50d8ae3SPaolo Bonzini } 1041c50d8ae3SPaolo Bonzini } 1042c50d8ae3SPaolo Bonzini desc = desc->more; 1043c50d8ae3SPaolo Bonzini } 104452e322edSSean Christopherson 104552e322edSSean Christopherson KVM_BUG_ON_DATA_CORRUPTION(true, kvm); 1046c50d8ae3SPaolo Bonzini } 1047c50d8ae3SPaolo Bonzini } 1048c50d8ae3SPaolo Bonzini 10499202aee8SSean Christopherson static void kvm_zap_one_rmap_spte(struct kvm *kvm, 10509202aee8SSean Christopherson struct kvm_rmap_head *rmap_head, u64 *sptep) 1051c50d8ae3SPaolo Bonzini { 105271f51d2cSMingwei Zhang mmu_spte_clear_track_bits(kvm, sptep); 1053069f30c6SMingwei Zhang pte_list_remove(kvm, sptep, rmap_head); 1054c50d8ae3SPaolo Bonzini } 1055c50d8ae3SPaolo Bonzini 10569202aee8SSean Christopherson /* Return true if at least one SPTE was zapped, false otherwise */ 10579202aee8SSean Christopherson static bool kvm_zap_all_rmap_sptes(struct kvm *kvm, 10589202aee8SSean Christopherson struct kvm_rmap_head *rmap_head) 1059a75b5404SPeter Xu { 1060a75b5404SPeter Xu struct pte_list_desc *desc, *next; 1061a75b5404SPeter Xu int i; 1062a75b5404SPeter Xu 1063a75b5404SPeter Xu if (!rmap_head->val) 1064a75b5404SPeter Xu return false; 1065a75b5404SPeter Xu 1066a75b5404SPeter Xu if (!(rmap_head->val & 1)) { 106771f51d2cSMingwei Zhang mmu_spte_clear_track_bits(kvm, (u64 *)rmap_head->val); 1068a75b5404SPeter Xu goto out; 1069a75b5404SPeter Xu } 1070a75b5404SPeter Xu 1071a75b5404SPeter Xu desc = (struct pte_list_desc *)(rmap_head->val & ~1ul); 1072a75b5404SPeter Xu 1073a75b5404SPeter Xu for (; desc; desc = next) { 1074a75b5404SPeter Xu for (i = 0; i < desc->spte_count; i++) 107571f51d2cSMingwei Zhang mmu_spte_clear_track_bits(kvm, desc->sptes[i]); 1076a75b5404SPeter Xu next = desc->more; 1077a75b5404SPeter Xu mmu_free_pte_list_desc(desc); 1078a75b5404SPeter Xu } 1079a75b5404SPeter Xu out: 1080a75b5404SPeter Xu /* rmap_head is meaningless now, remember to reset it */ 1081a75b5404SPeter Xu rmap_head->val = 0; 1082a75b5404SPeter Xu return true; 1083a75b5404SPeter Xu } 1084a75b5404SPeter Xu 10853bcd0662SPeter Xu unsigned int pte_list_count(struct kvm_rmap_head *rmap_head) 10863bcd0662SPeter Xu { 10873bcd0662SPeter Xu struct pte_list_desc *desc; 10883bcd0662SPeter Xu 10893bcd0662SPeter Xu if (!rmap_head->val) 10903bcd0662SPeter Xu return 0; 10913bcd0662SPeter Xu else if (!(rmap_head->val & 1)) 10923bcd0662SPeter Xu return 1; 10933bcd0662SPeter Xu 10943bcd0662SPeter Xu desc = (struct pte_list_desc *)(rmap_head->val & ~1ul); 1095141705b7SLai Jiangshan return desc->tail_count + desc->spte_count; 10963bcd0662SPeter Xu } 10973bcd0662SPeter Xu 109893e083d4SDavid Matlack static struct kvm_rmap_head *gfn_to_rmap(gfn_t gfn, int level, 1099269e9552SHamza Mahfooz const struct kvm_memory_slot *slot) 1100c50d8ae3SPaolo Bonzini { 1101c50d8ae3SPaolo Bonzini unsigned long idx; 1102c50d8ae3SPaolo Bonzini 1103c50d8ae3SPaolo Bonzini idx = gfn_to_index(gfn, slot->base_gfn, level); 11043bae0459SSean Christopherson return &slot->arch.rmap[level - PG_LEVEL_4K][idx]; 1105c50d8ae3SPaolo Bonzini } 1106c50d8ae3SPaolo Bonzini 1107c50d8ae3SPaolo Bonzini static void rmap_remove(struct kvm *kvm, u64 *spte) 1108c50d8ae3SPaolo Bonzini { 1109601f8af0SDavid Matlack struct kvm_memslots *slots; 1110601f8af0SDavid Matlack struct kvm_memory_slot *slot; 1111c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 1112c50d8ae3SPaolo Bonzini gfn_t gfn; 1113c50d8ae3SPaolo Bonzini struct kvm_rmap_head *rmap_head; 1114c50d8ae3SPaolo Bonzini 111557354682SSean Christopherson sp = sptep_to_sp(spte); 111679e48cecSSean Christopherson gfn = kvm_mmu_page_get_gfn(sp, spte_index(spte)); 1117601f8af0SDavid Matlack 1118601f8af0SDavid Matlack /* 111968be1306SDavid Matlack * Unlike rmap_add, rmap_remove does not run in the context of a vCPU 112068be1306SDavid Matlack * so we have to determine which memslots to use based on context 112168be1306SDavid Matlack * information in sp->role. 1122601f8af0SDavid Matlack */ 1123601f8af0SDavid Matlack slots = kvm_memslots_for_spte_role(kvm, sp->role); 1124601f8af0SDavid Matlack 1125601f8af0SDavid Matlack slot = __gfn_to_memslot(slots, gfn); 112693e083d4SDavid Matlack rmap_head = gfn_to_rmap(gfn, sp->role.level, slot); 1127601f8af0SDavid Matlack 1128069f30c6SMingwei Zhang pte_list_remove(kvm, spte, rmap_head); 1129c50d8ae3SPaolo Bonzini } 1130c50d8ae3SPaolo Bonzini 1131c50d8ae3SPaolo Bonzini /* 1132c50d8ae3SPaolo Bonzini * Used by the following functions to iterate through the sptes linked by a 1133c50d8ae3SPaolo Bonzini * rmap. All fields are private and not assumed to be used outside. 1134c50d8ae3SPaolo Bonzini */ 1135c50d8ae3SPaolo Bonzini struct rmap_iterator { 1136c50d8ae3SPaolo Bonzini /* private fields */ 1137c50d8ae3SPaolo Bonzini struct pte_list_desc *desc; /* holds the sptep if not NULL */ 1138c50d8ae3SPaolo Bonzini int pos; /* index of the sptep */ 1139c50d8ae3SPaolo Bonzini }; 1140c50d8ae3SPaolo Bonzini 1141c50d8ae3SPaolo Bonzini /* 1142c50d8ae3SPaolo Bonzini * Iteration must be started by this function. This should also be used after 1143c50d8ae3SPaolo Bonzini * removing/dropping sptes from the rmap link because in such cases the 11440a03cbdaSMiaohe Lin * information in the iterator may not be valid. 1145c50d8ae3SPaolo Bonzini * 1146c50d8ae3SPaolo Bonzini * Returns sptep if found, NULL otherwise. 1147c50d8ae3SPaolo Bonzini */ 1148c50d8ae3SPaolo Bonzini static u64 *rmap_get_first(struct kvm_rmap_head *rmap_head, 1149c50d8ae3SPaolo Bonzini struct rmap_iterator *iter) 1150c50d8ae3SPaolo Bonzini { 1151c50d8ae3SPaolo Bonzini u64 *sptep; 1152c50d8ae3SPaolo Bonzini 1153c50d8ae3SPaolo Bonzini if (!rmap_head->val) 1154c50d8ae3SPaolo Bonzini return NULL; 1155c50d8ae3SPaolo Bonzini 1156c50d8ae3SPaolo Bonzini if (!(rmap_head->val & 1)) { 1157c50d8ae3SPaolo Bonzini iter->desc = NULL; 1158c50d8ae3SPaolo Bonzini sptep = (u64 *)rmap_head->val; 1159c50d8ae3SPaolo Bonzini goto out; 1160c50d8ae3SPaolo Bonzini } 1161c50d8ae3SPaolo Bonzini 1162c50d8ae3SPaolo Bonzini iter->desc = (struct pte_list_desc *)(rmap_head->val & ~1ul); 1163c50d8ae3SPaolo Bonzini iter->pos = 0; 1164c50d8ae3SPaolo Bonzini sptep = iter->desc->sptes[iter->pos]; 1165c50d8ae3SPaolo Bonzini out: 1166c50d8ae3SPaolo Bonzini BUG_ON(!is_shadow_present_pte(*sptep)); 1167c50d8ae3SPaolo Bonzini return sptep; 1168c50d8ae3SPaolo Bonzini } 1169c50d8ae3SPaolo Bonzini 1170c50d8ae3SPaolo Bonzini /* 1171c50d8ae3SPaolo Bonzini * Must be used with a valid iterator: e.g. after rmap_get_first(). 1172c50d8ae3SPaolo Bonzini * 1173c50d8ae3SPaolo Bonzini * Returns sptep if found, NULL otherwise. 1174c50d8ae3SPaolo Bonzini */ 1175c50d8ae3SPaolo Bonzini static u64 *rmap_get_next(struct rmap_iterator *iter) 1176c50d8ae3SPaolo Bonzini { 1177c50d8ae3SPaolo Bonzini u64 *sptep; 1178c50d8ae3SPaolo Bonzini 1179c50d8ae3SPaolo Bonzini if (iter->desc) { 1180c50d8ae3SPaolo Bonzini if (iter->pos < PTE_LIST_EXT - 1) { 1181c50d8ae3SPaolo Bonzini ++iter->pos; 1182c50d8ae3SPaolo Bonzini sptep = iter->desc->sptes[iter->pos]; 1183c50d8ae3SPaolo Bonzini if (sptep) 1184c50d8ae3SPaolo Bonzini goto out; 1185c50d8ae3SPaolo Bonzini } 1186c50d8ae3SPaolo Bonzini 1187c50d8ae3SPaolo Bonzini iter->desc = iter->desc->more; 1188c50d8ae3SPaolo Bonzini 1189c50d8ae3SPaolo Bonzini if (iter->desc) { 1190c50d8ae3SPaolo Bonzini iter->pos = 0; 1191c50d8ae3SPaolo Bonzini /* desc->sptes[0] cannot be NULL */ 1192c50d8ae3SPaolo Bonzini sptep = iter->desc->sptes[iter->pos]; 1193c50d8ae3SPaolo Bonzini goto out; 1194c50d8ae3SPaolo Bonzini } 1195c50d8ae3SPaolo Bonzini } 1196c50d8ae3SPaolo Bonzini 1197c50d8ae3SPaolo Bonzini return NULL; 1198c50d8ae3SPaolo Bonzini out: 1199c50d8ae3SPaolo Bonzini BUG_ON(!is_shadow_present_pte(*sptep)); 1200c50d8ae3SPaolo Bonzini return sptep; 1201c50d8ae3SPaolo Bonzini } 1202c50d8ae3SPaolo Bonzini 1203c50d8ae3SPaolo Bonzini #define for_each_rmap_spte(_rmap_head_, _iter_, _spte_) \ 1204c50d8ae3SPaolo Bonzini for (_spte_ = rmap_get_first(_rmap_head_, _iter_); \ 1205c50d8ae3SPaolo Bonzini _spte_; _spte_ = rmap_get_next(_iter_)) 1206c50d8ae3SPaolo Bonzini 1207c50d8ae3SPaolo Bonzini static void drop_spte(struct kvm *kvm, u64 *sptep) 1208c50d8ae3SPaolo Bonzini { 120971f51d2cSMingwei Zhang u64 old_spte = mmu_spte_clear_track_bits(kvm, sptep); 12107fa2a347SSean Christopherson 12117fa2a347SSean Christopherson if (is_shadow_present_pte(old_spte)) 1212c50d8ae3SPaolo Bonzini rmap_remove(kvm, sptep); 1213c50d8ae3SPaolo Bonzini } 1214c50d8ae3SPaolo Bonzini 121503787394SPaolo Bonzini static void drop_large_spte(struct kvm *kvm, u64 *sptep, bool flush) 1216c50d8ae3SPaolo Bonzini { 12170cd8dc73SPaolo Bonzini struct kvm_mmu_page *sp; 12180cd8dc73SPaolo Bonzini 12190cd8dc73SPaolo Bonzini sp = sptep_to_sp(sptep); 122020ba462dSSean Christopherson WARN_ON_ONCE(sp->role.level == PG_LEVEL_4K); 12210cd8dc73SPaolo Bonzini 1222c50d8ae3SPaolo Bonzini drop_spte(kvm, sptep); 122303787394SPaolo Bonzini 122403787394SPaolo Bonzini if (flush) 12251b2dc736SHou Wenlong kvm_flush_remote_tlbs_sptep(kvm, sptep); 1226c50d8ae3SPaolo Bonzini } 1227c50d8ae3SPaolo Bonzini 1228c50d8ae3SPaolo Bonzini /* 1229c50d8ae3SPaolo Bonzini * Write-protect on the specified @sptep, @pt_protect indicates whether 1230c50d8ae3SPaolo Bonzini * spte write-protection is caused by protecting shadow page table. 1231c50d8ae3SPaolo Bonzini * 1232c50d8ae3SPaolo Bonzini * Note: write protection is difference between dirty logging and spte 1233c50d8ae3SPaolo Bonzini * protection: 1234c50d8ae3SPaolo Bonzini * - for dirty logging, the spte can be set to writable at anytime if 1235c50d8ae3SPaolo Bonzini * its dirty bitmap is properly set. 1236c50d8ae3SPaolo Bonzini * - for spte protection, the spte can be writable only after unsync-ing 1237c50d8ae3SPaolo Bonzini * shadow page. 1238c50d8ae3SPaolo Bonzini * 1239c50d8ae3SPaolo Bonzini * Return true if tlb need be flushed. 1240c50d8ae3SPaolo Bonzini */ 1241c50d8ae3SPaolo Bonzini static bool spte_write_protect(u64 *sptep, bool pt_protect) 1242c50d8ae3SPaolo Bonzini { 1243c50d8ae3SPaolo Bonzini u64 spte = *sptep; 1244c50d8ae3SPaolo Bonzini 1245c50d8ae3SPaolo Bonzini if (!is_writable_pte(spte) && 1246706c9c55SSean Christopherson !(pt_protect && is_mmu_writable_spte(spte))) 1247c50d8ae3SPaolo Bonzini return false; 1248c50d8ae3SPaolo Bonzini 1249c50d8ae3SPaolo Bonzini if (pt_protect) 12505fc3424fSSean Christopherson spte &= ~shadow_mmu_writable_mask; 1251c50d8ae3SPaolo Bonzini spte = spte & ~PT_WRITABLE_MASK; 1252c50d8ae3SPaolo Bonzini 1253c50d8ae3SPaolo Bonzini return mmu_spte_update(sptep, spte); 1254c50d8ae3SPaolo Bonzini } 1255c50d8ae3SPaolo Bonzini 12561346bbb6SDavid Matlack static bool rmap_write_protect(struct kvm_rmap_head *rmap_head, 1257c50d8ae3SPaolo Bonzini bool pt_protect) 1258c50d8ae3SPaolo Bonzini { 1259c50d8ae3SPaolo Bonzini u64 *sptep; 1260c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 1261c50d8ae3SPaolo Bonzini bool flush = false; 1262c50d8ae3SPaolo Bonzini 1263c50d8ae3SPaolo Bonzini for_each_rmap_spte(rmap_head, &iter, sptep) 1264c50d8ae3SPaolo Bonzini flush |= spte_write_protect(sptep, pt_protect); 1265c50d8ae3SPaolo Bonzini 1266c50d8ae3SPaolo Bonzini return flush; 1267c50d8ae3SPaolo Bonzini } 1268c50d8ae3SPaolo Bonzini 1269c50d8ae3SPaolo Bonzini static bool spte_clear_dirty(u64 *sptep) 1270c50d8ae3SPaolo Bonzini { 1271c50d8ae3SPaolo Bonzini u64 spte = *sptep; 1272c50d8ae3SPaolo Bonzini 12730fe6370eSSean Christopherson KVM_MMU_WARN_ON(!spte_ad_enabled(spte)); 1274c50d8ae3SPaolo Bonzini spte &= ~shadow_dirty_mask; 1275c50d8ae3SPaolo Bonzini return mmu_spte_update(sptep, spte); 1276c50d8ae3SPaolo Bonzini } 1277c50d8ae3SPaolo Bonzini 1278c50d8ae3SPaolo Bonzini static bool spte_wrprot_for_clear_dirty(u64 *sptep) 1279c50d8ae3SPaolo Bonzini { 1280c50d8ae3SPaolo Bonzini bool was_writable = test_and_clear_bit(PT_WRITABLE_SHIFT, 1281c50d8ae3SPaolo Bonzini (unsigned long *)sptep); 1282c50d8ae3SPaolo Bonzini if (was_writable && !spte_ad_enabled(*sptep)) 1283c50d8ae3SPaolo Bonzini kvm_set_pfn_dirty(spte_to_pfn(*sptep)); 1284c50d8ae3SPaolo Bonzini 1285c50d8ae3SPaolo Bonzini return was_writable; 1286c50d8ae3SPaolo Bonzini } 1287c50d8ae3SPaolo Bonzini 1288c50d8ae3SPaolo Bonzini /* 1289c50d8ae3SPaolo Bonzini * Gets the GFN ready for another round of dirty logging by clearing the 1290c50d8ae3SPaolo Bonzini * - D bit on ad-enabled SPTEs, and 1291c50d8ae3SPaolo Bonzini * - W bit on ad-disabled SPTEs. 1292c50d8ae3SPaolo Bonzini * Returns true iff any D or W bits were cleared. 1293c50d8ae3SPaolo Bonzini */ 12940a234f5dSSean Christopherson static bool __rmap_clear_dirty(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1295269e9552SHamza Mahfooz const struct kvm_memory_slot *slot) 1296c50d8ae3SPaolo Bonzini { 1297c50d8ae3SPaolo Bonzini u64 *sptep; 1298c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 1299c50d8ae3SPaolo Bonzini bool flush = false; 1300c50d8ae3SPaolo Bonzini 1301c50d8ae3SPaolo Bonzini for_each_rmap_spte(rmap_head, &iter, sptep) 1302c50d8ae3SPaolo Bonzini if (spte_ad_need_write_protect(*sptep)) 1303c50d8ae3SPaolo Bonzini flush |= spte_wrprot_for_clear_dirty(sptep); 1304c50d8ae3SPaolo Bonzini else 1305c50d8ae3SPaolo Bonzini flush |= spte_clear_dirty(sptep); 1306c50d8ae3SPaolo Bonzini 1307c50d8ae3SPaolo Bonzini return flush; 1308c50d8ae3SPaolo Bonzini } 1309c50d8ae3SPaolo Bonzini 1310c50d8ae3SPaolo Bonzini /** 1311c50d8ae3SPaolo Bonzini * kvm_mmu_write_protect_pt_masked - write protect selected PT level pages 1312c50d8ae3SPaolo Bonzini * @kvm: kvm instance 1313c50d8ae3SPaolo Bonzini * @slot: slot to protect 1314c50d8ae3SPaolo Bonzini * @gfn_offset: start of the BITS_PER_LONG pages we care about 1315c50d8ae3SPaolo Bonzini * @mask: indicates which pages we should protect 1316c50d8ae3SPaolo Bonzini * 131789212919SKeqian Zhu * Used when we do not need to care about huge page mappings. 1318c50d8ae3SPaolo Bonzini */ 1319c50d8ae3SPaolo Bonzini static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm, 1320c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot, 1321c50d8ae3SPaolo Bonzini gfn_t gfn_offset, unsigned long mask) 1322c50d8ae3SPaolo Bonzini { 1323c50d8ae3SPaolo Bonzini struct kvm_rmap_head *rmap_head; 1324c50d8ae3SPaolo Bonzini 13251f98f2bdSDavid Matlack if (tdp_mmu_enabled) 1326a6a0b05dSBen Gardon kvm_tdp_mmu_clear_dirty_pt_masked(kvm, slot, 1327a6a0b05dSBen Gardon slot->base_gfn + gfn_offset, mask, true); 1328e2209710SBen Gardon 1329e2209710SBen Gardon if (!kvm_memslots_have_rmaps(kvm)) 1330e2209710SBen Gardon return; 1331e2209710SBen Gardon 1332c50d8ae3SPaolo Bonzini while (mask) { 133393e083d4SDavid Matlack rmap_head = gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask), 13343bae0459SSean Christopherson PG_LEVEL_4K, slot); 13351346bbb6SDavid Matlack rmap_write_protect(rmap_head, false); 1336c50d8ae3SPaolo Bonzini 1337c50d8ae3SPaolo Bonzini /* clear the first set bit */ 1338c50d8ae3SPaolo Bonzini mask &= mask - 1; 1339c50d8ae3SPaolo Bonzini } 1340c50d8ae3SPaolo Bonzini } 1341c50d8ae3SPaolo Bonzini 1342c50d8ae3SPaolo Bonzini /** 1343c50d8ae3SPaolo Bonzini * kvm_mmu_clear_dirty_pt_masked - clear MMU D-bit for PT level pages, or write 1344c50d8ae3SPaolo Bonzini * protect the page if the D-bit isn't supported. 1345c50d8ae3SPaolo Bonzini * @kvm: kvm instance 1346c50d8ae3SPaolo Bonzini * @slot: slot to clear D-bit 1347c50d8ae3SPaolo Bonzini * @gfn_offset: start of the BITS_PER_LONG pages we care about 1348c50d8ae3SPaolo Bonzini * @mask: indicates which pages we should clear D-bit 1349c50d8ae3SPaolo Bonzini * 1350c50d8ae3SPaolo Bonzini * Used for PML to re-log the dirty GPAs after userspace querying dirty_bitmap. 1351c50d8ae3SPaolo Bonzini */ 1352a018eba5SSean Christopherson static void kvm_mmu_clear_dirty_pt_masked(struct kvm *kvm, 1353c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot, 1354c50d8ae3SPaolo Bonzini gfn_t gfn_offset, unsigned long mask) 1355c50d8ae3SPaolo Bonzini { 1356c50d8ae3SPaolo Bonzini struct kvm_rmap_head *rmap_head; 1357c50d8ae3SPaolo Bonzini 13581f98f2bdSDavid Matlack if (tdp_mmu_enabled) 1359a6a0b05dSBen Gardon kvm_tdp_mmu_clear_dirty_pt_masked(kvm, slot, 1360a6a0b05dSBen Gardon slot->base_gfn + gfn_offset, mask, false); 1361e2209710SBen Gardon 1362e2209710SBen Gardon if (!kvm_memslots_have_rmaps(kvm)) 1363e2209710SBen Gardon return; 1364e2209710SBen Gardon 1365c50d8ae3SPaolo Bonzini while (mask) { 136693e083d4SDavid Matlack rmap_head = gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask), 13673bae0459SSean Christopherson PG_LEVEL_4K, slot); 13680a234f5dSSean Christopherson __rmap_clear_dirty(kvm, rmap_head, slot); 1369c50d8ae3SPaolo Bonzini 1370c50d8ae3SPaolo Bonzini /* clear the first set bit */ 1371c50d8ae3SPaolo Bonzini mask &= mask - 1; 1372c50d8ae3SPaolo Bonzini } 1373c50d8ae3SPaolo Bonzini } 1374c50d8ae3SPaolo Bonzini 1375c50d8ae3SPaolo Bonzini /** 1376c50d8ae3SPaolo Bonzini * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected 1377c50d8ae3SPaolo Bonzini * PT level pages. 1378c50d8ae3SPaolo Bonzini * 1379c50d8ae3SPaolo Bonzini * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to 1380c50d8ae3SPaolo Bonzini * enable dirty logging for them. 1381c50d8ae3SPaolo Bonzini * 138289212919SKeqian Zhu * We need to care about huge page mappings: e.g. during dirty logging we may 138389212919SKeqian Zhu * have such mappings. 1384c50d8ae3SPaolo Bonzini */ 1385c50d8ae3SPaolo Bonzini void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm, 1386c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot, 1387c50d8ae3SPaolo Bonzini gfn_t gfn_offset, unsigned long mask) 1388c50d8ae3SPaolo Bonzini { 138989212919SKeqian Zhu /* 139089212919SKeqian Zhu * Huge pages are NOT write protected when we start dirty logging in 139189212919SKeqian Zhu * initially-all-set mode; must write protect them here so that they 139289212919SKeqian Zhu * are split to 4K on the first write. 139389212919SKeqian Zhu * 139489212919SKeqian Zhu * The gfn_offset is guaranteed to be aligned to 64, but the base_gfn 139589212919SKeqian Zhu * of memslot has no such restriction, so the range can cross two large 139689212919SKeqian Zhu * pages. 139789212919SKeqian Zhu */ 139889212919SKeqian Zhu if (kvm_dirty_log_manual_protect_and_init_set(kvm)) { 139989212919SKeqian Zhu gfn_t start = slot->base_gfn + gfn_offset + __ffs(mask); 140089212919SKeqian Zhu gfn_t end = slot->base_gfn + gfn_offset + __fls(mask); 140189212919SKeqian Zhu 1402cb00a70bSDavid Matlack if (READ_ONCE(eager_page_split)) 14031aa4bb91SDavid Matlack kvm_mmu_try_split_huge_pages(kvm, slot, start, end + 1, PG_LEVEL_4K); 1404cb00a70bSDavid Matlack 140589212919SKeqian Zhu kvm_mmu_slot_gfn_write_protect(kvm, slot, start, PG_LEVEL_2M); 140689212919SKeqian Zhu 140789212919SKeqian Zhu /* Cross two large pages? */ 140889212919SKeqian Zhu if (ALIGN(start << PAGE_SHIFT, PMD_SIZE) != 140989212919SKeqian Zhu ALIGN(end << PAGE_SHIFT, PMD_SIZE)) 141089212919SKeqian Zhu kvm_mmu_slot_gfn_write_protect(kvm, slot, end, 141189212919SKeqian Zhu PG_LEVEL_2M); 141289212919SKeqian Zhu } 141389212919SKeqian Zhu 141489212919SKeqian Zhu /* Now handle 4K PTEs. */ 1415a018eba5SSean Christopherson if (kvm_x86_ops.cpu_dirty_log_size) 1416a018eba5SSean Christopherson kvm_mmu_clear_dirty_pt_masked(kvm, slot, gfn_offset, mask); 1417c50d8ae3SPaolo Bonzini else 1418c50d8ae3SPaolo Bonzini kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask); 1419c50d8ae3SPaolo Bonzini } 1420c50d8ae3SPaolo Bonzini 1421fb04a1edSPeter Xu int kvm_cpu_dirty_log_size(void) 1422fb04a1edSPeter Xu { 14236dd03800SSean Christopherson return kvm_x86_ops.cpu_dirty_log_size; 1424fb04a1edSPeter Xu } 1425fb04a1edSPeter Xu 1426c50d8ae3SPaolo Bonzini bool kvm_mmu_slot_gfn_write_protect(struct kvm *kvm, 14273ad93562SKeqian Zhu struct kvm_memory_slot *slot, u64 gfn, 14283ad93562SKeqian Zhu int min_level) 1429c50d8ae3SPaolo Bonzini { 1430c50d8ae3SPaolo Bonzini struct kvm_rmap_head *rmap_head; 1431c50d8ae3SPaolo Bonzini int i; 1432c50d8ae3SPaolo Bonzini bool write_protected = false; 1433c50d8ae3SPaolo Bonzini 1434e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) { 14353ad93562SKeqian Zhu for (i = min_level; i <= KVM_MAX_HUGEPAGE_LEVEL; ++i) { 143693e083d4SDavid Matlack rmap_head = gfn_to_rmap(gfn, i, slot); 14371346bbb6SDavid Matlack write_protected |= rmap_write_protect(rmap_head, true); 1438c50d8ae3SPaolo Bonzini } 1439e2209710SBen Gardon } 1440c50d8ae3SPaolo Bonzini 14411f98f2bdSDavid Matlack if (tdp_mmu_enabled) 144246044f72SBen Gardon write_protected |= 14433ad93562SKeqian Zhu kvm_tdp_mmu_write_protect_gfn(kvm, slot, gfn, min_level); 144446044f72SBen Gardon 1445c50d8ae3SPaolo Bonzini return write_protected; 1446c50d8ae3SPaolo Bonzini } 1447c50d8ae3SPaolo Bonzini 1448cf48f9e2SDavid Matlack static bool kvm_vcpu_write_protect_gfn(struct kvm_vcpu *vcpu, u64 gfn) 1449c50d8ae3SPaolo Bonzini { 1450c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot; 1451c50d8ae3SPaolo Bonzini 1452c50d8ae3SPaolo Bonzini slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn); 14533ad93562SKeqian Zhu return kvm_mmu_slot_gfn_write_protect(vcpu->kvm, slot, gfn, PG_LEVEL_4K); 1454c50d8ae3SPaolo Bonzini } 1455c50d8ae3SPaolo Bonzini 1456f8480721SSean Christopherson static bool __kvm_zap_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1457269e9552SHamza Mahfooz const struct kvm_memory_slot *slot) 1458c50d8ae3SPaolo Bonzini { 14599202aee8SSean Christopherson return kvm_zap_all_rmap_sptes(kvm, rmap_head); 1460c50d8ae3SPaolo Bonzini } 1461c50d8ae3SPaolo Bonzini 1462f8480721SSean Christopherson static bool kvm_zap_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1463f3b65bbaSPaolo Bonzini struct kvm_memory_slot *slot, gfn_t gfn, int level) 1464c50d8ae3SPaolo Bonzini { 1465f8480721SSean Christopherson return __kvm_zap_rmap(kvm, rmap_head, slot); 1466c50d8ae3SPaolo Bonzini } 1467c50d8ae3SPaolo Bonzini 1468c50d8ae3SPaolo Bonzini struct slot_rmap_walk_iterator { 1469c50d8ae3SPaolo Bonzini /* input fields. */ 1470269e9552SHamza Mahfooz const struct kvm_memory_slot *slot; 1471c50d8ae3SPaolo Bonzini gfn_t start_gfn; 1472c50d8ae3SPaolo Bonzini gfn_t end_gfn; 1473c50d8ae3SPaolo Bonzini int start_level; 1474c50d8ae3SPaolo Bonzini int end_level; 1475c50d8ae3SPaolo Bonzini 1476c50d8ae3SPaolo Bonzini /* output fields. */ 1477c50d8ae3SPaolo Bonzini gfn_t gfn; 1478c50d8ae3SPaolo Bonzini struct kvm_rmap_head *rmap; 1479c50d8ae3SPaolo Bonzini int level; 1480c50d8ae3SPaolo Bonzini 1481c50d8ae3SPaolo Bonzini /* private field. */ 1482c50d8ae3SPaolo Bonzini struct kvm_rmap_head *end_rmap; 1483c50d8ae3SPaolo Bonzini }; 1484c50d8ae3SPaolo Bonzini 1485f3d90f90SSean Christopherson static void rmap_walk_init_level(struct slot_rmap_walk_iterator *iterator, 1486f3d90f90SSean Christopherson int level) 1487c50d8ae3SPaolo Bonzini { 1488c50d8ae3SPaolo Bonzini iterator->level = level; 1489c50d8ae3SPaolo Bonzini iterator->gfn = iterator->start_gfn; 149093e083d4SDavid Matlack iterator->rmap = gfn_to_rmap(iterator->gfn, level, iterator->slot); 149193e083d4SDavid Matlack iterator->end_rmap = gfn_to_rmap(iterator->end_gfn, level, iterator->slot); 1492c50d8ae3SPaolo Bonzini } 1493c50d8ae3SPaolo Bonzini 1494f3d90f90SSean Christopherson static void slot_rmap_walk_init(struct slot_rmap_walk_iterator *iterator, 1495f3d90f90SSean Christopherson const struct kvm_memory_slot *slot, 1496f3d90f90SSean Christopherson int start_level, int end_level, 1497f3d90f90SSean Christopherson gfn_t start_gfn, gfn_t end_gfn) 1498c50d8ae3SPaolo Bonzini { 1499c50d8ae3SPaolo Bonzini iterator->slot = slot; 1500c50d8ae3SPaolo Bonzini iterator->start_level = start_level; 1501c50d8ae3SPaolo Bonzini iterator->end_level = end_level; 1502c50d8ae3SPaolo Bonzini iterator->start_gfn = start_gfn; 1503c50d8ae3SPaolo Bonzini iterator->end_gfn = end_gfn; 1504c50d8ae3SPaolo Bonzini 1505c50d8ae3SPaolo Bonzini rmap_walk_init_level(iterator, iterator->start_level); 1506c50d8ae3SPaolo Bonzini } 1507c50d8ae3SPaolo Bonzini 1508c50d8ae3SPaolo Bonzini static bool slot_rmap_walk_okay(struct slot_rmap_walk_iterator *iterator) 1509c50d8ae3SPaolo Bonzini { 1510c50d8ae3SPaolo Bonzini return !!iterator->rmap; 1511c50d8ae3SPaolo Bonzini } 1512c50d8ae3SPaolo Bonzini 1513c50d8ae3SPaolo Bonzini static void slot_rmap_walk_next(struct slot_rmap_walk_iterator *iterator) 1514c50d8ae3SPaolo Bonzini { 15156ba1e04fSVipin Sharma while (++iterator->rmap <= iterator->end_rmap) { 1516c50d8ae3SPaolo Bonzini iterator->gfn += (1UL << KVM_HPAGE_GFN_SHIFT(iterator->level)); 15176ba1e04fSVipin Sharma 15186ba1e04fSVipin Sharma if (iterator->rmap->val) 1519c50d8ae3SPaolo Bonzini return; 1520c50d8ae3SPaolo Bonzini } 1521c50d8ae3SPaolo Bonzini 1522c50d8ae3SPaolo Bonzini if (++iterator->level > iterator->end_level) { 1523c50d8ae3SPaolo Bonzini iterator->rmap = NULL; 1524c50d8ae3SPaolo Bonzini return; 1525c50d8ae3SPaolo Bonzini } 1526c50d8ae3SPaolo Bonzini 1527c50d8ae3SPaolo Bonzini rmap_walk_init_level(iterator, iterator->level); 1528c50d8ae3SPaolo Bonzini } 1529c50d8ae3SPaolo Bonzini 1530c50d8ae3SPaolo Bonzini #define for_each_slot_rmap_range(_slot_, _start_level_, _end_level_, \ 1531c50d8ae3SPaolo Bonzini _start_gfn, _end_gfn, _iter_) \ 1532c50d8ae3SPaolo Bonzini for (slot_rmap_walk_init(_iter_, _slot_, _start_level_, \ 1533c50d8ae3SPaolo Bonzini _end_level_, _start_gfn, _end_gfn); \ 1534c50d8ae3SPaolo Bonzini slot_rmap_walk_okay(_iter_); \ 1535c50d8ae3SPaolo Bonzini slot_rmap_walk_next(_iter_)) 1536c50d8ae3SPaolo Bonzini 15373039bcc7SSean Christopherson typedef bool (*rmap_handler_t)(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1538c1b91493SSean Christopherson struct kvm_memory_slot *slot, gfn_t gfn, 1539f3b65bbaSPaolo Bonzini int level); 1540c1b91493SSean Christopherson 15413039bcc7SSean Christopherson static __always_inline bool kvm_handle_gfn_range(struct kvm *kvm, 15423039bcc7SSean Christopherson struct kvm_gfn_range *range, 1543c1b91493SSean Christopherson rmap_handler_t handler) 1544c50d8ae3SPaolo Bonzini { 1545c50d8ae3SPaolo Bonzini struct slot_rmap_walk_iterator iterator; 15463039bcc7SSean Christopherson bool ret = false; 1547c50d8ae3SPaolo Bonzini 15483039bcc7SSean Christopherson for_each_slot_rmap_range(range->slot, PG_LEVEL_4K, KVM_MAX_HUGEPAGE_LEVEL, 15493039bcc7SSean Christopherson range->start, range->end - 1, &iterator) 15503039bcc7SSean Christopherson ret |= handler(kvm, iterator.rmap, range->slot, iterator.gfn, 1551f3b65bbaSPaolo Bonzini iterator.level); 1552c50d8ae3SPaolo Bonzini 1553c50d8ae3SPaolo Bonzini return ret; 1554c50d8ae3SPaolo Bonzini } 1555c50d8ae3SPaolo Bonzini 15563039bcc7SSean Christopherson bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range) 1557c50d8ae3SPaolo Bonzini { 1558e2209710SBen Gardon bool flush = false; 1559c50d8ae3SPaolo Bonzini 1560e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) 1561f8480721SSean Christopherson flush = kvm_handle_gfn_range(kvm, range, kvm_zap_rmap); 1562063afacdSBen Gardon 15631f98f2bdSDavid Matlack if (tdp_mmu_enabled) 1564c7785d85SHou Wenlong flush = kvm_tdp_mmu_unmap_gfn_range(kvm, range, flush); 1565063afacdSBen Gardon 15660a3869e1SSean Christopherson if (kvm_x86_ops.set_apic_access_page_addr && 15670a3869e1SSean Christopherson range->slot->id == APIC_ACCESS_PAGE_PRIVATE_MEMSLOT) 15680a8a5f2cSSean Christopherson kvm_make_all_cpus_request(kvm, KVM_REQ_APIC_PAGE_RELOAD); 15690a8a5f2cSSean Christopherson 15703039bcc7SSean Christopherson return flush; 1571c50d8ae3SPaolo Bonzini } 1572c50d8ae3SPaolo Bonzini 1573aed02fe3SSean Christopherson static bool kvm_age_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1574f3b65bbaSPaolo Bonzini struct kvm_memory_slot *slot, gfn_t gfn, int level) 1575c50d8ae3SPaolo Bonzini { 1576c50d8ae3SPaolo Bonzini u64 *sptep; 15773f649ab7SKees Cook struct rmap_iterator iter; 1578c50d8ae3SPaolo Bonzini int young = 0; 1579c50d8ae3SPaolo Bonzini 1580c50d8ae3SPaolo Bonzini for_each_rmap_spte(rmap_head, &iter, sptep) 1581c50d8ae3SPaolo Bonzini young |= mmu_spte_age(sptep); 1582c50d8ae3SPaolo Bonzini 1583c50d8ae3SPaolo Bonzini return young; 1584c50d8ae3SPaolo Bonzini } 1585c50d8ae3SPaolo Bonzini 1586aed02fe3SSean Christopherson static bool kvm_test_age_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1587f3b65bbaSPaolo Bonzini struct kvm_memory_slot *slot, gfn_t gfn, int level) 1588c50d8ae3SPaolo Bonzini { 1589c50d8ae3SPaolo Bonzini u64 *sptep; 1590c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 1591c50d8ae3SPaolo Bonzini 1592c50d8ae3SPaolo Bonzini for_each_rmap_spte(rmap_head, &iter, sptep) 1593c50d8ae3SPaolo Bonzini if (is_accessed_spte(*sptep)) 159498a26b69SVihas Mak return true; 159598a26b69SVihas Mak return false; 1596c50d8ae3SPaolo Bonzini } 1597c50d8ae3SPaolo Bonzini 1598c50d8ae3SPaolo Bonzini #define RMAP_RECYCLE_THRESHOLD 1000 1599c50d8ae3SPaolo Bonzini 16002ff9039aSDavid Matlack static void __rmap_add(struct kvm *kvm, 16012ff9039aSDavid Matlack struct kvm_mmu_memory_cache *cache, 16022ff9039aSDavid Matlack const struct kvm_memory_slot *slot, 160372ae5822SSean Christopherson u64 *spte, gfn_t gfn, unsigned int access) 1604c50d8ae3SPaolo Bonzini { 1605c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 160668be1306SDavid Matlack struct kvm_rmap_head *rmap_head; 160768be1306SDavid Matlack int rmap_count; 1608c50d8ae3SPaolo Bonzini 160957354682SSean Christopherson sp = sptep_to_sp(spte); 161079e48cecSSean Christopherson kvm_mmu_page_set_translation(sp, spte_index(spte), gfn, access); 161181cb4657SDavid Matlack kvm_update_page_stats(kvm, sp->role.level, 1); 161281cb4657SDavid Matlack 161393e083d4SDavid Matlack rmap_head = gfn_to_rmap(gfn, sp->role.level, slot); 16142ff9039aSDavid Matlack rmap_count = pte_list_add(cache, spte, rmap_head); 1615c50d8ae3SPaolo Bonzini 1616604f5332SMiaohe Lin if (rmap_count > kvm->stat.max_mmu_rmap_size) 1617604f5332SMiaohe Lin kvm->stat.max_mmu_rmap_size = rmap_count; 161868be1306SDavid Matlack if (rmap_count > RMAP_RECYCLE_THRESHOLD) { 16199202aee8SSean Christopherson kvm_zap_all_rmap_sptes(kvm, rmap_head); 16201b2dc736SHou Wenlong kvm_flush_remote_tlbs_gfn(kvm, gfn, sp->role.level); 162168be1306SDavid Matlack } 1622c50d8ae3SPaolo Bonzini } 1623c50d8ae3SPaolo Bonzini 16242ff9039aSDavid Matlack static void rmap_add(struct kvm_vcpu *vcpu, const struct kvm_memory_slot *slot, 162572ae5822SSean Christopherson u64 *spte, gfn_t gfn, unsigned int access) 16262ff9039aSDavid Matlack { 16272ff9039aSDavid Matlack struct kvm_mmu_memory_cache *cache = &vcpu->arch.mmu_pte_list_desc_cache; 16282ff9039aSDavid Matlack 16296a97575dSDavid Matlack __rmap_add(vcpu->kvm, cache, slot, spte, gfn, access); 16302ff9039aSDavid Matlack } 16312ff9039aSDavid Matlack 16323039bcc7SSean Christopherson bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range) 1633c50d8ae3SPaolo Bonzini { 1634e2209710SBen Gardon bool young = false; 1635f8e14497SBen Gardon 1636e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) 1637aed02fe3SSean Christopherson young = kvm_handle_gfn_range(kvm, range, kvm_age_rmap); 16383039bcc7SSean Christopherson 16391f98f2bdSDavid Matlack if (tdp_mmu_enabled) 16403039bcc7SSean Christopherson young |= kvm_tdp_mmu_age_gfn_range(kvm, range); 1641f8e14497SBen Gardon 1642f8e14497SBen Gardon return young; 1643c50d8ae3SPaolo Bonzini } 1644c50d8ae3SPaolo Bonzini 16453039bcc7SSean Christopherson bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range) 1646c50d8ae3SPaolo Bonzini { 1647e2209710SBen Gardon bool young = false; 1648f8e14497SBen Gardon 1649e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) 1650aed02fe3SSean Christopherson young = kvm_handle_gfn_range(kvm, range, kvm_test_age_rmap); 16513039bcc7SSean Christopherson 16521f98f2bdSDavid Matlack if (tdp_mmu_enabled) 16533039bcc7SSean Christopherson young |= kvm_tdp_mmu_test_age_gfn(kvm, range); 1654f8e14497SBen Gardon 1655f8e14497SBen Gardon return young; 1656c50d8ae3SPaolo Bonzini } 1657c50d8ae3SPaolo Bonzini 165858da926cSSean Christopherson static void kvm_mmu_check_sptes_at_free(struct kvm_mmu_page *sp) 1659c50d8ae3SPaolo Bonzini { 1660870d4d4eSSean Christopherson #ifdef CONFIG_KVM_PROVE_MMU 1661242a6dd8SSean Christopherson int i; 1662c50d8ae3SPaolo Bonzini 1663242a6dd8SSean Christopherson for (i = 0; i < SPTE_ENT_PER_PAGE; i++) { 16640fe6370eSSean Christopherson if (KVM_MMU_WARN_ON(is_shadow_present_pte(sp->spt[i]))) 166558da926cSSean Christopherson pr_err_ratelimited("SPTE %llx (@ %p) for gfn %llx shadow-present at free", 166658da926cSSean Christopherson sp->spt[i], &sp->spt[i], 166758da926cSSean Christopherson kvm_mmu_page_get_gfn(sp, i)); 1668c50d8ae3SPaolo Bonzini } 1669c50d8ae3SPaolo Bonzini #endif 167058da926cSSean Christopherson } 1671c50d8ae3SPaolo Bonzini 1672c50d8ae3SPaolo Bonzini /* 1673c50d8ae3SPaolo Bonzini * This value is the sum of all of the kvm instances's 1674c50d8ae3SPaolo Bonzini * kvm->arch.n_used_mmu_pages values. We need a global, 1675c50d8ae3SPaolo Bonzini * aggregate version in order to make the slab shrinker 1676c50d8ae3SPaolo Bonzini * faster 1677c50d8ae3SPaolo Bonzini */ 1678d5aaad6fSSean Christopherson static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, long nr) 1679c50d8ae3SPaolo Bonzini { 1680c50d8ae3SPaolo Bonzini kvm->arch.n_used_mmu_pages += nr; 1681c50d8ae3SPaolo Bonzini percpu_counter_add(&kvm_total_used_mmu_pages, nr); 1682c50d8ae3SPaolo Bonzini } 1683c50d8ae3SPaolo Bonzini 168443a063caSYosry Ahmed static void kvm_account_mmu_page(struct kvm *kvm, struct kvm_mmu_page *sp) 168543a063caSYosry Ahmed { 168643a063caSYosry Ahmed kvm_mod_used_mmu_pages(kvm, +1); 168743a063caSYosry Ahmed kvm_account_pgtable_pages((void *)sp->spt, +1); 168843a063caSYosry Ahmed } 168943a063caSYosry Ahmed 169043a063caSYosry Ahmed static void kvm_unaccount_mmu_page(struct kvm *kvm, struct kvm_mmu_page *sp) 169143a063caSYosry Ahmed { 169243a063caSYosry Ahmed kvm_mod_used_mmu_pages(kvm, -1); 169343a063caSYosry Ahmed kvm_account_pgtable_pages((void *)sp->spt, -1); 169443a063caSYosry Ahmed } 169543a063caSYosry Ahmed 169687654643SDavid Matlack static void kvm_mmu_free_shadow_page(struct kvm_mmu_page *sp) 1697c50d8ae3SPaolo Bonzini { 169858da926cSSean Christopherson kvm_mmu_check_sptes_at_free(sp); 169958da926cSSean Christopherson 1700c50d8ae3SPaolo Bonzini hlist_del(&sp->hash_link); 1701c50d8ae3SPaolo Bonzini list_del(&sp->link); 1702c50d8ae3SPaolo Bonzini free_page((unsigned long)sp->spt); 1703c50d8ae3SPaolo Bonzini if (!sp->role.direct) 17046a97575dSDavid Matlack free_page((unsigned long)sp->shadowed_translation); 1705c50d8ae3SPaolo Bonzini kmem_cache_free(mmu_page_header_cache, sp); 1706c50d8ae3SPaolo Bonzini } 1707c50d8ae3SPaolo Bonzini 1708c50d8ae3SPaolo Bonzini static unsigned kvm_page_table_hashfn(gfn_t gfn) 1709c50d8ae3SPaolo Bonzini { 1710c50d8ae3SPaolo Bonzini return hash_64(gfn, KVM_MMU_HASH_SHIFT); 1711c50d8ae3SPaolo Bonzini } 1712c50d8ae3SPaolo Bonzini 17132ff9039aSDavid Matlack static void mmu_page_add_parent_pte(struct kvm_mmu_memory_cache *cache, 1714c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, u64 *parent_pte) 1715c50d8ae3SPaolo Bonzini { 1716c50d8ae3SPaolo Bonzini if (!parent_pte) 1717c50d8ae3SPaolo Bonzini return; 1718c50d8ae3SPaolo Bonzini 17192ff9039aSDavid Matlack pte_list_add(cache, parent_pte, &sp->parent_ptes); 1720c50d8ae3SPaolo Bonzini } 1721c50d8ae3SPaolo Bonzini 1722069f30c6SMingwei Zhang static void mmu_page_remove_parent_pte(struct kvm *kvm, struct kvm_mmu_page *sp, 1723c50d8ae3SPaolo Bonzini u64 *parent_pte) 1724c50d8ae3SPaolo Bonzini { 1725069f30c6SMingwei Zhang pte_list_remove(kvm, parent_pte, &sp->parent_ptes); 1726c50d8ae3SPaolo Bonzini } 1727c50d8ae3SPaolo Bonzini 1728069f30c6SMingwei Zhang static void drop_parent_pte(struct kvm *kvm, struct kvm_mmu_page *sp, 1729c50d8ae3SPaolo Bonzini u64 *parent_pte) 1730c50d8ae3SPaolo Bonzini { 1731069f30c6SMingwei Zhang mmu_page_remove_parent_pte(kvm, sp, parent_pte); 1732c50d8ae3SPaolo Bonzini mmu_spte_clear_no_track(parent_pte); 1733c50d8ae3SPaolo Bonzini } 1734c50d8ae3SPaolo Bonzini 1735c50d8ae3SPaolo Bonzini static void mark_unsync(u64 *spte); 1736c50d8ae3SPaolo Bonzini static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp) 1737c50d8ae3SPaolo Bonzini { 1738c50d8ae3SPaolo Bonzini u64 *sptep; 1739c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 1740c50d8ae3SPaolo Bonzini 1741c50d8ae3SPaolo Bonzini for_each_rmap_spte(&sp->parent_ptes, &iter, sptep) { 1742c50d8ae3SPaolo Bonzini mark_unsync(sptep); 1743c50d8ae3SPaolo Bonzini } 1744c50d8ae3SPaolo Bonzini } 1745c50d8ae3SPaolo Bonzini 1746c50d8ae3SPaolo Bonzini static void mark_unsync(u64 *spte) 1747c50d8ae3SPaolo Bonzini { 1748c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 1749c50d8ae3SPaolo Bonzini 175057354682SSean Christopherson sp = sptep_to_sp(spte); 175179e48cecSSean Christopherson if (__test_and_set_bit(spte_index(spte), sp->unsync_child_bitmap)) 1752c50d8ae3SPaolo Bonzini return; 1753c50d8ae3SPaolo Bonzini if (sp->unsync_children++) 1754c50d8ae3SPaolo Bonzini return; 1755c50d8ae3SPaolo Bonzini kvm_mmu_mark_parents_unsync(sp); 1756c50d8ae3SPaolo Bonzini } 1757c50d8ae3SPaolo Bonzini 1758c50d8ae3SPaolo Bonzini #define KVM_PAGE_ARRAY_NR 16 1759c50d8ae3SPaolo Bonzini 1760c50d8ae3SPaolo Bonzini struct kvm_mmu_pages { 1761c50d8ae3SPaolo Bonzini struct mmu_page_and_offset { 1762c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 1763c50d8ae3SPaolo Bonzini unsigned int idx; 1764c50d8ae3SPaolo Bonzini } page[KVM_PAGE_ARRAY_NR]; 1765c50d8ae3SPaolo Bonzini unsigned int nr; 1766c50d8ae3SPaolo Bonzini }; 1767c50d8ae3SPaolo Bonzini 1768c50d8ae3SPaolo Bonzini static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp, 1769c50d8ae3SPaolo Bonzini int idx) 1770c50d8ae3SPaolo Bonzini { 1771c50d8ae3SPaolo Bonzini int i; 1772c50d8ae3SPaolo Bonzini 1773c50d8ae3SPaolo Bonzini if (sp->unsync) 1774c50d8ae3SPaolo Bonzini for (i=0; i < pvec->nr; i++) 1775c50d8ae3SPaolo Bonzini if (pvec->page[i].sp == sp) 1776c50d8ae3SPaolo Bonzini return 0; 1777c50d8ae3SPaolo Bonzini 1778c50d8ae3SPaolo Bonzini pvec->page[pvec->nr].sp = sp; 1779c50d8ae3SPaolo Bonzini pvec->page[pvec->nr].idx = idx; 1780c50d8ae3SPaolo Bonzini pvec->nr++; 1781c50d8ae3SPaolo Bonzini return (pvec->nr == KVM_PAGE_ARRAY_NR); 1782c50d8ae3SPaolo Bonzini } 1783c50d8ae3SPaolo Bonzini 1784c50d8ae3SPaolo Bonzini static inline void clear_unsync_child_bit(struct kvm_mmu_page *sp, int idx) 1785c50d8ae3SPaolo Bonzini { 1786c50d8ae3SPaolo Bonzini --sp->unsync_children; 178720ba462dSSean Christopherson WARN_ON_ONCE((int)sp->unsync_children < 0); 1788c50d8ae3SPaolo Bonzini __clear_bit(idx, sp->unsync_child_bitmap); 1789c50d8ae3SPaolo Bonzini } 1790c50d8ae3SPaolo Bonzini 1791c50d8ae3SPaolo Bonzini static int __mmu_unsync_walk(struct kvm_mmu_page *sp, 1792c50d8ae3SPaolo Bonzini struct kvm_mmu_pages *pvec) 1793c50d8ae3SPaolo Bonzini { 1794c50d8ae3SPaolo Bonzini int i, ret, nr_unsync_leaf = 0; 1795c50d8ae3SPaolo Bonzini 1796c50d8ae3SPaolo Bonzini for_each_set_bit(i, sp->unsync_child_bitmap, 512) { 1797c50d8ae3SPaolo Bonzini struct kvm_mmu_page *child; 1798c50d8ae3SPaolo Bonzini u64 ent = sp->spt[i]; 1799c50d8ae3SPaolo Bonzini 1800c50d8ae3SPaolo Bonzini if (!is_shadow_present_pte(ent) || is_large_pte(ent)) { 1801c50d8ae3SPaolo Bonzini clear_unsync_child_bit(sp, i); 1802c50d8ae3SPaolo Bonzini continue; 1803c50d8ae3SPaolo Bonzini } 1804c50d8ae3SPaolo Bonzini 18055e3edd7eSSean Christopherson child = spte_to_child_sp(ent); 1806c50d8ae3SPaolo Bonzini 1807c50d8ae3SPaolo Bonzini if (child->unsync_children) { 1808c50d8ae3SPaolo Bonzini if (mmu_pages_add(pvec, child, i)) 1809c50d8ae3SPaolo Bonzini return -ENOSPC; 1810c50d8ae3SPaolo Bonzini 1811c50d8ae3SPaolo Bonzini ret = __mmu_unsync_walk(child, pvec); 1812c50d8ae3SPaolo Bonzini if (!ret) { 1813c50d8ae3SPaolo Bonzini clear_unsync_child_bit(sp, i); 1814c50d8ae3SPaolo Bonzini continue; 1815c50d8ae3SPaolo Bonzini } else if (ret > 0) { 1816c50d8ae3SPaolo Bonzini nr_unsync_leaf += ret; 1817c50d8ae3SPaolo Bonzini } else 1818c50d8ae3SPaolo Bonzini return ret; 1819c50d8ae3SPaolo Bonzini } else if (child->unsync) { 1820c50d8ae3SPaolo Bonzini nr_unsync_leaf++; 1821c50d8ae3SPaolo Bonzini if (mmu_pages_add(pvec, child, i)) 1822c50d8ae3SPaolo Bonzini return -ENOSPC; 1823c50d8ae3SPaolo Bonzini } else 1824c50d8ae3SPaolo Bonzini clear_unsync_child_bit(sp, i); 1825c50d8ae3SPaolo Bonzini } 1826c50d8ae3SPaolo Bonzini 1827c50d8ae3SPaolo Bonzini return nr_unsync_leaf; 1828c50d8ae3SPaolo Bonzini } 1829c50d8ae3SPaolo Bonzini 1830c50d8ae3SPaolo Bonzini #define INVALID_INDEX (-1) 1831c50d8ae3SPaolo Bonzini 1832c50d8ae3SPaolo Bonzini static int mmu_unsync_walk(struct kvm_mmu_page *sp, 1833c50d8ae3SPaolo Bonzini struct kvm_mmu_pages *pvec) 1834c50d8ae3SPaolo Bonzini { 1835c50d8ae3SPaolo Bonzini pvec->nr = 0; 1836c50d8ae3SPaolo Bonzini if (!sp->unsync_children) 1837c50d8ae3SPaolo Bonzini return 0; 1838c50d8ae3SPaolo Bonzini 1839c50d8ae3SPaolo Bonzini mmu_pages_add(pvec, sp, INVALID_INDEX); 1840c50d8ae3SPaolo Bonzini return __mmu_unsync_walk(sp, pvec); 1841c50d8ae3SPaolo Bonzini } 1842c50d8ae3SPaolo Bonzini 1843c50d8ae3SPaolo Bonzini static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp) 1844c50d8ae3SPaolo Bonzini { 184520ba462dSSean Christopherson WARN_ON_ONCE(!sp->unsync); 1846c50d8ae3SPaolo Bonzini trace_kvm_mmu_sync_page(sp); 1847c50d8ae3SPaolo Bonzini sp->unsync = 0; 1848c50d8ae3SPaolo Bonzini --kvm->stat.mmu_unsync; 1849c50d8ae3SPaolo Bonzini } 1850c50d8ae3SPaolo Bonzini 1851c50d8ae3SPaolo Bonzini static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp, 1852c50d8ae3SPaolo Bonzini struct list_head *invalid_list); 1853c50d8ae3SPaolo Bonzini static void kvm_mmu_commit_zap_page(struct kvm *kvm, 1854c50d8ae3SPaolo Bonzini struct list_head *invalid_list); 1855c50d8ae3SPaolo Bonzini 1856767d8d8dSLai Jiangshan static bool sp_has_gptes(struct kvm_mmu_page *sp) 1857767d8d8dSLai Jiangshan { 1858767d8d8dSLai Jiangshan if (sp->role.direct) 1859767d8d8dSLai Jiangshan return false; 1860767d8d8dSLai Jiangshan 186184e5ffd0SLai Jiangshan if (sp->role.passthrough) 186284e5ffd0SLai Jiangshan return false; 186384e5ffd0SLai Jiangshan 1864767d8d8dSLai Jiangshan return true; 1865767d8d8dSLai Jiangshan } 1866767d8d8dSLai Jiangshan 1867ac101b7cSSean Christopherson #define for_each_valid_sp(_kvm, _sp, _list) \ 1868ac101b7cSSean Christopherson hlist_for_each_entry(_sp, _list, hash_link) \ 1869c50d8ae3SPaolo Bonzini if (is_obsolete_sp((_kvm), (_sp))) { \ 1870c50d8ae3SPaolo Bonzini } else 1871c50d8ae3SPaolo Bonzini 1872767d8d8dSLai Jiangshan #define for_each_gfn_valid_sp_with_gptes(_kvm, _sp, _gfn) \ 1873ac101b7cSSean Christopherson for_each_valid_sp(_kvm, _sp, \ 1874ac101b7cSSean Christopherson &(_kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(_gfn)]) \ 1875767d8d8dSLai Jiangshan if ((_sp)->gfn != (_gfn) || !sp_has_gptes(_sp)) {} else 1876c50d8ae3SPaolo Bonzini 187790e44470SLai Jiangshan static bool kvm_sync_page_check(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp) 187890e44470SLai Jiangshan { 187990e44470SLai Jiangshan union kvm_mmu_page_role root_role = vcpu->arch.mmu->root_role; 188090e44470SLai Jiangshan 188190e44470SLai Jiangshan /* 188290e44470SLai Jiangshan * Ignore various flags when verifying that it's safe to sync a shadow 188390e44470SLai Jiangshan * page using the current MMU context. 188490e44470SLai Jiangshan * 188590e44470SLai Jiangshan * - level: not part of the overall MMU role and will never match as the MMU's 188690e44470SLai Jiangshan * level tracks the root level 188790e44470SLai Jiangshan * - access: updated based on the new guest PTE 188890e44470SLai Jiangshan * - quadrant: not part of the overall MMU role (similar to level) 188990e44470SLai Jiangshan */ 189090e44470SLai Jiangshan const union kvm_mmu_page_role sync_role_ign = { 189190e44470SLai Jiangshan .level = 0xf, 189290e44470SLai Jiangshan .access = 0x7, 189390e44470SLai Jiangshan .quadrant = 0x3, 189490e44470SLai Jiangshan .passthrough = 0x1, 189590e44470SLai Jiangshan }; 189690e44470SLai Jiangshan 189790e44470SLai Jiangshan /* 189890e44470SLai Jiangshan * Direct pages can never be unsync, and KVM should never attempt to 189990e44470SLai Jiangshan * sync a shadow page for a different MMU context, e.g. if the role 190090e44470SLai Jiangshan * differs then the memslot lookup (SMM vs. non-SMM) will be bogus, the 190190e44470SLai Jiangshan * reserved bits checks will be wrong, etc... 190290e44470SLai Jiangshan */ 1903c3c6c9fcSLai Jiangshan if (WARN_ON_ONCE(sp->role.direct || !vcpu->arch.mmu->sync_spte || 190490e44470SLai Jiangshan (sp->role.word ^ root_role.word) & ~sync_role_ign.word)) 190590e44470SLai Jiangshan return false; 190690e44470SLai Jiangshan 190790e44470SLai Jiangshan return true; 190890e44470SLai Jiangshan } 190990e44470SLai Jiangshan 191019ace7d6SLai Jiangshan static int kvm_sync_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp, int i) 191119ace7d6SLai Jiangshan { 1912d8fa2031SSean Christopherson /* sp->spt[i] has initial value of shadow page table allocation */ 1913d8fa2031SSean Christopherson if (sp->spt[i] == SHADOW_NONPRESENT_VALUE) 191419ace7d6SLai Jiangshan return 0; 191519ace7d6SLai Jiangshan 191619ace7d6SLai Jiangshan return vcpu->arch.mmu->sync_spte(vcpu, sp, i); 191719ace7d6SLai Jiangshan } 191819ace7d6SLai Jiangshan 191990e44470SLai Jiangshan static int __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp) 192090e44470SLai Jiangshan { 1921c3c6c9fcSLai Jiangshan int flush = 0; 1922c3c6c9fcSLai Jiangshan int i; 1923c3c6c9fcSLai Jiangshan 192490e44470SLai Jiangshan if (!kvm_sync_page_check(vcpu, sp)) 192590e44470SLai Jiangshan return -1; 192690e44470SLai Jiangshan 1927c3c6c9fcSLai Jiangshan for (i = 0; i < SPTE_ENT_PER_PAGE; i++) { 192819ace7d6SLai Jiangshan int ret = kvm_sync_spte(vcpu, sp, i); 1929c3c6c9fcSLai Jiangshan 1930c3c6c9fcSLai Jiangshan if (ret < -1) 1931c3c6c9fcSLai Jiangshan return -1; 1932c3c6c9fcSLai Jiangshan flush |= ret; 1933c3c6c9fcSLai Jiangshan } 1934c3c6c9fcSLai Jiangshan 1935c3c6c9fcSLai Jiangshan /* 1936c3c6c9fcSLai Jiangshan * Note, any flush is purely for KVM's correctness, e.g. when dropping 1937c3c6c9fcSLai Jiangshan * an existing SPTE or clearing W/A/D bits to ensure an mmu_notifier 1938c3c6c9fcSLai Jiangshan * unmap or dirty logging event doesn't fail to flush. The guest is 1939c3c6c9fcSLai Jiangshan * responsible for flushing the TLB to ensure any changes in protection 1940c3c6c9fcSLai Jiangshan * bits are recognized, i.e. until the guest flushes or page faults on 1941c3c6c9fcSLai Jiangshan * a relevant address, KVM is architecturally allowed to let vCPUs use 1942c3c6c9fcSLai Jiangshan * cached translations with the old protection bits. 1943c3c6c9fcSLai Jiangshan */ 1944c3c6c9fcSLai Jiangshan return flush; 194590e44470SLai Jiangshan } 194690e44470SLai Jiangshan 19478d5678a7SHou Wenlong static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp, 1948c50d8ae3SPaolo Bonzini struct list_head *invalid_list) 1949c50d8ae3SPaolo Bonzini { 195090e44470SLai Jiangshan int ret = __kvm_sync_page(vcpu, sp); 1951c3e5e415SLai Jiangshan 19528d5678a7SHou Wenlong if (ret < 0) 1953c50d8ae3SPaolo Bonzini kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list); 19548d5678a7SHou Wenlong return ret; 1955c50d8ae3SPaolo Bonzini } 1956c50d8ae3SPaolo Bonzini 1957c50d8ae3SPaolo Bonzini static bool kvm_mmu_remote_flush_or_zap(struct kvm *kvm, 1958c50d8ae3SPaolo Bonzini struct list_head *invalid_list, 1959c50d8ae3SPaolo Bonzini bool remote_flush) 1960c50d8ae3SPaolo Bonzini { 1961c50d8ae3SPaolo Bonzini if (!remote_flush && list_empty(invalid_list)) 1962c50d8ae3SPaolo Bonzini return false; 1963c50d8ae3SPaolo Bonzini 1964c50d8ae3SPaolo Bonzini if (!list_empty(invalid_list)) 1965c50d8ae3SPaolo Bonzini kvm_mmu_commit_zap_page(kvm, invalid_list); 1966c50d8ae3SPaolo Bonzini else 1967c50d8ae3SPaolo Bonzini kvm_flush_remote_tlbs(kvm); 1968c50d8ae3SPaolo Bonzini return true; 1969c50d8ae3SPaolo Bonzini } 1970c50d8ae3SPaolo Bonzini 1971c50d8ae3SPaolo Bonzini static bool is_obsolete_sp(struct kvm *kvm, struct kvm_mmu_page *sp) 1972c50d8ae3SPaolo Bonzini { 1973a955cad8SSean Christopherson if (sp->role.invalid) 1974a955cad8SSean Christopherson return true; 1975a955cad8SSean Christopherson 1976fa3e4203SMiaohe Lin /* TDP MMU pages do not use the MMU generation. */ 1977de0322f5SSean Christopherson return !is_tdp_mmu_page(sp) && 1978c50d8ae3SPaolo Bonzini unlikely(sp->mmu_valid_gen != kvm->arch.mmu_valid_gen); 1979c50d8ae3SPaolo Bonzini } 1980c50d8ae3SPaolo Bonzini 1981c50d8ae3SPaolo Bonzini struct mmu_page_path { 1982c50d8ae3SPaolo Bonzini struct kvm_mmu_page *parent[PT64_ROOT_MAX_LEVEL]; 1983c50d8ae3SPaolo Bonzini unsigned int idx[PT64_ROOT_MAX_LEVEL]; 1984c50d8ae3SPaolo Bonzini }; 1985c50d8ae3SPaolo Bonzini 1986c50d8ae3SPaolo Bonzini #define for_each_sp(pvec, sp, parents, i) \ 1987c50d8ae3SPaolo Bonzini for (i = mmu_pages_first(&pvec, &parents); \ 1988c50d8ae3SPaolo Bonzini i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \ 1989c50d8ae3SPaolo Bonzini i = mmu_pages_next(&pvec, &parents, i)) 1990c50d8ae3SPaolo Bonzini 1991c50d8ae3SPaolo Bonzini static int mmu_pages_next(struct kvm_mmu_pages *pvec, 1992c50d8ae3SPaolo Bonzini struct mmu_page_path *parents, 1993c50d8ae3SPaolo Bonzini int i) 1994c50d8ae3SPaolo Bonzini { 1995c50d8ae3SPaolo Bonzini int n; 1996c50d8ae3SPaolo Bonzini 1997c50d8ae3SPaolo Bonzini for (n = i+1; n < pvec->nr; n++) { 1998c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp = pvec->page[n].sp; 1999c50d8ae3SPaolo Bonzini unsigned idx = pvec->page[n].idx; 2000c50d8ae3SPaolo Bonzini int level = sp->role.level; 2001c50d8ae3SPaolo Bonzini 2002c50d8ae3SPaolo Bonzini parents->idx[level-1] = idx; 20033bae0459SSean Christopherson if (level == PG_LEVEL_4K) 2004c50d8ae3SPaolo Bonzini break; 2005c50d8ae3SPaolo Bonzini 2006c50d8ae3SPaolo Bonzini parents->parent[level-2] = sp; 2007c50d8ae3SPaolo Bonzini } 2008c50d8ae3SPaolo Bonzini 2009c50d8ae3SPaolo Bonzini return n; 2010c50d8ae3SPaolo Bonzini } 2011c50d8ae3SPaolo Bonzini 2012c50d8ae3SPaolo Bonzini static int mmu_pages_first(struct kvm_mmu_pages *pvec, 2013c50d8ae3SPaolo Bonzini struct mmu_page_path *parents) 2014c50d8ae3SPaolo Bonzini { 2015c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2016c50d8ae3SPaolo Bonzini int level; 2017c50d8ae3SPaolo Bonzini 2018c50d8ae3SPaolo Bonzini if (pvec->nr == 0) 2019c50d8ae3SPaolo Bonzini return 0; 2020c50d8ae3SPaolo Bonzini 202120ba462dSSean Christopherson WARN_ON_ONCE(pvec->page[0].idx != INVALID_INDEX); 2022c50d8ae3SPaolo Bonzini 2023c50d8ae3SPaolo Bonzini sp = pvec->page[0].sp; 2024c50d8ae3SPaolo Bonzini level = sp->role.level; 202520ba462dSSean Christopherson WARN_ON_ONCE(level == PG_LEVEL_4K); 2026c50d8ae3SPaolo Bonzini 2027c50d8ae3SPaolo Bonzini parents->parent[level-2] = sp; 2028c50d8ae3SPaolo Bonzini 2029c50d8ae3SPaolo Bonzini /* Also set up a sentinel. Further entries in pvec are all 2030c50d8ae3SPaolo Bonzini * children of sp, so this element is never overwritten. 2031c50d8ae3SPaolo Bonzini */ 2032c50d8ae3SPaolo Bonzini parents->parent[level-1] = NULL; 2033c50d8ae3SPaolo Bonzini return mmu_pages_next(pvec, parents, 0); 2034c50d8ae3SPaolo Bonzini } 2035c50d8ae3SPaolo Bonzini 2036c50d8ae3SPaolo Bonzini static void mmu_pages_clear_parents(struct mmu_page_path *parents) 2037c50d8ae3SPaolo Bonzini { 2038c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2039c50d8ae3SPaolo Bonzini unsigned int level = 0; 2040c50d8ae3SPaolo Bonzini 2041c50d8ae3SPaolo Bonzini do { 2042c50d8ae3SPaolo Bonzini unsigned int idx = parents->idx[level]; 2043c50d8ae3SPaolo Bonzini sp = parents->parent[level]; 2044c50d8ae3SPaolo Bonzini if (!sp) 2045c50d8ae3SPaolo Bonzini return; 2046c50d8ae3SPaolo Bonzini 204720ba462dSSean Christopherson WARN_ON_ONCE(idx == INVALID_INDEX); 2048c50d8ae3SPaolo Bonzini clear_unsync_child_bit(sp, idx); 2049c50d8ae3SPaolo Bonzini level++; 2050c50d8ae3SPaolo Bonzini } while (!sp->unsync_children); 2051c50d8ae3SPaolo Bonzini } 2052c50d8ae3SPaolo Bonzini 205365855ed8SLai Jiangshan static int mmu_sync_children(struct kvm_vcpu *vcpu, 205465855ed8SLai Jiangshan struct kvm_mmu_page *parent, bool can_yield) 2055c50d8ae3SPaolo Bonzini { 2056c50d8ae3SPaolo Bonzini int i; 2057c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2058c50d8ae3SPaolo Bonzini struct mmu_page_path parents; 2059c50d8ae3SPaolo Bonzini struct kvm_mmu_pages pages; 2060c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 2061c50d8ae3SPaolo Bonzini bool flush = false; 2062c50d8ae3SPaolo Bonzini 2063c50d8ae3SPaolo Bonzini while (mmu_unsync_walk(parent, &pages)) { 2064c50d8ae3SPaolo Bonzini bool protected = false; 2065c50d8ae3SPaolo Bonzini 2066c50d8ae3SPaolo Bonzini for_each_sp(pages, sp, parents, i) 2067cf48f9e2SDavid Matlack protected |= kvm_vcpu_write_protect_gfn(vcpu, sp->gfn); 2068c50d8ae3SPaolo Bonzini 2069c50d8ae3SPaolo Bonzini if (protected) { 20705591c069SLai Jiangshan kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, true); 2071c50d8ae3SPaolo Bonzini flush = false; 2072c50d8ae3SPaolo Bonzini } 2073c50d8ae3SPaolo Bonzini 2074c50d8ae3SPaolo Bonzini for_each_sp(pages, sp, parents, i) { 2075479a1efcSSean Christopherson kvm_unlink_unsync_page(vcpu->kvm, sp); 20768d5678a7SHou Wenlong flush |= kvm_sync_page(vcpu, sp, &invalid_list) > 0; 2077c50d8ae3SPaolo Bonzini mmu_pages_clear_parents(&parents); 2078c50d8ae3SPaolo Bonzini } 2079531810caSBen Gardon if (need_resched() || rwlock_needbreak(&vcpu->kvm->mmu_lock)) { 2080c3e5e415SLai Jiangshan kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, flush); 208165855ed8SLai Jiangshan if (!can_yield) { 208265855ed8SLai Jiangshan kvm_make_request(KVM_REQ_MMU_SYNC, vcpu); 208365855ed8SLai Jiangshan return -EINTR; 208465855ed8SLai Jiangshan } 208565855ed8SLai Jiangshan 2086531810caSBen Gardon cond_resched_rwlock_write(&vcpu->kvm->mmu_lock); 2087c50d8ae3SPaolo Bonzini flush = false; 2088c50d8ae3SPaolo Bonzini } 2089c50d8ae3SPaolo Bonzini } 2090c50d8ae3SPaolo Bonzini 2091c3e5e415SLai Jiangshan kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, flush); 209265855ed8SLai Jiangshan return 0; 2093c50d8ae3SPaolo Bonzini } 2094c50d8ae3SPaolo Bonzini 2095c50d8ae3SPaolo Bonzini static void __clear_sp_write_flooding_count(struct kvm_mmu_page *sp) 2096c50d8ae3SPaolo Bonzini { 2097c50d8ae3SPaolo Bonzini atomic_set(&sp->write_flooding_count, 0); 2098c50d8ae3SPaolo Bonzini } 2099c50d8ae3SPaolo Bonzini 2100c50d8ae3SPaolo Bonzini static void clear_sp_write_flooding_count(u64 *spte) 2101c50d8ae3SPaolo Bonzini { 210257354682SSean Christopherson __clear_sp_write_flooding_count(sptep_to_sp(spte)); 2103c50d8ae3SPaolo Bonzini } 2104c50d8ae3SPaolo Bonzini 2105cbd858b1SDavid Matlack /* 2106cbd858b1SDavid Matlack * The vCPU is required when finding indirect shadow pages; the shadow 2107cbd858b1SDavid Matlack * page may already exist and syncing it needs the vCPU pointer in 2108cbd858b1SDavid Matlack * order to read guest page tables. Direct shadow pages are never 2109cbd858b1SDavid Matlack * unsync, thus @vcpu can be NULL if @role.direct is true. 2110cbd858b1SDavid Matlack */ 21113cc736b3SDavid Matlack static struct kvm_mmu_page *kvm_mmu_find_shadow_page(struct kvm *kvm, 21123cc736b3SDavid Matlack struct kvm_vcpu *vcpu, 211394c81364SDavid Matlack gfn_t gfn, 211494c81364SDavid Matlack struct hlist_head *sp_list, 21152e65e842SDavid Matlack union kvm_mmu_page_role role) 2116c50d8ae3SPaolo Bonzini { 2117c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 21188d5678a7SHou Wenlong int ret; 2119c50d8ae3SPaolo Bonzini int collisions = 0; 2120c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 2121c50d8ae3SPaolo Bonzini 21223cc736b3SDavid Matlack for_each_valid_sp(kvm, sp, sp_list) { 2123c50d8ae3SPaolo Bonzini if (sp->gfn != gfn) { 2124c50d8ae3SPaolo Bonzini collisions++; 2125c50d8ae3SPaolo Bonzini continue; 2126c50d8ae3SPaolo Bonzini } 2127c50d8ae3SPaolo Bonzini 2128ddc16abbSSean Christopherson if (sp->role.word != role.word) { 2129ddc16abbSSean Christopherson /* 2130ddc16abbSSean Christopherson * If the guest is creating an upper-level page, zap 2131ddc16abbSSean Christopherson * unsync pages for the same gfn. While it's possible 2132ddc16abbSSean Christopherson * the guest is using recursive page tables, in all 2133ddc16abbSSean Christopherson * likelihood the guest has stopped using the unsync 2134ddc16abbSSean Christopherson * page and is installing a completely unrelated page. 2135ddc16abbSSean Christopherson * Unsync pages must not be left as is, because the new 2136ddc16abbSSean Christopherson * upper-level page will be write-protected. 2137ddc16abbSSean Christopherson */ 21382e65e842SDavid Matlack if (role.level > PG_LEVEL_4K && sp->unsync) 21393cc736b3SDavid Matlack kvm_mmu_prepare_zap_page(kvm, sp, 2140ddc16abbSSean Christopherson &invalid_list); 2141c50d8ae3SPaolo Bonzini continue; 2142ddc16abbSSean Christopherson } 2143c50d8ae3SPaolo Bonzini 2144bb924ca6SDavid Matlack /* unsync and write-flooding only apply to indirect SPs. */ 2145bb924ca6SDavid Matlack if (sp->role.direct) 214694c81364SDavid Matlack goto out; 2147fb58a9c3SSean Christopherson 2148c50d8ae3SPaolo Bonzini if (sp->unsync) { 2149cbd858b1SDavid Matlack if (KVM_BUG_ON(!vcpu, kvm)) 2150cbd858b1SDavid Matlack break; 2151cbd858b1SDavid Matlack 215207dc4f35SSean Christopherson /* 2153479a1efcSSean Christopherson * The page is good, but is stale. kvm_sync_page does 215407dc4f35SSean Christopherson * get the latest guest state, but (unlike mmu_unsync_children) 215507dc4f35SSean Christopherson * it doesn't write-protect the page or mark it synchronized! 215607dc4f35SSean Christopherson * This way the validity of the mapping is ensured, but the 215707dc4f35SSean Christopherson * overhead of write protection is not incurred until the 215807dc4f35SSean Christopherson * guest invalidates the TLB mapping. This allows multiple 215907dc4f35SSean Christopherson * SPs for a single gfn to be unsync. 216007dc4f35SSean Christopherson * 216107dc4f35SSean Christopherson * If the sync fails, the page is zapped. If so, break 216207dc4f35SSean Christopherson * in order to rebuild it. 2163c50d8ae3SPaolo Bonzini */ 21648d5678a7SHou Wenlong ret = kvm_sync_page(vcpu, sp, &invalid_list); 21658d5678a7SHou Wenlong if (ret < 0) 2166c50d8ae3SPaolo Bonzini break; 2167c50d8ae3SPaolo Bonzini 216820ba462dSSean Christopherson WARN_ON_ONCE(!list_empty(&invalid_list)); 21698d5678a7SHou Wenlong if (ret > 0) 21703cc736b3SDavid Matlack kvm_flush_remote_tlbs(kvm); 2171c50d8ae3SPaolo Bonzini } 2172c50d8ae3SPaolo Bonzini 2173c50d8ae3SPaolo Bonzini __clear_sp_write_flooding_count(sp); 2174fb58a9c3SSean Christopherson 2175c50d8ae3SPaolo Bonzini goto out; 2176c50d8ae3SPaolo Bonzini } 2177c50d8ae3SPaolo Bonzini 217894c81364SDavid Matlack sp = NULL; 21793cc736b3SDavid Matlack ++kvm->stat.mmu_cache_miss; 2180c50d8ae3SPaolo Bonzini 218194c81364SDavid Matlack out: 21823cc736b3SDavid Matlack kvm_mmu_commit_zap_page(kvm, &invalid_list); 218394c81364SDavid Matlack 21843cc736b3SDavid Matlack if (collisions > kvm->stat.max_mmu_page_hash_collisions) 21853cc736b3SDavid Matlack kvm->stat.max_mmu_page_hash_collisions = collisions; 218694c81364SDavid Matlack return sp; 218794c81364SDavid Matlack } 218894c81364SDavid Matlack 21892f8b1b53SDavid Matlack /* Caches used when allocating a new shadow page. */ 21902f8b1b53SDavid Matlack struct shadow_page_caches { 21912f8b1b53SDavid Matlack struct kvm_mmu_memory_cache *page_header_cache; 21922f8b1b53SDavid Matlack struct kvm_mmu_memory_cache *shadow_page_cache; 21936a97575dSDavid Matlack struct kvm_mmu_memory_cache *shadowed_info_cache; 21942f8b1b53SDavid Matlack }; 21952f8b1b53SDavid Matlack 2196336081fbSDavid Matlack static struct kvm_mmu_page *kvm_mmu_alloc_shadow_page(struct kvm *kvm, 21972f8b1b53SDavid Matlack struct shadow_page_caches *caches, 219894c81364SDavid Matlack gfn_t gfn, 219994c81364SDavid Matlack struct hlist_head *sp_list, 220094c81364SDavid Matlack union kvm_mmu_page_role role) 220194c81364SDavid Matlack { 2202c306aec8SDavid Matlack struct kvm_mmu_page *sp; 2203c306aec8SDavid Matlack 22042f8b1b53SDavid Matlack sp = kvm_mmu_memory_cache_alloc(caches->page_header_cache); 22052f8b1b53SDavid Matlack sp->spt = kvm_mmu_memory_cache_alloc(caches->shadow_page_cache); 2206c306aec8SDavid Matlack if (!role.direct) 22076a97575dSDavid Matlack sp->shadowed_translation = kvm_mmu_memory_cache_alloc(caches->shadowed_info_cache); 2208c306aec8SDavid Matlack 2209c306aec8SDavid Matlack set_page_private(virt_to_page(sp->spt), (unsigned long)sp); 2210c306aec8SDavid Matlack 221155c510e2SSean Christopherson INIT_LIST_HEAD(&sp->possible_nx_huge_page_link); 2212428e9216SSean Christopherson 2213c306aec8SDavid Matlack /* 2214c306aec8SDavid Matlack * active_mmu_pages must be a FIFO list, as kvm_zap_obsolete_pages() 2215c306aec8SDavid Matlack * depends on valid pages being added to the head of the list. See 2216c306aec8SDavid Matlack * comments in kvm_zap_obsolete_pages(). 2217c306aec8SDavid Matlack */ 2218336081fbSDavid Matlack sp->mmu_valid_gen = kvm->arch.mmu_valid_gen; 2219336081fbSDavid Matlack list_add(&sp->link, &kvm->arch.active_mmu_pages); 222043a063caSYosry Ahmed kvm_account_mmu_page(kvm, sp); 2221c50d8ae3SPaolo Bonzini 2222c50d8ae3SPaolo Bonzini sp->gfn = gfn; 2223c50d8ae3SPaolo Bonzini sp->role = role; 2224ac101b7cSSean Christopherson hlist_add_head(&sp->hash_link, sp_list); 2225be911771SDavid Matlack if (sp_has_gptes(sp)) 2226336081fbSDavid Matlack account_shadowed(kvm, sp); 2227ddc16abbSSean Christopherson 222894c81364SDavid Matlack return sp; 222994c81364SDavid Matlack } 223094c81364SDavid Matlack 2231cbd858b1SDavid Matlack /* Note, @vcpu may be NULL if @role.direct is true; see kvm_mmu_find_shadow_page. */ 22323cc736b3SDavid Matlack static struct kvm_mmu_page *__kvm_mmu_get_shadow_page(struct kvm *kvm, 22333cc736b3SDavid Matlack struct kvm_vcpu *vcpu, 22342f8b1b53SDavid Matlack struct shadow_page_caches *caches, 223587654643SDavid Matlack gfn_t gfn, 223694c81364SDavid Matlack union kvm_mmu_page_role role) 223794c81364SDavid Matlack { 223894c81364SDavid Matlack struct hlist_head *sp_list; 223994c81364SDavid Matlack struct kvm_mmu_page *sp; 224094c81364SDavid Matlack bool created = false; 224194c81364SDavid Matlack 22423cc736b3SDavid Matlack sp_list = &kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]; 224394c81364SDavid Matlack 22443cc736b3SDavid Matlack sp = kvm_mmu_find_shadow_page(kvm, vcpu, gfn, sp_list, role); 224594c81364SDavid Matlack if (!sp) { 224694c81364SDavid Matlack created = true; 22473cc736b3SDavid Matlack sp = kvm_mmu_alloc_shadow_page(kvm, caches, gfn, sp_list, role); 224894c81364SDavid Matlack } 224994c81364SDavid Matlack 225094c81364SDavid Matlack trace_kvm_mmu_get_page(sp, created); 2251c50d8ae3SPaolo Bonzini return sp; 2252c50d8ae3SPaolo Bonzini } 2253c50d8ae3SPaolo Bonzini 22542f8b1b53SDavid Matlack static struct kvm_mmu_page *kvm_mmu_get_shadow_page(struct kvm_vcpu *vcpu, 22552f8b1b53SDavid Matlack gfn_t gfn, 22562f8b1b53SDavid Matlack union kvm_mmu_page_role role) 22572f8b1b53SDavid Matlack { 22582f8b1b53SDavid Matlack struct shadow_page_caches caches = { 22592f8b1b53SDavid Matlack .page_header_cache = &vcpu->arch.mmu_page_header_cache, 22602f8b1b53SDavid Matlack .shadow_page_cache = &vcpu->arch.mmu_shadow_page_cache, 22616a97575dSDavid Matlack .shadowed_info_cache = &vcpu->arch.mmu_shadowed_info_cache, 22622f8b1b53SDavid Matlack }; 22632f8b1b53SDavid Matlack 22643cc736b3SDavid Matlack return __kvm_mmu_get_shadow_page(vcpu->kvm, vcpu, &caches, gfn, role); 22652f8b1b53SDavid Matlack } 22662f8b1b53SDavid Matlack 226739944ab9SSean Christopherson static union kvm_mmu_page_role kvm_mmu_child_role(u64 *sptep, bool direct, 226839944ab9SSean Christopherson unsigned int access) 22692e65e842SDavid Matlack { 22702e65e842SDavid Matlack struct kvm_mmu_page *parent_sp = sptep_to_sp(sptep); 22712e65e842SDavid Matlack union kvm_mmu_page_role role; 22722e65e842SDavid Matlack 22732e65e842SDavid Matlack role = parent_sp->role; 22742e65e842SDavid Matlack role.level--; 22752e65e842SDavid Matlack role.access = access; 22762e65e842SDavid Matlack role.direct = direct; 22772e65e842SDavid Matlack role.passthrough = 0; 22782e65e842SDavid Matlack 22792e65e842SDavid Matlack /* 22802e65e842SDavid Matlack * If the guest has 4-byte PTEs then that means it's using 32-bit, 22812e65e842SDavid Matlack * 2-level, non-PAE paging. KVM shadows such guests with PAE paging 22822e65e842SDavid Matlack * (i.e. 8-byte PTEs). The difference in PTE size means that KVM must 22832e65e842SDavid Matlack * shadow each guest page table with multiple shadow page tables, which 22842e65e842SDavid Matlack * requires extra bookkeeping in the role. 22852e65e842SDavid Matlack * 22862e65e842SDavid Matlack * Specifically, to shadow the guest's page directory (which covers a 22872e65e842SDavid Matlack * 4GiB address space), KVM uses 4 PAE page directories, each mapping 22882e65e842SDavid Matlack * 1GiB of the address space. @role.quadrant encodes which quarter of 22892e65e842SDavid Matlack * the address space each maps. 22902e65e842SDavid Matlack * 22912e65e842SDavid Matlack * To shadow the guest's page tables (which each map a 4MiB region), KVM 22922e65e842SDavid Matlack * uses 2 PAE page tables, each mapping a 2MiB region. For these, 22932e65e842SDavid Matlack * @role.quadrant encodes which half of the region they map. 22942e65e842SDavid Matlack * 229539944ab9SSean Christopherson * Concretely, a 4-byte PDE consumes bits 31:22, while an 8-byte PDE 229639944ab9SSean Christopherson * consumes bits 29:21. To consume bits 31:30, KVM's uses 4 shadow 229739944ab9SSean Christopherson * PDPTEs; those 4 PAE page directories are pre-allocated and their 229839944ab9SSean Christopherson * quadrant is assigned in mmu_alloc_root(). A 4-byte PTE consumes 229939944ab9SSean Christopherson * bits 21:12, while an 8-byte PTE consumes bits 20:12. To consume 230039944ab9SSean Christopherson * bit 21 in the PTE (the child here), KVM propagates that bit to the 230139944ab9SSean Christopherson * quadrant, i.e. sets quadrant to '0' or '1'. The parent 8-byte PDE 230239944ab9SSean Christopherson * covers bit 21 (see above), thus the quadrant is calculated from the 230339944ab9SSean Christopherson * _least_ significant bit of the PDE index. 23042e65e842SDavid Matlack */ 23052e65e842SDavid Matlack if (role.has_4_byte_gpte) { 23062e65e842SDavid Matlack WARN_ON_ONCE(role.level != PG_LEVEL_4K); 230779e48cecSSean Christopherson role.quadrant = spte_index(sptep) & 1; 23082e65e842SDavid Matlack } 23092e65e842SDavid Matlack 23102e65e842SDavid Matlack return role; 23112e65e842SDavid Matlack } 23122e65e842SDavid Matlack 23132e65e842SDavid Matlack static struct kvm_mmu_page *kvm_mmu_get_child_sp(struct kvm_vcpu *vcpu, 23142e65e842SDavid Matlack u64 *sptep, gfn_t gfn, 23152e65e842SDavid Matlack bool direct, unsigned int access) 23162e65e842SDavid Matlack { 23172e65e842SDavid Matlack union kvm_mmu_page_role role; 23182e65e842SDavid Matlack 23190cd8dc73SPaolo Bonzini if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) 23200cd8dc73SPaolo Bonzini return ERR_PTR(-EEXIST); 23210cd8dc73SPaolo Bonzini 23222e65e842SDavid Matlack role = kvm_mmu_child_role(sptep, direct, access); 232387654643SDavid Matlack return kvm_mmu_get_shadow_page(vcpu, gfn, role); 23242e65e842SDavid Matlack } 23252e65e842SDavid Matlack 2326c50d8ae3SPaolo Bonzini static void shadow_walk_init_using_root(struct kvm_shadow_walk_iterator *iterator, 2327c50d8ae3SPaolo Bonzini struct kvm_vcpu *vcpu, hpa_t root, 2328c50d8ae3SPaolo Bonzini u64 addr) 2329c50d8ae3SPaolo Bonzini { 2330c50d8ae3SPaolo Bonzini iterator->addr = addr; 2331c50d8ae3SPaolo Bonzini iterator->shadow_addr = root; 2332a972e29cSPaolo Bonzini iterator->level = vcpu->arch.mmu->root_role.level; 2333c50d8ae3SPaolo Bonzini 233412ec33a7SLai Jiangshan if (iterator->level >= PT64_ROOT_4LEVEL && 23354d25502aSPaolo Bonzini vcpu->arch.mmu->cpu_role.base.level < PT64_ROOT_4LEVEL && 2336347a0d0dSPaolo Bonzini !vcpu->arch.mmu->root_role.direct) 233712ec33a7SLai Jiangshan iterator->level = PT32E_ROOT_LEVEL; 2338c50d8ae3SPaolo Bonzini 2339c50d8ae3SPaolo Bonzini if (iterator->level == PT32E_ROOT_LEVEL) { 2340c50d8ae3SPaolo Bonzini /* 2341c50d8ae3SPaolo Bonzini * prev_root is currently only used for 64-bit hosts. So only 2342c50d8ae3SPaolo Bonzini * the active root_hpa is valid here. 2343c50d8ae3SPaolo Bonzini */ 2344b9e5603cSPaolo Bonzini BUG_ON(root != vcpu->arch.mmu->root.hpa); 2345c50d8ae3SPaolo Bonzini 2346c50d8ae3SPaolo Bonzini iterator->shadow_addr 2347c50d8ae3SPaolo Bonzini = vcpu->arch.mmu->pae_root[(addr >> 30) & 3]; 23482ca3129eSSean Christopherson iterator->shadow_addr &= SPTE_BASE_ADDR_MASK; 2349c50d8ae3SPaolo Bonzini --iterator->level; 2350c50d8ae3SPaolo Bonzini if (!iterator->shadow_addr) 2351c50d8ae3SPaolo Bonzini iterator->level = 0; 2352c50d8ae3SPaolo Bonzini } 2353c50d8ae3SPaolo Bonzini } 2354c50d8ae3SPaolo Bonzini 2355c50d8ae3SPaolo Bonzini static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator, 2356c50d8ae3SPaolo Bonzini struct kvm_vcpu *vcpu, u64 addr) 2357c50d8ae3SPaolo Bonzini { 2358b9e5603cSPaolo Bonzini shadow_walk_init_using_root(iterator, vcpu, vcpu->arch.mmu->root.hpa, 2359c50d8ae3SPaolo Bonzini addr); 2360c50d8ae3SPaolo Bonzini } 2361c50d8ae3SPaolo Bonzini 2362c50d8ae3SPaolo Bonzini static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator) 2363c50d8ae3SPaolo Bonzini { 23643bae0459SSean Christopherson if (iterator->level < PG_LEVEL_4K) 2365c50d8ae3SPaolo Bonzini return false; 2366c50d8ae3SPaolo Bonzini 23672ca3129eSSean Christopherson iterator->index = SPTE_INDEX(iterator->addr, iterator->level); 2368c50d8ae3SPaolo Bonzini iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index; 2369c50d8ae3SPaolo Bonzini return true; 2370c50d8ae3SPaolo Bonzini } 2371c50d8ae3SPaolo Bonzini 2372c50d8ae3SPaolo Bonzini static void __shadow_walk_next(struct kvm_shadow_walk_iterator *iterator, 2373c50d8ae3SPaolo Bonzini u64 spte) 2374c50d8ae3SPaolo Bonzini { 23753e44dce4SLai Jiangshan if (!is_shadow_present_pte(spte) || is_last_spte(spte, iterator->level)) { 2376c50d8ae3SPaolo Bonzini iterator->level = 0; 2377c50d8ae3SPaolo Bonzini return; 2378c50d8ae3SPaolo Bonzini } 2379c50d8ae3SPaolo Bonzini 23802ca3129eSSean Christopherson iterator->shadow_addr = spte & SPTE_BASE_ADDR_MASK; 2381c50d8ae3SPaolo Bonzini --iterator->level; 2382c50d8ae3SPaolo Bonzini } 2383c50d8ae3SPaolo Bonzini 2384c50d8ae3SPaolo Bonzini static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator) 2385c50d8ae3SPaolo Bonzini { 2386c50d8ae3SPaolo Bonzini __shadow_walk_next(iterator, *iterator->sptep); 2387c50d8ae3SPaolo Bonzini } 2388c50d8ae3SPaolo Bonzini 23890cd8dc73SPaolo Bonzini static void __link_shadow_page(struct kvm *kvm, 23900cd8dc73SPaolo Bonzini struct kvm_mmu_memory_cache *cache, u64 *sptep, 239103787394SPaolo Bonzini struct kvm_mmu_page *sp, bool flush) 2392c50d8ae3SPaolo Bonzini { 2393c50d8ae3SPaolo Bonzini u64 spte; 2394c50d8ae3SPaolo Bonzini 2395c50d8ae3SPaolo Bonzini BUILD_BUG_ON(VMX_EPT_WRITABLE_MASK != PT_WRITABLE_MASK); 2396c50d8ae3SPaolo Bonzini 23970cd8dc73SPaolo Bonzini /* 23980cd8dc73SPaolo Bonzini * If an SPTE is present already, it must be a leaf and therefore 239903787394SPaolo Bonzini * a large one. Drop it, and flush the TLB if needed, before 240003787394SPaolo Bonzini * installing sp. 24010cd8dc73SPaolo Bonzini */ 24020cd8dc73SPaolo Bonzini if (is_shadow_present_pte(*sptep)) 240303787394SPaolo Bonzini drop_large_spte(kvm, sptep, flush); 24040cd8dc73SPaolo Bonzini 2405cc4674d0SBen Gardon spte = make_nonleaf_spte(sp->spt, sp_ad_disabled(sp)); 2406c50d8ae3SPaolo Bonzini 2407c50d8ae3SPaolo Bonzini mmu_spte_set(sptep, spte); 2408c50d8ae3SPaolo Bonzini 24092ff9039aSDavid Matlack mmu_page_add_parent_pte(cache, sp, sptep); 2410c50d8ae3SPaolo Bonzini 2411c4a48868SLai Jiangshan /* 2412c4a48868SLai Jiangshan * The non-direct sub-pagetable must be updated before linking. For 2413c4a48868SLai Jiangshan * L1 sp, the pagetable is updated via kvm_sync_page() in 2414c4a48868SLai Jiangshan * kvm_mmu_find_shadow_page() without write-protecting the gfn, 2415c4a48868SLai Jiangshan * so sp->unsync can be true or false. For higher level non-direct 2416c4a48868SLai Jiangshan * sp, the pagetable is updated/synced via mmu_sync_children() in 2417c4a48868SLai Jiangshan * FNAME(fetch)(), so sp->unsync_children can only be false. 2418c4a48868SLai Jiangshan * WARN_ON_ONCE() if anything happens unexpectedly. 2419c4a48868SLai Jiangshan */ 2420c4a48868SLai Jiangshan if (WARN_ON_ONCE(sp->unsync_children) || sp->unsync) 2421c50d8ae3SPaolo Bonzini mark_unsync(sptep); 2422c50d8ae3SPaolo Bonzini } 2423c50d8ae3SPaolo Bonzini 24242ff9039aSDavid Matlack static void link_shadow_page(struct kvm_vcpu *vcpu, u64 *sptep, 24252ff9039aSDavid Matlack struct kvm_mmu_page *sp) 24262ff9039aSDavid Matlack { 242703787394SPaolo Bonzini __link_shadow_page(vcpu->kvm, &vcpu->arch.mmu_pte_list_desc_cache, sptep, sp, true); 24282ff9039aSDavid Matlack } 24292ff9039aSDavid Matlack 2430c50d8ae3SPaolo Bonzini static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep, 2431c50d8ae3SPaolo Bonzini unsigned direct_access) 2432c50d8ae3SPaolo Bonzini { 2433c50d8ae3SPaolo Bonzini if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) { 2434c50d8ae3SPaolo Bonzini struct kvm_mmu_page *child; 2435c50d8ae3SPaolo Bonzini 2436c50d8ae3SPaolo Bonzini /* 2437c50d8ae3SPaolo Bonzini * For the direct sp, if the guest pte's dirty bit 2438c50d8ae3SPaolo Bonzini * changed form clean to dirty, it will corrupt the 2439c50d8ae3SPaolo Bonzini * sp's access: allow writable in the read-only sp, 2440c50d8ae3SPaolo Bonzini * so we should update the spte at this point to get 2441c50d8ae3SPaolo Bonzini * a new sp with the correct access. 2442c50d8ae3SPaolo Bonzini */ 24435e3edd7eSSean Christopherson child = spte_to_child_sp(*sptep); 2444c50d8ae3SPaolo Bonzini if (child->role.access == direct_access) 2445c50d8ae3SPaolo Bonzini return; 2446c50d8ae3SPaolo Bonzini 2447069f30c6SMingwei Zhang drop_parent_pte(vcpu->kvm, child, sptep); 24483cdf9374SHou Wenlong kvm_flush_remote_tlbs_sptep(vcpu->kvm, sptep); 2449c50d8ae3SPaolo Bonzini } 2450c50d8ae3SPaolo Bonzini } 2451c50d8ae3SPaolo Bonzini 24522de4085cSBen Gardon /* Returns the number of zapped non-leaf child shadow pages. */ 24532de4085cSBen Gardon static int mmu_page_zap_pte(struct kvm *kvm, struct kvm_mmu_page *sp, 24542de4085cSBen Gardon u64 *spte, struct list_head *invalid_list) 2455c50d8ae3SPaolo Bonzini { 2456c50d8ae3SPaolo Bonzini u64 pte; 2457c50d8ae3SPaolo Bonzini struct kvm_mmu_page *child; 2458c50d8ae3SPaolo Bonzini 2459c50d8ae3SPaolo Bonzini pte = *spte; 2460c50d8ae3SPaolo Bonzini if (is_shadow_present_pte(pte)) { 2461c50d8ae3SPaolo Bonzini if (is_last_spte(pte, sp->role.level)) { 2462c50d8ae3SPaolo Bonzini drop_spte(kvm, spte); 2463c50d8ae3SPaolo Bonzini } else { 24645e3edd7eSSean Christopherson child = spte_to_child_sp(pte); 2465069f30c6SMingwei Zhang drop_parent_pte(kvm, child, spte); 24662de4085cSBen Gardon 24672de4085cSBen Gardon /* 24682de4085cSBen Gardon * Recursively zap nested TDP SPs, parentless SPs are 24692de4085cSBen Gardon * unlikely to be used again in the near future. This 24702de4085cSBen Gardon * avoids retaining a large number of stale nested SPs. 24712de4085cSBen Gardon */ 24722de4085cSBen Gardon if (tdp_enabled && invalid_list && 24732de4085cSBen Gardon child->role.guest_mode && !child->parent_ptes.val) 24742de4085cSBen Gardon return kvm_mmu_prepare_zap_page(kvm, child, 24752de4085cSBen Gardon invalid_list); 2476c50d8ae3SPaolo Bonzini } 2477949019b9SSean Christopherson } else if (is_mmio_spte(kvm, pte)) { 2478c50d8ae3SPaolo Bonzini mmu_spte_clear_no_track(spte); 2479ace569e0SSean Christopherson } 24802de4085cSBen Gardon return 0; 2481c50d8ae3SPaolo Bonzini } 2482c50d8ae3SPaolo Bonzini 24832de4085cSBen Gardon static int kvm_mmu_page_unlink_children(struct kvm *kvm, 24842de4085cSBen Gardon struct kvm_mmu_page *sp, 24852de4085cSBen Gardon struct list_head *invalid_list) 2486c50d8ae3SPaolo Bonzini { 24872de4085cSBen Gardon int zapped = 0; 2488c50d8ae3SPaolo Bonzini unsigned i; 2489c50d8ae3SPaolo Bonzini 24902ca3129eSSean Christopherson for (i = 0; i < SPTE_ENT_PER_PAGE; ++i) 24912de4085cSBen Gardon zapped += mmu_page_zap_pte(kvm, sp, sp->spt + i, invalid_list); 24922de4085cSBen Gardon 24932de4085cSBen Gardon return zapped; 2494c50d8ae3SPaolo Bonzini } 2495c50d8ae3SPaolo Bonzini 2496069f30c6SMingwei Zhang static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp) 2497c50d8ae3SPaolo Bonzini { 2498c50d8ae3SPaolo Bonzini u64 *sptep; 2499c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 2500c50d8ae3SPaolo Bonzini 2501c50d8ae3SPaolo Bonzini while ((sptep = rmap_get_first(&sp->parent_ptes, &iter))) 2502069f30c6SMingwei Zhang drop_parent_pte(kvm, sp, sptep); 2503c50d8ae3SPaolo Bonzini } 2504c50d8ae3SPaolo Bonzini 2505c50d8ae3SPaolo Bonzini static int mmu_zap_unsync_children(struct kvm *kvm, 2506c50d8ae3SPaolo Bonzini struct kvm_mmu_page *parent, 2507c50d8ae3SPaolo Bonzini struct list_head *invalid_list) 2508c50d8ae3SPaolo Bonzini { 2509c50d8ae3SPaolo Bonzini int i, zapped = 0; 2510c50d8ae3SPaolo Bonzini struct mmu_page_path parents; 2511c50d8ae3SPaolo Bonzini struct kvm_mmu_pages pages; 2512c50d8ae3SPaolo Bonzini 25133bae0459SSean Christopherson if (parent->role.level == PG_LEVEL_4K) 2514c50d8ae3SPaolo Bonzini return 0; 2515c50d8ae3SPaolo Bonzini 2516c50d8ae3SPaolo Bonzini while (mmu_unsync_walk(parent, &pages)) { 2517c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2518c50d8ae3SPaolo Bonzini 2519c50d8ae3SPaolo Bonzini for_each_sp(pages, sp, parents, i) { 2520c50d8ae3SPaolo Bonzini kvm_mmu_prepare_zap_page(kvm, sp, invalid_list); 2521c50d8ae3SPaolo Bonzini mmu_pages_clear_parents(&parents); 2522c50d8ae3SPaolo Bonzini zapped++; 2523c50d8ae3SPaolo Bonzini } 2524c50d8ae3SPaolo Bonzini } 2525c50d8ae3SPaolo Bonzini 2526c50d8ae3SPaolo Bonzini return zapped; 2527c50d8ae3SPaolo Bonzini } 2528c50d8ae3SPaolo Bonzini 2529c50d8ae3SPaolo Bonzini static bool __kvm_mmu_prepare_zap_page(struct kvm *kvm, 2530c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, 2531c50d8ae3SPaolo Bonzini struct list_head *invalid_list, 2532c50d8ae3SPaolo Bonzini int *nr_zapped) 2533c50d8ae3SPaolo Bonzini { 2534527d5cd7SSean Christopherson bool list_unstable, zapped_root = false; 2535c50d8ae3SPaolo Bonzini 253647b0c2e4SKazuki Takiguchi lockdep_assert_held_write(&kvm->mmu_lock); 2537c50d8ae3SPaolo Bonzini trace_kvm_mmu_prepare_zap_page(sp); 2538c50d8ae3SPaolo Bonzini ++kvm->stat.mmu_shadow_zapped; 2539c50d8ae3SPaolo Bonzini *nr_zapped = mmu_zap_unsync_children(kvm, sp, invalid_list); 25402de4085cSBen Gardon *nr_zapped += kvm_mmu_page_unlink_children(kvm, sp, invalid_list); 2541069f30c6SMingwei Zhang kvm_mmu_unlink_parents(kvm, sp); 2542c50d8ae3SPaolo Bonzini 2543c50d8ae3SPaolo Bonzini /* Zapping children means active_mmu_pages has become unstable. */ 2544c50d8ae3SPaolo Bonzini list_unstable = *nr_zapped; 2545c50d8ae3SPaolo Bonzini 2546767d8d8dSLai Jiangshan if (!sp->role.invalid && sp_has_gptes(sp)) 2547c50d8ae3SPaolo Bonzini unaccount_shadowed(kvm, sp); 2548c50d8ae3SPaolo Bonzini 2549c50d8ae3SPaolo Bonzini if (sp->unsync) 2550c50d8ae3SPaolo Bonzini kvm_unlink_unsync_page(kvm, sp); 2551c50d8ae3SPaolo Bonzini if (!sp->root_count) { 2552c50d8ae3SPaolo Bonzini /* Count self */ 2553c50d8ae3SPaolo Bonzini (*nr_zapped)++; 2554f95eec9bSSean Christopherson 2555f95eec9bSSean Christopherson /* 2556f95eec9bSSean Christopherson * Already invalid pages (previously active roots) are not on 2557f95eec9bSSean Christopherson * the active page list. See list_del() in the "else" case of 2558f95eec9bSSean Christopherson * !sp->root_count. 2559f95eec9bSSean Christopherson */ 2560f95eec9bSSean Christopherson if (sp->role.invalid) 2561f95eec9bSSean Christopherson list_add(&sp->link, invalid_list); 2562f95eec9bSSean Christopherson else 2563c50d8ae3SPaolo Bonzini list_move(&sp->link, invalid_list); 256443a063caSYosry Ahmed kvm_unaccount_mmu_page(kvm, sp); 2565c50d8ae3SPaolo Bonzini } else { 2566f95eec9bSSean Christopherson /* 2567f95eec9bSSean Christopherson * Remove the active root from the active page list, the root 2568f95eec9bSSean Christopherson * will be explicitly freed when the root_count hits zero. 2569f95eec9bSSean Christopherson */ 2570f95eec9bSSean Christopherson list_del(&sp->link); 2571c50d8ae3SPaolo Bonzini 2572c50d8ae3SPaolo Bonzini /* 2573c50d8ae3SPaolo Bonzini * Obsolete pages cannot be used on any vCPUs, see the comment 2574c50d8ae3SPaolo Bonzini * in kvm_mmu_zap_all_fast(). Note, is_obsolete_sp() also 2575c50d8ae3SPaolo Bonzini * treats invalid shadow pages as being obsolete. 2576c50d8ae3SPaolo Bonzini */ 2577527d5cd7SSean Christopherson zapped_root = !is_obsolete_sp(kvm, sp); 2578c50d8ae3SPaolo Bonzini } 2579c50d8ae3SPaolo Bonzini 258055c510e2SSean Christopherson if (sp->nx_huge_page_disallowed) 258155c510e2SSean Christopherson unaccount_nx_huge_page(kvm, sp); 2582c50d8ae3SPaolo Bonzini 2583c50d8ae3SPaolo Bonzini sp->role.invalid = 1; 2584527d5cd7SSean Christopherson 2585527d5cd7SSean Christopherson /* 2586527d5cd7SSean Christopherson * Make the request to free obsolete roots after marking the root 2587527d5cd7SSean Christopherson * invalid, otherwise other vCPUs may not see it as invalid. 2588527d5cd7SSean Christopherson */ 2589527d5cd7SSean Christopherson if (zapped_root) 2590527d5cd7SSean Christopherson kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_FREE_OBSOLETE_ROOTS); 2591c50d8ae3SPaolo Bonzini return list_unstable; 2592c50d8ae3SPaolo Bonzini } 2593c50d8ae3SPaolo Bonzini 2594c50d8ae3SPaolo Bonzini static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp, 2595c50d8ae3SPaolo Bonzini struct list_head *invalid_list) 2596c50d8ae3SPaolo Bonzini { 2597c50d8ae3SPaolo Bonzini int nr_zapped; 2598c50d8ae3SPaolo Bonzini 2599c50d8ae3SPaolo Bonzini __kvm_mmu_prepare_zap_page(kvm, sp, invalid_list, &nr_zapped); 2600c50d8ae3SPaolo Bonzini return nr_zapped; 2601c50d8ae3SPaolo Bonzini } 2602c50d8ae3SPaolo Bonzini 2603c50d8ae3SPaolo Bonzini static void kvm_mmu_commit_zap_page(struct kvm *kvm, 2604c50d8ae3SPaolo Bonzini struct list_head *invalid_list) 2605c50d8ae3SPaolo Bonzini { 2606c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, *nsp; 2607c50d8ae3SPaolo Bonzini 2608c50d8ae3SPaolo Bonzini if (list_empty(invalid_list)) 2609c50d8ae3SPaolo Bonzini return; 2610c50d8ae3SPaolo Bonzini 2611c50d8ae3SPaolo Bonzini /* 2612c50d8ae3SPaolo Bonzini * We need to make sure everyone sees our modifications to 2613c50d8ae3SPaolo Bonzini * the page tables and see changes to vcpu->mode here. The barrier 2614c50d8ae3SPaolo Bonzini * in the kvm_flush_remote_tlbs() achieves this. This pairs 2615c50d8ae3SPaolo Bonzini * with vcpu_enter_guest and walk_shadow_page_lockless_begin/end. 2616c50d8ae3SPaolo Bonzini * 2617c50d8ae3SPaolo Bonzini * In addition, kvm_flush_remote_tlbs waits for all vcpus to exit 2618c50d8ae3SPaolo Bonzini * guest mode and/or lockless shadow page table walks. 2619c50d8ae3SPaolo Bonzini */ 2620c50d8ae3SPaolo Bonzini kvm_flush_remote_tlbs(kvm); 2621c50d8ae3SPaolo Bonzini 2622c50d8ae3SPaolo Bonzini list_for_each_entry_safe(sp, nsp, invalid_list, link) { 262320ba462dSSean Christopherson WARN_ON_ONCE(!sp->role.invalid || sp->root_count); 262487654643SDavid Matlack kvm_mmu_free_shadow_page(sp); 2625c50d8ae3SPaolo Bonzini } 2626c50d8ae3SPaolo Bonzini } 2627c50d8ae3SPaolo Bonzini 26286b82ef2cSSean Christopherson static unsigned long kvm_mmu_zap_oldest_mmu_pages(struct kvm *kvm, 26296b82ef2cSSean Christopherson unsigned long nr_to_zap) 2630c50d8ae3SPaolo Bonzini { 26316b82ef2cSSean Christopherson unsigned long total_zapped = 0; 26326b82ef2cSSean Christopherson struct kvm_mmu_page *sp, *tmp; 2633ba7888ddSSean Christopherson LIST_HEAD(invalid_list); 26346b82ef2cSSean Christopherson bool unstable; 26356b82ef2cSSean Christopherson int nr_zapped; 2636c50d8ae3SPaolo Bonzini 2637c50d8ae3SPaolo Bonzini if (list_empty(&kvm->arch.active_mmu_pages)) 2638ba7888ddSSean Christopherson return 0; 2639c50d8ae3SPaolo Bonzini 26406b82ef2cSSean Christopherson restart: 26418fc51726SSean Christopherson list_for_each_entry_safe_reverse(sp, tmp, &kvm->arch.active_mmu_pages, link) { 26426b82ef2cSSean Christopherson /* 26436b82ef2cSSean Christopherson * Don't zap active root pages, the page itself can't be freed 26446b82ef2cSSean Christopherson * and zapping it will just force vCPUs to realloc and reload. 26456b82ef2cSSean Christopherson */ 26466b82ef2cSSean Christopherson if (sp->root_count) 26476b82ef2cSSean Christopherson continue; 26486b82ef2cSSean Christopherson 26496b82ef2cSSean Christopherson unstable = __kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list, 26506b82ef2cSSean Christopherson &nr_zapped); 26516b82ef2cSSean Christopherson total_zapped += nr_zapped; 26526b82ef2cSSean Christopherson if (total_zapped >= nr_to_zap) 2653ba7888ddSSean Christopherson break; 2654ba7888ddSSean Christopherson 26556b82ef2cSSean Christopherson if (unstable) 26566b82ef2cSSean Christopherson goto restart; 2657ba7888ddSSean Christopherson } 26586b82ef2cSSean Christopherson 26596b82ef2cSSean Christopherson kvm_mmu_commit_zap_page(kvm, &invalid_list); 26606b82ef2cSSean Christopherson 26616b82ef2cSSean Christopherson kvm->stat.mmu_recycled += total_zapped; 26626b82ef2cSSean Christopherson return total_zapped; 26636b82ef2cSSean Christopherson } 26646b82ef2cSSean Christopherson 2665afe8d7e6SSean Christopherson static inline unsigned long kvm_mmu_available_pages(struct kvm *kvm) 2666afe8d7e6SSean Christopherson { 2667afe8d7e6SSean Christopherson if (kvm->arch.n_max_mmu_pages > kvm->arch.n_used_mmu_pages) 2668afe8d7e6SSean Christopherson return kvm->arch.n_max_mmu_pages - 2669afe8d7e6SSean Christopherson kvm->arch.n_used_mmu_pages; 2670afe8d7e6SSean Christopherson 2671afe8d7e6SSean Christopherson return 0; 2672c50d8ae3SPaolo Bonzini } 2673c50d8ae3SPaolo Bonzini 2674ba7888ddSSean Christopherson static int make_mmu_pages_available(struct kvm_vcpu *vcpu) 2675ba7888ddSSean Christopherson { 26766b82ef2cSSean Christopherson unsigned long avail = kvm_mmu_available_pages(vcpu->kvm); 2677ba7888ddSSean Christopherson 26786b82ef2cSSean Christopherson if (likely(avail >= KVM_MIN_FREE_MMU_PAGES)) 2679ba7888ddSSean Christopherson return 0; 2680ba7888ddSSean Christopherson 26816b82ef2cSSean Christopherson kvm_mmu_zap_oldest_mmu_pages(vcpu->kvm, KVM_REFILL_PAGES - avail); 2682ba7888ddSSean Christopherson 26836e6ec584SSean Christopherson /* 26846e6ec584SSean Christopherson * Note, this check is intentionally soft, it only guarantees that one 26856e6ec584SSean Christopherson * page is available, while the caller may end up allocating as many as 26866e6ec584SSean Christopherson * four pages, e.g. for PAE roots or for 5-level paging. Temporarily 26876e6ec584SSean Christopherson * exceeding the (arbitrary by default) limit will not harm the host, 2688c4342633SIngo Molnar * being too aggressive may unnecessarily kill the guest, and getting an 26896e6ec584SSean Christopherson * exact count is far more trouble than it's worth, especially in the 26906e6ec584SSean Christopherson * page fault paths. 26916e6ec584SSean Christopherson */ 2692ba7888ddSSean Christopherson if (!kvm_mmu_available_pages(vcpu->kvm)) 2693ba7888ddSSean Christopherson return -ENOSPC; 2694ba7888ddSSean Christopherson return 0; 2695ba7888ddSSean Christopherson } 2696ba7888ddSSean Christopherson 2697c50d8ae3SPaolo Bonzini /* 2698c50d8ae3SPaolo Bonzini * Changing the number of mmu pages allocated to the vm 2699c50d8ae3SPaolo Bonzini * Note: if goal_nr_mmu_pages is too small, you will get dead lock 2700c50d8ae3SPaolo Bonzini */ 2701c50d8ae3SPaolo Bonzini void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned long goal_nr_mmu_pages) 2702c50d8ae3SPaolo Bonzini { 2703531810caSBen Gardon write_lock(&kvm->mmu_lock); 2704c50d8ae3SPaolo Bonzini 2705c50d8ae3SPaolo Bonzini if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) { 27066b82ef2cSSean Christopherson kvm_mmu_zap_oldest_mmu_pages(kvm, kvm->arch.n_used_mmu_pages - 27076b82ef2cSSean Christopherson goal_nr_mmu_pages); 2708c50d8ae3SPaolo Bonzini 2709c50d8ae3SPaolo Bonzini goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages; 2710c50d8ae3SPaolo Bonzini } 2711c50d8ae3SPaolo Bonzini 2712c50d8ae3SPaolo Bonzini kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages; 2713c50d8ae3SPaolo Bonzini 2714531810caSBen Gardon write_unlock(&kvm->mmu_lock); 2715c50d8ae3SPaolo Bonzini } 2716c50d8ae3SPaolo Bonzini 2717c50d8ae3SPaolo Bonzini int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn) 2718c50d8ae3SPaolo Bonzini { 2719c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2720c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 2721c50d8ae3SPaolo Bonzini int r; 2722c50d8ae3SPaolo Bonzini 2723c50d8ae3SPaolo Bonzini r = 0; 2724531810caSBen Gardon write_lock(&kvm->mmu_lock); 2725767d8d8dSLai Jiangshan for_each_gfn_valid_sp_with_gptes(kvm, sp, gfn) { 2726c50d8ae3SPaolo Bonzini r = 1; 2727c50d8ae3SPaolo Bonzini kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list); 2728c50d8ae3SPaolo Bonzini } 2729c50d8ae3SPaolo Bonzini kvm_mmu_commit_zap_page(kvm, &invalid_list); 2730531810caSBen Gardon write_unlock(&kvm->mmu_lock); 2731c50d8ae3SPaolo Bonzini 2732c50d8ae3SPaolo Bonzini return r; 2733c50d8ae3SPaolo Bonzini } 273496ad91aeSSean Christopherson 273596ad91aeSSean Christopherson static int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva) 273696ad91aeSSean Christopherson { 273796ad91aeSSean Christopherson gpa_t gpa; 273896ad91aeSSean Christopherson int r; 273996ad91aeSSean Christopherson 2740347a0d0dSPaolo Bonzini if (vcpu->arch.mmu->root_role.direct) 274196ad91aeSSean Christopherson return 0; 274296ad91aeSSean Christopherson 274396ad91aeSSean Christopherson gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL); 274496ad91aeSSean Christopherson 274596ad91aeSSean Christopherson r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT); 274696ad91aeSSean Christopherson 274796ad91aeSSean Christopherson return r; 274896ad91aeSSean Christopherson } 2749c50d8ae3SPaolo Bonzini 27504d78d0b3SBen Gardon static void kvm_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp) 2751c50d8ae3SPaolo Bonzini { 2752c50d8ae3SPaolo Bonzini trace_kvm_mmu_unsync_page(sp); 27534d78d0b3SBen Gardon ++kvm->stat.mmu_unsync; 2754c50d8ae3SPaolo Bonzini sp->unsync = 1; 2755c50d8ae3SPaolo Bonzini 2756c50d8ae3SPaolo Bonzini kvm_mmu_mark_parents_unsync(sp); 2757c50d8ae3SPaolo Bonzini } 2758c50d8ae3SPaolo Bonzini 27590337f585SSean Christopherson /* 27600337f585SSean Christopherson * Attempt to unsync any shadow pages that can be reached by the specified gfn, 27610337f585SSean Christopherson * KVM is creating a writable mapping for said gfn. Returns 0 if all pages 27620337f585SSean Christopherson * were marked unsync (or if there is no shadow page), -EPERM if the SPTE must 27630337f585SSean Christopherson * be write-protected. 27640337f585SSean Christopherson */ 27658283e36aSBen Gardon int mmu_try_to_unsync_pages(struct kvm *kvm, const struct kvm_memory_slot *slot, 27662839180cSPaolo Bonzini gfn_t gfn, bool can_unsync, bool prefetch) 2767c50d8ae3SPaolo Bonzini { 2768c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2769ce25681dSSean Christopherson bool locked = false; 2770c50d8ae3SPaolo Bonzini 27710337f585SSean Christopherson /* 27720337f585SSean Christopherson * Force write-protection if the page is being tracked. Note, the page 27730337f585SSean Christopherson * track machinery is used to write-protect upper-level shadow pages, 27740337f585SSean Christopherson * i.e. this guards the role.level == 4K assertion below! 27750337f585SSean Christopherson */ 27767b574863SSean Christopherson if (kvm_gfn_is_write_tracked(kvm, slot, gfn)) 27770337f585SSean Christopherson return -EPERM; 2778c50d8ae3SPaolo Bonzini 27790337f585SSean Christopherson /* 27800337f585SSean Christopherson * The page is not write-tracked, mark existing shadow pages unsync 27810337f585SSean Christopherson * unless KVM is synchronizing an unsync SP (can_unsync = false). In 27820337f585SSean Christopherson * that case, KVM must complete emulation of the guest TLB flush before 27830337f585SSean Christopherson * allowing shadow pages to become unsync (writable by the guest). 27840337f585SSean Christopherson */ 2785767d8d8dSLai Jiangshan for_each_gfn_valid_sp_with_gptes(kvm, sp, gfn) { 2786c50d8ae3SPaolo Bonzini if (!can_unsync) 27870337f585SSean Christopherson return -EPERM; 2788c50d8ae3SPaolo Bonzini 2789c50d8ae3SPaolo Bonzini if (sp->unsync) 2790c50d8ae3SPaolo Bonzini continue; 2791c50d8ae3SPaolo Bonzini 27922839180cSPaolo Bonzini if (prefetch) 2793f1c4a88cSLai Jiangshan return -EEXIST; 2794f1c4a88cSLai Jiangshan 2795ce25681dSSean Christopherson /* 2796ce25681dSSean Christopherson * TDP MMU page faults require an additional spinlock as they 2797ce25681dSSean Christopherson * run with mmu_lock held for read, not write, and the unsync 2798ce25681dSSean Christopherson * logic is not thread safe. Take the spinklock regardless of 2799ce25681dSSean Christopherson * the MMU type to avoid extra conditionals/parameters, there's 2800ce25681dSSean Christopherson * no meaningful penalty if mmu_lock is held for write. 2801ce25681dSSean Christopherson */ 2802ce25681dSSean Christopherson if (!locked) { 2803ce25681dSSean Christopherson locked = true; 28044d78d0b3SBen Gardon spin_lock(&kvm->arch.mmu_unsync_pages_lock); 2805ce25681dSSean Christopherson 2806ce25681dSSean Christopherson /* 2807ce25681dSSean Christopherson * Recheck after taking the spinlock, a different vCPU 2808ce25681dSSean Christopherson * may have since marked the page unsync. A false 2809e59f75deSPaolo Bonzini * negative on the unprotected check above is not 2810ce25681dSSean Christopherson * possible as clearing sp->unsync _must_ hold mmu_lock 2811e59f75deSPaolo Bonzini * for write, i.e. unsync cannot transition from 1->0 2812ce25681dSSean Christopherson * while this CPU holds mmu_lock for read (or write). 2813ce25681dSSean Christopherson */ 2814ce25681dSSean Christopherson if (READ_ONCE(sp->unsync)) 2815ce25681dSSean Christopherson continue; 2816ce25681dSSean Christopherson } 2817ce25681dSSean Christopherson 281820ba462dSSean Christopherson WARN_ON_ONCE(sp->role.level != PG_LEVEL_4K); 28194d78d0b3SBen Gardon kvm_unsync_page(kvm, sp); 2820c50d8ae3SPaolo Bonzini } 2821ce25681dSSean Christopherson if (locked) 28224d78d0b3SBen Gardon spin_unlock(&kvm->arch.mmu_unsync_pages_lock); 2823c50d8ae3SPaolo Bonzini 2824c50d8ae3SPaolo Bonzini /* 2825c50d8ae3SPaolo Bonzini * We need to ensure that the marking of unsync pages is visible 2826c50d8ae3SPaolo Bonzini * before the SPTE is updated to allow writes because 2827c50d8ae3SPaolo Bonzini * kvm_mmu_sync_roots() checks the unsync flags without holding 2828c50d8ae3SPaolo Bonzini * the MMU lock and so can race with this. If the SPTE was updated 2829c50d8ae3SPaolo Bonzini * before the page had been marked as unsync-ed, something like the 2830c50d8ae3SPaolo Bonzini * following could happen: 2831c50d8ae3SPaolo Bonzini * 2832c50d8ae3SPaolo Bonzini * CPU 1 CPU 2 2833c50d8ae3SPaolo Bonzini * --------------------------------------------------------------------- 2834c50d8ae3SPaolo Bonzini * 1.2 Host updates SPTE 2835c50d8ae3SPaolo Bonzini * to be writable 2836c50d8ae3SPaolo Bonzini * 2.1 Guest writes a GPTE for GVA X. 2837c50d8ae3SPaolo Bonzini * (GPTE being in the guest page table shadowed 2838c50d8ae3SPaolo Bonzini * by the SP from CPU 1.) 2839c50d8ae3SPaolo Bonzini * This reads SPTE during the page table walk. 2840c50d8ae3SPaolo Bonzini * Since SPTE.W is read as 1, there is no 2841c50d8ae3SPaolo Bonzini * fault. 2842c50d8ae3SPaolo Bonzini * 2843c50d8ae3SPaolo Bonzini * 2.2 Guest issues TLB flush. 2844c50d8ae3SPaolo Bonzini * That causes a VM Exit. 2845c50d8ae3SPaolo Bonzini * 28460337f585SSean Christopherson * 2.3 Walking of unsync pages sees sp->unsync is 28470337f585SSean Christopherson * false and skips the page. 2848c50d8ae3SPaolo Bonzini * 2849c50d8ae3SPaolo Bonzini * 2.4 Guest accesses GVA X. 2850c50d8ae3SPaolo Bonzini * Since the mapping in the SP was not updated, 2851c50d8ae3SPaolo Bonzini * so the old mapping for GVA X incorrectly 2852c50d8ae3SPaolo Bonzini * gets used. 2853c50d8ae3SPaolo Bonzini * 1.1 Host marks SP 2854c50d8ae3SPaolo Bonzini * as unsync 2855c50d8ae3SPaolo Bonzini * (sp->unsync = true) 2856c50d8ae3SPaolo Bonzini * 2857c50d8ae3SPaolo Bonzini * The write barrier below ensures that 1.1 happens before 1.2 and thus 2858264d3dc1SLai Jiangshan * the situation in 2.4 does not arise. It pairs with the read barrier 2859264d3dc1SLai Jiangshan * in is_unsync_root(), placed between 2.1's load of SPTE.W and 2.3. 2860c50d8ae3SPaolo Bonzini */ 2861c50d8ae3SPaolo Bonzini smp_wmb(); 2862c50d8ae3SPaolo Bonzini 28630337f585SSean Christopherson return 0; 2864c50d8ae3SPaolo Bonzini } 2865c50d8ae3SPaolo Bonzini 28668a9f566aSDavid Matlack static int mmu_set_spte(struct kvm_vcpu *vcpu, struct kvm_memory_slot *slot, 28678a9f566aSDavid Matlack u64 *sptep, unsigned int pte_access, gfn_t gfn, 2868a12f4381SPaolo Bonzini kvm_pfn_t pfn, struct kvm_page_fault *fault) 2869799a4190SBen Gardon { 2870d786c778SPaolo Bonzini struct kvm_mmu_page *sp = sptep_to_sp(sptep); 2871eb5cd7ffSPaolo Bonzini int level = sp->role.level; 2872c50d8ae3SPaolo Bonzini int was_rmapped = 0; 2873c4371c2aSSean Christopherson int ret = RET_PF_FIXED; 2874c50d8ae3SPaolo Bonzini bool flush = false; 2875ad67e480SPaolo Bonzini bool wrprot; 2876d786c778SPaolo Bonzini u64 spte; 2877c50d8ae3SPaolo Bonzini 2878a12f4381SPaolo Bonzini /* Prefetching always gets a writable pfn. */ 2879a12f4381SPaolo Bonzini bool host_writable = !fault || fault->map_writable; 28802839180cSPaolo Bonzini bool prefetch = !fault || fault->prefetch; 2881a12f4381SPaolo Bonzini bool write_fault = fault && fault->write; 2882c50d8ae3SPaolo Bonzini 2883a54aa15cSSean Christopherson if (unlikely(is_noslot_pfn(pfn))) { 28841075d41eSSean Christopherson vcpu->stat.pf_mmio_spte_created++; 2885a54aa15cSSean Christopherson mark_mmio_spte(vcpu, sptep, gfn, pte_access); 2886a54aa15cSSean Christopherson return RET_PF_EMULATE; 2887a54aa15cSSean Christopherson } 2888a54aa15cSSean Christopherson 2889c50d8ae3SPaolo Bonzini if (is_shadow_present_pte(*sptep)) { 2890c50d8ae3SPaolo Bonzini /* 2891c50d8ae3SPaolo Bonzini * If we overwrite a PTE page pointer with a 2MB PMD, unlink 2892c50d8ae3SPaolo Bonzini * the parent of the now unreachable PTE. 2893c50d8ae3SPaolo Bonzini */ 28943bae0459SSean Christopherson if (level > PG_LEVEL_4K && !is_large_pte(*sptep)) { 2895c50d8ae3SPaolo Bonzini struct kvm_mmu_page *child; 2896c50d8ae3SPaolo Bonzini u64 pte = *sptep; 2897c50d8ae3SPaolo Bonzini 28985e3edd7eSSean Christopherson child = spte_to_child_sp(pte); 2899069f30c6SMingwei Zhang drop_parent_pte(vcpu->kvm, child, sptep); 2900c50d8ae3SPaolo Bonzini flush = true; 2901c50d8ae3SPaolo Bonzini } else if (pfn != spte_to_pfn(*sptep)) { 2902c50d8ae3SPaolo Bonzini drop_spte(vcpu->kvm, sptep); 2903c50d8ae3SPaolo Bonzini flush = true; 2904c50d8ae3SPaolo Bonzini } else 2905c50d8ae3SPaolo Bonzini was_rmapped = 1; 2906c50d8ae3SPaolo Bonzini } 2907c50d8ae3SPaolo Bonzini 29082839180cSPaolo Bonzini wrprot = make_spte(vcpu, sp, slot, pte_access, gfn, pfn, *sptep, prefetch, 29097158bee4SPaolo Bonzini true, host_writable, &spte); 2910d786c778SPaolo Bonzini 2911d786c778SPaolo Bonzini if (*sptep == spte) { 2912d786c778SPaolo Bonzini ret = RET_PF_SPURIOUS; 2913d786c778SPaolo Bonzini } else { 2914d786c778SPaolo Bonzini flush |= mmu_spte_update(sptep, spte); 29155959ff4aSMaxim Levitsky trace_kvm_mmu_set_spte(level, gfn, sptep); 2916c50d8ae3SPaolo Bonzini } 2917c50d8ae3SPaolo Bonzini 2918ad67e480SPaolo Bonzini if (wrprot) { 2919c50d8ae3SPaolo Bonzini if (write_fault) 2920c50d8ae3SPaolo Bonzini ret = RET_PF_EMULATE; 2921c50d8ae3SPaolo Bonzini } 2922c50d8ae3SPaolo Bonzini 2923d786c778SPaolo Bonzini if (flush) 29244ad980aeSHou Wenlong kvm_flush_remote_tlbs_gfn(vcpu->kvm, gfn, level); 2925c50d8ae3SPaolo Bonzini 2926c50d8ae3SPaolo Bonzini if (!was_rmapped) { 2927d786c778SPaolo Bonzini WARN_ON_ONCE(ret == RET_PF_SPURIOUS); 29286a97575dSDavid Matlack rmap_add(vcpu, slot, sptep, gfn, pte_access); 29296a97575dSDavid Matlack } else { 29306a97575dSDavid Matlack /* Already rmapped but the pte_access bits may have changed. */ 293179e48cecSSean Christopherson kvm_mmu_page_set_access(sp, spte_index(sptep), pte_access); 2932c50d8ae3SPaolo Bonzini } 2933c50d8ae3SPaolo Bonzini 2934c50d8ae3SPaolo Bonzini return ret; 2935c50d8ae3SPaolo Bonzini } 2936c50d8ae3SPaolo Bonzini 2937c50d8ae3SPaolo Bonzini static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu, 2938c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, 2939c50d8ae3SPaolo Bonzini u64 *start, u64 *end) 2940c50d8ae3SPaolo Bonzini { 2941c50d8ae3SPaolo Bonzini struct page *pages[PTE_PREFETCH_NUM]; 2942c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot; 29430a2b64c5SBen Gardon unsigned int access = sp->role.access; 2944c50d8ae3SPaolo Bonzini int i, ret; 2945c50d8ae3SPaolo Bonzini gfn_t gfn; 2946c50d8ae3SPaolo Bonzini 294779e48cecSSean Christopherson gfn = kvm_mmu_page_get_gfn(sp, spte_index(start)); 2948c50d8ae3SPaolo Bonzini slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, access & ACC_WRITE_MASK); 2949c50d8ae3SPaolo Bonzini if (!slot) 2950c50d8ae3SPaolo Bonzini return -1; 2951c50d8ae3SPaolo Bonzini 2952c50d8ae3SPaolo Bonzini ret = gfn_to_page_many_atomic(slot, gfn, pages, end - start); 2953c50d8ae3SPaolo Bonzini if (ret <= 0) 2954c50d8ae3SPaolo Bonzini return -1; 2955c50d8ae3SPaolo Bonzini 2956c50d8ae3SPaolo Bonzini for (i = 0; i < ret; i++, gfn++, start++) { 29578a9f566aSDavid Matlack mmu_set_spte(vcpu, slot, start, access, gfn, 2958a12f4381SPaolo Bonzini page_to_pfn(pages[i]), NULL); 2959c50d8ae3SPaolo Bonzini put_page(pages[i]); 2960c50d8ae3SPaolo Bonzini } 2961c50d8ae3SPaolo Bonzini 2962c50d8ae3SPaolo Bonzini return 0; 2963c50d8ae3SPaolo Bonzini } 2964c50d8ae3SPaolo Bonzini 2965c50d8ae3SPaolo Bonzini static void __direct_pte_prefetch(struct kvm_vcpu *vcpu, 2966c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, u64 *sptep) 2967c50d8ae3SPaolo Bonzini { 2968c50d8ae3SPaolo Bonzini u64 *spte, *start = NULL; 2969c50d8ae3SPaolo Bonzini int i; 2970c50d8ae3SPaolo Bonzini 297120ba462dSSean Christopherson WARN_ON_ONCE(!sp->role.direct); 2972c50d8ae3SPaolo Bonzini 297379e48cecSSean Christopherson i = spte_index(sptep) & ~(PTE_PREFETCH_NUM - 1); 2974c50d8ae3SPaolo Bonzini spte = sp->spt + i; 2975c50d8ae3SPaolo Bonzini 2976c50d8ae3SPaolo Bonzini for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) { 2977c50d8ae3SPaolo Bonzini if (is_shadow_present_pte(*spte) || spte == sptep) { 2978c50d8ae3SPaolo Bonzini if (!start) 2979c50d8ae3SPaolo Bonzini continue; 2980c50d8ae3SPaolo Bonzini if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0) 2981c6cecc4bSSean Christopherson return; 2982c50d8ae3SPaolo Bonzini start = NULL; 2983c50d8ae3SPaolo Bonzini } else if (!start) 2984c50d8ae3SPaolo Bonzini start = spte; 2985c50d8ae3SPaolo Bonzini } 2986c6cecc4bSSean Christopherson if (start) 2987c6cecc4bSSean Christopherson direct_pte_prefetch_many(vcpu, sp, start, spte); 2988c50d8ae3SPaolo Bonzini } 2989c50d8ae3SPaolo Bonzini 2990c50d8ae3SPaolo Bonzini static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep) 2991c50d8ae3SPaolo Bonzini { 2992c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2993c50d8ae3SPaolo Bonzini 299457354682SSean Christopherson sp = sptep_to_sp(sptep); 2995c50d8ae3SPaolo Bonzini 2996c50d8ae3SPaolo Bonzini /* 2997c50d8ae3SPaolo Bonzini * Without accessed bits, there's no way to distinguish between 2998c50d8ae3SPaolo Bonzini * actually accessed translations and prefetched, so disable pte 2999c50d8ae3SPaolo Bonzini * prefetch if accessed bits aren't available. 3000c50d8ae3SPaolo Bonzini */ 3001c50d8ae3SPaolo Bonzini if (sp_ad_disabled(sp)) 3002c50d8ae3SPaolo Bonzini return; 3003c50d8ae3SPaolo Bonzini 30043bae0459SSean Christopherson if (sp->role.level > PG_LEVEL_4K) 3005c50d8ae3SPaolo Bonzini return; 3006c50d8ae3SPaolo Bonzini 30074a42d848SDavid Stevens /* 30084a42d848SDavid Stevens * If addresses are being invalidated, skip prefetching to avoid 30094a42d848SDavid Stevens * accidentally prefetching those addresses. 30104a42d848SDavid Stevens */ 301120ec3ebdSChao Peng if (unlikely(vcpu->kvm->mmu_invalidate_in_progress)) 30124a42d848SDavid Stevens return; 30134a42d848SDavid Stevens 3014c50d8ae3SPaolo Bonzini __direct_pte_prefetch(vcpu, sp, sptep); 3015c50d8ae3SPaolo Bonzini } 3016c50d8ae3SPaolo Bonzini 301765e3b446SSean Christopherson /* 301865e3b446SSean Christopherson * Lookup the mapping level for @gfn in the current mm. 301965e3b446SSean Christopherson * 302065e3b446SSean Christopherson * WARNING! Use of host_pfn_mapping_level() requires the caller and the end 302165e3b446SSean Christopherson * consumer to be tied into KVM's handlers for MMU notifier events! 302265e3b446SSean Christopherson * 302365e3b446SSean Christopherson * There are several ways to safely use this helper: 302465e3b446SSean Christopherson * 30258569992dSChao Peng * - Check mmu_invalidate_retry_gfn() after grabbing the mapping level, before 302665e3b446SSean Christopherson * consuming it. In this case, mmu_lock doesn't need to be held during the 302765e3b446SSean Christopherson * lookup, but it does need to be held while checking the MMU notifier. 302865e3b446SSean Christopherson * 302965e3b446SSean Christopherson * - Hold mmu_lock AND ensure there is no in-progress MMU notifier invalidation 303065e3b446SSean Christopherson * event for the hva. This can be done by explicit checking the MMU notifier 303165e3b446SSean Christopherson * or by ensuring that KVM already has a valid mapping that covers the hva. 303265e3b446SSean Christopherson * 303365e3b446SSean Christopherson * - Do not use the result to install new mappings, e.g. use the host mapping 303465e3b446SSean Christopherson * level only to decide whether or not to zap an entry. In this case, it's 303565e3b446SSean Christopherson * not required to hold mmu_lock (though it's highly likely the caller will 303665e3b446SSean Christopherson * want to hold mmu_lock anyways, e.g. to modify SPTEs). 303765e3b446SSean Christopherson * 303865e3b446SSean Christopherson * Note! The lookup can still race with modifications to host page tables, but 303965e3b446SSean Christopherson * the above "rules" ensure KVM will not _consume_ the result of the walk if a 304065e3b446SSean Christopherson * race with the primary MMU occurs. 304165e3b446SSean Christopherson */ 3042a8ac499bSSean Christopherson static int host_pfn_mapping_level(struct kvm *kvm, gfn_t gfn, 30438ca6f063SBen Gardon const struct kvm_memory_slot *slot) 3044db543216SSean Christopherson { 3045284dc493SSean Christopherson int level = PG_LEVEL_4K; 3046db543216SSean Christopherson unsigned long hva; 304744187235SMingwei Zhang unsigned long flags; 304844187235SMingwei Zhang pgd_t pgd; 304944187235SMingwei Zhang p4d_t p4d; 305044187235SMingwei Zhang pud_t pud; 305144187235SMingwei Zhang pmd_t pmd; 3052db543216SSean Christopherson 30535d49f08cSSean Christopherson /* 3054293e306eSSean Christopherson * Note, using the already-retrieved memslot and __gfn_to_hva_memslot() 3055293e306eSSean Christopherson * is not solely for performance, it's also necessary to avoid the 3056293e306eSSean Christopherson * "writable" check in __gfn_to_hva_many(), which will always fail on 3057293e306eSSean Christopherson * read-only memslots due to gfn_to_hva() assuming writes. Earlier 3058293e306eSSean Christopherson * page fault steps have already verified the guest isn't writing a 3059293e306eSSean Christopherson * read-only memslot. 3060293e306eSSean Christopherson */ 3061db543216SSean Christopherson hva = __gfn_to_hva_memslot(slot, gfn); 3062db543216SSean Christopherson 306344187235SMingwei Zhang /* 306465e3b446SSean Christopherson * Disable IRQs to prevent concurrent tear down of host page tables, 306565e3b446SSean Christopherson * e.g. if the primary MMU promotes a P*D to a huge page and then frees 306665e3b446SSean Christopherson * the original page table. 306744187235SMingwei Zhang */ 306844187235SMingwei Zhang local_irq_save(flags); 3069db543216SSean Christopherson 307065e3b446SSean Christopherson /* 307165e3b446SSean Christopherson * Read each entry once. As above, a non-leaf entry can be promoted to 307265e3b446SSean Christopherson * a huge page _during_ this walk. Re-reading the entry could send the 3073e72c7c2bSPeter Xu * walk into the weeks, e.g. p*d_leaf() returns false (sees the old 307465e3b446SSean Christopherson * value) and then p*d_offset() walks into the target huge page instead 307565e3b446SSean Christopherson * of the old page table (sees the new value). 307665e3b446SSean Christopherson */ 307744187235SMingwei Zhang pgd = READ_ONCE(*pgd_offset(kvm->mm, hva)); 307844187235SMingwei Zhang if (pgd_none(pgd)) 307944187235SMingwei Zhang goto out; 308044187235SMingwei Zhang 308144187235SMingwei Zhang p4d = READ_ONCE(*p4d_offset(&pgd, hva)); 308244187235SMingwei Zhang if (p4d_none(p4d) || !p4d_present(p4d)) 308344187235SMingwei Zhang goto out; 308444187235SMingwei Zhang 308544187235SMingwei Zhang pud = READ_ONCE(*pud_offset(&p4d, hva)); 308644187235SMingwei Zhang if (pud_none(pud) || !pud_present(pud)) 308744187235SMingwei Zhang goto out; 308844187235SMingwei Zhang 30890a845e0fSPeter Xu if (pud_leaf(pud)) { 309044187235SMingwei Zhang level = PG_LEVEL_1G; 309144187235SMingwei Zhang goto out; 309244187235SMingwei Zhang } 309344187235SMingwei Zhang 309444187235SMingwei Zhang pmd = READ_ONCE(*pmd_offset(&pud, hva)); 309544187235SMingwei Zhang if (pmd_none(pmd) || !pmd_present(pmd)) 309644187235SMingwei Zhang goto out; 309744187235SMingwei Zhang 30982f709f7bSPeter Xu if (pmd_leaf(pmd)) 309944187235SMingwei Zhang level = PG_LEVEL_2M; 310044187235SMingwei Zhang 310144187235SMingwei Zhang out: 310244187235SMingwei Zhang local_irq_restore(flags); 3103db543216SSean Christopherson return level; 3104db543216SSean Christopherson } 3105db543216SSean Christopherson 31068dd2eee9SChao Peng static int __kvm_mmu_max_mapping_level(struct kvm *kvm, 31078dd2eee9SChao Peng const struct kvm_memory_slot *slot, 31088dd2eee9SChao Peng gfn_t gfn, int max_level, bool is_private) 31091b6d9d9eSSean Christopherson { 31101b6d9d9eSSean Christopherson struct kvm_lpage_info *linfo; 3111ec607a56SPaolo Bonzini int host_level; 31121b6d9d9eSSean Christopherson 31131b6d9d9eSSean Christopherson max_level = min(max_level, max_huge_page_level); 31141b6d9d9eSSean Christopherson for ( ; max_level > PG_LEVEL_4K; max_level--) { 31151b6d9d9eSSean Christopherson linfo = lpage_info_slot(gfn, slot, max_level); 31161b6d9d9eSSean Christopherson if (!linfo->disallow_lpage) 31171b6d9d9eSSean Christopherson break; 31181b6d9d9eSSean Christopherson } 31191b6d9d9eSSean Christopherson 31208dd2eee9SChao Peng if (is_private) 31218dd2eee9SChao Peng return max_level; 31228dd2eee9SChao Peng 31231b6d9d9eSSean Christopherson if (max_level == PG_LEVEL_4K) 31241b6d9d9eSSean Christopherson return PG_LEVEL_4K; 31251b6d9d9eSSean Christopherson 3126a8ac499bSSean Christopherson host_level = host_pfn_mapping_level(kvm, gfn, slot); 3127ec607a56SPaolo Bonzini return min(host_level, max_level); 31281b6d9d9eSSean Christopherson } 31291b6d9d9eSSean Christopherson 31308dd2eee9SChao Peng int kvm_mmu_max_mapping_level(struct kvm *kvm, 31318dd2eee9SChao Peng const struct kvm_memory_slot *slot, gfn_t gfn, 31328dd2eee9SChao Peng int max_level) 31338dd2eee9SChao Peng { 31348dd2eee9SChao Peng bool is_private = kvm_slot_can_be_private(slot) && 31358dd2eee9SChao Peng kvm_mem_is_private(kvm, gfn); 31368dd2eee9SChao Peng 31378dd2eee9SChao Peng return __kvm_mmu_max_mapping_level(kvm, slot, gfn, max_level, is_private); 31388dd2eee9SChao Peng } 31398dd2eee9SChao Peng 314073a3c659SPaolo Bonzini void kvm_mmu_hugepage_adjust(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 31410885904dSSean Christopherson { 3142e710c5f6SDavid Matlack struct kvm_memory_slot *slot = fault->slot; 314317eff019SSean Christopherson kvm_pfn_t mask; 31440885904dSSean Christopherson 314573a3c659SPaolo Bonzini fault->huge_page_disallowed = fault->exec && fault->nx_huge_page_workaround_enabled; 31463cf06612SSean Christopherson 314773a3c659SPaolo Bonzini if (unlikely(fault->max_level == PG_LEVEL_4K)) 314873a3c659SPaolo Bonzini return; 314917eff019SSean Christopherson 31505d49f08cSSean Christopherson if (is_error_noslot_pfn(fault->pfn)) 315173a3c659SPaolo Bonzini return; 315217eff019SSean Christopherson 3153e710c5f6SDavid Matlack if (kvm_slot_dirty_track_enabled(slot)) 315473a3c659SPaolo Bonzini return; 3155293e306eSSean Christopherson 31563cf06612SSean Christopherson /* 31573cf06612SSean Christopherson * Enforce the iTLB multihit workaround after capturing the requested 31583cf06612SSean Christopherson * level, which will be used to do precise, accurate accounting. 31593cf06612SSean Christopherson */ 31608dd2eee9SChao Peng fault->req_level = __kvm_mmu_max_mapping_level(vcpu->kvm, slot, 31618dd2eee9SChao Peng fault->gfn, fault->max_level, 31628dd2eee9SChao Peng fault->is_private); 316373a3c659SPaolo Bonzini if (fault->req_level == PG_LEVEL_4K || fault->huge_page_disallowed) 316473a3c659SPaolo Bonzini return; 31654cd071d1SSean Christopherson 31660885904dSSean Christopherson /* 316720ec3ebdSChao Peng * mmu_invalidate_retry() was successful and mmu_lock is held, so 31684cd071d1SSean Christopherson * the pmd can't be split from under us. 31690885904dSSean Christopherson */ 317073a3c659SPaolo Bonzini fault->goal_level = fault->req_level; 317173a3c659SPaolo Bonzini mask = KVM_PAGES_PER_HPAGE(fault->goal_level) - 1; 317273a3c659SPaolo Bonzini VM_BUG_ON((fault->gfn & mask) != (fault->pfn & mask)); 317373a3c659SPaolo Bonzini fault->pfn &= ~mask; 31740885904dSSean Christopherson } 31750885904dSSean Christopherson 3176536f0e6aSPaolo Bonzini void disallowed_hugepage_adjust(struct kvm_page_fault *fault, u64 spte, int cur_level) 3177c50d8ae3SPaolo Bonzini { 3178536f0e6aSPaolo Bonzini if (cur_level > PG_LEVEL_4K && 3179536f0e6aSPaolo Bonzini cur_level == fault->goal_level && 3180c50d8ae3SPaolo Bonzini is_shadow_present_pte(spte) && 318176901e56SMingwei Zhang !is_large_pte(spte) && 318276901e56SMingwei Zhang spte_to_child_sp(spte)->nx_huge_page_disallowed) { 3183c50d8ae3SPaolo Bonzini /* 31846c882ef4SDavid Matlack * A small SPTE exists for this pfn, but FNAME(fetch), 31856c882ef4SDavid Matlack * direct_map(), or kvm_tdp_mmu_map() would like to create a 31866c882ef4SDavid Matlack * large PTE instead: just force them to go down another level, 31876c882ef4SDavid Matlack * patching back for them into pfn the next 9 bits of the 31886c882ef4SDavid Matlack * address. 3189c50d8ae3SPaolo Bonzini */ 3190536f0e6aSPaolo Bonzini u64 page_mask = KVM_PAGES_PER_HPAGE(cur_level) - 3191536f0e6aSPaolo Bonzini KVM_PAGES_PER_HPAGE(cur_level - 1); 3192536f0e6aSPaolo Bonzini fault->pfn |= fault->gfn & page_mask; 3193536f0e6aSPaolo Bonzini fault->goal_level--; 3194c50d8ae3SPaolo Bonzini } 3195c50d8ae3SPaolo Bonzini } 3196c50d8ae3SPaolo Bonzini 31976c882ef4SDavid Matlack static int direct_map(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 3198c50d8ae3SPaolo Bonzini { 3199c50d8ae3SPaolo Bonzini struct kvm_shadow_walk_iterator it; 3200c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 320173a3c659SPaolo Bonzini int ret; 320243b74355SPaolo Bonzini gfn_t base_gfn = fault->gfn; 3203c50d8ae3SPaolo Bonzini 320473a3c659SPaolo Bonzini kvm_mmu_hugepage_adjust(vcpu, fault); 32054cd071d1SSean Christopherson 3206f0066d94SPaolo Bonzini trace_kvm_mmu_spte_requested(fault); 320743b74355SPaolo Bonzini for_each_shadow_entry(vcpu, fault->addr, it) { 3208c50d8ae3SPaolo Bonzini /* 3209c50d8ae3SPaolo Bonzini * We cannot overwrite existing page tables with an NX 3210c50d8ae3SPaolo Bonzini * large page, as the leaf could be executable. 3211c50d8ae3SPaolo Bonzini */ 321273a3c659SPaolo Bonzini if (fault->nx_huge_page_workaround_enabled) 3213536f0e6aSPaolo Bonzini disallowed_hugepage_adjust(fault, *it.sptep, it.level); 3214c50d8ae3SPaolo Bonzini 3215c667a3baSHou Wenlong base_gfn = gfn_round_for_level(fault->gfn, it.level); 321673a3c659SPaolo Bonzini if (it.level == fault->goal_level) 3217c50d8ae3SPaolo Bonzini break; 3218c50d8ae3SPaolo Bonzini 32192e65e842SDavid Matlack sp = kvm_mmu_get_child_sp(vcpu, it.sptep, base_gfn, true, ACC_ALL); 32200cd8dc73SPaolo Bonzini if (sp == ERR_PTR(-EEXIST)) 32210cd8dc73SPaolo Bonzini continue; 3222c50d8ae3SPaolo Bonzini 3223c50d8ae3SPaolo Bonzini link_shadow_page(vcpu, it.sptep, sp); 3224b5b0977fSSean Christopherson if (fault->huge_page_disallowed) 322555c510e2SSean Christopherson account_nx_huge_page(vcpu->kvm, sp, 3226428e9216SSean Christopherson fault->req_level >= it.level); 3227c50d8ae3SPaolo Bonzini } 3228c50d8ae3SPaolo Bonzini 3229b1a429fbSSean Christopherson if (WARN_ON_ONCE(it.level != fault->goal_level)) 3230b1a429fbSSean Christopherson return -EFAULT; 3231b1a429fbSSean Christopherson 32328a9f566aSDavid Matlack ret = mmu_set_spte(vcpu, fault->slot, it.sptep, ACC_ALL, 3233a12f4381SPaolo Bonzini base_gfn, fault->pfn, fault); 323412703759SSean Christopherson if (ret == RET_PF_SPURIOUS) 323512703759SSean Christopherson return ret; 323612703759SSean Christopherson 3237c50d8ae3SPaolo Bonzini direct_pte_prefetch(vcpu, it.sptep); 3238c50d8ae3SPaolo Bonzini return ret; 3239c50d8ae3SPaolo Bonzini } 3240c50d8ae3SPaolo Bonzini 3241cd08d178SDavid Matlack static void kvm_send_hwpoison_signal(struct kvm_memory_slot *slot, gfn_t gfn) 3242c50d8ae3SPaolo Bonzini { 3243cd08d178SDavid Matlack unsigned long hva = gfn_to_hva_memslot(slot, gfn); 3244cd08d178SDavid Matlack 3245cd08d178SDavid Matlack send_sig_mceerr(BUS_MCEERR_AR, (void __user *)hva, PAGE_SHIFT, current); 3246c50d8ae3SPaolo Bonzini } 3247c50d8ae3SPaolo Bonzini 3248cd08d178SDavid Matlack static int kvm_handle_error_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 3249c50d8ae3SPaolo Bonzini { 3250cd08d178SDavid Matlack if (is_sigpending_pfn(fault->pfn)) { 325176657687SPeter Xu kvm_handle_signal_exit(vcpu); 325276657687SPeter Xu return -EINTR; 325376657687SPeter Xu } 325476657687SPeter Xu 3255c50d8ae3SPaolo Bonzini /* 3256c50d8ae3SPaolo Bonzini * Do not cache the mmio info caused by writing the readonly gfn 3257c50d8ae3SPaolo Bonzini * into the spte otherwise read access on readonly gfn also can 3258c50d8ae3SPaolo Bonzini * caused mmio page fault and treat it as mmio access. 3259c50d8ae3SPaolo Bonzini */ 3260cd08d178SDavid Matlack if (fault->pfn == KVM_PFN_ERR_RO_FAULT) 3261c50d8ae3SPaolo Bonzini return RET_PF_EMULATE; 3262c50d8ae3SPaolo Bonzini 3263cd08d178SDavid Matlack if (fault->pfn == KVM_PFN_ERR_HWPOISON) { 3264cd08d178SDavid Matlack kvm_send_hwpoison_signal(fault->slot, fault->gfn); 3265c50d8ae3SPaolo Bonzini return RET_PF_RETRY; 3266c50d8ae3SPaolo Bonzini } 3267c50d8ae3SPaolo Bonzini 3268c50d8ae3SPaolo Bonzini return -EFAULT; 3269c50d8ae3SPaolo Bonzini } 3270c50d8ae3SPaolo Bonzini 3271354c908cSDavid Matlack static int kvm_handle_noslot_fault(struct kvm_vcpu *vcpu, 3272354c908cSDavid Matlack struct kvm_page_fault *fault, 32735276c616SSean Christopherson unsigned int access) 3274c50d8ae3SPaolo Bonzini { 32753a13f4feSPaolo Bonzini gva_t gva = fault->is_tdp ? 0 : fault->addr; 32763a13f4feSPaolo Bonzini 3277bde9f9d2SSean Christopherson if (fault->is_private) { 3278bde9f9d2SSean Christopherson kvm_mmu_prepare_memory_fault_exit(vcpu, fault); 3279bde9f9d2SSean Christopherson return -EFAULT; 3280bde9f9d2SSean Christopherson } 3281bde9f9d2SSean Christopherson 32823a13f4feSPaolo Bonzini vcpu_cache_mmio_info(vcpu, gva, fault->gfn, 3283c50d8ae3SPaolo Bonzini access & shadow_mmio_access_mask); 3284354c908cSDavid Matlack 3285f6adeae8SSean Christopherson fault->slot = NULL; 3286f6adeae8SSean Christopherson fault->pfn = KVM_PFN_NOSLOT; 3287f6adeae8SSean Christopherson fault->map_writable = false; 328836d44927SSean Christopherson fault->hva = KVM_HVA_ERR_BAD; 3289f6adeae8SSean Christopherson 329030ab5901SSean Christopherson /* 329130ab5901SSean Christopherson * If MMIO caching is disabled, emulate immediately without 329230ab5901SSean Christopherson * touching the shadow page tables as attempting to install an 3293354c908cSDavid Matlack * MMIO SPTE will just be an expensive nop. 329430ab5901SSean Christopherson */ 3295354c908cSDavid Matlack if (unlikely(!enable_mmio_caching)) 32965276c616SSean Christopherson return RET_PF_EMULATE; 3297354c908cSDavid Matlack 3298354c908cSDavid Matlack /* 3299354c908cSDavid Matlack * Do not create an MMIO SPTE for a gfn greater than host.MAXPHYADDR, 3300354c908cSDavid Matlack * any guest that generates such gfns is running nested and is being 3301354c908cSDavid Matlack * tricked by L0 userspace (you can observe gfn > L1.MAXPHYADDR if and 3302354c908cSDavid Matlack * only if L1's MAXPHYADDR is inaccurate with respect to the 3303354c908cSDavid Matlack * hardware's). 3304354c908cSDavid Matlack */ 3305354c908cSDavid Matlack if (unlikely(fault->gfn > kvm_mmu_max_gfn())) 3306354c908cSDavid Matlack return RET_PF_EMULATE; 3307c50d8ae3SPaolo Bonzini 33085276c616SSean Christopherson return RET_PF_CONTINUE; 3309c50d8ae3SPaolo Bonzini } 3310c50d8ae3SPaolo Bonzini 3311b74d002dSMichael Roth static bool page_fault_can_be_fast(struct kvm *kvm, struct kvm_page_fault *fault) 3312c50d8ae3SPaolo Bonzini { 3313c50d8ae3SPaolo Bonzini /* 33145c64aba5SSean Christopherson * Page faults with reserved bits set, i.e. faults on MMIO SPTEs, only 33155c64aba5SSean Christopherson * reach the common page fault handler if the SPTE has an invalid MMIO 33165c64aba5SSean Christopherson * generation number. Refreshing the MMIO generation needs to go down 33175c64aba5SSean Christopherson * the slow path. Note, EPT Misconfigs do NOT set the PRESENT flag! 3318c50d8ae3SPaolo Bonzini */ 33193c8ad5a6SPaolo Bonzini if (fault->rsvd) 3320c50d8ae3SPaolo Bonzini return false; 3321c50d8ae3SPaolo Bonzini 3322c50d8ae3SPaolo Bonzini /* 3323b74d002dSMichael Roth * For hardware-protected VMs, certain conditions like attempting to 3324b74d002dSMichael Roth * perform a write to a page which is not in the state that the guest 3325b74d002dSMichael Roth * expects it to be in can result in a nested/extended #PF. In this 3326b74d002dSMichael Roth * case, the below code might misconstrue this situation as being the 3327b74d002dSMichael Roth * result of a write-protected access, and treat it as a spurious case 3328b74d002dSMichael Roth * rather than taking any action to satisfy the real source of the #PF 3329b74d002dSMichael Roth * such as generating a KVM_EXIT_MEMORY_FAULT. This can lead to the 3330b74d002dSMichael Roth * guest spinning on a #PF indefinitely, so don't attempt the fast path 3331b74d002dSMichael Roth * in this case. 3332b74d002dSMichael Roth * 3333b74d002dSMichael Roth * Note that the kvm_mem_is_private() check might race with an 3334b74d002dSMichael Roth * attribute update, but this will either result in the guest spinning 3335b74d002dSMichael Roth * on RET_PF_SPURIOUS until the update completes, or an actual spurious 3336b74d002dSMichael Roth * case might go down the slow path. Either case will resolve itself. 3337b74d002dSMichael Roth */ 3338b74d002dSMichael Roth if (kvm->arch.has_private_mem && 3339b74d002dSMichael Roth fault->is_private != kvm_mem_is_private(kvm, fault->gfn)) 3340b74d002dSMichael Roth return false; 3341b74d002dSMichael Roth 3342b74d002dSMichael Roth /* 3343c50d8ae3SPaolo Bonzini * #PF can be fast if: 3344c50d8ae3SPaolo Bonzini * 334554275f74SSean Christopherson * 1. The shadow page table entry is not present and A/D bits are 334654275f74SSean Christopherson * disabled _by KVM_, which could mean that the fault is potentially 334754275f74SSean Christopherson * caused by access tracking (if enabled). If A/D bits are enabled 334854275f74SSean Christopherson * by KVM, but disabled by L1 for L2, KVM is forced to disable A/D 334954275f74SSean Christopherson * bits for L2 and employ access tracking, but the fast page fault 335054275f74SSean Christopherson * mechanism only supports direct MMUs. 335154275f74SSean Christopherson * 2. The shadow page table entry is present, the access is a write, 335254275f74SSean Christopherson * and no reserved bits are set (MMIO SPTEs cannot be "fixed"), i.e. 335354275f74SSean Christopherson * the fault was caused by a write-protection violation. If the 335454275f74SSean Christopherson * SPTE is MMU-writable (determined later), the fault can be fixed 335554275f74SSean Christopherson * by setting the Writable bit, which can be done out of mmu_lock. 3356c50d8ae3SPaolo Bonzini */ 33575c64aba5SSean Christopherson if (!fault->present) 33585c64aba5SSean Christopherson return !kvm_ad_enabled(); 33595c64aba5SSean Christopherson 33605c64aba5SSean Christopherson /* 33615c64aba5SSean Christopherson * Note, instruction fetches and writes are mutually exclusive, ignore 33625c64aba5SSean Christopherson * the "exec" flag. 33635c64aba5SSean Christopherson */ 33645c64aba5SSean Christopherson return fault->write; 3365c50d8ae3SPaolo Bonzini } 3366c50d8ae3SPaolo Bonzini 3367c50d8ae3SPaolo Bonzini /* 3368c50d8ae3SPaolo Bonzini * Returns true if the SPTE was fixed successfully. Otherwise, 3369c50d8ae3SPaolo Bonzini * someone else modified the SPTE from its original value. 3370c50d8ae3SPaolo Bonzini */ 3371f3d90f90SSean Christopherson static bool fast_pf_fix_direct_spte(struct kvm_vcpu *vcpu, 3372f3d90f90SSean Christopherson struct kvm_page_fault *fault, 3373c50d8ae3SPaolo Bonzini u64 *sptep, u64 old_spte, u64 new_spte) 3374c50d8ae3SPaolo Bonzini { 3375c50d8ae3SPaolo Bonzini /* 3376c50d8ae3SPaolo Bonzini * Theoretically we could also set dirty bit (and flush TLB) here in 3377c50d8ae3SPaolo Bonzini * order to eliminate unnecessary PML logging. See comments in 3378c50d8ae3SPaolo Bonzini * set_spte. But fast_page_fault is very unlikely to happen with PML 3379c50d8ae3SPaolo Bonzini * enabled, so we do not do this. This might result in the same GPA 3380c50d8ae3SPaolo Bonzini * to be logged in PML buffer again when the write really happens, and 3381c50d8ae3SPaolo Bonzini * eventually to be called by mark_page_dirty twice. But it's also no 3382c50d8ae3SPaolo Bonzini * harm. This also avoids the TLB flush needed after setting dirty bit 3383c50d8ae3SPaolo Bonzini * so non-PML cases won't be impacted. 3384c50d8ae3SPaolo Bonzini * 3385c50d8ae3SPaolo Bonzini * Compare with set_spte where instead shadow_dirty_mask is set. 3386c50d8ae3SPaolo Bonzini */ 33872db2f46fSUros Bizjak if (!try_cmpxchg64(sptep, &old_spte, new_spte)) 3388c50d8ae3SPaolo Bonzini return false; 3389c50d8ae3SPaolo Bonzini 3390e710c5f6SDavid Matlack if (is_writable_pte(new_spte) && !is_writable_pte(old_spte)) 3391e710c5f6SDavid Matlack mark_page_dirty_in_slot(vcpu->kvm, fault->slot, fault->gfn); 3392c50d8ae3SPaolo Bonzini 3393c50d8ae3SPaolo Bonzini return true; 3394c50d8ae3SPaolo Bonzini } 3395c50d8ae3SPaolo Bonzini 33963c8ad5a6SPaolo Bonzini static bool is_access_allowed(struct kvm_page_fault *fault, u64 spte) 3397c50d8ae3SPaolo Bonzini { 33983c8ad5a6SPaolo Bonzini if (fault->exec) 3399c50d8ae3SPaolo Bonzini return is_executable_pte(spte); 3400c50d8ae3SPaolo Bonzini 34013c8ad5a6SPaolo Bonzini if (fault->write) 3402c50d8ae3SPaolo Bonzini return is_writable_pte(spte); 3403c50d8ae3SPaolo Bonzini 3404c50d8ae3SPaolo Bonzini /* Fault was on Read access */ 3405c50d8ae3SPaolo Bonzini return spte & PT_PRESENT_MASK; 3406c50d8ae3SPaolo Bonzini } 3407c50d8ae3SPaolo Bonzini 3408c50d8ae3SPaolo Bonzini /* 34096e8eb206SDavid Matlack * Returns the last level spte pointer of the shadow page walk for the given 34106e8eb206SDavid Matlack * gpa, and sets *spte to the spte value. This spte may be non-preset. If no 34116e8eb206SDavid Matlack * walk could be performed, returns NULL and *spte does not contain valid data. 34126e8eb206SDavid Matlack * 34136e8eb206SDavid Matlack * Contract: 34146e8eb206SDavid Matlack * - Must be called between walk_shadow_page_lockless_{begin,end}. 34156e8eb206SDavid Matlack * - The returned sptep must not be used after walk_shadow_page_lockless_end. 34166e8eb206SDavid Matlack */ 34176e8eb206SDavid Matlack static u64 *fast_pf_get_last_sptep(struct kvm_vcpu *vcpu, gpa_t gpa, u64 *spte) 34186e8eb206SDavid Matlack { 34196e8eb206SDavid Matlack struct kvm_shadow_walk_iterator iterator; 34206e8eb206SDavid Matlack u64 old_spte; 34216e8eb206SDavid Matlack u64 *sptep = NULL; 34226e8eb206SDavid Matlack 34236e8eb206SDavid Matlack for_each_shadow_entry_lockless(vcpu, gpa, iterator, old_spte) { 34246e8eb206SDavid Matlack sptep = iterator.sptep; 34256e8eb206SDavid Matlack *spte = old_spte; 34266e8eb206SDavid Matlack } 34276e8eb206SDavid Matlack 34286e8eb206SDavid Matlack return sptep; 34296e8eb206SDavid Matlack } 34306e8eb206SDavid Matlack 34316e8eb206SDavid Matlack /* 3432c4371c2aSSean Christopherson * Returns one of RET_PF_INVALID, RET_PF_FIXED or RET_PF_SPURIOUS. 3433c50d8ae3SPaolo Bonzini */ 34343c8ad5a6SPaolo Bonzini static int fast_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 3435c50d8ae3SPaolo Bonzini { 3436c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 3437c4371c2aSSean Christopherson int ret = RET_PF_INVALID; 34381de9992fSLi zeming u64 spte; 34391de9992fSLi zeming u64 *sptep; 3440c50d8ae3SPaolo Bonzini uint retry_count = 0; 3441c50d8ae3SPaolo Bonzini 3442b74d002dSMichael Roth if (!page_fault_can_be_fast(vcpu->kvm, fault)) 3443c4371c2aSSean Christopherson return ret; 3444c50d8ae3SPaolo Bonzini 3445c50d8ae3SPaolo Bonzini walk_shadow_page_lockless_begin(vcpu); 3446c50d8ae3SPaolo Bonzini 3447c50d8ae3SPaolo Bonzini do { 3448c50d8ae3SPaolo Bonzini u64 new_spte; 3449c50d8ae3SPaolo Bonzini 3450dfe0ecc6SSean Christopherson if (tdp_mmu_enabled) 34513c8ad5a6SPaolo Bonzini sptep = kvm_tdp_mmu_fast_pf_get_last_sptep(vcpu, fault->addr, &spte); 34526e8eb206SDavid Matlack else 34533c8ad5a6SPaolo Bonzini sptep = fast_pf_get_last_sptep(vcpu, fault->addr, &spte); 3454c50d8ae3SPaolo Bonzini 34551de9992fSLi zeming /* 34561de9992fSLi zeming * It's entirely possible for the mapping to have been zapped 34571de9992fSLi zeming * by a different task, but the root page should always be 34581de9992fSLi zeming * available as the vCPU holds a reference to its root(s). 34591de9992fSLi zeming */ 34601de9992fSLi zeming if (WARN_ON_ONCE(!sptep)) 34611de9992fSLi zeming spte = REMOVED_SPTE; 34621de9992fSLi zeming 3463ec89e643SSean Christopherson if (!is_shadow_present_pte(spte)) 3464ec89e643SSean Christopherson break; 3465ec89e643SSean Christopherson 34666e8eb206SDavid Matlack sp = sptep_to_sp(sptep); 3467c50d8ae3SPaolo Bonzini if (!is_last_spte(spte, sp->role.level)) 3468c50d8ae3SPaolo Bonzini break; 3469c50d8ae3SPaolo Bonzini 3470c50d8ae3SPaolo Bonzini /* 3471c50d8ae3SPaolo Bonzini * Check whether the memory access that caused the fault would 3472c50d8ae3SPaolo Bonzini * still cause it if it were to be performed right now. If not, 3473c50d8ae3SPaolo Bonzini * then this is a spurious fault caused by TLB lazily flushed, 3474c50d8ae3SPaolo Bonzini * or some other CPU has already fixed the PTE after the 3475c50d8ae3SPaolo Bonzini * current CPU took the fault. 3476c50d8ae3SPaolo Bonzini * 3477c50d8ae3SPaolo Bonzini * Need not check the access of upper level table entries since 3478c50d8ae3SPaolo Bonzini * they are always ACC_ALL. 3479c50d8ae3SPaolo Bonzini */ 34803c8ad5a6SPaolo Bonzini if (is_access_allowed(fault, spte)) { 3481c4371c2aSSean Christopherson ret = RET_PF_SPURIOUS; 3482c50d8ae3SPaolo Bonzini break; 3483c50d8ae3SPaolo Bonzini } 3484c50d8ae3SPaolo Bonzini 3485c50d8ae3SPaolo Bonzini new_spte = spte; 3486c50d8ae3SPaolo Bonzini 348754275f74SSean Christopherson /* 348854275f74SSean Christopherson * KVM only supports fixing page faults outside of MMU lock for 348954275f74SSean Christopherson * direct MMUs, nested MMUs are always indirect, and KVM always 349054275f74SSean Christopherson * uses A/D bits for non-nested MMUs. Thus, if A/D bits are 349154275f74SSean Christopherson * enabled, the SPTE can't be an access-tracked SPTE. 349254275f74SSean Christopherson */ 349354275f74SSean Christopherson if (unlikely(!kvm_ad_enabled()) && is_access_track_spte(spte)) 3494c50d8ae3SPaolo Bonzini new_spte = restore_acc_track_spte(new_spte); 3495c50d8ae3SPaolo Bonzini 3496c50d8ae3SPaolo Bonzini /* 349754275f74SSean Christopherson * To keep things simple, only SPTEs that are MMU-writable can 349854275f74SSean Christopherson * be made fully writable outside of mmu_lock, e.g. only SPTEs 349954275f74SSean Christopherson * that were write-protected for dirty-logging or access 350054275f74SSean Christopherson * tracking are handled here. Don't bother checking if the 350154275f74SSean Christopherson * SPTE is writable to prioritize running with A/D bits enabled. 350254275f74SSean Christopherson * The is_access_allowed() check above handles the common case 350354275f74SSean Christopherson * of the fault being spurious, and the SPTE is known to be 350454275f74SSean Christopherson * shadow-present, i.e. except for access tracking restoration 350554275f74SSean Christopherson * making the new SPTE writable, the check is wasteful. 3506c50d8ae3SPaolo Bonzini */ 3507706c9c55SSean Christopherson if (fault->write && is_mmu_writable_spte(spte)) { 3508c50d8ae3SPaolo Bonzini new_spte |= PT_WRITABLE_MASK; 3509c50d8ae3SPaolo Bonzini 3510c50d8ae3SPaolo Bonzini /* 351110c30de0SJunaid Shahid * Do not fix write-permission on the large spte when 351210c30de0SJunaid Shahid * dirty logging is enabled. Since we only dirty the 351310c30de0SJunaid Shahid * first page into the dirty-bitmap in 3514c50d8ae3SPaolo Bonzini * fast_pf_fix_direct_spte(), other pages are missed 3515c50d8ae3SPaolo Bonzini * if its slot has dirty logging enabled. 3516c50d8ae3SPaolo Bonzini * 3517c50d8ae3SPaolo Bonzini * Instead, we let the slow page fault path create a 3518c50d8ae3SPaolo Bonzini * normal spte to fix the access. 3519c50d8ae3SPaolo Bonzini */ 352010c30de0SJunaid Shahid if (sp->role.level > PG_LEVEL_4K && 352110c30de0SJunaid Shahid kvm_slot_dirty_track_enabled(fault->slot)) 3522c50d8ae3SPaolo Bonzini break; 3523c50d8ae3SPaolo Bonzini } 3524c50d8ae3SPaolo Bonzini 3525c50d8ae3SPaolo Bonzini /* Verify that the fault can be handled in the fast path */ 3526c50d8ae3SPaolo Bonzini if (new_spte == spte || 35273c8ad5a6SPaolo Bonzini !is_access_allowed(fault, new_spte)) 3528c50d8ae3SPaolo Bonzini break; 3529c50d8ae3SPaolo Bonzini 3530c50d8ae3SPaolo Bonzini /* 3531c50d8ae3SPaolo Bonzini * Currently, fast page fault only works for direct mapping 3532c50d8ae3SPaolo Bonzini * since the gfn is not stable for indirect shadow page. See 35333ecad8c2SMauro Carvalho Chehab * Documentation/virt/kvm/locking.rst to get more detail. 3534c50d8ae3SPaolo Bonzini */ 3535e710c5f6SDavid Matlack if (fast_pf_fix_direct_spte(vcpu, fault, sptep, spte, new_spte)) { 3536c4371c2aSSean Christopherson ret = RET_PF_FIXED; 3537c50d8ae3SPaolo Bonzini break; 3538c4371c2aSSean Christopherson } 3539c50d8ae3SPaolo Bonzini 3540c50d8ae3SPaolo Bonzini if (++retry_count > 4) { 35418d20bd63SSean Christopherson pr_warn_once("Fast #PF retrying more than 4 times.\n"); 3542c50d8ae3SPaolo Bonzini break; 3543c50d8ae3SPaolo Bonzini } 3544c50d8ae3SPaolo Bonzini 3545c50d8ae3SPaolo Bonzini } while (true); 3546c50d8ae3SPaolo Bonzini 3547f0066d94SPaolo Bonzini trace_fast_page_fault(vcpu, fault, sptep, spte, ret); 3548c50d8ae3SPaolo Bonzini walk_shadow_page_lockless_end(vcpu); 3549c50d8ae3SPaolo Bonzini 35501075d41eSSean Christopherson if (ret != RET_PF_INVALID) 35511075d41eSSean Christopherson vcpu->stat.pf_fast++; 35521075d41eSSean Christopherson 3553c4371c2aSSean Christopherson return ret; 3554c50d8ae3SPaolo Bonzini } 3555c50d8ae3SPaolo Bonzini 3556c50d8ae3SPaolo Bonzini static void mmu_free_root_page(struct kvm *kvm, hpa_t *root_hpa, 3557c50d8ae3SPaolo Bonzini struct list_head *invalid_list) 3558c50d8ae3SPaolo Bonzini { 3559c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 3560c50d8ae3SPaolo Bonzini 3561c50d8ae3SPaolo Bonzini if (!VALID_PAGE(*root_hpa)) 3562c50d8ae3SPaolo Bonzini return; 3563c50d8ae3SPaolo Bonzini 3564c5f2d564SSean Christopherson sp = root_to_sp(*root_hpa); 356520ba462dSSean Christopherson if (WARN_ON_ONCE(!sp)) 35669191b8f0SPaolo Bonzini return; 356702c00b3aSBen Gardon 3568576a15deSSean Christopherson if (is_tdp_mmu_page(sp)) { 3569576a15deSSean Christopherson lockdep_assert_held_read(&kvm->mmu_lock); 35705f3c8c91SPaolo Bonzini kvm_tdp_mmu_put_root(kvm, sp); 3571576a15deSSean Christopherson } else { 3572576a15deSSean Christopherson lockdep_assert_held_write(&kvm->mmu_lock); 3573576a15deSSean Christopherson if (!--sp->root_count && sp->role.invalid) 3574c50d8ae3SPaolo Bonzini kvm_mmu_prepare_zap_page(kvm, sp, invalid_list); 3575576a15deSSean Christopherson } 3576c50d8ae3SPaolo Bonzini 3577c50d8ae3SPaolo Bonzini *root_hpa = INVALID_PAGE; 3578c50d8ae3SPaolo Bonzini } 3579c50d8ae3SPaolo Bonzini 3580c50d8ae3SPaolo Bonzini /* roots_to_free must be some combination of the KVM_MMU_ROOT_* flags */ 35810c1c92f1SPaolo Bonzini void kvm_mmu_free_roots(struct kvm *kvm, struct kvm_mmu *mmu, 3582c50d8ae3SPaolo Bonzini ulong roots_to_free) 3583c50d8ae3SPaolo Bonzini { 3584576a15deSSean Christopherson bool is_tdp_mmu = tdp_mmu_enabled && mmu->root_role.direct; 3585c50d8ae3SPaolo Bonzini int i; 3586c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 3587594bef79SPaolo Bonzini bool free_active_root; 3588c50d8ae3SPaolo Bonzini 3589f94db0c8SSean Christopherson WARN_ON_ONCE(roots_to_free & ~KVM_MMU_ROOTS_ALL); 3590f94db0c8SSean Christopherson 3591c50d8ae3SPaolo Bonzini BUILD_BUG_ON(KVM_MMU_NUM_PREV_ROOTS >= BITS_PER_LONG); 3592c50d8ae3SPaolo Bonzini 3593c50d8ae3SPaolo Bonzini /* Before acquiring the MMU lock, see if we need to do any real work. */ 3594594bef79SPaolo Bonzini free_active_root = (roots_to_free & KVM_MMU_ROOT_CURRENT) 3595594bef79SPaolo Bonzini && VALID_PAGE(mmu->root.hpa); 3596594bef79SPaolo Bonzini 3597594bef79SPaolo Bonzini if (!free_active_root) { 3598c50d8ae3SPaolo Bonzini for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) 3599c50d8ae3SPaolo Bonzini if ((roots_to_free & KVM_MMU_ROOT_PREVIOUS(i)) && 3600c50d8ae3SPaolo Bonzini VALID_PAGE(mmu->prev_roots[i].hpa)) 3601c50d8ae3SPaolo Bonzini break; 3602c50d8ae3SPaolo Bonzini 3603c50d8ae3SPaolo Bonzini if (i == KVM_MMU_NUM_PREV_ROOTS) 3604c50d8ae3SPaolo Bonzini return; 3605c50d8ae3SPaolo Bonzini } 3606c50d8ae3SPaolo Bonzini 3607576a15deSSean Christopherson if (is_tdp_mmu) 3608576a15deSSean Christopherson read_lock(&kvm->mmu_lock); 3609576a15deSSean Christopherson else 3610531810caSBen Gardon write_lock(&kvm->mmu_lock); 3611c50d8ae3SPaolo Bonzini 3612c50d8ae3SPaolo Bonzini for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) 3613c50d8ae3SPaolo Bonzini if (roots_to_free & KVM_MMU_ROOT_PREVIOUS(i)) 36144d710de9SSean Christopherson mmu_free_root_page(kvm, &mmu->prev_roots[i].hpa, 3615c50d8ae3SPaolo Bonzini &invalid_list); 3616c50d8ae3SPaolo Bonzini 3617c50d8ae3SPaolo Bonzini if (free_active_root) { 36180e3223d8SSean Christopherson if (kvm_mmu_is_dummy_root(mmu->root.hpa)) { 36190e3223d8SSean Christopherson /* Nothing to cleanup for dummy roots. */ 36200e3223d8SSean Christopherson } else if (root_to_sp(mmu->root.hpa)) { 3621b9e5603cSPaolo Bonzini mmu_free_root_page(kvm, &mmu->root.hpa, &invalid_list); 362204d45551SSean Christopherson } else if (mmu->pae_root) { 3623c834e5e4SSean Christopherson for (i = 0; i < 4; ++i) { 3624c834e5e4SSean Christopherson if (!IS_VALID_PAE_ROOT(mmu->pae_root[i])) 3625c834e5e4SSean Christopherson continue; 3626c834e5e4SSean Christopherson 3627c834e5e4SSean Christopherson mmu_free_root_page(kvm, &mmu->pae_root[i], 3628c50d8ae3SPaolo Bonzini &invalid_list); 3629c834e5e4SSean Christopherson mmu->pae_root[i] = INVALID_PAE_ROOT; 3630c50d8ae3SPaolo Bonzini } 3631c50d8ae3SPaolo Bonzini } 3632b9e5603cSPaolo Bonzini mmu->root.hpa = INVALID_PAGE; 3633b9e5603cSPaolo Bonzini mmu->root.pgd = 0; 3634c50d8ae3SPaolo Bonzini } 3635c50d8ae3SPaolo Bonzini 3636576a15deSSean Christopherson if (is_tdp_mmu) { 3637576a15deSSean Christopherson read_unlock(&kvm->mmu_lock); 3638576a15deSSean Christopherson WARN_ON_ONCE(!list_empty(&invalid_list)); 3639576a15deSSean Christopherson } else { 36404d710de9SSean Christopherson kvm_mmu_commit_zap_page(kvm, &invalid_list); 3641531810caSBen Gardon write_unlock(&kvm->mmu_lock); 3642c50d8ae3SPaolo Bonzini } 3643576a15deSSean Christopherson } 3644c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_mmu_free_roots); 3645c50d8ae3SPaolo Bonzini 36460c1c92f1SPaolo Bonzini void kvm_mmu_free_guest_mode_roots(struct kvm *kvm, struct kvm_mmu *mmu) 364725b62c62SSean Christopherson { 364825b62c62SSean Christopherson unsigned long roots_to_free = 0; 3649c5f2d564SSean Christopherson struct kvm_mmu_page *sp; 365025b62c62SSean Christopherson hpa_t root_hpa; 365125b62c62SSean Christopherson int i; 365225b62c62SSean Christopherson 365325b62c62SSean Christopherson /* 365425b62c62SSean Christopherson * This should not be called while L2 is active, L2 can't invalidate 365525b62c62SSean Christopherson * _only_ its own roots, e.g. INVVPID unconditionally exits. 365625b62c62SSean Christopherson */ 36577a458f0eSPaolo Bonzini WARN_ON_ONCE(mmu->root_role.guest_mode); 365825b62c62SSean Christopherson 365925b62c62SSean Christopherson for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) { 366025b62c62SSean Christopherson root_hpa = mmu->prev_roots[i].hpa; 366125b62c62SSean Christopherson if (!VALID_PAGE(root_hpa)) 366225b62c62SSean Christopherson continue; 366325b62c62SSean Christopherson 3664c5f2d564SSean Christopherson sp = root_to_sp(root_hpa); 3665c5f2d564SSean Christopherson if (!sp || sp->role.guest_mode) 366625b62c62SSean Christopherson roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i); 366725b62c62SSean Christopherson } 366825b62c62SSean Christopherson 36690c1c92f1SPaolo Bonzini kvm_mmu_free_roots(kvm, mmu, roots_to_free); 367025b62c62SSean Christopherson } 367125b62c62SSean Christopherson EXPORT_SYMBOL_GPL(kvm_mmu_free_guest_mode_roots); 367225b62c62SSean Christopherson 36732e65e842SDavid Matlack static hpa_t mmu_alloc_root(struct kvm_vcpu *vcpu, gfn_t gfn, int quadrant, 367486938ab6SDavid Matlack u8 level) 3675c50d8ae3SPaolo Bonzini { 36762e65e842SDavid Matlack union kvm_mmu_page_role role = vcpu->arch.mmu->root_role; 3677c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 36788123f265SSean Christopherson 36792e65e842SDavid Matlack role.level = level; 36802e65e842SDavid Matlack role.quadrant = quadrant; 36812e65e842SDavid Matlack 36827f497775SDavid Matlack WARN_ON_ONCE(quadrant && !role.has_4_byte_gpte); 36837f497775SDavid Matlack WARN_ON_ONCE(role.direct && role.has_4_byte_gpte); 36847f497775SDavid Matlack 368587654643SDavid Matlack sp = kvm_mmu_get_shadow_page(vcpu, gfn, role); 36868123f265SSean Christopherson ++sp->root_count; 36878123f265SSean Christopherson 36888123f265SSean Christopherson return __pa(sp->spt); 36898123f265SSean Christopherson } 36908123f265SSean Christopherson 36918123f265SSean Christopherson static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu) 36928123f265SSean Christopherson { 3693b37233c9SSean Christopherson struct kvm_mmu *mmu = vcpu->arch.mmu; 3694a972e29cSPaolo Bonzini u8 shadow_root_level = mmu->root_role.level; 36958123f265SSean Christopherson hpa_t root; 3696c50d8ae3SPaolo Bonzini unsigned i; 36974a38162eSPaolo Bonzini int r; 36984a38162eSPaolo Bonzini 3699f5238c2aSSean Christopherson if (tdp_mmu_enabled) 3700f5238c2aSSean Christopherson return kvm_tdp_mmu_alloc_root(vcpu); 3701f5238c2aSSean Christopherson 37024a38162eSPaolo Bonzini write_lock(&vcpu->kvm->mmu_lock); 37034a38162eSPaolo Bonzini r = make_mmu_pages_available(vcpu); 37044a38162eSPaolo Bonzini if (r < 0) 37054a38162eSPaolo Bonzini goto out_unlock; 3706c50d8ae3SPaolo Bonzini 3707f5238c2aSSean Christopherson if (shadow_root_level >= PT64_ROOT_4LEVEL) { 370886938ab6SDavid Matlack root = mmu_alloc_root(vcpu, 0, 0, shadow_root_level); 3709b9e5603cSPaolo Bonzini mmu->root.hpa = root; 37108123f265SSean Christopherson } else if (shadow_root_level == PT32E_ROOT_LEVEL) { 37114a38162eSPaolo Bonzini if (WARN_ON_ONCE(!mmu->pae_root)) { 37124a38162eSPaolo Bonzini r = -EIO; 37134a38162eSPaolo Bonzini goto out_unlock; 37144a38162eSPaolo Bonzini } 371573ad1606SSean Christopherson 3716c50d8ae3SPaolo Bonzini for (i = 0; i < 4; ++i) { 3717c834e5e4SSean Christopherson WARN_ON_ONCE(IS_VALID_PAE_ROOT(mmu->pae_root[i])); 3718c50d8ae3SPaolo Bonzini 37197f497775SDavid Matlack root = mmu_alloc_root(vcpu, i << (30 - PAGE_SHIFT), 0, 37202e65e842SDavid Matlack PT32_ROOT_LEVEL); 372117e368d9SSean Christopherson mmu->pae_root[i] = root | PT_PRESENT_MASK | 3722d2263de1SYuan Yao shadow_me_value; 3723c50d8ae3SPaolo Bonzini } 3724b9e5603cSPaolo Bonzini mmu->root.hpa = __pa(mmu->pae_root); 372573ad1606SSean Christopherson } else { 372673ad1606SSean Christopherson WARN_ONCE(1, "Bad TDP root level = %d\n", shadow_root_level); 37274a38162eSPaolo Bonzini r = -EIO; 37284a38162eSPaolo Bonzini goto out_unlock; 372973ad1606SSean Christopherson } 37303651c7fcSSean Christopherson 3731b9e5603cSPaolo Bonzini /* root.pgd is ignored for direct MMUs. */ 3732b9e5603cSPaolo Bonzini mmu->root.pgd = 0; 37334a38162eSPaolo Bonzini out_unlock: 37344a38162eSPaolo Bonzini write_unlock(&vcpu->kvm->mmu_lock); 37354a38162eSPaolo Bonzini return r; 3736c50d8ae3SPaolo Bonzini } 3737c50d8ae3SPaolo Bonzini 37381e76a3ceSDavid Stevens static int mmu_first_shadow_root_alloc(struct kvm *kvm) 37391e76a3ceSDavid Stevens { 37401e76a3ceSDavid Stevens struct kvm_memslots *slots; 37411e76a3ceSDavid Stevens struct kvm_memory_slot *slot; 3742a54d8066SMaciej S. Szmigiero int r = 0, i, bkt; 37431e76a3ceSDavid Stevens 37441e76a3ceSDavid Stevens /* 37451e76a3ceSDavid Stevens * Check if this is the first shadow root being allocated before 37461e76a3ceSDavid Stevens * taking the lock. 37471e76a3ceSDavid Stevens */ 37481e76a3ceSDavid Stevens if (kvm_shadow_root_allocated(kvm)) 37491e76a3ceSDavid Stevens return 0; 37501e76a3ceSDavid Stevens 37511e76a3ceSDavid Stevens mutex_lock(&kvm->slots_arch_lock); 37521e76a3ceSDavid Stevens 37531e76a3ceSDavid Stevens /* Recheck, under the lock, whether this is the first shadow root. */ 37541e76a3ceSDavid Stevens if (kvm_shadow_root_allocated(kvm)) 37551e76a3ceSDavid Stevens goto out_unlock; 37561e76a3ceSDavid Stevens 37571e76a3ceSDavid Stevens /* 37581e76a3ceSDavid Stevens * Check if anything actually needs to be allocated, e.g. all metadata 37591e76a3ceSDavid Stevens * will be allocated upfront if TDP is disabled. 37601e76a3ceSDavid Stevens */ 37611e76a3ceSDavid Stevens if (kvm_memslots_have_rmaps(kvm) && 37621e76a3ceSDavid Stevens kvm_page_track_write_tracking_enabled(kvm)) 37631e76a3ceSDavid Stevens goto out_success; 37641e76a3ceSDavid Stevens 3765eed52e43SSean Christopherson for (i = 0; i < kvm_arch_nr_memslot_as_ids(kvm); i++) { 37661e76a3ceSDavid Stevens slots = __kvm_memslots(kvm, i); 3767a54d8066SMaciej S. Szmigiero kvm_for_each_memslot(slot, bkt, slots) { 37681e76a3ceSDavid Stevens /* 37691e76a3ceSDavid Stevens * Both of these functions are no-ops if the target is 37701e76a3ceSDavid Stevens * already allocated, so unconditionally calling both 37711e76a3ceSDavid Stevens * is safe. Intentionally do NOT free allocations on 37721e76a3ceSDavid Stevens * failure to avoid having to track which allocations 37731e76a3ceSDavid Stevens * were made now versus when the memslot was created. 37741e76a3ceSDavid Stevens * The metadata is guaranteed to be freed when the slot 37751e76a3ceSDavid Stevens * is freed, and will be kept/used if userspace retries 37761e76a3ceSDavid Stevens * KVM_RUN instead of killing the VM. 37771e76a3ceSDavid Stevens */ 37781e76a3ceSDavid Stevens r = memslot_rmap_alloc(slot, slot->npages); 37791e76a3ceSDavid Stevens if (r) 37801e76a3ceSDavid Stevens goto out_unlock; 37811e76a3ceSDavid Stevens r = kvm_page_track_write_tracking_alloc(slot); 37821e76a3ceSDavid Stevens if (r) 37831e76a3ceSDavid Stevens goto out_unlock; 37841e76a3ceSDavid Stevens } 37851e76a3ceSDavid Stevens } 37861e76a3ceSDavid Stevens 37871e76a3ceSDavid Stevens /* 37881e76a3ceSDavid Stevens * Ensure that shadow_root_allocated becomes true strictly after 37891e76a3ceSDavid Stevens * all the related pointers are set. 37901e76a3ceSDavid Stevens */ 37911e76a3ceSDavid Stevens out_success: 37921e76a3ceSDavid Stevens smp_store_release(&kvm->arch.shadow_root_allocated, true); 37931e76a3ceSDavid Stevens 37941e76a3ceSDavid Stevens out_unlock: 37951e76a3ceSDavid Stevens mutex_unlock(&kvm->slots_arch_lock); 37961e76a3ceSDavid Stevens return r; 37971e76a3ceSDavid Stevens } 37981e76a3ceSDavid Stevens 3799c50d8ae3SPaolo Bonzini static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu) 3800c50d8ae3SPaolo Bonzini { 3801b37233c9SSean Christopherson struct kvm_mmu *mmu = vcpu->arch.mmu; 38026e0918aeSSean Christopherson u64 pdptrs[4], pm_mask; 3803be01e8e2SSean Christopherson gfn_t root_gfn, root_pgd; 38047f497775SDavid Matlack int quadrant, i, r; 38058123f265SSean Christopherson hpa_t root; 3806c50d8ae3SPaolo Bonzini 38072fdcc1b3SPaolo Bonzini root_pgd = kvm_mmu_get_guest_pgd(vcpu, mmu); 3808a130066fSBinbin Wu root_gfn = (root_pgd & __PT_BASE_ADDR_MASK) >> PAGE_SHIFT; 3809c50d8ae3SPaolo Bonzini 38100e3223d8SSean Christopherson if (!kvm_vcpu_is_visible_gfn(vcpu, root_gfn)) { 38110e3223d8SSean Christopherson mmu->root.hpa = kvm_mmu_get_dummy_root(); 38120e3223d8SSean Christopherson return 0; 38130e3223d8SSean Christopherson } 3814c50d8ae3SPaolo Bonzini 3815c50d8ae3SPaolo Bonzini /* 38164a38162eSPaolo Bonzini * On SVM, reading PDPTRs might access guest memory, which might fault 38174a38162eSPaolo Bonzini * and thus might sleep. Grab the PDPTRs before acquiring mmu_lock. 38184a38162eSPaolo Bonzini */ 38194d25502aSPaolo Bonzini if (mmu->cpu_role.base.level == PT32E_ROOT_LEVEL) { 38206e0918aeSSean Christopherson for (i = 0; i < 4; ++i) { 38216e0918aeSSean Christopherson pdptrs[i] = mmu->get_pdptr(vcpu, i); 38226e0918aeSSean Christopherson if (!(pdptrs[i] & PT_PRESENT_MASK)) 38236e0918aeSSean Christopherson continue; 38246e0918aeSSean Christopherson 38250e3223d8SSean Christopherson if (!kvm_vcpu_is_visible_gfn(vcpu, pdptrs[i] >> PAGE_SHIFT)) 38260e3223d8SSean Christopherson pdptrs[i] = 0; 38276e0918aeSSean Christopherson } 38286e0918aeSSean Christopherson } 38296e0918aeSSean Christopherson 38301e76a3ceSDavid Stevens r = mmu_first_shadow_root_alloc(vcpu->kvm); 3831d501f747SBen Gardon if (r) 3832d501f747SBen Gardon return r; 3833d501f747SBen Gardon 38344a38162eSPaolo Bonzini write_lock(&vcpu->kvm->mmu_lock); 38354a38162eSPaolo Bonzini r = make_mmu_pages_available(vcpu); 38364a38162eSPaolo Bonzini if (r < 0) 38374a38162eSPaolo Bonzini goto out_unlock; 38384a38162eSPaolo Bonzini 3839c50d8ae3SPaolo Bonzini /* 3840c50d8ae3SPaolo Bonzini * Do we shadow a long mode page table? If so we need to 3841c50d8ae3SPaolo Bonzini * write-protect the guests page table root. 3842c50d8ae3SPaolo Bonzini */ 38434d25502aSPaolo Bonzini if (mmu->cpu_role.base.level >= PT64_ROOT_4LEVEL) { 38448123f265SSean Christopherson root = mmu_alloc_root(vcpu, root_gfn, 0, 384586938ab6SDavid Matlack mmu->root_role.level); 3846b9e5603cSPaolo Bonzini mmu->root.hpa = root; 3847be01e8e2SSean Christopherson goto set_root_pgd; 3848c50d8ae3SPaolo Bonzini } 3849c50d8ae3SPaolo Bonzini 38504a38162eSPaolo Bonzini if (WARN_ON_ONCE(!mmu->pae_root)) { 38514a38162eSPaolo Bonzini r = -EIO; 38524a38162eSPaolo Bonzini goto out_unlock; 38534a38162eSPaolo Bonzini } 385473ad1606SSean Christopherson 3855c50d8ae3SPaolo Bonzini /* 3856c50d8ae3SPaolo Bonzini * We shadow a 32 bit page table. This may be a legacy 2-level 3857c50d8ae3SPaolo Bonzini * or a PAE 3-level page table. In either case we need to be aware that 3858c50d8ae3SPaolo Bonzini * the shadow page table may be a PAE or a long mode page table. 3859c50d8ae3SPaolo Bonzini */ 3860e54f1ff2SKai Huang pm_mask = PT_PRESENT_MASK | shadow_me_value; 3861a972e29cSPaolo Bonzini if (mmu->root_role.level >= PT64_ROOT_4LEVEL) { 3862c50d8ae3SPaolo Bonzini pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK; 3863c50d8ae3SPaolo Bonzini 386403ca4589SSean Christopherson if (WARN_ON_ONCE(!mmu->pml4_root)) { 38654a38162eSPaolo Bonzini r = -EIO; 38664a38162eSPaolo Bonzini goto out_unlock; 38674a38162eSPaolo Bonzini } 386803ca4589SSean Christopherson mmu->pml4_root[0] = __pa(mmu->pae_root) | pm_mask; 3869cb0f722aSWei Huang 3870a972e29cSPaolo Bonzini if (mmu->root_role.level == PT64_ROOT_5LEVEL) { 3871cb0f722aSWei Huang if (WARN_ON_ONCE(!mmu->pml5_root)) { 3872cb0f722aSWei Huang r = -EIO; 3873cb0f722aSWei Huang goto out_unlock; 3874cb0f722aSWei Huang } 3875cb0f722aSWei Huang mmu->pml5_root[0] = __pa(mmu->pml4_root) | pm_mask; 3876cb0f722aSWei Huang } 387704d45551SSean Christopherson } 387804d45551SSean Christopherson 3879c50d8ae3SPaolo Bonzini for (i = 0; i < 4; ++i) { 3880c834e5e4SSean Christopherson WARN_ON_ONCE(IS_VALID_PAE_ROOT(mmu->pae_root[i])); 38816e6ec584SSean Christopherson 38824d25502aSPaolo Bonzini if (mmu->cpu_role.base.level == PT32E_ROOT_LEVEL) { 38836e0918aeSSean Christopherson if (!(pdptrs[i] & PT_PRESENT_MASK)) { 3884c834e5e4SSean Christopherson mmu->pae_root[i] = INVALID_PAE_ROOT; 3885c50d8ae3SPaolo Bonzini continue; 3886c50d8ae3SPaolo Bonzini } 38876e0918aeSSean Christopherson root_gfn = pdptrs[i] >> PAGE_SHIFT; 3888c50d8ae3SPaolo Bonzini } 3889c50d8ae3SPaolo Bonzini 38907f497775SDavid Matlack /* 38917f497775SDavid Matlack * If shadowing 32-bit non-PAE page tables, each PAE page 38927f497775SDavid Matlack * directory maps one quarter of the guest's non-PAE page 38937f497775SDavid Matlack * directory. Othwerise each PAE page direct shadows one guest 38947f497775SDavid Matlack * PAE page directory so that quadrant should be 0. 38957f497775SDavid Matlack */ 38967f497775SDavid Matlack quadrant = (mmu->cpu_role.base.level == PT32_ROOT_LEVEL) ? i : 0; 38977f497775SDavid Matlack 38987f497775SDavid Matlack root = mmu_alloc_root(vcpu, root_gfn, quadrant, PT32_ROOT_LEVEL); 3899b37233c9SSean Christopherson mmu->pae_root[i] = root | pm_mask; 3900c50d8ae3SPaolo Bonzini } 3901c50d8ae3SPaolo Bonzini 3902a972e29cSPaolo Bonzini if (mmu->root_role.level == PT64_ROOT_5LEVEL) 3903b9e5603cSPaolo Bonzini mmu->root.hpa = __pa(mmu->pml5_root); 3904a972e29cSPaolo Bonzini else if (mmu->root_role.level == PT64_ROOT_4LEVEL) 3905b9e5603cSPaolo Bonzini mmu->root.hpa = __pa(mmu->pml4_root); 3906ba0a194fSSean Christopherson else 3907b9e5603cSPaolo Bonzini mmu->root.hpa = __pa(mmu->pae_root); 3908c50d8ae3SPaolo Bonzini 3909be01e8e2SSean Christopherson set_root_pgd: 3910b9e5603cSPaolo Bonzini mmu->root.pgd = root_pgd; 39114a38162eSPaolo Bonzini out_unlock: 39124a38162eSPaolo Bonzini write_unlock(&vcpu->kvm->mmu_lock); 3913c50d8ae3SPaolo Bonzini 3914c6c937d6SLike Xu return r; 3915c50d8ae3SPaolo Bonzini } 3916c50d8ae3SPaolo Bonzini 3917748e52b9SSean Christopherson static int mmu_alloc_special_roots(struct kvm_vcpu *vcpu) 3918c50d8ae3SPaolo Bonzini { 3919748e52b9SSean Christopherson struct kvm_mmu *mmu = vcpu->arch.mmu; 3920a972e29cSPaolo Bonzini bool need_pml5 = mmu->root_role.level > PT64_ROOT_4LEVEL; 3921cb0f722aSWei Huang u64 *pml5_root = NULL; 3922cb0f722aSWei Huang u64 *pml4_root = NULL; 3923cb0f722aSWei Huang u64 *pae_root; 3924748e52b9SSean Christopherson 3925748e52b9SSean Christopherson /* 3926748e52b9SSean Christopherson * When shadowing 32-bit or PAE NPT with 64-bit NPT, the PML4 and PDP 3927748e52b9SSean Christopherson * tables are allocated and initialized at root creation as there is no 3928748e52b9SSean Christopherson * equivalent level in the guest's NPT to shadow. Allocate the tables 3929748e52b9SSean Christopherson * on demand, as running a 32-bit L1 VMM on 64-bit KVM is very rare. 3930748e52b9SSean Christopherson */ 3931347a0d0dSPaolo Bonzini if (mmu->root_role.direct || 3932347a0d0dSPaolo Bonzini mmu->cpu_role.base.level >= PT64_ROOT_4LEVEL || 3933a972e29cSPaolo Bonzini mmu->root_role.level < PT64_ROOT_4LEVEL) 3934748e52b9SSean Christopherson return 0; 3935748e52b9SSean Christopherson 3936a717a780SSean Christopherson /* 3937a717a780SSean Christopherson * NPT, the only paging mode that uses this horror, uses a fixed number 3938a717a780SSean Christopherson * of levels for the shadow page tables, e.g. all MMUs are 4-level or 3939a717a780SSean Christopherson * all MMus are 5-level. Thus, this can safely require that pml5_root 3940a717a780SSean Christopherson * is allocated if the other roots are valid and pml5 is needed, as any 3941a717a780SSean Christopherson * prior MMU would also have required pml5. 3942a717a780SSean Christopherson */ 3943a717a780SSean Christopherson if (mmu->pae_root && mmu->pml4_root && (!need_pml5 || mmu->pml5_root)) 3944748e52b9SSean Christopherson return 0; 3945748e52b9SSean Christopherson 3946748e52b9SSean Christopherson /* 3947748e52b9SSean Christopherson * The special roots should always be allocated in concert. Yell and 3948748e52b9SSean Christopherson * bail if KVM ends up in a state where only one of the roots is valid. 3949748e52b9SSean Christopherson */ 3950cb0f722aSWei Huang if (WARN_ON_ONCE(!tdp_enabled || mmu->pae_root || mmu->pml4_root || 3951a717a780SSean Christopherson (need_pml5 && mmu->pml5_root))) 3952748e52b9SSean Christopherson return -EIO; 3953748e52b9SSean Christopherson 39544a98623dSSean Christopherson /* 39554a98623dSSean Christopherson * Unlike 32-bit NPT, the PDP table doesn't need to be in low mem, and 39564a98623dSSean Christopherson * doesn't need to be decrypted. 39574a98623dSSean Christopherson */ 3958748e52b9SSean Christopherson pae_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT); 3959748e52b9SSean Christopherson if (!pae_root) 3960748e52b9SSean Christopherson return -ENOMEM; 3961748e52b9SSean Christopherson 3962cb0f722aSWei Huang #ifdef CONFIG_X86_64 396303ca4589SSean Christopherson pml4_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT); 3964cb0f722aSWei Huang if (!pml4_root) 3965cb0f722aSWei Huang goto err_pml4; 3966cb0f722aSWei Huang 3967a717a780SSean Christopherson if (need_pml5) { 3968cb0f722aSWei Huang pml5_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT); 3969cb0f722aSWei Huang if (!pml5_root) 3970cb0f722aSWei Huang goto err_pml5; 3971748e52b9SSean Christopherson } 3972cb0f722aSWei Huang #endif 3973748e52b9SSean Christopherson 3974748e52b9SSean Christopherson mmu->pae_root = pae_root; 397503ca4589SSean Christopherson mmu->pml4_root = pml4_root; 3976cb0f722aSWei Huang mmu->pml5_root = pml5_root; 3977748e52b9SSean Christopherson 3978748e52b9SSean Christopherson return 0; 3979cb0f722aSWei Huang 3980cb0f722aSWei Huang #ifdef CONFIG_X86_64 3981cb0f722aSWei Huang err_pml5: 3982cb0f722aSWei Huang free_page((unsigned long)pml4_root); 3983cb0f722aSWei Huang err_pml4: 3984cb0f722aSWei Huang free_page((unsigned long)pae_root); 3985cb0f722aSWei Huang return -ENOMEM; 3986cb0f722aSWei Huang #endif 3987c50d8ae3SPaolo Bonzini } 3988c50d8ae3SPaolo Bonzini 3989264d3dc1SLai Jiangshan static bool is_unsync_root(hpa_t root) 3990264d3dc1SLai Jiangshan { 3991264d3dc1SLai Jiangshan struct kvm_mmu_page *sp; 3992264d3dc1SLai Jiangshan 39930e3223d8SSean Christopherson if (!VALID_PAGE(root) || kvm_mmu_is_dummy_root(root)) 399461b05a9fSLai Jiangshan return false; 399561b05a9fSLai Jiangshan 3996264d3dc1SLai Jiangshan /* 3997264d3dc1SLai Jiangshan * The read barrier orders the CPU's read of SPTE.W during the page table 3998264d3dc1SLai Jiangshan * walk before the reads of sp->unsync/sp->unsync_children here. 3999264d3dc1SLai Jiangshan * 4000264d3dc1SLai Jiangshan * Even if another CPU was marking the SP as unsync-ed simultaneously, 4001264d3dc1SLai Jiangshan * any guest page table changes are not guaranteed to be visible anyway 4002264d3dc1SLai Jiangshan * until this VCPU issues a TLB flush strictly after those changes are 4003264d3dc1SLai Jiangshan * made. We only need to ensure that the other CPU sets these flags 4004264d3dc1SLai Jiangshan * before any actual changes to the page tables are made. The comments 4005264d3dc1SLai Jiangshan * in mmu_try_to_unsync_pages() describe what could go wrong if this 4006264d3dc1SLai Jiangshan * requirement isn't satisfied. 4007264d3dc1SLai Jiangshan */ 4008264d3dc1SLai Jiangshan smp_rmb(); 4009c5f2d564SSean Christopherson sp = root_to_sp(root); 40105d6a3221SSean Christopherson 40115d6a3221SSean Christopherson /* 40125d6a3221SSean Christopherson * PAE roots (somewhat arbitrarily) aren't backed by shadow pages, the 40135d6a3221SSean Christopherson * PDPTEs for a given PAE root need to be synchronized individually. 40145d6a3221SSean Christopherson */ 40155d6a3221SSean Christopherson if (WARN_ON_ONCE(!sp)) 40165d6a3221SSean Christopherson return false; 40175d6a3221SSean Christopherson 4018264d3dc1SLai Jiangshan if (sp->unsync || sp->unsync_children) 4019264d3dc1SLai Jiangshan return true; 4020264d3dc1SLai Jiangshan 4021264d3dc1SLai Jiangshan return false; 4022264d3dc1SLai Jiangshan } 4023264d3dc1SLai Jiangshan 4024c50d8ae3SPaolo Bonzini void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu) 4025c50d8ae3SPaolo Bonzini { 4026c50d8ae3SPaolo Bonzini int i; 4027c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 4028c50d8ae3SPaolo Bonzini 4029347a0d0dSPaolo Bonzini if (vcpu->arch.mmu->root_role.direct) 4030c50d8ae3SPaolo Bonzini return; 4031c50d8ae3SPaolo Bonzini 4032b9e5603cSPaolo Bonzini if (!VALID_PAGE(vcpu->arch.mmu->root.hpa)) 4033c50d8ae3SPaolo Bonzini return; 4034c50d8ae3SPaolo Bonzini 4035c50d8ae3SPaolo Bonzini vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY); 4036c50d8ae3SPaolo Bonzini 40374d25502aSPaolo Bonzini if (vcpu->arch.mmu->cpu_role.base.level >= PT64_ROOT_4LEVEL) { 4038b9e5603cSPaolo Bonzini hpa_t root = vcpu->arch.mmu->root.hpa; 4039c50d8ae3SPaolo Bonzini 4040264d3dc1SLai Jiangshan if (!is_unsync_root(root)) 4041c50d8ae3SPaolo Bonzini return; 4042c50d8ae3SPaolo Bonzini 4043c5f2d564SSean Christopherson sp = root_to_sp(root); 4044c5f2d564SSean Christopherson 4045531810caSBen Gardon write_lock(&vcpu->kvm->mmu_lock); 404665855ed8SLai Jiangshan mmu_sync_children(vcpu, sp, true); 4047531810caSBen Gardon write_unlock(&vcpu->kvm->mmu_lock); 4048c50d8ae3SPaolo Bonzini return; 4049c50d8ae3SPaolo Bonzini } 4050c50d8ae3SPaolo Bonzini 4051531810caSBen Gardon write_lock(&vcpu->kvm->mmu_lock); 4052c50d8ae3SPaolo Bonzini 4053c50d8ae3SPaolo Bonzini for (i = 0; i < 4; ++i) { 4054c50d8ae3SPaolo Bonzini hpa_t root = vcpu->arch.mmu->pae_root[i]; 4055c50d8ae3SPaolo Bonzini 4056c834e5e4SSean Christopherson if (IS_VALID_PAE_ROOT(root)) { 40575e3edd7eSSean Christopherson sp = spte_to_child_sp(root); 405865855ed8SLai Jiangshan mmu_sync_children(vcpu, sp, true); 4059c50d8ae3SPaolo Bonzini } 4060c50d8ae3SPaolo Bonzini } 4061c50d8ae3SPaolo Bonzini 4062531810caSBen Gardon write_unlock(&vcpu->kvm->mmu_lock); 4063c50d8ae3SPaolo Bonzini } 4064c50d8ae3SPaolo Bonzini 406561b05a9fSLai Jiangshan void kvm_mmu_sync_prev_roots(struct kvm_vcpu *vcpu) 406661b05a9fSLai Jiangshan { 406761b05a9fSLai Jiangshan unsigned long roots_to_free = 0; 406861b05a9fSLai Jiangshan int i; 406961b05a9fSLai Jiangshan 407061b05a9fSLai Jiangshan for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) 407161b05a9fSLai Jiangshan if (is_unsync_root(vcpu->arch.mmu->prev_roots[i].hpa)) 407261b05a9fSLai Jiangshan roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i); 407361b05a9fSLai Jiangshan 407461b05a9fSLai Jiangshan /* sync prev_roots by simply freeing them */ 40750c1c92f1SPaolo Bonzini kvm_mmu_free_roots(vcpu->kvm, vcpu->arch.mmu, roots_to_free); 407661b05a9fSLai Jiangshan } 407761b05a9fSLai Jiangshan 40781f5a21eeSLai Jiangshan static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu, 40795b22bbe7SLai Jiangshan gpa_t vaddr, u64 access, 4080c50d8ae3SPaolo Bonzini struct x86_exception *exception) 4081c50d8ae3SPaolo Bonzini { 4082c50d8ae3SPaolo Bonzini if (exception) 4083c50d8ae3SPaolo Bonzini exception->error_code = 0; 4084c59a0f57SLai Jiangshan return kvm_translate_gpa(vcpu, mmu, vaddr, access, exception); 4085c50d8ae3SPaolo Bonzini } 4086c50d8ae3SPaolo Bonzini 4087c50d8ae3SPaolo Bonzini static bool mmio_info_in_cache(struct kvm_vcpu *vcpu, u64 addr, bool direct) 4088c50d8ae3SPaolo Bonzini { 4089c50d8ae3SPaolo Bonzini /* 4090c50d8ae3SPaolo Bonzini * A nested guest cannot use the MMIO cache if it is using nested 4091c50d8ae3SPaolo Bonzini * page tables, because cr2 is a nGPA while the cache stores GPAs. 4092c50d8ae3SPaolo Bonzini */ 4093c50d8ae3SPaolo Bonzini if (mmu_is_nested(vcpu)) 4094c50d8ae3SPaolo Bonzini return false; 4095c50d8ae3SPaolo Bonzini 4096c50d8ae3SPaolo Bonzini if (direct) 4097c50d8ae3SPaolo Bonzini return vcpu_match_mmio_gpa(vcpu, addr); 4098c50d8ae3SPaolo Bonzini 4099c50d8ae3SPaolo Bonzini return vcpu_match_mmio_gva(vcpu, addr); 4100c50d8ae3SPaolo Bonzini } 4101c50d8ae3SPaolo Bonzini 410295fb5b02SBen Gardon /* 410395fb5b02SBen Gardon * Return the level of the lowest level SPTE added to sptes. 410495fb5b02SBen Gardon * That SPTE may be non-present. 4105c5c8c7c5SDavid Matlack * 4106c5c8c7c5SDavid Matlack * Must be called between walk_shadow_page_lockless_{begin,end}. 410795fb5b02SBen Gardon */ 410839b4d43eSSean Christopherson static int get_walk(struct kvm_vcpu *vcpu, u64 addr, u64 *sptes, int *root_level) 4109c50d8ae3SPaolo Bonzini { 4110c50d8ae3SPaolo Bonzini struct kvm_shadow_walk_iterator iterator; 41112aa07893SSean Christopherson int leaf = -1; 411295fb5b02SBen Gardon u64 spte; 4113c50d8ae3SPaolo Bonzini 411439b4d43eSSean Christopherson for (shadow_walk_init(&iterator, vcpu, addr), 411539b4d43eSSean Christopherson *root_level = iterator.level; 4116c50d8ae3SPaolo Bonzini shadow_walk_okay(&iterator); 4117c50d8ae3SPaolo Bonzini __shadow_walk_next(&iterator, spte)) { 411895fb5b02SBen Gardon leaf = iterator.level; 4119c50d8ae3SPaolo Bonzini spte = mmu_spte_get_lockless(iterator.sptep); 4120c50d8ae3SPaolo Bonzini 4121dde81f94SSean Christopherson sptes[leaf] = spte; 412295fb5b02SBen Gardon } 412395fb5b02SBen Gardon 412495fb5b02SBen Gardon return leaf; 412595fb5b02SBen Gardon } 412695fb5b02SBen Gardon 4127bca99c03SSean Christopherson static int get_sptes_lockless(struct kvm_vcpu *vcpu, u64 addr, u64 *sptes, 4128bca99c03SSean Christopherson int *root_level) 4129bca99c03SSean Christopherson { 4130bca99c03SSean Christopherson int leaf; 4131bca99c03SSean Christopherson 4132bca99c03SSean Christopherson walk_shadow_page_lockless_begin(vcpu); 4133bca99c03SSean Christopherson 4134bca99c03SSean Christopherson if (is_tdp_mmu_active(vcpu)) 4135bca99c03SSean Christopherson leaf = kvm_tdp_mmu_get_walk(vcpu, addr, sptes, root_level); 4136bca99c03SSean Christopherson else 4137bca99c03SSean Christopherson leaf = get_walk(vcpu, addr, sptes, root_level); 4138bca99c03SSean Christopherson 4139bca99c03SSean Christopherson walk_shadow_page_lockless_end(vcpu); 4140bca99c03SSean Christopherson return leaf; 4141bca99c03SSean Christopherson } 4142bca99c03SSean Christopherson 41439aa41879SSean Christopherson /* return true if reserved bit(s) are detected on a valid, non-MMIO SPTE. */ 414495fb5b02SBen Gardon static bool get_mmio_spte(struct kvm_vcpu *vcpu, u64 addr, u64 *sptep) 414595fb5b02SBen Gardon { 4146dde81f94SSean Christopherson u64 sptes[PT64_ROOT_MAX_LEVEL + 1]; 414795fb5b02SBen Gardon struct rsvd_bits_validate *rsvd_check; 414839b4d43eSSean Christopherson int root, leaf, level; 414995fb5b02SBen Gardon bool reserved = false; 415095fb5b02SBen Gardon 4151bca99c03SSean Christopherson leaf = get_sptes_lockless(vcpu, addr, sptes, &root); 41522aa07893SSean Christopherson if (unlikely(leaf < 0)) { 41532aa07893SSean Christopherson *sptep = 0ull; 41542aa07893SSean Christopherson return reserved; 41552aa07893SSean Christopherson } 41562aa07893SSean Christopherson 41579aa41879SSean Christopherson *sptep = sptes[leaf]; 41589aa41879SSean Christopherson 41599aa41879SSean Christopherson /* 41609aa41879SSean Christopherson * Skip reserved bits checks on the terminal leaf if it's not a valid 41619aa41879SSean Christopherson * SPTE. Note, this also (intentionally) skips MMIO SPTEs, which, by 41629aa41879SSean Christopherson * design, always have reserved bits set. The purpose of the checks is 41639aa41879SSean Christopherson * to detect reserved bits on non-MMIO SPTEs. i.e. buggy SPTEs. 41649aa41879SSean Christopherson */ 41659aa41879SSean Christopherson if (!is_shadow_present_pte(sptes[leaf])) 41669aa41879SSean Christopherson leaf++; 416795fb5b02SBen Gardon 416895fb5b02SBen Gardon rsvd_check = &vcpu->arch.mmu->shadow_zero_check; 416995fb5b02SBen Gardon 41709aa41879SSean Christopherson for (level = root; level >= leaf; level--) 4171961f8445SSean Christopherson reserved |= is_rsvd_spte(rsvd_check, sptes[level], level); 4172c50d8ae3SPaolo Bonzini 4173c50d8ae3SPaolo Bonzini if (reserved) { 4174bb4cdf3aSSean Christopherson pr_err("%s: reserved bits set on MMU-present spte, addr 0x%llx, hierarchy:\n", 4175c50d8ae3SPaolo Bonzini __func__, addr); 417695fb5b02SBen Gardon for (level = root; level >= leaf; level--) 4177bb4cdf3aSSean Christopherson pr_err("------ spte = 0x%llx level = %d, rsvd bits = 0x%llx", 4178bb4cdf3aSSean Christopherson sptes[level], level, 4179961f8445SSean Christopherson get_rsvd_bits(rsvd_check, sptes[level], level)); 4180c50d8ae3SPaolo Bonzini } 4181ddce6208SSean Christopherson 4182c50d8ae3SPaolo Bonzini return reserved; 4183c50d8ae3SPaolo Bonzini } 4184c50d8ae3SPaolo Bonzini 4185c50d8ae3SPaolo Bonzini static int handle_mmio_page_fault(struct kvm_vcpu *vcpu, u64 addr, bool direct) 4186c50d8ae3SPaolo Bonzini { 4187c50d8ae3SPaolo Bonzini u64 spte; 4188c50d8ae3SPaolo Bonzini bool reserved; 4189c50d8ae3SPaolo Bonzini 4190c50d8ae3SPaolo Bonzini if (mmio_info_in_cache(vcpu, addr, direct)) 4191c50d8ae3SPaolo Bonzini return RET_PF_EMULATE; 4192c50d8ae3SPaolo Bonzini 419395fb5b02SBen Gardon reserved = get_mmio_spte(vcpu, addr, &spte); 419420ba462dSSean Christopherson if (WARN_ON_ONCE(reserved)) 4195c50d8ae3SPaolo Bonzini return -EINVAL; 4196c50d8ae3SPaolo Bonzini 4197949019b9SSean Christopherson if (is_mmio_spte(vcpu->kvm, spte)) { 4198c50d8ae3SPaolo Bonzini gfn_t gfn = get_mmio_spte_gfn(spte); 41990a2b64c5SBen Gardon unsigned int access = get_mmio_spte_access(spte); 4200c50d8ae3SPaolo Bonzini 4201c50d8ae3SPaolo Bonzini if (!check_mmio_spte(vcpu, spte)) 4202c50d8ae3SPaolo Bonzini return RET_PF_INVALID; 4203c50d8ae3SPaolo Bonzini 4204c50d8ae3SPaolo Bonzini if (direct) 4205c50d8ae3SPaolo Bonzini addr = 0; 4206c50d8ae3SPaolo Bonzini 4207c50d8ae3SPaolo Bonzini trace_handle_mmio_page_fault(addr, gfn, access); 4208c50d8ae3SPaolo Bonzini vcpu_cache_mmio_info(vcpu, addr, gfn, access); 4209c50d8ae3SPaolo Bonzini return RET_PF_EMULATE; 4210c50d8ae3SPaolo Bonzini } 4211c50d8ae3SPaolo Bonzini 4212c50d8ae3SPaolo Bonzini /* 4213c50d8ae3SPaolo Bonzini * If the page table is zapped by other cpus, let CPU fault again on 4214c50d8ae3SPaolo Bonzini * the address. 4215c50d8ae3SPaolo Bonzini */ 4216c50d8ae3SPaolo Bonzini return RET_PF_RETRY; 4217c50d8ae3SPaolo Bonzini } 4218c50d8ae3SPaolo Bonzini 4219c50d8ae3SPaolo Bonzini static bool page_fault_handle_page_track(struct kvm_vcpu *vcpu, 4220b8a5d551SPaolo Bonzini struct kvm_page_fault *fault) 4221c50d8ae3SPaolo Bonzini { 4222b8a5d551SPaolo Bonzini if (unlikely(fault->rsvd)) 4223c50d8ae3SPaolo Bonzini return false; 4224c50d8ae3SPaolo Bonzini 4225b8a5d551SPaolo Bonzini if (!fault->present || !fault->write) 4226c50d8ae3SPaolo Bonzini return false; 4227c50d8ae3SPaolo Bonzini 4228c50d8ae3SPaolo Bonzini /* 4229c50d8ae3SPaolo Bonzini * guest is writing the page which is write tracked which can 4230c50d8ae3SPaolo Bonzini * not be fixed by page fault handler. 4231c50d8ae3SPaolo Bonzini */ 42327b574863SSean Christopherson if (kvm_gfn_is_write_tracked(vcpu->kvm, fault->slot, fault->gfn)) 4233c50d8ae3SPaolo Bonzini return true; 4234c50d8ae3SPaolo Bonzini 4235c50d8ae3SPaolo Bonzini return false; 4236c50d8ae3SPaolo Bonzini } 4237c50d8ae3SPaolo Bonzini 4238c50d8ae3SPaolo Bonzini static void shadow_page_table_clear_flood(struct kvm_vcpu *vcpu, gva_t addr) 4239c50d8ae3SPaolo Bonzini { 4240c50d8ae3SPaolo Bonzini struct kvm_shadow_walk_iterator iterator; 4241c50d8ae3SPaolo Bonzini u64 spte; 4242c50d8ae3SPaolo Bonzini 4243c50d8ae3SPaolo Bonzini walk_shadow_page_lockless_begin(vcpu); 42443e44dce4SLai Jiangshan for_each_shadow_entry_lockless(vcpu, addr, iterator, spte) 4245c50d8ae3SPaolo Bonzini clear_sp_write_flooding_count(iterator.sptep); 4246c50d8ae3SPaolo Bonzini walk_shadow_page_lockless_end(vcpu); 4247c50d8ae3SPaolo Bonzini } 4248c50d8ae3SPaolo Bonzini 42496f3c1fc5SLiang Zhang static u32 alloc_apf_token(struct kvm_vcpu *vcpu) 42506f3c1fc5SLiang Zhang { 42516f3c1fc5SLiang Zhang /* make sure the token value is not 0 */ 42526f3c1fc5SLiang Zhang u32 id = vcpu->arch.apf.id; 42536f3c1fc5SLiang Zhang 42546f3c1fc5SLiang Zhang if (id << 12 == 0) 42556f3c1fc5SLiang Zhang vcpu->arch.apf.id = 1; 42566f3c1fc5SLiang Zhang 42576f3c1fc5SLiang Zhang return (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id; 42586f3c1fc5SLiang Zhang } 42596f3c1fc5SLiang Zhang 4260cd389f50SPaolo Bonzini static bool kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, 4261cd389f50SPaolo Bonzini struct kvm_page_fault *fault) 4262c50d8ae3SPaolo Bonzini { 4263c50d8ae3SPaolo Bonzini struct kvm_arch_async_pf arch; 4264c50d8ae3SPaolo Bonzini 42656f3c1fc5SLiang Zhang arch.token = alloc_apf_token(vcpu); 4266cd389f50SPaolo Bonzini arch.gfn = fault->gfn; 4267cd389f50SPaolo Bonzini arch.error_code = fault->error_code; 4268347a0d0dSPaolo Bonzini arch.direct_map = vcpu->arch.mmu->root_role.direct; 42692fdcc1b3SPaolo Bonzini arch.cr3 = kvm_mmu_get_guest_pgd(vcpu, vcpu->arch.mmu); 4270c50d8ae3SPaolo Bonzini 4271cd389f50SPaolo Bonzini return kvm_setup_async_pf(vcpu, fault->addr, 4272cd389f50SPaolo Bonzini kvm_vcpu_gfn_to_hva(vcpu, fault->gfn), &arch); 4273c50d8ae3SPaolo Bonzini } 4274c50d8ae3SPaolo Bonzini 42758a009d5bSSean Christopherson void kvm_arch_async_page_ready(struct kvm_vcpu *vcpu, struct kvm_async_pf *work) 42768a009d5bSSean Christopherson { 42778a009d5bSSean Christopherson int r; 42788a009d5bSSean Christopherson 4279cd389f50SPaolo Bonzini if (WARN_ON_ONCE(work->arch.error_code & PFERR_PRIVATE_ACCESS)) 4280cd389f50SPaolo Bonzini return; 4281cd389f50SPaolo Bonzini 42828a009d5bSSean Christopherson if ((vcpu->arch.mmu->root_role.direct != work->arch.direct_map) || 42838a009d5bSSean Christopherson work->wakeup_all) 42848a009d5bSSean Christopherson return; 42858a009d5bSSean Christopherson 42868a009d5bSSean Christopherson r = kvm_mmu_reload(vcpu); 42878a009d5bSSean Christopherson if (unlikely(r)) 42888a009d5bSSean Christopherson return; 42898a009d5bSSean Christopherson 42908a009d5bSSean Christopherson if (!vcpu->arch.mmu->root_role.direct && 42912fdcc1b3SPaolo Bonzini work->arch.cr3 != kvm_mmu_get_guest_pgd(vcpu, vcpu->arch.mmu)) 42928a009d5bSSean Christopherson return; 42938a009d5bSSean Christopherson 4294f5e7f00cSSean Christopherson r = kvm_mmu_do_page_fault(vcpu, work->cr2_or_gpa, work->arch.error_code, 429558ef2469SPaolo Bonzini true, NULL, NULL); 4296f5e7f00cSSean Christopherson 4297f5e7f00cSSean Christopherson /* 4298f5e7f00cSSean Christopherson * Account fixed page faults, otherwise they'll never be counted, but 4299f5e7f00cSSean Christopherson * ignore stats for all other return times. Page-ready "faults" aren't 4300f5e7f00cSSean Christopherson * truly spurious and never trigger emulation 4301f5e7f00cSSean Christopherson */ 4302f5e7f00cSSean Christopherson if (r == RET_PF_FIXED) 4303f5e7f00cSSean Christopherson vcpu->stat.pf_fixed++; 43048a009d5bSSean Christopherson } 43058a009d5bSSean Christopherson 43068dd2eee9SChao Peng static inline u8 kvm_max_level_for_order(int order) 43078dd2eee9SChao Peng { 43088dd2eee9SChao Peng BUILD_BUG_ON(KVM_MAX_HUGEPAGE_LEVEL > PG_LEVEL_1G); 43098dd2eee9SChao Peng 43108dd2eee9SChao Peng KVM_MMU_WARN_ON(order != KVM_HPAGE_GFN_SHIFT(PG_LEVEL_1G) && 43118dd2eee9SChao Peng order != KVM_HPAGE_GFN_SHIFT(PG_LEVEL_2M) && 43128dd2eee9SChao Peng order != KVM_HPAGE_GFN_SHIFT(PG_LEVEL_4K)); 43138dd2eee9SChao Peng 43148dd2eee9SChao Peng if (order >= KVM_HPAGE_GFN_SHIFT(PG_LEVEL_1G)) 43158dd2eee9SChao Peng return PG_LEVEL_1G; 43168dd2eee9SChao Peng 43178dd2eee9SChao Peng if (order >= KVM_HPAGE_GFN_SHIFT(PG_LEVEL_2M)) 43188dd2eee9SChao Peng return PG_LEVEL_2M; 43198dd2eee9SChao Peng 43208dd2eee9SChao Peng return PG_LEVEL_4K; 43218dd2eee9SChao Peng } 43228dd2eee9SChao Peng 4323f32fb328SMichael Roth static u8 kvm_max_private_mapping_level(struct kvm *kvm, kvm_pfn_t pfn, 4324f32fb328SMichael Roth u8 max_level, int gmem_order) 4325f32fb328SMichael Roth { 4326f32fb328SMichael Roth u8 req_max_level; 4327f32fb328SMichael Roth 4328f32fb328SMichael Roth if (max_level == PG_LEVEL_4K) 4329f32fb328SMichael Roth return PG_LEVEL_4K; 4330f32fb328SMichael Roth 4331f32fb328SMichael Roth max_level = min(kvm_max_level_for_order(gmem_order), max_level); 4332f32fb328SMichael Roth if (max_level == PG_LEVEL_4K) 4333f32fb328SMichael Roth return PG_LEVEL_4K; 4334f32fb328SMichael Roth 4335f32fb328SMichael Roth req_max_level = static_call(kvm_x86_private_max_mapping_level)(kvm, pfn); 4336f32fb328SMichael Roth if (req_max_level) 4337f32fb328SMichael Roth max_level = min(max_level, req_max_level); 4338f32fb328SMichael Roth 4339f32fb328SMichael Roth return req_max_level; 4340f32fb328SMichael Roth } 4341f32fb328SMichael Roth 43428dd2eee9SChao Peng static int kvm_faultin_pfn_private(struct kvm_vcpu *vcpu, 43438dd2eee9SChao Peng struct kvm_page_fault *fault) 43448dd2eee9SChao Peng { 43458dd2eee9SChao Peng int max_order, r; 43468dd2eee9SChao Peng 43478dd2eee9SChao Peng if (!kvm_slot_can_be_private(fault->slot)) { 43488dd2eee9SChao Peng kvm_mmu_prepare_memory_fault_exit(vcpu, fault); 43498dd2eee9SChao Peng return -EFAULT; 43508dd2eee9SChao Peng } 43518dd2eee9SChao Peng 43528dd2eee9SChao Peng r = kvm_gmem_get_pfn(vcpu->kvm, fault->slot, fault->gfn, &fault->pfn, 43538dd2eee9SChao Peng &max_order); 43548dd2eee9SChao Peng if (r) { 43558dd2eee9SChao Peng kvm_mmu_prepare_memory_fault_exit(vcpu, fault); 43568dd2eee9SChao Peng return r; 43578dd2eee9SChao Peng } 43588dd2eee9SChao Peng 43598dd2eee9SChao Peng fault->map_writable = !(fault->slot->flags & KVM_MEM_READONLY); 4360f32fb328SMichael Roth fault->max_level = kvm_max_private_mapping_level(vcpu->kvm, fault->pfn, 4361f32fb328SMichael Roth fault->max_level, max_order); 43628dd2eee9SChao Peng 43638dd2eee9SChao Peng return RET_PF_CONTINUE; 43648dd2eee9SChao Peng } 43658dd2eee9SChao Peng 4366ba6e3fe2SDavid Matlack static int __kvm_faultin_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 4367c50d8ae3SPaolo Bonzini { 4368c50d8ae3SPaolo Bonzini bool async; 4369c50d8ae3SPaolo Bonzini 43708dd2eee9SChao Peng if (fault->is_private) 43718dd2eee9SChao Peng return kvm_faultin_pfn_private(vcpu, fault); 43728dd2eee9SChao Peng 4373c50d8ae3SPaolo Bonzini async = false; 4374cd272fc4SSean Christopherson fault->pfn = __gfn_to_pfn_memslot(fault->slot, fault->gfn, false, false, 4375cd272fc4SSean Christopherson &async, fault->write, 4376cd272fc4SSean Christopherson &fault->map_writable, &fault->hva); 4377c50d8ae3SPaolo Bonzini if (!async) 43785276c616SSean Christopherson return RET_PF_CONTINUE; /* *pfn has correct page already */ 4379c50d8ae3SPaolo Bonzini 43802839180cSPaolo Bonzini if (!fault->prefetch && kvm_can_do_async_pf(vcpu)) { 43813647cd04SPaolo Bonzini trace_kvm_try_async_get_page(fault->addr, fault->gfn); 43823647cd04SPaolo Bonzini if (kvm_find_async_pf_gfn(vcpu, fault->gfn)) { 43831685c0f3SMingwei Zhang trace_kvm_async_pf_repeated_fault(fault->addr, fault->gfn); 4384c50d8ae3SPaolo Bonzini kvm_make_request(KVM_REQ_APF_HALT, vcpu); 43855276c616SSean Christopherson return RET_PF_RETRY; 4386cd272fc4SSean Christopherson } else if (kvm_arch_setup_async_pf(vcpu, fault)) { 43875276c616SSean Christopherson return RET_PF_RETRY; 43885276c616SSean Christopherson } 4389c50d8ae3SPaolo Bonzini } 4390c50d8ae3SPaolo Bonzini 439176657687SPeter Xu /* 439276657687SPeter Xu * Allow gup to bail on pending non-fatal signals when it's also allowed 439376657687SPeter Xu * to wait for IO. Note, gup always bails if it is unable to quickly 439476657687SPeter Xu * get a page and a fatal signal, i.e. SIGKILL, is pending. 439576657687SPeter Xu */ 4396cd272fc4SSean Christopherson fault->pfn = __gfn_to_pfn_memslot(fault->slot, fault->gfn, false, true, 4397cd272fc4SSean Christopherson NULL, fault->write, 4398cd272fc4SSean Christopherson &fault->map_writable, &fault->hva); 43995276c616SSean Christopherson return RET_PF_CONTINUE; 4400c50d8ae3SPaolo Bonzini } 4401c50d8ae3SPaolo Bonzini 4402354c908cSDavid Matlack static int kvm_faultin_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault, 4403354c908cSDavid Matlack unsigned int access) 4404ba6e3fe2SDavid Matlack { 4405cd272fc4SSean Christopherson struct kvm_memory_slot *slot = fault->slot; 440656c3a4e4SDavid Matlack int ret; 440756c3a4e4SDavid Matlack 4408cd272fc4SSean Christopherson /* 4409cd272fc4SSean Christopherson * Note that the mmu_invalidate_seq also serves to detect a concurrent 4410cd272fc4SSean Christopherson * change in attributes. is_page_fault_stale() will detect an 4411cd272fc4SSean Christopherson * invalidation relate to fault->fn and resume the guest without 4412cd272fc4SSean Christopherson * installing a mapping in the page tables. 4413cd272fc4SSean Christopherson */ 4414cd272fc4SSean Christopherson fault->mmu_seq = vcpu->kvm->mmu_invalidate_seq; 4415cd272fc4SSean Christopherson smp_rmb(); 4416cd272fc4SSean Christopherson 4417cd272fc4SSean Christopherson /* 4418cd272fc4SSean Christopherson * Now that we have a snapshot of mmu_invalidate_seq we can check for a 4419cd272fc4SSean Christopherson * private vs. shared mismatch. 4420cd272fc4SSean Christopherson */ 4421cd272fc4SSean Christopherson if (fault->is_private != kvm_mem_is_private(vcpu->kvm, fault->gfn)) { 4422cd272fc4SSean Christopherson kvm_mmu_prepare_memory_fault_exit(vcpu, fault); 4423cd272fc4SSean Christopherson return -EFAULT; 4424cd272fc4SSean Christopherson } 4425cd272fc4SSean Christopherson 4426f6adeae8SSean Christopherson if (unlikely(!slot)) 4427f6adeae8SSean Christopherson return kvm_handle_noslot_fault(vcpu, fault, access); 4428f6adeae8SSean Christopherson 4429c50d8ae3SPaolo Bonzini /* 4430c50d8ae3SPaolo Bonzini * Retry the page fault if the gfn hit a memslot that is being deleted 4431c50d8ae3SPaolo Bonzini * or moved. This ensures any existing SPTEs for the old memslot will 4432c50d8ae3SPaolo Bonzini * be zapped before KVM inserts a new MMIO SPTE for the gfn. 4433c50d8ae3SPaolo Bonzini */ 4434f6adeae8SSean Christopherson if (slot->flags & KVM_MEMSLOT_INVALID) 4435c50d8ae3SPaolo Bonzini return RET_PF_RETRY; 4436c50d8ae3SPaolo Bonzini 4437f6adeae8SSean Christopherson if (slot->id == APIC_ACCESS_PAGE_PRIVATE_MEMSLOT) { 44385bd74f6eSSean Christopherson /* 44395bd74f6eSSean Christopherson * Don't map L1's APIC access page into L2, KVM doesn't support 44405bd74f6eSSean Christopherson * using APICv/AVIC to accelerate L2 accesses to L1's APIC, 44415bd74f6eSSean Christopherson * i.e. the access needs to be emulated. Emulating access to 44425bd74f6eSSean Christopherson * L1's APIC is also correct if L1 is accelerating L2's own 44435bd74f6eSSean Christopherson * virtual APIC, but for some reason L1 also maps _L1's_ APIC 44445bd74f6eSSean Christopherson * into L2. Note, vcpu_is_mmio_gpa() always treats access to 44455bd74f6eSSean Christopherson * the APIC as MMIO. Allow an MMIO SPTE to be created, as KVM 44465bd74f6eSSean Christopherson * uses different roots for L1 vs. L2, i.e. there is no danger 44475bd74f6eSSean Christopherson * of breaking APICv/AVIC for L1. 44485bd74f6eSSean Christopherson */ 4449f6adeae8SSean Christopherson if (is_guest_mode(vcpu)) 4450f6adeae8SSean Christopherson return kvm_handle_noslot_fault(vcpu, fault, access); 4451f6adeae8SSean Christopherson 4452c50d8ae3SPaolo Bonzini /* 4453c50d8ae3SPaolo Bonzini * If the APIC access page exists but is disabled, go directly 4454c50d8ae3SPaolo Bonzini * to emulation without caching the MMIO access or creating a 4455c50d8ae3SPaolo Bonzini * MMIO SPTE. That way the cache doesn't need to be purged 4456c50d8ae3SPaolo Bonzini * when the AVIC is re-enabled. 4457c50d8ae3SPaolo Bonzini */ 44585bd74f6eSSean Christopherson if (!kvm_apicv_activated(vcpu->kvm)) 4459c50d8ae3SPaolo Bonzini return RET_PF_EMULATE; 4460c50d8ae3SPaolo Bonzini } 4461c50d8ae3SPaolo Bonzini 4462d02c357eSSean Christopherson /* 4463d02c357eSSean Christopherson * Check for a relevant mmu_notifier invalidation event before getting 4464d02c357eSSean Christopherson * the pfn from the primary MMU, and before acquiring mmu_lock. 4465d02c357eSSean Christopherson * 4466d02c357eSSean Christopherson * For mmu_lock, if there is an in-progress invalidation and the kernel 4467d02c357eSSean Christopherson * allows preemption, the invalidation task may drop mmu_lock and yield 4468d02c357eSSean Christopherson * in response to mmu_lock being contended, which is *very* counter- 4469d02c357eSSean Christopherson * productive as this vCPU can't actually make forward progress until 4470d02c357eSSean Christopherson * the invalidation completes. 4471d02c357eSSean Christopherson * 4472d02c357eSSean Christopherson * Retrying now can also avoid unnessary lock contention in the primary 4473d02c357eSSean Christopherson * MMU, as the primary MMU doesn't necessarily hold a single lock for 4474d02c357eSSean Christopherson * the duration of the invalidation, i.e. faulting in a conflicting pfn 4475d02c357eSSean Christopherson * can cause the invalidation to take longer by holding locks that are 4476d02c357eSSean Christopherson * needed to complete the invalidation. 4477d02c357eSSean Christopherson * 4478d02c357eSSean Christopherson * Do the pre-check even for non-preemtible kernels, i.e. even if KVM 4479d02c357eSSean Christopherson * will never yield mmu_lock in response to contention, as this vCPU is 4480d02c357eSSean Christopherson * *guaranteed* to need to retry, i.e. waiting until mmu_lock is held 4481d02c357eSSean Christopherson * to detect retry guarantees the worst case latency for the vCPU. 4482d02c357eSSean Christopherson */ 4483f6adeae8SSean Christopherson if (mmu_invalidate_retry_gfn_unsafe(vcpu->kvm, fault->mmu_seq, fault->gfn)) 4484d02c357eSSean Christopherson return RET_PF_RETRY; 4485d02c357eSSean Christopherson 448656c3a4e4SDavid Matlack ret = __kvm_faultin_pfn(vcpu, fault); 448756c3a4e4SDavid Matlack if (ret != RET_PF_CONTINUE) 448856c3a4e4SDavid Matlack return ret; 448956c3a4e4SDavid Matlack 449056c3a4e4SDavid Matlack if (unlikely(is_error_pfn(fault->pfn))) 4491cd08d178SDavid Matlack return kvm_handle_error_pfn(vcpu, fault); 449256c3a4e4SDavid Matlack 44932b1f4355SSean Christopherson if (WARN_ON_ONCE(!fault->slot || is_noslot_pfn(fault->pfn))) 4494354c908cSDavid Matlack return kvm_handle_noslot_fault(vcpu, fault, access); 4495354c908cSDavid Matlack 4496d02c357eSSean Christopherson /* 4497d02c357eSSean Christopherson * Check again for a relevant mmu_notifier invalidation event purely to 4498d02c357eSSean Christopherson * avoid contending mmu_lock. Most invalidations will be detected by 4499d02c357eSSean Christopherson * the previous check, but checking is extremely cheap relative to the 4500d02c357eSSean Christopherson * overall cost of failing to detect the invalidation until after 4501d02c357eSSean Christopherson * mmu_lock is acquired. 4502d02c357eSSean Christopherson */ 4503d02c357eSSean Christopherson if (mmu_invalidate_retry_gfn_unsafe(vcpu->kvm, fault->mmu_seq, fault->gfn)) { 4504d02c357eSSean Christopherson kvm_release_pfn_clean(fault->pfn); 4505d02c357eSSean Christopherson return RET_PF_RETRY; 4506d02c357eSSean Christopherson } 4507d02c357eSSean Christopherson 450856c3a4e4SDavid Matlack return RET_PF_CONTINUE; 4509ba6e3fe2SDavid Matlack } 4510ba6e3fe2SDavid Matlack 4511a955cad8SSean Christopherson /* 4512a955cad8SSean Christopherson * Returns true if the page fault is stale and needs to be retried, i.e. if the 4513a955cad8SSean Christopherson * root was invalidated by a memslot update or a relevant mmu_notifier fired. 4514a955cad8SSean Christopherson */ 4515a955cad8SSean Christopherson static bool is_page_fault_stale(struct kvm_vcpu *vcpu, 4516ba6e3fe2SDavid Matlack struct kvm_page_fault *fault) 4517a955cad8SSean Christopherson { 4518c5f2d564SSean Christopherson struct kvm_mmu_page *sp = root_to_sp(vcpu->arch.mmu->root.hpa); 451918c841e1SSean Christopherson 452018c841e1SSean Christopherson /* Special roots, e.g. pae_root, are not backed by shadow pages. */ 452118c841e1SSean Christopherson if (sp && is_obsolete_sp(vcpu->kvm, sp)) 452218c841e1SSean Christopherson return true; 452318c841e1SSean Christopherson 452418c841e1SSean Christopherson /* 452518c841e1SSean Christopherson * Roots without an associated shadow page are considered invalid if 452618c841e1SSean Christopherson * there is a pending request to free obsolete roots. The request is 452718c841e1SSean Christopherson * only a hint that the current root _may_ be obsolete and needs to be 452818c841e1SSean Christopherson * reloaded, e.g. if the guest frees a PGD that KVM is tracking as a 452918c841e1SSean Christopherson * previous root, then __kvm_mmu_prepare_zap_page() signals all vCPUs 453018c841e1SSean Christopherson * to reload even if no vCPU is actively using the root. 453118c841e1SSean Christopherson */ 4532527d5cd7SSean Christopherson if (!sp && kvm_test_request(KVM_REQ_MMU_FREE_OBSOLETE_ROOTS, vcpu)) 4533a955cad8SSean Christopherson return true; 4534a955cad8SSean Christopherson 4535d02c357eSSean Christopherson /* 4536d02c357eSSean Christopherson * Check for a relevant mmu_notifier invalidation event one last time 4537d02c357eSSean Christopherson * now that mmu_lock is held, as the "unsafe" checks performed without 4538d02c357eSSean Christopherson * holding mmu_lock can get false negatives. 4539d02c357eSSean Christopherson */ 4540a955cad8SSean Christopherson return fault->slot && 45418569992dSChao Peng mmu_invalidate_retry_gfn(vcpu->kvm, fault->mmu_seq, fault->gfn); 4542a955cad8SSean Christopherson } 4543a955cad8SSean Christopherson 45444326e57eSPaolo Bonzini static int direct_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 4545c50d8ae3SPaolo Bonzini { 454683f06fa7SSean Christopherson int r; 4547c50d8ae3SPaolo Bonzini 45480e3223d8SSean Christopherson /* Dummy roots are used only for shadowing bad guest roots. */ 45490e3223d8SSean Christopherson if (WARN_ON_ONCE(kvm_mmu_is_dummy_root(vcpu->arch.mmu->root.hpa))) 45500e3223d8SSean Christopherson return RET_PF_RETRY; 45510e3223d8SSean Christopherson 4552b8a5d551SPaolo Bonzini if (page_fault_handle_page_track(vcpu, fault)) 4553c50d8ae3SPaolo Bonzini return RET_PF_EMULATE; 4554c50d8ae3SPaolo Bonzini 45553c8ad5a6SPaolo Bonzini r = fast_page_fault(vcpu, fault); 4556c4371c2aSSean Christopherson if (r != RET_PF_INVALID) 4557c4371c2aSSean Christopherson return r; 455883291445SSean Christopherson 4559378f5cd6SSean Christopherson r = mmu_topup_memory_caches(vcpu, false); 4560c50d8ae3SPaolo Bonzini if (r) 4561c50d8ae3SPaolo Bonzini return r; 4562c50d8ae3SPaolo Bonzini 4563354c908cSDavid Matlack r = kvm_faultin_pfn(vcpu, fault, ACC_ALL); 45645276c616SSean Christopherson if (r != RET_PF_CONTINUE) 4565367fd790SSean Christopherson return r; 4566367fd790SSean Christopherson 4567367fd790SSean Christopherson r = RET_PF_RETRY; 4568531810caSBen Gardon write_lock(&vcpu->kvm->mmu_lock); 4569a2855afcSBen Gardon 4570ba6e3fe2SDavid Matlack if (is_page_fault_stale(vcpu, fault)) 4571367fd790SSean Christopherson goto out_unlock; 4572a955cad8SSean Christopherson 45737bd7ded6SSean Christopherson r = make_mmu_pages_available(vcpu); 45747bd7ded6SSean Christopherson if (r) 4575367fd790SSean Christopherson goto out_unlock; 4576bb18842eSBen Gardon 45776c882ef4SDavid Matlack r = direct_map(vcpu, fault); 45780f90e1c1SSean Christopherson 4579367fd790SSean Christopherson out_unlock: 4580531810caSBen Gardon write_unlock(&vcpu->kvm->mmu_lock); 45813647cd04SPaolo Bonzini kvm_release_pfn_clean(fault->pfn); 4582367fd790SSean Christopherson return r; 4583c50d8ae3SPaolo Bonzini } 4584c50d8ae3SPaolo Bonzini 4585c501040aSPaolo Bonzini static int nonpaging_page_fault(struct kvm_vcpu *vcpu, 4586c501040aSPaolo Bonzini struct kvm_page_fault *fault) 45870f90e1c1SSean Christopherson { 45880f90e1c1SSean Christopherson /* This path builds a PAE pagetable, we can map 2mb pages at maximum. */ 45894326e57eSPaolo Bonzini fault->max_level = PG_LEVEL_2M; 45904326e57eSPaolo Bonzini return direct_page_fault(vcpu, fault); 45910f90e1c1SSean Christopherson } 45920f90e1c1SSean Christopherson 4593c50d8ae3SPaolo Bonzini int kvm_handle_page_fault(struct kvm_vcpu *vcpu, u64 error_code, 4594c50d8ae3SPaolo Bonzini u64 fault_address, char *insn, int insn_len) 4595c50d8ae3SPaolo Bonzini { 4596c50d8ae3SPaolo Bonzini int r = 1; 45979ce372b3SVitaly Kuznetsov u32 flags = vcpu->arch.apf.host_apf_flags; 4598c50d8ae3SPaolo Bonzini 4599736c291cSSean Christopherson #ifndef CONFIG_X86_64 4600736c291cSSean Christopherson /* A 64-bit CR2 should be impossible on 32-bit KVM. */ 4601736c291cSSean Christopherson if (WARN_ON_ONCE(fault_address >> 32)) 4602736c291cSSean Christopherson return -EFAULT; 4603736c291cSSean Christopherson #endif 46047bdbb820SSean Christopherson /* 46057bdbb820SSean Christopherson * Legacy #PF exception only have a 32-bit error code. Simply drop the 46067bdbb820SSean Christopherson * upper bits as KVM doesn't use them for #PF (because they are never 46077bdbb820SSean Christopherson * set), and to ensure there are no collisions with KVM-defined bits. 46087bdbb820SSean Christopherson */ 46097bdbb820SSean Christopherson if (WARN_ON_ONCE(error_code >> 32)) 46107bdbb820SSean Christopherson error_code = lower_32_bits(error_code); 4611736c291cSSean Christopherson 4612dee281e4SSean Christopherson /* Ensure the above sanity check also covers KVM-defined flags. */ 4613dee281e4SSean Christopherson BUILD_BUG_ON(lower_32_bits(PFERR_SYNTHETIC_MASK)); 4614c50d8ae3SPaolo Bonzini 4615c50d8ae3SPaolo Bonzini vcpu->arch.l1tf_flush_l1d = true; 46169ce372b3SVitaly Kuznetsov if (!flags) { 4617faa03b39SWonhyuk Yang trace_kvm_page_fault(vcpu, fault_address, error_code); 4618c50d8ae3SPaolo Bonzini 4619c50d8ae3SPaolo Bonzini if (kvm_event_needs_reinjection(vcpu)) 4620c50d8ae3SPaolo Bonzini kvm_mmu_unprotect_page_virt(vcpu, fault_address); 4621c50d8ae3SPaolo Bonzini r = kvm_mmu_page_fault(vcpu, fault_address, error_code, insn, 4622c50d8ae3SPaolo Bonzini insn_len); 46239ce372b3SVitaly Kuznetsov } else if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) { 462468fd66f1SVitaly Kuznetsov vcpu->arch.apf.host_apf_flags = 0; 4625c50d8ae3SPaolo Bonzini local_irq_disable(); 46266bca69adSThomas Gleixner kvm_async_pf_task_wait_schedule(fault_address); 4627c50d8ae3SPaolo Bonzini local_irq_enable(); 46289ce372b3SVitaly Kuznetsov } else { 46299ce372b3SVitaly Kuznetsov WARN_ONCE(1, "Unexpected host async PF flags: %x\n", flags); 4630c50d8ae3SPaolo Bonzini } 46319ce372b3SVitaly Kuznetsov 4632c50d8ae3SPaolo Bonzini return r; 4633c50d8ae3SPaolo Bonzini } 4634c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_handle_page_fault); 4635c50d8ae3SPaolo Bonzini 46369aa8ab43SDavid Matlack #ifdef CONFIG_X86_64 46379aa8ab43SDavid Matlack static int kvm_tdp_mmu_page_fault(struct kvm_vcpu *vcpu, 46389aa8ab43SDavid Matlack struct kvm_page_fault *fault) 46399aa8ab43SDavid Matlack { 46409aa8ab43SDavid Matlack int r; 46419aa8ab43SDavid Matlack 46429aa8ab43SDavid Matlack if (page_fault_handle_page_track(vcpu, fault)) 46439aa8ab43SDavid Matlack return RET_PF_EMULATE; 46449aa8ab43SDavid Matlack 46459aa8ab43SDavid Matlack r = fast_page_fault(vcpu, fault); 46469aa8ab43SDavid Matlack if (r != RET_PF_INVALID) 46479aa8ab43SDavid Matlack return r; 46489aa8ab43SDavid Matlack 46499aa8ab43SDavid Matlack r = mmu_topup_memory_caches(vcpu, false); 46509aa8ab43SDavid Matlack if (r) 46519aa8ab43SDavid Matlack return r; 46529aa8ab43SDavid Matlack 46539aa8ab43SDavid Matlack r = kvm_faultin_pfn(vcpu, fault, ACC_ALL); 46549aa8ab43SDavid Matlack if (r != RET_PF_CONTINUE) 46559aa8ab43SDavid Matlack return r; 46569aa8ab43SDavid Matlack 46579aa8ab43SDavid Matlack r = RET_PF_RETRY; 46589aa8ab43SDavid Matlack read_lock(&vcpu->kvm->mmu_lock); 46599aa8ab43SDavid Matlack 46609aa8ab43SDavid Matlack if (is_page_fault_stale(vcpu, fault)) 46619aa8ab43SDavid Matlack goto out_unlock; 46629aa8ab43SDavid Matlack 46639aa8ab43SDavid Matlack r = kvm_tdp_mmu_map(vcpu, fault); 46649aa8ab43SDavid Matlack 46659aa8ab43SDavid Matlack out_unlock: 46669aa8ab43SDavid Matlack read_unlock(&vcpu->kvm->mmu_lock); 46679aa8ab43SDavid Matlack kvm_release_pfn_clean(fault->pfn); 46689aa8ab43SDavid Matlack return r; 46699aa8ab43SDavid Matlack } 46709aa8ab43SDavid Matlack #endif 46719aa8ab43SDavid Matlack 46721affe455SYan Zhao bool __kvm_mmu_honors_guest_mtrrs(bool vm_has_noncoherent_dma) 46731affe455SYan Zhao { 46741affe455SYan Zhao /* 46751affe455SYan Zhao * If host MTRRs are ignored (shadow_memtype_mask is non-zero), and the 46761affe455SYan Zhao * VM has non-coherent DMA (DMA doesn't snoop CPU caches), KVM's ABI is 46771affe455SYan Zhao * to honor the memtype from the guest's MTRRs so that guest accesses 46781affe455SYan Zhao * to memory that is DMA'd aren't cached against the guest's wishes. 46791affe455SYan Zhao * 46801affe455SYan Zhao * Note, KVM may still ultimately ignore guest MTRRs for certain PFNs, 46811affe455SYan Zhao * e.g. KVM will force UC memtype for host MMIO. 46821affe455SYan Zhao */ 46831affe455SYan Zhao return vm_has_noncoherent_dma && shadow_memtype_mask; 46841affe455SYan Zhao } 46851affe455SYan Zhao 4686c501040aSPaolo Bonzini int kvm_tdp_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 4687c50d8ae3SPaolo Bonzini { 4688d5e90a69SSean Christopherson /* 4689d5e90a69SSean Christopherson * If the guest's MTRRs may be used to compute the "real" memtype, 4690d5e90a69SSean Christopherson * restrict the mapping level to ensure KVM uses a consistent memtype 46911affe455SYan Zhao * across the entire mapping. 4692d5e90a69SSean Christopherson */ 46931affe455SYan Zhao if (kvm_mmu_honors_guest_mtrrs(vcpu->kvm)) { 4694d5e90a69SSean Christopherson for ( ; fault->max_level > PG_LEVEL_4K; --fault->max_level) { 46954326e57eSPaolo Bonzini int page_num = KVM_PAGES_PER_HPAGE(fault->max_level); 4696c667a3baSHou Wenlong gfn_t base = gfn_round_for_level(fault->gfn, 4697c667a3baSHou Wenlong fault->max_level); 4698c50d8ae3SPaolo Bonzini 4699cb9b88c6SSean Christopherson if (kvm_mtrr_check_gfn_range_consistency(vcpu, base, page_num)) 4700cb9b88c6SSean Christopherson break; 4701d5e90a69SSean Christopherson } 4702c50d8ae3SPaolo Bonzini } 4703c50d8ae3SPaolo Bonzini 47049aa8ab43SDavid Matlack #ifdef CONFIG_X86_64 47059aa8ab43SDavid Matlack if (tdp_mmu_enabled) 47069aa8ab43SDavid Matlack return kvm_tdp_mmu_page_fault(vcpu, fault); 47079aa8ab43SDavid Matlack #endif 47089aa8ab43SDavid Matlack 47094326e57eSPaolo Bonzini return direct_page_fault(vcpu, fault); 4710c50d8ae3SPaolo Bonzini } 4711c50d8ae3SPaolo Bonzini 4712*6e01b760SPaolo Bonzini static int kvm_tdp_map_page(struct kvm_vcpu *vcpu, gpa_t gpa, u64 error_code, 4713*6e01b760SPaolo Bonzini u8 *level) 4714*6e01b760SPaolo Bonzini { 4715*6e01b760SPaolo Bonzini int r; 4716*6e01b760SPaolo Bonzini 4717*6e01b760SPaolo Bonzini /* 4718*6e01b760SPaolo Bonzini * Restrict to TDP page fault, since that's the only case where the MMU 4719*6e01b760SPaolo Bonzini * is indexed by GPA. 4720*6e01b760SPaolo Bonzini */ 4721*6e01b760SPaolo Bonzini if (vcpu->arch.mmu->page_fault != kvm_tdp_page_fault) 4722*6e01b760SPaolo Bonzini return -EOPNOTSUPP; 4723*6e01b760SPaolo Bonzini 4724*6e01b760SPaolo Bonzini do { 4725*6e01b760SPaolo Bonzini if (signal_pending(current)) 4726*6e01b760SPaolo Bonzini return -EINTR; 4727*6e01b760SPaolo Bonzini cond_resched(); 4728*6e01b760SPaolo Bonzini r = kvm_mmu_do_page_fault(vcpu, gpa, error_code, true, NULL, level); 4729*6e01b760SPaolo Bonzini } while (r == RET_PF_RETRY); 4730*6e01b760SPaolo Bonzini 4731*6e01b760SPaolo Bonzini if (r < 0) 4732*6e01b760SPaolo Bonzini return r; 4733*6e01b760SPaolo Bonzini 4734*6e01b760SPaolo Bonzini switch (r) { 4735*6e01b760SPaolo Bonzini case RET_PF_FIXED: 4736*6e01b760SPaolo Bonzini case RET_PF_SPURIOUS: 4737*6e01b760SPaolo Bonzini return 0; 4738*6e01b760SPaolo Bonzini 4739*6e01b760SPaolo Bonzini case RET_PF_EMULATE: 4740*6e01b760SPaolo Bonzini return -ENOENT; 4741*6e01b760SPaolo Bonzini 4742*6e01b760SPaolo Bonzini case RET_PF_RETRY: 4743*6e01b760SPaolo Bonzini case RET_PF_CONTINUE: 4744*6e01b760SPaolo Bonzini case RET_PF_INVALID: 4745*6e01b760SPaolo Bonzini default: 4746*6e01b760SPaolo Bonzini WARN_ONCE(1, "could not fix page fault during prefault"); 4747*6e01b760SPaolo Bonzini return -EIO; 4748*6e01b760SPaolo Bonzini } 4749*6e01b760SPaolo Bonzini } 4750*6e01b760SPaolo Bonzini 4751*6e01b760SPaolo Bonzini long kvm_arch_vcpu_pre_fault_memory(struct kvm_vcpu *vcpu, 4752*6e01b760SPaolo Bonzini struct kvm_pre_fault_memory *range) 4753*6e01b760SPaolo Bonzini { 4754*6e01b760SPaolo Bonzini u64 error_code = PFERR_GUEST_FINAL_MASK; 4755*6e01b760SPaolo Bonzini u8 level = PG_LEVEL_4K; 4756*6e01b760SPaolo Bonzini u64 end; 4757*6e01b760SPaolo Bonzini int r; 4758*6e01b760SPaolo Bonzini 4759*6e01b760SPaolo Bonzini /* 4760*6e01b760SPaolo Bonzini * reload is efficient when called repeatedly, so we can do it on 4761*6e01b760SPaolo Bonzini * every iteration. 4762*6e01b760SPaolo Bonzini */ 4763*6e01b760SPaolo Bonzini kvm_mmu_reload(vcpu); 4764*6e01b760SPaolo Bonzini 4765*6e01b760SPaolo Bonzini if (kvm_arch_has_private_mem(vcpu->kvm) && 4766*6e01b760SPaolo Bonzini kvm_mem_is_private(vcpu->kvm, gpa_to_gfn(range->gpa))) 4767*6e01b760SPaolo Bonzini error_code |= PFERR_PRIVATE_ACCESS; 4768*6e01b760SPaolo Bonzini 4769*6e01b760SPaolo Bonzini /* 4770*6e01b760SPaolo Bonzini * Shadow paging uses GVA for kvm page fault, so restrict to 4771*6e01b760SPaolo Bonzini * two-dimensional paging. 4772*6e01b760SPaolo Bonzini */ 4773*6e01b760SPaolo Bonzini r = kvm_tdp_map_page(vcpu, range->gpa, error_code, &level); 4774*6e01b760SPaolo Bonzini if (r < 0) 4775*6e01b760SPaolo Bonzini return r; 4776*6e01b760SPaolo Bonzini 4777*6e01b760SPaolo Bonzini /* 4778*6e01b760SPaolo Bonzini * If the mapping that covers range->gpa can use a huge page, it 4779*6e01b760SPaolo Bonzini * may start below it or end after range->gpa + range->size. 4780*6e01b760SPaolo Bonzini */ 4781*6e01b760SPaolo Bonzini end = (range->gpa & KVM_HPAGE_MASK(level)) + KVM_HPAGE_SIZE(level); 4782*6e01b760SPaolo Bonzini return min(range->size, end - range->gpa); 4783*6e01b760SPaolo Bonzini } 4784*6e01b760SPaolo Bonzini 478584a16226SSean Christopherson static void nonpaging_init_context(struct kvm_mmu *context) 4786c50d8ae3SPaolo Bonzini { 4787c50d8ae3SPaolo Bonzini context->page_fault = nonpaging_page_fault; 4788c50d8ae3SPaolo Bonzini context->gva_to_gpa = nonpaging_gva_to_gpa; 4789c3c6c9fcSLai Jiangshan context->sync_spte = NULL; 4790c50d8ae3SPaolo Bonzini } 4791c50d8ae3SPaolo Bonzini 4792be01e8e2SSean Christopherson static inline bool is_root_usable(struct kvm_mmu_root_info *root, gpa_t pgd, 47930be44352SSean Christopherson union kvm_mmu_page_role role) 47940be44352SSean Christopherson { 4795c30e000eSSean Christopherson struct kvm_mmu_page *sp; 4796c30e000eSSean Christopherson 4797c30e000eSSean Christopherson if (!VALID_PAGE(root->hpa)) 4798c30e000eSSean Christopherson return false; 4799c30e000eSSean Christopherson 4800c30e000eSSean Christopherson if (!role.direct && pgd != root->pgd) 4801c30e000eSSean Christopherson return false; 4802c30e000eSSean Christopherson 4803c30e000eSSean Christopherson sp = root_to_sp(root->hpa); 4804c30e000eSSean Christopherson if (WARN_ON_ONCE(!sp)) 4805c30e000eSSean Christopherson return false; 4806c30e000eSSean Christopherson 4807c30e000eSSean Christopherson return role.word == sp->role.word; 48080be44352SSean Christopherson } 48090be44352SSean Christopherson 4810c50d8ae3SPaolo Bonzini /* 48115499ea73SPaolo Bonzini * Find out if a previously cached root matching the new pgd/role is available, 48125499ea73SPaolo Bonzini * and insert the current root as the MRU in the cache. 48135499ea73SPaolo Bonzini * If a matching root is found, it is assigned to kvm_mmu->root and 48145499ea73SPaolo Bonzini * true is returned. 48155499ea73SPaolo Bonzini * If no match is found, kvm_mmu->root is left invalid, the LRU root is 48165499ea73SPaolo Bonzini * evicted to make room for the current root, and false is returned. 4817c50d8ae3SPaolo Bonzini */ 48185499ea73SPaolo Bonzini static bool cached_root_find_and_keep_current(struct kvm *kvm, struct kvm_mmu *mmu, 48195499ea73SPaolo Bonzini gpa_t new_pgd, 4820c50d8ae3SPaolo Bonzini union kvm_mmu_page_role new_role) 4821c50d8ae3SPaolo Bonzini { 4822c50d8ae3SPaolo Bonzini uint i; 4823c50d8ae3SPaolo Bonzini 4824b9e5603cSPaolo Bonzini if (is_root_usable(&mmu->root, new_pgd, new_role)) 48250be44352SSean Christopherson return true; 48260be44352SSean Christopherson 4827c50d8ae3SPaolo Bonzini for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) { 48285499ea73SPaolo Bonzini /* 48295499ea73SPaolo Bonzini * The swaps end up rotating the cache like this: 48305499ea73SPaolo Bonzini * C 0 1 2 3 (on entry to the function) 48315499ea73SPaolo Bonzini * 0 C 1 2 3 48325499ea73SPaolo Bonzini * 1 C 0 2 3 48335499ea73SPaolo Bonzini * 2 C 0 1 3 48345499ea73SPaolo Bonzini * 3 C 0 1 2 (on exit from the loop) 48355499ea73SPaolo Bonzini */ 4836b9e5603cSPaolo Bonzini swap(mmu->root, mmu->prev_roots[i]); 4837b9e5603cSPaolo Bonzini if (is_root_usable(&mmu->root, new_pgd, new_role)) 48385499ea73SPaolo Bonzini return true; 4839c50d8ae3SPaolo Bonzini } 4840c50d8ae3SPaolo Bonzini 48415499ea73SPaolo Bonzini kvm_mmu_free_roots(kvm, mmu, KVM_MMU_ROOT_CURRENT); 48425499ea73SPaolo Bonzini return false; 4843c50d8ae3SPaolo Bonzini } 4844c50d8ae3SPaolo Bonzini 4845c50d8ae3SPaolo Bonzini /* 48465499ea73SPaolo Bonzini * Find out if a previously cached root matching the new pgd/role is available. 48475499ea73SPaolo Bonzini * On entry, mmu->root is invalid. 48485499ea73SPaolo Bonzini * If a matching root is found, it is assigned to kvm_mmu->root, the LRU entry 48495499ea73SPaolo Bonzini * of the cache becomes invalid, and true is returned. 48505499ea73SPaolo Bonzini * If no match is found, kvm_mmu->root is left invalid and false is returned. 48515499ea73SPaolo Bonzini */ 48525499ea73SPaolo Bonzini static bool cached_root_find_without_current(struct kvm *kvm, struct kvm_mmu *mmu, 48535499ea73SPaolo Bonzini gpa_t new_pgd, 48545499ea73SPaolo Bonzini union kvm_mmu_page_role new_role) 48555499ea73SPaolo Bonzini { 48565499ea73SPaolo Bonzini uint i; 48575499ea73SPaolo Bonzini 48585499ea73SPaolo Bonzini for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) 48595499ea73SPaolo Bonzini if (is_root_usable(&mmu->prev_roots[i], new_pgd, new_role)) 48605499ea73SPaolo Bonzini goto hit; 48615499ea73SPaolo Bonzini 48625499ea73SPaolo Bonzini return false; 48635499ea73SPaolo Bonzini 48645499ea73SPaolo Bonzini hit: 48655499ea73SPaolo Bonzini swap(mmu->root, mmu->prev_roots[i]); 48665499ea73SPaolo Bonzini /* Bubble up the remaining roots. */ 48675499ea73SPaolo Bonzini for (; i < KVM_MMU_NUM_PREV_ROOTS - 1; i++) 48685499ea73SPaolo Bonzini mmu->prev_roots[i] = mmu->prev_roots[i + 1]; 48695499ea73SPaolo Bonzini mmu->prev_roots[i].hpa = INVALID_PAGE; 48705499ea73SPaolo Bonzini return true; 48715499ea73SPaolo Bonzini } 48725499ea73SPaolo Bonzini 48735499ea73SPaolo Bonzini static bool fast_pgd_switch(struct kvm *kvm, struct kvm_mmu *mmu, 48745499ea73SPaolo Bonzini gpa_t new_pgd, union kvm_mmu_page_role new_role) 48755499ea73SPaolo Bonzini { 48765499ea73SPaolo Bonzini /* 48770e3223d8SSean Christopherson * Limit reuse to 64-bit hosts+VMs without "special" roots in order to 48780e3223d8SSean Christopherson * avoid having to deal with PDPTEs and other complexities. 4879c50d8ae3SPaolo Bonzini */ 4880c5f2d564SSean Christopherson if (VALID_PAGE(mmu->root.hpa) && !root_to_sp(mmu->root.hpa)) 48815499ea73SPaolo Bonzini kvm_mmu_free_roots(kvm, mmu, KVM_MMU_ROOT_CURRENT); 4882c50d8ae3SPaolo Bonzini 48835499ea73SPaolo Bonzini if (VALID_PAGE(mmu->root.hpa)) 48845499ea73SPaolo Bonzini return cached_root_find_and_keep_current(kvm, mmu, new_pgd, new_role); 48855499ea73SPaolo Bonzini else 48865499ea73SPaolo Bonzini return cached_root_find_without_current(kvm, mmu, new_pgd, new_role); 4887c50d8ae3SPaolo Bonzini } 4888c50d8ae3SPaolo Bonzini 4889d2e5f333SPaolo Bonzini void kvm_mmu_new_pgd(struct kvm_vcpu *vcpu, gpa_t new_pgd) 4890c50d8ae3SPaolo Bonzini { 48910c1c92f1SPaolo Bonzini struct kvm_mmu *mmu = vcpu->arch.mmu; 48927a458f0eSPaolo Bonzini union kvm_mmu_page_role new_role = mmu->root_role; 48930c1c92f1SPaolo Bonzini 4894a7e48ef7SWei Liu /* 4895a7e48ef7SWei Liu * Return immediately if no usable root was found, kvm_mmu_reload() 4896a7e48ef7SWei Liu * will establish a valid root prior to the next VM-Enter. 4897a7e48ef7SWei Liu */ 4898a7e48ef7SWei Liu if (!fast_pgd_switch(vcpu->kvm, mmu, new_pgd, new_role)) 4899b869855bSSean Christopherson return; 4900c50d8ae3SPaolo Bonzini 4901c50d8ae3SPaolo Bonzini /* 4902b869855bSSean Christopherson * It's possible that the cached previous root page is obsolete because 4903b869855bSSean Christopherson * of a change in the MMU generation number. However, changing the 4904527d5cd7SSean Christopherson * generation number is accompanied by KVM_REQ_MMU_FREE_OBSOLETE_ROOTS, 4905527d5cd7SSean Christopherson * which will free the root set here and allocate a new one. 4906b869855bSSean Christopherson */ 4907b869855bSSean Christopherson kvm_make_request(KVM_REQ_LOAD_MMU_PGD, vcpu); 4908b869855bSSean Christopherson 4909b5129100SSean Christopherson if (force_flush_and_sync_on_reuse) { 4910b869855bSSean Christopherson kvm_make_request(KVM_REQ_MMU_SYNC, vcpu); 4911b869855bSSean Christopherson kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu); 4912b5129100SSean Christopherson } 4913b869855bSSean Christopherson 4914b869855bSSean Christopherson /* 4915b869855bSSean Christopherson * The last MMIO access's GVA and GPA are cached in the VCPU. When 4916b869855bSSean Christopherson * switching to a new CR3, that GVA->GPA mapping may no longer be 4917b869855bSSean Christopherson * valid. So clear any cached MMIO info even when we don't need to sync 4918b869855bSSean Christopherson * the shadow page tables. 4919c50d8ae3SPaolo Bonzini */ 4920c50d8ae3SPaolo Bonzini vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY); 4921c50d8ae3SPaolo Bonzini 4922daa5b6c1SBen Gardon /* 4923daa5b6c1SBen Gardon * If this is a direct root page, it doesn't have a write flooding 4924daa5b6c1SBen Gardon * count. Otherwise, clear the write flooding count. 4925daa5b6c1SBen Gardon */ 4926c30e000eSSean Christopherson if (!new_role.direct) { 4927c30e000eSSean Christopherson struct kvm_mmu_page *sp = root_to_sp(vcpu->arch.mmu->root.hpa); 4928c30e000eSSean Christopherson 4929c30e000eSSean Christopherson if (!WARN_ON_ONCE(!sp)) 4930c30e000eSSean Christopherson __clear_sp_write_flooding_count(sp); 4931c30e000eSSean Christopherson } 4932c50d8ae3SPaolo Bonzini } 4933be01e8e2SSean Christopherson EXPORT_SYMBOL_GPL(kvm_mmu_new_pgd); 4934c50d8ae3SPaolo Bonzini 4935c50d8ae3SPaolo Bonzini static bool sync_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn, 4936c3e5e415SLai Jiangshan unsigned int access) 4937c50d8ae3SPaolo Bonzini { 4938949019b9SSean Christopherson if (unlikely(is_mmio_spte(vcpu->kvm, *sptep))) { 4939c50d8ae3SPaolo Bonzini if (gfn != get_mmio_spte_gfn(*sptep)) { 4940c50d8ae3SPaolo Bonzini mmu_spte_clear_no_track(sptep); 4941c50d8ae3SPaolo Bonzini return true; 4942c50d8ae3SPaolo Bonzini } 4943c50d8ae3SPaolo Bonzini 4944c50d8ae3SPaolo Bonzini mark_mmio_spte(vcpu, sptep, gfn, access); 4945c50d8ae3SPaolo Bonzini return true; 4946c50d8ae3SPaolo Bonzini } 4947c50d8ae3SPaolo Bonzini 4948c50d8ae3SPaolo Bonzini return false; 4949c50d8ae3SPaolo Bonzini } 4950c50d8ae3SPaolo Bonzini 4951c50d8ae3SPaolo Bonzini #define PTTYPE_EPT 18 /* arbitrary */ 4952c50d8ae3SPaolo Bonzini #define PTTYPE PTTYPE_EPT 4953c50d8ae3SPaolo Bonzini #include "paging_tmpl.h" 4954c50d8ae3SPaolo Bonzini #undef PTTYPE 4955c50d8ae3SPaolo Bonzini 4956c50d8ae3SPaolo Bonzini #define PTTYPE 64 4957c50d8ae3SPaolo Bonzini #include "paging_tmpl.h" 4958c50d8ae3SPaolo Bonzini #undef PTTYPE 4959c50d8ae3SPaolo Bonzini 4960c50d8ae3SPaolo Bonzini #define PTTYPE 32 4961c50d8ae3SPaolo Bonzini #include "paging_tmpl.h" 4962c50d8ae3SPaolo Bonzini #undef PTTYPE 4963c50d8ae3SPaolo Bonzini 4964f3d90f90SSean Christopherson static void __reset_rsvds_bits_mask(struct rsvd_bits_validate *rsvd_check, 4965f3d90f90SSean Christopherson u64 pa_bits_rsvd, int level, bool nx, 4966f3d90f90SSean Christopherson bool gbpages, bool pse, bool amd) 4967c50d8ae3SPaolo Bonzini { 4968c50d8ae3SPaolo Bonzini u64 gbpages_bit_rsvd = 0; 4969c50d8ae3SPaolo Bonzini u64 nonleaf_bit8_rsvd = 0; 49705b7f575cSSean Christopherson u64 high_bits_rsvd; 4971c50d8ae3SPaolo Bonzini 4972c50d8ae3SPaolo Bonzini rsvd_check->bad_mt_xwr = 0; 4973c50d8ae3SPaolo Bonzini 4974c50d8ae3SPaolo Bonzini if (!gbpages) 4975c50d8ae3SPaolo Bonzini gbpages_bit_rsvd = rsvd_bits(7, 7); 4976c50d8ae3SPaolo Bonzini 49775b7f575cSSean Christopherson if (level == PT32E_ROOT_LEVEL) 49785b7f575cSSean Christopherson high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 62); 49795b7f575cSSean Christopherson else 49805b7f575cSSean Christopherson high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 51); 49815b7f575cSSean Christopherson 49825b7f575cSSean Christopherson /* Note, NX doesn't exist in PDPTEs, this is handled below. */ 49835b7f575cSSean Christopherson if (!nx) 49845b7f575cSSean Christopherson high_bits_rsvd |= rsvd_bits(63, 63); 49855b7f575cSSean Christopherson 4986c50d8ae3SPaolo Bonzini /* 4987c50d8ae3SPaolo Bonzini * Non-leaf PML4Es and PDPEs reserve bit 8 (which would be the G bit for 4988c50d8ae3SPaolo Bonzini * leaf entries) on AMD CPUs only. 4989c50d8ae3SPaolo Bonzini */ 4990c50d8ae3SPaolo Bonzini if (amd) 4991c50d8ae3SPaolo Bonzini nonleaf_bit8_rsvd = rsvd_bits(8, 8); 4992c50d8ae3SPaolo Bonzini 4993c50d8ae3SPaolo Bonzini switch (level) { 4994c50d8ae3SPaolo Bonzini case PT32_ROOT_LEVEL: 4995c50d8ae3SPaolo Bonzini /* no rsvd bits for 2 level 4K page table entries */ 4996c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][1] = 0; 4997c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][0] = 0; 4998c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][0] = 4999c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][0]; 5000c50d8ae3SPaolo Bonzini 5001c50d8ae3SPaolo Bonzini if (!pse) { 5002c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][1] = 0; 5003c50d8ae3SPaolo Bonzini break; 5004c50d8ae3SPaolo Bonzini } 5005c50d8ae3SPaolo Bonzini 5006c50d8ae3SPaolo Bonzini if (is_cpuid_PSE36()) 5007c50d8ae3SPaolo Bonzini /* 36bits PSE 4MB page */ 5008c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(17, 21); 5009c50d8ae3SPaolo Bonzini else 5010c50d8ae3SPaolo Bonzini /* 32 bits PSE 4MB page */ 5011c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(13, 21); 5012c50d8ae3SPaolo Bonzini break; 5013c50d8ae3SPaolo Bonzini case PT32E_ROOT_LEVEL: 50145b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][2] = rsvd_bits(63, 63) | 50155b7f575cSSean Christopherson high_bits_rsvd | 50165b7f575cSSean Christopherson rsvd_bits(5, 8) | 50175b7f575cSSean Christopherson rsvd_bits(1, 2); /* PDPTE */ 50185b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd; /* PDE */ 50195b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd; /* PTE */ 50205b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd | 5021c50d8ae3SPaolo Bonzini rsvd_bits(13, 20); /* large page */ 5022c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][0] = 5023c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][0]; 5024c50d8ae3SPaolo Bonzini break; 5025c50d8ae3SPaolo Bonzini case PT64_ROOT_5LEVEL: 50265b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][4] = high_bits_rsvd | 50275b7f575cSSean Christopherson nonleaf_bit8_rsvd | 50285b7f575cSSean Christopherson rsvd_bits(7, 7); 5029c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][4] = 5030c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][4]; 5031df561f66SGustavo A. R. Silva fallthrough; 5032c50d8ae3SPaolo Bonzini case PT64_ROOT_4LEVEL: 50335b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][3] = high_bits_rsvd | 50345b7f575cSSean Christopherson nonleaf_bit8_rsvd | 50355b7f575cSSean Christopherson rsvd_bits(7, 7); 50365b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][2] = high_bits_rsvd | 50375b7f575cSSean Christopherson gbpages_bit_rsvd; 50385b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd; 50395b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd; 5040c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][3] = 5041c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][3]; 50425b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[1][2] = high_bits_rsvd | 50435b7f575cSSean Christopherson gbpages_bit_rsvd | 5044c50d8ae3SPaolo Bonzini rsvd_bits(13, 29); 50455b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd | 5046c50d8ae3SPaolo Bonzini rsvd_bits(13, 20); /* large page */ 5047c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][0] = 5048c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][0]; 5049c50d8ae3SPaolo Bonzini break; 5050c50d8ae3SPaolo Bonzini } 5051c50d8ae3SPaolo Bonzini } 5052c50d8ae3SPaolo Bonzini 5053c919e881SKai Huang static void reset_guest_rsvds_bits_mask(struct kvm_vcpu *vcpu, 5054c50d8ae3SPaolo Bonzini struct kvm_mmu *context) 5055c50d8ae3SPaolo Bonzini { 5056b705a277SSean Christopherson __reset_rsvds_bits_mask(&context->guest_rsvd_check, 50575b7f575cSSean Christopherson vcpu->arch.reserved_gpa_bits, 50584d25502aSPaolo Bonzini context->cpu_role.base.level, is_efer_nx(context), 5059ccf31d6eSSean Christopherson guest_can_use(vcpu, X86_FEATURE_GBPAGES), 50604e9c0d80SSean Christopherson is_cr4_pse(context), 5061fd706c9bSSean Christopherson guest_cpuid_is_amd_compatible(vcpu)); 5062c50d8ae3SPaolo Bonzini } 5063c50d8ae3SPaolo Bonzini 5064f3d90f90SSean Christopherson static void __reset_rsvds_bits_mask_ept(struct rsvd_bits_validate *rsvd_check, 5065f3d90f90SSean Christopherson u64 pa_bits_rsvd, bool execonly, 5066f3d90f90SSean Christopherson int huge_page_level) 5067c50d8ae3SPaolo Bonzini { 50685b7f575cSSean Christopherson u64 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 51); 506984ea5c09SLai Jiangshan u64 large_1g_rsvd = 0, large_2m_rsvd = 0; 5070c50d8ae3SPaolo Bonzini u64 bad_mt_xwr; 5071c50d8ae3SPaolo Bonzini 507284ea5c09SLai Jiangshan if (huge_page_level < PG_LEVEL_1G) 507384ea5c09SLai Jiangshan large_1g_rsvd = rsvd_bits(7, 7); 507484ea5c09SLai Jiangshan if (huge_page_level < PG_LEVEL_2M) 507584ea5c09SLai Jiangshan large_2m_rsvd = rsvd_bits(7, 7); 507684ea5c09SLai Jiangshan 50775b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][4] = high_bits_rsvd | rsvd_bits(3, 7); 50785b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][3] = high_bits_rsvd | rsvd_bits(3, 7); 507984ea5c09SLai Jiangshan rsvd_check->rsvd_bits_mask[0][2] = high_bits_rsvd | rsvd_bits(3, 6) | large_1g_rsvd; 508084ea5c09SLai Jiangshan rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd | rsvd_bits(3, 6) | large_2m_rsvd; 50815b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd; 5082c50d8ae3SPaolo Bonzini 5083c50d8ae3SPaolo Bonzini /* large page */ 5084c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][4] = rsvd_check->rsvd_bits_mask[0][4]; 5085c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][3] = rsvd_check->rsvd_bits_mask[0][3]; 508684ea5c09SLai Jiangshan rsvd_check->rsvd_bits_mask[1][2] = high_bits_rsvd | rsvd_bits(12, 29) | large_1g_rsvd; 508784ea5c09SLai Jiangshan rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd | rsvd_bits(12, 20) | large_2m_rsvd; 5088c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][0] = rsvd_check->rsvd_bits_mask[0][0]; 5089c50d8ae3SPaolo Bonzini 5090c50d8ae3SPaolo Bonzini bad_mt_xwr = 0xFFull << (2 * 8); /* bits 3..5 must not be 2 */ 5091c50d8ae3SPaolo Bonzini bad_mt_xwr |= 0xFFull << (3 * 8); /* bits 3..5 must not be 3 */ 5092c50d8ae3SPaolo Bonzini bad_mt_xwr |= 0xFFull << (7 * 8); /* bits 3..5 must not be 7 */ 5093c50d8ae3SPaolo Bonzini bad_mt_xwr |= REPEAT_BYTE(1ull << 2); /* bits 0..2 must not be 010 */ 5094c50d8ae3SPaolo Bonzini bad_mt_xwr |= REPEAT_BYTE(1ull << 6); /* bits 0..2 must not be 110 */ 5095c50d8ae3SPaolo Bonzini if (!execonly) { 5096c50d8ae3SPaolo Bonzini /* bits 0..2 must not be 100 unless VMX capabilities allow it */ 5097c50d8ae3SPaolo Bonzini bad_mt_xwr |= REPEAT_BYTE(1ull << 4); 5098c50d8ae3SPaolo Bonzini } 5099c50d8ae3SPaolo Bonzini rsvd_check->bad_mt_xwr = bad_mt_xwr; 5100c50d8ae3SPaolo Bonzini } 5101c50d8ae3SPaolo Bonzini 5102c50d8ae3SPaolo Bonzini static void reset_rsvds_bits_mask_ept(struct kvm_vcpu *vcpu, 510384ea5c09SLai Jiangshan struct kvm_mmu *context, bool execonly, int huge_page_level) 5104c50d8ae3SPaolo Bonzini { 5105c50d8ae3SPaolo Bonzini __reset_rsvds_bits_mask_ept(&context->guest_rsvd_check, 510684ea5c09SLai Jiangshan vcpu->arch.reserved_gpa_bits, execonly, 510784ea5c09SLai Jiangshan huge_page_level); 5108c50d8ae3SPaolo Bonzini } 5109c50d8ae3SPaolo Bonzini 51106f8e65a6SSean Christopherson static inline u64 reserved_hpa_bits(void) 51116f8e65a6SSean Christopherson { 51126f8e65a6SSean Christopherson return rsvd_bits(shadow_phys_bits, 63); 51136f8e65a6SSean Christopherson } 51146f8e65a6SSean Christopherson 5115c50d8ae3SPaolo Bonzini /* 5116c50d8ae3SPaolo Bonzini * the page table on host is the shadow page table for the page 5117c50d8ae3SPaolo Bonzini * table in guest or amd nested guest, its mmu features completely 5118c50d8ae3SPaolo Bonzini * follow the features in guest. 5119c50d8ae3SPaolo Bonzini */ 512016be1d12SSean Christopherson static void reset_shadow_zero_bits_mask(struct kvm_vcpu *vcpu, 512116be1d12SSean Christopherson struct kvm_mmu *context) 5122c50d8ae3SPaolo Bonzini { 51238c985b2dSSean Christopherson /* @amd adds a check on bit of SPTEs, which KVM shouldn't use anyways. */ 51248c985b2dSSean Christopherson bool is_amd = true; 51258c985b2dSSean Christopherson /* KVM doesn't use 2-level page tables for the shadow MMU. */ 51268c985b2dSSean Christopherson bool is_pse = false; 5127c50d8ae3SPaolo Bonzini struct rsvd_bits_validate *shadow_zero_check; 5128c50d8ae3SPaolo Bonzini int i; 5129c50d8ae3SPaolo Bonzini 5130a972e29cSPaolo Bonzini WARN_ON_ONCE(context->root_role.level < PT32E_ROOT_LEVEL); 51318c985b2dSSean Christopherson 5132c50d8ae3SPaolo Bonzini shadow_zero_check = &context->shadow_zero_check; 5133b705a277SSean Christopherson __reset_rsvds_bits_mask(shadow_zero_check, reserved_hpa_bits(), 5134a972e29cSPaolo Bonzini context->root_role.level, 51357a458f0eSPaolo Bonzini context->root_role.efer_nx, 5136ccf31d6eSSean Christopherson guest_can_use(vcpu, X86_FEATURE_GBPAGES), 5137ccf31d6eSSean Christopherson is_pse, is_amd); 5138c50d8ae3SPaolo Bonzini 5139c50d8ae3SPaolo Bonzini if (!shadow_me_mask) 5140c50d8ae3SPaolo Bonzini return; 5141c50d8ae3SPaolo Bonzini 5142a972e29cSPaolo Bonzini for (i = context->root_role.level; --i >= 0;) { 5143e54f1ff2SKai Huang /* 5144e54f1ff2SKai Huang * So far shadow_me_value is a constant during KVM's life 5145e54f1ff2SKai Huang * time. Bits in shadow_me_value are allowed to be set. 5146e54f1ff2SKai Huang * Bits in shadow_me_mask but not in shadow_me_value are 5147e54f1ff2SKai Huang * not allowed to be set. 5148e54f1ff2SKai Huang */ 5149e54f1ff2SKai Huang shadow_zero_check->rsvd_bits_mask[0][i] |= shadow_me_mask; 5150e54f1ff2SKai Huang shadow_zero_check->rsvd_bits_mask[1][i] |= shadow_me_mask; 5151e54f1ff2SKai Huang shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_value; 5152e54f1ff2SKai Huang shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_value; 5153c50d8ae3SPaolo Bonzini } 5154c50d8ae3SPaolo Bonzini 5155c50d8ae3SPaolo Bonzini } 5156c50d8ae3SPaolo Bonzini 5157c50d8ae3SPaolo Bonzini static inline bool boot_cpu_is_amd(void) 5158c50d8ae3SPaolo Bonzini { 5159c50d8ae3SPaolo Bonzini WARN_ON_ONCE(!tdp_enabled); 5160c50d8ae3SPaolo Bonzini return shadow_x_mask == 0; 5161c50d8ae3SPaolo Bonzini } 5162c50d8ae3SPaolo Bonzini 5163c50d8ae3SPaolo Bonzini /* 5164c50d8ae3SPaolo Bonzini * the direct page table on host, use as much mmu features as 5165c50d8ae3SPaolo Bonzini * possible, however, kvm currently does not do execution-protection. 5166c50d8ae3SPaolo Bonzini */ 5167f3d90f90SSean Christopherson static void reset_tdp_shadow_zero_bits_mask(struct kvm_mmu *context) 5168c50d8ae3SPaolo Bonzini { 5169c50d8ae3SPaolo Bonzini struct rsvd_bits_validate *shadow_zero_check; 5170c50d8ae3SPaolo Bonzini int i; 5171c50d8ae3SPaolo Bonzini 5172c50d8ae3SPaolo Bonzini shadow_zero_check = &context->shadow_zero_check; 5173c50d8ae3SPaolo Bonzini 5174c50d8ae3SPaolo Bonzini if (boot_cpu_is_amd()) 5175b705a277SSean Christopherson __reset_rsvds_bits_mask(shadow_zero_check, reserved_hpa_bits(), 51766c6ab524SSean Christopherson context->root_role.level, true, 5177c50d8ae3SPaolo Bonzini boot_cpu_has(X86_FEATURE_GBPAGES), 51788c985b2dSSean Christopherson false, true); 5179c50d8ae3SPaolo Bonzini else 5180c50d8ae3SPaolo Bonzini __reset_rsvds_bits_mask_ept(shadow_zero_check, 518184ea5c09SLai Jiangshan reserved_hpa_bits(), false, 518284ea5c09SLai Jiangshan max_huge_page_level); 5183c50d8ae3SPaolo Bonzini 5184c50d8ae3SPaolo Bonzini if (!shadow_me_mask) 5185c50d8ae3SPaolo Bonzini return; 5186c50d8ae3SPaolo Bonzini 5187a972e29cSPaolo Bonzini for (i = context->root_role.level; --i >= 0;) { 5188c50d8ae3SPaolo Bonzini shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask; 5189c50d8ae3SPaolo Bonzini shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask; 5190c50d8ae3SPaolo Bonzini } 5191c50d8ae3SPaolo Bonzini } 5192c50d8ae3SPaolo Bonzini 5193c50d8ae3SPaolo Bonzini /* 5194c50d8ae3SPaolo Bonzini * as the comments in reset_shadow_zero_bits_mask() except it 5195c50d8ae3SPaolo Bonzini * is the shadow page table for intel nested guest. 5196c50d8ae3SPaolo Bonzini */ 5197c50d8ae3SPaolo Bonzini static void 5198e8f6e738SJinrong Liang reset_ept_shadow_zero_bits_mask(struct kvm_mmu *context, bool execonly) 5199c50d8ae3SPaolo Bonzini { 5200c50d8ae3SPaolo Bonzini __reset_rsvds_bits_mask_ept(&context->shadow_zero_check, 520184ea5c09SLai Jiangshan reserved_hpa_bits(), execonly, 520284ea5c09SLai Jiangshan max_huge_page_level); 5203c50d8ae3SPaolo Bonzini } 5204c50d8ae3SPaolo Bonzini 5205c50d8ae3SPaolo Bonzini #define BYTE_MASK(access) \ 5206c50d8ae3SPaolo Bonzini ((1 & (access) ? 2 : 0) | \ 5207c50d8ae3SPaolo Bonzini (2 & (access) ? 4 : 0) | \ 5208c50d8ae3SPaolo Bonzini (3 & (access) ? 8 : 0) | \ 5209c50d8ae3SPaolo Bonzini (4 & (access) ? 16 : 0) | \ 5210c50d8ae3SPaolo Bonzini (5 & (access) ? 32 : 0) | \ 5211c50d8ae3SPaolo Bonzini (6 & (access) ? 64 : 0) | \ 5212c50d8ae3SPaolo Bonzini (7 & (access) ? 128 : 0)) 5213c50d8ae3SPaolo Bonzini 5214c50d8ae3SPaolo Bonzini 5215c596f147SSean Christopherson static void update_permission_bitmask(struct kvm_mmu *mmu, bool ept) 5216c50d8ae3SPaolo Bonzini { 5217c50d8ae3SPaolo Bonzini unsigned byte; 5218c50d8ae3SPaolo Bonzini 5219c50d8ae3SPaolo Bonzini const u8 x = BYTE_MASK(ACC_EXEC_MASK); 5220c50d8ae3SPaolo Bonzini const u8 w = BYTE_MASK(ACC_WRITE_MASK); 5221c50d8ae3SPaolo Bonzini const u8 u = BYTE_MASK(ACC_USER_MASK); 5222c50d8ae3SPaolo Bonzini 5223c596f147SSean Christopherson bool cr4_smep = is_cr4_smep(mmu); 5224c596f147SSean Christopherson bool cr4_smap = is_cr4_smap(mmu); 5225c596f147SSean Christopherson bool cr0_wp = is_cr0_wp(mmu); 522690599c28SSean Christopherson bool efer_nx = is_efer_nx(mmu); 5227c50d8ae3SPaolo Bonzini 5228c50d8ae3SPaolo Bonzini for (byte = 0; byte < ARRAY_SIZE(mmu->permissions); ++byte) { 5229c50d8ae3SPaolo Bonzini unsigned pfec = byte << 1; 5230c50d8ae3SPaolo Bonzini 5231c50d8ae3SPaolo Bonzini /* 5232c50d8ae3SPaolo Bonzini * Each "*f" variable has a 1 bit for each UWX value 5233c50d8ae3SPaolo Bonzini * that causes a fault with the given PFEC. 5234c50d8ae3SPaolo Bonzini */ 5235c50d8ae3SPaolo Bonzini 5236c50d8ae3SPaolo Bonzini /* Faults from writes to non-writable pages */ 5237c50d8ae3SPaolo Bonzini u8 wf = (pfec & PFERR_WRITE_MASK) ? (u8)~w : 0; 5238c50d8ae3SPaolo Bonzini /* Faults from user mode accesses to supervisor pages */ 5239c50d8ae3SPaolo Bonzini u8 uf = (pfec & PFERR_USER_MASK) ? (u8)~u : 0; 5240c50d8ae3SPaolo Bonzini /* Faults from fetches of non-executable pages*/ 5241c50d8ae3SPaolo Bonzini u8 ff = (pfec & PFERR_FETCH_MASK) ? (u8)~x : 0; 5242c50d8ae3SPaolo Bonzini /* Faults from kernel mode fetches of user pages */ 5243c50d8ae3SPaolo Bonzini u8 smepf = 0; 5244c50d8ae3SPaolo Bonzini /* Faults from kernel mode accesses of user pages */ 5245c50d8ae3SPaolo Bonzini u8 smapf = 0; 5246c50d8ae3SPaolo Bonzini 5247c50d8ae3SPaolo Bonzini if (!ept) { 5248c50d8ae3SPaolo Bonzini /* Faults from kernel mode accesses to user pages */ 5249c50d8ae3SPaolo Bonzini u8 kf = (pfec & PFERR_USER_MASK) ? 0 : u; 5250c50d8ae3SPaolo Bonzini 5251c50d8ae3SPaolo Bonzini /* Not really needed: !nx will cause pte.nx to fault */ 525290599c28SSean Christopherson if (!efer_nx) 5253c50d8ae3SPaolo Bonzini ff = 0; 5254c50d8ae3SPaolo Bonzini 5255c50d8ae3SPaolo Bonzini /* Allow supervisor writes if !cr0.wp */ 5256c50d8ae3SPaolo Bonzini if (!cr0_wp) 5257c50d8ae3SPaolo Bonzini wf = (pfec & PFERR_USER_MASK) ? wf : 0; 5258c50d8ae3SPaolo Bonzini 5259c50d8ae3SPaolo Bonzini /* Disallow supervisor fetches of user code if cr4.smep */ 5260c50d8ae3SPaolo Bonzini if (cr4_smep) 5261c50d8ae3SPaolo Bonzini smepf = (pfec & PFERR_FETCH_MASK) ? kf : 0; 5262c50d8ae3SPaolo Bonzini 5263c50d8ae3SPaolo Bonzini /* 5264c50d8ae3SPaolo Bonzini * SMAP:kernel-mode data accesses from user-mode 5265c50d8ae3SPaolo Bonzini * mappings should fault. A fault is considered 5266c50d8ae3SPaolo Bonzini * as a SMAP violation if all of the following 5267c50d8ae3SPaolo Bonzini * conditions are true: 5268c50d8ae3SPaolo Bonzini * - X86_CR4_SMAP is set in CR4 5269c50d8ae3SPaolo Bonzini * - A user page is accessed 5270c50d8ae3SPaolo Bonzini * - The access is not a fetch 52714f4aa80eSLai Jiangshan * - The access is supervisor mode 52724f4aa80eSLai Jiangshan * - If implicit supervisor access or X86_EFLAGS_AC is clear 5273c50d8ae3SPaolo Bonzini * 527494b4a2f1SLai Jiangshan * Here, we cover the first four conditions. 527594b4a2f1SLai Jiangshan * The fifth is computed dynamically in permission_fault(); 5276c50d8ae3SPaolo Bonzini * PFERR_RSVD_MASK bit will be set in PFEC if the access is 5277c50d8ae3SPaolo Bonzini * *not* subject to SMAP restrictions. 5278c50d8ae3SPaolo Bonzini */ 5279c50d8ae3SPaolo Bonzini if (cr4_smap) 5280c50d8ae3SPaolo Bonzini smapf = (pfec & (PFERR_RSVD_MASK|PFERR_FETCH_MASK)) ? 0 : kf; 5281c50d8ae3SPaolo Bonzini } 5282c50d8ae3SPaolo Bonzini 5283c50d8ae3SPaolo Bonzini mmu->permissions[byte] = ff | uf | wf | smepf | smapf; 5284c50d8ae3SPaolo Bonzini } 5285c50d8ae3SPaolo Bonzini } 5286c50d8ae3SPaolo Bonzini 5287c50d8ae3SPaolo Bonzini /* 5288c50d8ae3SPaolo Bonzini * PKU is an additional mechanism by which the paging controls access to 5289c50d8ae3SPaolo Bonzini * user-mode addresses based on the value in the PKRU register. Protection 5290c50d8ae3SPaolo Bonzini * key violations are reported through a bit in the page fault error code. 5291c50d8ae3SPaolo Bonzini * Unlike other bits of the error code, the PK bit is not known at the 5292c50d8ae3SPaolo Bonzini * call site of e.g. gva_to_gpa; it must be computed directly in 5293c50d8ae3SPaolo Bonzini * permission_fault based on two bits of PKRU, on some machine state (CR4, 5294c50d8ae3SPaolo Bonzini * CR0, EFER, CPL), and on other bits of the error code and the page tables. 5295c50d8ae3SPaolo Bonzini * 5296c50d8ae3SPaolo Bonzini * In particular the following conditions come from the error code, the 5297c50d8ae3SPaolo Bonzini * page tables and the machine state: 5298c50d8ae3SPaolo Bonzini * - PK is always zero unless CR4.PKE=1 and EFER.LMA=1 5299c50d8ae3SPaolo Bonzini * - PK is always zero if RSVD=1 (reserved bit set) or F=1 (instruction fetch) 5300c50d8ae3SPaolo Bonzini * - PK is always zero if U=0 in the page tables 5301c50d8ae3SPaolo Bonzini * - PKRU.WD is ignored if CR0.WP=0 and the access is a supervisor access. 5302c50d8ae3SPaolo Bonzini * 5303c50d8ae3SPaolo Bonzini * The PKRU bitmask caches the result of these four conditions. The error 5304c50d8ae3SPaolo Bonzini * code (minus the P bit) and the page table's U bit form an index into the 5305c50d8ae3SPaolo Bonzini * PKRU bitmask. Two bits of the PKRU bitmask are then extracted and ANDed 5306c50d8ae3SPaolo Bonzini * with the two bits of the PKRU register corresponding to the protection key. 5307c50d8ae3SPaolo Bonzini * For the first three conditions above the bits will be 00, thus masking 5308c50d8ae3SPaolo Bonzini * away both AD and WD. For all reads or if the last condition holds, WD 5309c50d8ae3SPaolo Bonzini * only will be masked away. 5310c50d8ae3SPaolo Bonzini */ 53112e4c0661SSean Christopherson static void update_pkru_bitmask(struct kvm_mmu *mmu) 5312c50d8ae3SPaolo Bonzini { 5313c50d8ae3SPaolo Bonzini unsigned bit; 5314c50d8ae3SPaolo Bonzini bool wp; 5315c50d8ae3SPaolo Bonzini 5316c50d8ae3SPaolo Bonzini mmu->pkru_mask = 0; 5317a3ca5281SChenyi Qiang 5318a3ca5281SChenyi Qiang if (!is_cr4_pke(mmu)) 5319c50d8ae3SPaolo Bonzini return; 5320c50d8ae3SPaolo Bonzini 53212e4c0661SSean Christopherson wp = is_cr0_wp(mmu); 5322c50d8ae3SPaolo Bonzini 5323c50d8ae3SPaolo Bonzini for (bit = 0; bit < ARRAY_SIZE(mmu->permissions); ++bit) { 5324c50d8ae3SPaolo Bonzini unsigned pfec, pkey_bits; 5325c50d8ae3SPaolo Bonzini bool check_pkey, check_write, ff, uf, wf, pte_user; 5326c50d8ae3SPaolo Bonzini 5327c50d8ae3SPaolo Bonzini pfec = bit << 1; 5328c50d8ae3SPaolo Bonzini ff = pfec & PFERR_FETCH_MASK; 5329c50d8ae3SPaolo Bonzini uf = pfec & PFERR_USER_MASK; 5330c50d8ae3SPaolo Bonzini wf = pfec & PFERR_WRITE_MASK; 5331c50d8ae3SPaolo Bonzini 5332c50d8ae3SPaolo Bonzini /* PFEC.RSVD is replaced by ACC_USER_MASK. */ 5333c50d8ae3SPaolo Bonzini pte_user = pfec & PFERR_RSVD_MASK; 5334c50d8ae3SPaolo Bonzini 5335c50d8ae3SPaolo Bonzini /* 5336c50d8ae3SPaolo Bonzini * Only need to check the access which is not an 5337c50d8ae3SPaolo Bonzini * instruction fetch and is to a user page. 5338c50d8ae3SPaolo Bonzini */ 5339c50d8ae3SPaolo Bonzini check_pkey = (!ff && pte_user); 5340c50d8ae3SPaolo Bonzini /* 5341c50d8ae3SPaolo Bonzini * write access is controlled by PKRU if it is a 5342c50d8ae3SPaolo Bonzini * user access or CR0.WP = 1. 5343c50d8ae3SPaolo Bonzini */ 5344c50d8ae3SPaolo Bonzini check_write = check_pkey && wf && (uf || wp); 5345c50d8ae3SPaolo Bonzini 5346c50d8ae3SPaolo Bonzini /* PKRU.AD stops both read and write access. */ 5347c50d8ae3SPaolo Bonzini pkey_bits = !!check_pkey; 5348c50d8ae3SPaolo Bonzini /* PKRU.WD stops write access. */ 5349c50d8ae3SPaolo Bonzini pkey_bits |= (!!check_write) << 1; 5350c50d8ae3SPaolo Bonzini 5351c50d8ae3SPaolo Bonzini mmu->pkru_mask |= (pkey_bits & 3) << pfec; 5352c50d8ae3SPaolo Bonzini } 5353c50d8ae3SPaolo Bonzini } 5354c50d8ae3SPaolo Bonzini 5355533f9a4bSSean Christopherson static void reset_guest_paging_metadata(struct kvm_vcpu *vcpu, 5356533f9a4bSSean Christopherson struct kvm_mmu *mmu) 5357c50d8ae3SPaolo Bonzini { 5358533f9a4bSSean Christopherson if (!is_cr0_pg(mmu)) 5359533f9a4bSSean Christopherson return; 5360c50d8ae3SPaolo Bonzini 5361c919e881SKai Huang reset_guest_rsvds_bits_mask(vcpu, mmu); 5362533f9a4bSSean Christopherson update_permission_bitmask(mmu, false); 5363533f9a4bSSean Christopherson update_pkru_bitmask(mmu); 5364c50d8ae3SPaolo Bonzini } 5365c50d8ae3SPaolo Bonzini 5366fe660f72SSean Christopherson static void paging64_init_context(struct kvm_mmu *context) 5367c50d8ae3SPaolo Bonzini { 5368c50d8ae3SPaolo Bonzini context->page_fault = paging64_page_fault; 5369c50d8ae3SPaolo Bonzini context->gva_to_gpa = paging64_gva_to_gpa; 5370c3c6c9fcSLai Jiangshan context->sync_spte = paging64_sync_spte; 5371c50d8ae3SPaolo Bonzini } 5372c50d8ae3SPaolo Bonzini 537384a16226SSean Christopherson static void paging32_init_context(struct kvm_mmu *context) 5374c50d8ae3SPaolo Bonzini { 5375c50d8ae3SPaolo Bonzini context->page_fault = paging32_page_fault; 5376c50d8ae3SPaolo Bonzini context->gva_to_gpa = paging32_gva_to_gpa; 5377c3c6c9fcSLai Jiangshan context->sync_spte = paging32_sync_spte; 5378c50d8ae3SPaolo Bonzini } 5379c50d8ae3SPaolo Bonzini 5380f3d90f90SSean Christopherson static union kvm_cpu_role kvm_calc_cpu_role(struct kvm_vcpu *vcpu, 5381f3d90f90SSean Christopherson const struct kvm_mmu_role_regs *regs) 5382e5ed0fb0SPaolo Bonzini { 53837a7ae829SPaolo Bonzini union kvm_cpu_role role = {0}; 5384e5ed0fb0SPaolo Bonzini 5385e5ed0fb0SPaolo Bonzini role.base.access = ACC_ALL; 5386e5ed0fb0SPaolo Bonzini role.base.smm = is_smm(vcpu); 5387e5ed0fb0SPaolo Bonzini role.base.guest_mode = is_guest_mode(vcpu); 5388e5ed0fb0SPaolo Bonzini role.ext.valid = 1; 5389e5ed0fb0SPaolo Bonzini 5390e5ed0fb0SPaolo Bonzini if (!____is_cr0_pg(regs)) { 5391e5ed0fb0SPaolo Bonzini role.base.direct = 1; 5392e5ed0fb0SPaolo Bonzini return role; 5393e5ed0fb0SPaolo Bonzini } 5394e5ed0fb0SPaolo Bonzini 5395e5ed0fb0SPaolo Bonzini role.base.efer_nx = ____is_efer_nx(regs); 5396e5ed0fb0SPaolo Bonzini role.base.cr0_wp = ____is_cr0_wp(regs); 5397e5ed0fb0SPaolo Bonzini role.base.smep_andnot_wp = ____is_cr4_smep(regs) && !____is_cr0_wp(regs); 5398e5ed0fb0SPaolo Bonzini role.base.smap_andnot_wp = ____is_cr4_smap(regs) && !____is_cr0_wp(regs); 5399e5ed0fb0SPaolo Bonzini role.base.has_4_byte_gpte = !____is_cr4_pae(regs); 540060f3cb60SPaolo Bonzini 540160f3cb60SPaolo Bonzini if (____is_efer_lma(regs)) 540260f3cb60SPaolo Bonzini role.base.level = ____is_cr4_la57(regs) ? PT64_ROOT_5LEVEL 540360f3cb60SPaolo Bonzini : PT64_ROOT_4LEVEL; 540460f3cb60SPaolo Bonzini else if (____is_cr4_pae(regs)) 540560f3cb60SPaolo Bonzini role.base.level = PT32E_ROOT_LEVEL; 540660f3cb60SPaolo Bonzini else 540760f3cb60SPaolo Bonzini role.base.level = PT32_ROOT_LEVEL; 5408e5ed0fb0SPaolo Bonzini 5409e5ed0fb0SPaolo Bonzini role.ext.cr4_smep = ____is_cr4_smep(regs); 5410e5ed0fb0SPaolo Bonzini role.ext.cr4_smap = ____is_cr4_smap(regs); 5411e5ed0fb0SPaolo Bonzini role.ext.cr4_pse = ____is_cr4_pse(regs); 5412e5ed0fb0SPaolo Bonzini 5413e5ed0fb0SPaolo Bonzini /* PKEY and LA57 are active iff long mode is active. */ 5414e5ed0fb0SPaolo Bonzini role.ext.cr4_pke = ____is_efer_lma(regs) && ____is_cr4_pke(regs); 5415e5ed0fb0SPaolo Bonzini role.ext.cr4_la57 = ____is_efer_lma(regs) && ____is_cr4_la57(regs); 5416e5ed0fb0SPaolo Bonzini role.ext.efer_lma = ____is_efer_lma(regs); 5417e5ed0fb0SPaolo Bonzini return role; 5418e5ed0fb0SPaolo Bonzini } 5419e5ed0fb0SPaolo Bonzini 5420cf9f4c0eSSean Christopherson void __kvm_mmu_refresh_passthrough_bits(struct kvm_vcpu *vcpu, 5421cf9f4c0eSSean Christopherson struct kvm_mmu *mmu) 5422cf9f4c0eSSean Christopherson { 5423cf9f4c0eSSean Christopherson const bool cr0_wp = kvm_is_cr0_bit_set(vcpu, X86_CR0_WP); 5424cf9f4c0eSSean Christopherson 5425cf9f4c0eSSean Christopherson BUILD_BUG_ON((KVM_MMU_CR0_ROLE_BITS & KVM_POSSIBLE_CR0_GUEST_BITS) != X86_CR0_WP); 5426cf9f4c0eSSean Christopherson BUILD_BUG_ON((KVM_MMU_CR4_ROLE_BITS & KVM_POSSIBLE_CR4_GUEST_BITS)); 5427cf9f4c0eSSean Christopherson 5428cf9f4c0eSSean Christopherson if (is_cr0_wp(mmu) == cr0_wp) 5429cf9f4c0eSSean Christopherson return; 5430cf9f4c0eSSean Christopherson 5431cf9f4c0eSSean Christopherson mmu->cpu_role.base.cr0_wp = cr0_wp; 5432cf9f4c0eSSean Christopherson reset_guest_paging_metadata(vcpu, mmu); 5433cf9f4c0eSSean Christopherson } 5434cf9f4c0eSSean Christopherson 5435d468d94bSSean Christopherson static inline int kvm_mmu_get_tdp_level(struct kvm_vcpu *vcpu) 5436d468d94bSSean Christopherson { 5437746700d2SWei Huang /* tdp_root_level is architecture forced level, use it if nonzero */ 5438746700d2SWei Huang if (tdp_root_level) 5439746700d2SWei Huang return tdp_root_level; 5440746700d2SWei Huang 5441d468d94bSSean Christopherson /* Use 5-level TDP if and only if it's useful/necessary. */ 544283013059SSean Christopherson if (max_tdp_level == 5 && cpuid_maxphyaddr(vcpu) <= 48) 5443d468d94bSSean Christopherson return 4; 5444d468d94bSSean Christopherson 544583013059SSean Christopherson return max_tdp_level; 5446d468d94bSSean Christopherson } 5447d468d94bSSean Christopherson 5448b628cb52SGerd Hoffmann u8 kvm_mmu_get_max_tdp_level(void) 5449b628cb52SGerd Hoffmann { 5450b628cb52SGerd Hoffmann return tdp_root_level ? tdp_root_level : max_tdp_level; 5451b628cb52SGerd Hoffmann } 5452b628cb52SGerd Hoffmann 54537a458f0eSPaolo Bonzini static union kvm_mmu_page_role 54548626c120SSean Christopherson kvm_calc_tdp_mmu_root_page_role(struct kvm_vcpu *vcpu, 54557a7ae829SPaolo Bonzini union kvm_cpu_role cpu_role) 5456c50d8ae3SPaolo Bonzini { 54577a458f0eSPaolo Bonzini union kvm_mmu_page_role role = {0}; 5458c50d8ae3SPaolo Bonzini 54597a458f0eSPaolo Bonzini role.access = ACC_ALL; 54607a458f0eSPaolo Bonzini role.cr0_wp = true; 54617a458f0eSPaolo Bonzini role.efer_nx = true; 54627a458f0eSPaolo Bonzini role.smm = cpu_role.base.smm; 54637a458f0eSPaolo Bonzini role.guest_mode = cpu_role.base.guest_mode; 546454275f74SSean Christopherson role.ad_disabled = !kvm_ad_enabled(); 54657a458f0eSPaolo Bonzini role.level = kvm_mmu_get_tdp_level(vcpu); 54667a458f0eSPaolo Bonzini role.direct = true; 54677a458f0eSPaolo Bonzini role.has_4_byte_gpte = false; 5468c50d8ae3SPaolo Bonzini 5469c50d8ae3SPaolo Bonzini return role; 5470c50d8ae3SPaolo Bonzini } 5471c50d8ae3SPaolo Bonzini 547239e7e2bfSPaolo Bonzini static void init_kvm_tdp_mmu(struct kvm_vcpu *vcpu, 5473a7f1de9bSPaolo Bonzini union kvm_cpu_role cpu_role) 5474c50d8ae3SPaolo Bonzini { 54758c008659SPaolo Bonzini struct kvm_mmu *context = &vcpu->arch.root_mmu; 54767a458f0eSPaolo Bonzini union kvm_mmu_page_role root_role = kvm_calc_tdp_mmu_root_page_role(vcpu, cpu_role); 5477c50d8ae3SPaolo Bonzini 5478e5ed0fb0SPaolo Bonzini if (cpu_role.as_u64 == context->cpu_role.as_u64 && 54797a458f0eSPaolo Bonzini root_role.word == context->root_role.word) 5480c50d8ae3SPaolo Bonzini return; 5481c50d8ae3SPaolo Bonzini 5482e5ed0fb0SPaolo Bonzini context->cpu_role.as_u64 = cpu_role.as_u64; 54837a458f0eSPaolo Bonzini context->root_role.word = root_role.word; 54847a02674dSSean Christopherson context->page_fault = kvm_tdp_page_fault; 5485c3c6c9fcSLai Jiangshan context->sync_spte = NULL; 54862fdcc1b3SPaolo Bonzini context->get_guest_pgd = get_guest_cr3; 5487c50d8ae3SPaolo Bonzini context->get_pdptr = kvm_pdptr_read; 5488c50d8ae3SPaolo Bonzini context->inject_page_fault = kvm_inject_page_fault; 5489c50d8ae3SPaolo Bonzini 549036f26787SSean Christopherson if (!is_cr0_pg(context)) 5491c50d8ae3SPaolo Bonzini context->gva_to_gpa = nonpaging_gva_to_gpa; 549236f26787SSean Christopherson else if (is_cr4_pae(context)) 5493c50d8ae3SPaolo Bonzini context->gva_to_gpa = paging64_gva_to_gpa; 5494f4bd6f73SSean Christopherson else 5495c50d8ae3SPaolo Bonzini context->gva_to_gpa = paging32_gva_to_gpa; 5496c50d8ae3SPaolo Bonzini 5497533f9a4bSSean Christopherson reset_guest_paging_metadata(vcpu, context); 5498e8f6e738SJinrong Liang reset_tdp_shadow_zero_bits_mask(context); 5499c50d8ae3SPaolo Bonzini } 5500c50d8ae3SPaolo Bonzini 55018c008659SPaolo Bonzini static void shadow_mmu_init_context(struct kvm_vcpu *vcpu, struct kvm_mmu *context, 55027a7ae829SPaolo Bonzini union kvm_cpu_role cpu_role, 55037a458f0eSPaolo Bonzini union kvm_mmu_page_role root_role) 5504c50d8ae3SPaolo Bonzini { 5505e5ed0fb0SPaolo Bonzini if (cpu_role.as_u64 == context->cpu_role.as_u64 && 55067a458f0eSPaolo Bonzini root_role.word == context->root_role.word) 550718db1b17SSean Christopherson return; 5508c50d8ae3SPaolo Bonzini 5509e5ed0fb0SPaolo Bonzini context->cpu_role.as_u64 = cpu_role.as_u64; 55107a458f0eSPaolo Bonzini context->root_role.word = root_role.word; 551118db1b17SSean Christopherson 551236f26787SSean Christopherson if (!is_cr0_pg(context)) 551384a16226SSean Christopherson nonpaging_init_context(context); 551436f26787SSean Christopherson else if (is_cr4_pae(context)) 5515fe660f72SSean Christopherson paging64_init_context(context); 5516c50d8ae3SPaolo Bonzini else 551784a16226SSean Christopherson paging32_init_context(context); 5518c50d8ae3SPaolo Bonzini 5519533f9a4bSSean Christopherson reset_guest_paging_metadata(vcpu, context); 5520c50d8ae3SPaolo Bonzini reset_shadow_zero_bits_mask(vcpu, context); 5521c50d8ae3SPaolo Bonzini } 55220f04a2acSVitaly Kuznetsov 5523594e91a1SSean Christopherson static void kvm_init_shadow_mmu(struct kvm_vcpu *vcpu, 5524a7f1de9bSPaolo Bonzini union kvm_cpu_role cpu_role) 55250f04a2acSVitaly Kuznetsov { 55268c008659SPaolo Bonzini struct kvm_mmu *context = &vcpu->arch.root_mmu; 552756b321f9SPaolo Bonzini union kvm_mmu_page_role root_role; 552856b321f9SPaolo Bonzini 552956b321f9SPaolo Bonzini root_role = cpu_role.base; 553056b321f9SPaolo Bonzini 553156b321f9SPaolo Bonzini /* KVM uses PAE paging whenever the guest isn't using 64-bit paging. */ 553256b321f9SPaolo Bonzini root_role.level = max_t(u32, root_role.level, PT32E_ROOT_LEVEL); 553356b321f9SPaolo Bonzini 553456b321f9SPaolo Bonzini /* 553556b321f9SPaolo Bonzini * KVM forces EFER.NX=1 when TDP is disabled, reflect it in the MMU role. 553656b321f9SPaolo Bonzini * KVM uses NX when TDP is disabled to handle a variety of scenarios, 553756b321f9SPaolo Bonzini * notably for huge SPTEs if iTLB multi-hit mitigation is enabled and 553856b321f9SPaolo Bonzini * to generate correct permissions for CR0.WP=0/CR4.SMEP=1/EFER.NX=0. 553956b321f9SPaolo Bonzini * The iTLB multi-hit workaround can be toggled at any time, so assume 554056b321f9SPaolo Bonzini * NX can be used by any non-nested shadow MMU to avoid having to reset 554156b321f9SPaolo Bonzini * MMU contexts. 554256b321f9SPaolo Bonzini */ 554356b321f9SPaolo Bonzini root_role.efer_nx = true; 55440f04a2acSVitaly Kuznetsov 55457a458f0eSPaolo Bonzini shadow_mmu_init_context(vcpu, context, cpu_role, root_role); 55460f04a2acSVitaly Kuznetsov } 55470f04a2acSVitaly Kuznetsov 5548dbc4739bSSean Christopherson void kvm_init_shadow_npt_mmu(struct kvm_vcpu *vcpu, unsigned long cr0, 5549dbc4739bSSean Christopherson unsigned long cr4, u64 efer, gpa_t nested_cr3) 55500f04a2acSVitaly Kuznetsov { 55518c008659SPaolo Bonzini struct kvm_mmu *context = &vcpu->arch.guest_mmu; 5552594e91a1SSean Christopherson struct kvm_mmu_role_regs regs = { 5553594e91a1SSean Christopherson .cr0 = cr0, 555428f091bcSPaolo Bonzini .cr4 = cr4 & ~X86_CR4_PKE, 5555594e91a1SSean Christopherson .efer = efer, 5556594e91a1SSean Christopherson }; 55577a7ae829SPaolo Bonzini union kvm_cpu_role cpu_role = kvm_calc_cpu_role(vcpu, ®s); 555856b321f9SPaolo Bonzini union kvm_mmu_page_role root_role; 555956b321f9SPaolo Bonzini 556056b321f9SPaolo Bonzini /* NPT requires CR0.PG=1. */ 556156b321f9SPaolo Bonzini WARN_ON_ONCE(cpu_role.base.direct); 556256b321f9SPaolo Bonzini 556356b321f9SPaolo Bonzini root_role = cpu_role.base; 556456b321f9SPaolo Bonzini root_role.level = kvm_mmu_get_tdp_level(vcpu); 556584e5ffd0SLai Jiangshan if (root_role.level == PT64_ROOT_5LEVEL && 556684e5ffd0SLai Jiangshan cpu_role.base.level == PT64_ROOT_4LEVEL) 556784e5ffd0SLai Jiangshan root_role.passthrough = 1; 55680f04a2acSVitaly Kuznetsov 55697a458f0eSPaolo Bonzini shadow_mmu_init_context(vcpu, context, cpu_role, root_role); 5570d2e5f333SPaolo Bonzini kvm_mmu_new_pgd(vcpu, nested_cr3); 55710f04a2acSVitaly Kuznetsov } 55720f04a2acSVitaly Kuznetsov EXPORT_SYMBOL_GPL(kvm_init_shadow_npt_mmu); 5573c50d8ae3SPaolo Bonzini 55747a7ae829SPaolo Bonzini static union kvm_cpu_role 5575c50d8ae3SPaolo Bonzini kvm_calc_shadow_ept_root_page_role(struct kvm_vcpu *vcpu, bool accessed_dirty, 5576bb1fcc70SSean Christopherson bool execonly, u8 level) 5577c50d8ae3SPaolo Bonzini { 55787a7ae829SPaolo Bonzini union kvm_cpu_role role = {0}; 5579c50d8ae3SPaolo Bonzini 5580daed87b8SPaolo Bonzini /* 5581daed87b8SPaolo Bonzini * KVM does not support SMM transfer monitors, and consequently does not 5582daed87b8SPaolo Bonzini * support the "entry to SMM" control either. role.base.smm is always 0. 5583daed87b8SPaolo Bonzini */ 5584daed87b8SPaolo Bonzini WARN_ON_ONCE(is_smm(vcpu)); 5585bb1fcc70SSean Christopherson role.base.level = level; 5586bb3b394dSLai Jiangshan role.base.has_4_byte_gpte = false; 5587c50d8ae3SPaolo Bonzini role.base.direct = false; 5588c50d8ae3SPaolo Bonzini role.base.ad_disabled = !accessed_dirty; 5589c50d8ae3SPaolo Bonzini role.base.guest_mode = true; 5590c50d8ae3SPaolo Bonzini role.base.access = ACC_ALL; 5591c50d8ae3SPaolo Bonzini 5592cd6767c3SSean Christopherson role.ext.word = 0; 5593c50d8ae3SPaolo Bonzini role.ext.execonly = execonly; 5594cd6767c3SSean Christopherson role.ext.valid = 1; 5595c50d8ae3SPaolo Bonzini 5596c50d8ae3SPaolo Bonzini return role; 5597c50d8ae3SPaolo Bonzini } 5598c50d8ae3SPaolo Bonzini 5599c50d8ae3SPaolo Bonzini void kvm_init_shadow_ept_mmu(struct kvm_vcpu *vcpu, bool execonly, 5600cc022ae1SLai Jiangshan int huge_page_level, bool accessed_dirty, 5601cc022ae1SLai Jiangshan gpa_t new_eptp) 5602c50d8ae3SPaolo Bonzini { 56038c008659SPaolo Bonzini struct kvm_mmu *context = &vcpu->arch.guest_mmu; 5604bb1fcc70SSean Christopherson u8 level = vmx_eptp_page_walk_level(new_eptp); 56057a7ae829SPaolo Bonzini union kvm_cpu_role new_mode = 5606c50d8ae3SPaolo Bonzini kvm_calc_shadow_ept_root_page_role(vcpu, accessed_dirty, 5607bb1fcc70SSean Christopherson execonly, level); 5608c50d8ae3SPaolo Bonzini 5609e5ed0fb0SPaolo Bonzini if (new_mode.as_u64 != context->cpu_role.as_u64) { 5610e5ed0fb0SPaolo Bonzini /* EPT, and thus nested EPT, does not consume CR0, CR4, nor EFER. */ 5611e5ed0fb0SPaolo Bonzini context->cpu_role.as_u64 = new_mode.as_u64; 56127a458f0eSPaolo Bonzini context->root_role.word = new_mode.base.word; 561318db1b17SSean Christopherson 5614c50d8ae3SPaolo Bonzini context->page_fault = ept_page_fault; 5615c50d8ae3SPaolo Bonzini context->gva_to_gpa = ept_gva_to_gpa; 5616c3c6c9fcSLai Jiangshan context->sync_spte = ept_sync_spte; 5617347a0d0dSPaolo Bonzini 5618c596f147SSean Christopherson update_permission_bitmask(context, true); 561928f091bcSPaolo Bonzini context->pkru_mask = 0; 5620cc022ae1SLai Jiangshan reset_rsvds_bits_mask_ept(vcpu, context, execonly, huge_page_level); 5621e8f6e738SJinrong Liang reset_ept_shadow_zero_bits_mask(context, execonly); 5622c50d8ae3SPaolo Bonzini } 56233cffc89dSPaolo Bonzini 5624d2e5f333SPaolo Bonzini kvm_mmu_new_pgd(vcpu, new_eptp); 56253cffc89dSPaolo Bonzini } 5626c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_init_shadow_ept_mmu); 5627c50d8ae3SPaolo Bonzini 562839e7e2bfSPaolo Bonzini static void init_kvm_softmmu(struct kvm_vcpu *vcpu, 5629a7f1de9bSPaolo Bonzini union kvm_cpu_role cpu_role) 5630c50d8ae3SPaolo Bonzini { 56318c008659SPaolo Bonzini struct kvm_mmu *context = &vcpu->arch.root_mmu; 5632c50d8ae3SPaolo Bonzini 5633a7f1de9bSPaolo Bonzini kvm_init_shadow_mmu(vcpu, cpu_role); 5634929d1cfaSPaolo Bonzini 56352fdcc1b3SPaolo Bonzini context->get_guest_pgd = get_guest_cr3; 5636c50d8ae3SPaolo Bonzini context->get_pdptr = kvm_pdptr_read; 5637c50d8ae3SPaolo Bonzini context->inject_page_fault = kvm_inject_page_fault; 5638c50d8ae3SPaolo Bonzini } 5639c50d8ae3SPaolo Bonzini 564039e7e2bfSPaolo Bonzini static void init_kvm_nested_mmu(struct kvm_vcpu *vcpu, 5641a7f1de9bSPaolo Bonzini union kvm_cpu_role new_mode) 5642c50d8ae3SPaolo Bonzini { 5643c50d8ae3SPaolo Bonzini struct kvm_mmu *g_context = &vcpu->arch.nested_mmu; 5644c50d8ae3SPaolo Bonzini 5645e5ed0fb0SPaolo Bonzini if (new_mode.as_u64 == g_context->cpu_role.as_u64) 5646c50d8ae3SPaolo Bonzini return; 5647c50d8ae3SPaolo Bonzini 5648e5ed0fb0SPaolo Bonzini g_context->cpu_role.as_u64 = new_mode.as_u64; 56492fdcc1b3SPaolo Bonzini g_context->get_guest_pgd = get_guest_cr3; 5650c50d8ae3SPaolo Bonzini g_context->get_pdptr = kvm_pdptr_read; 5651c50d8ae3SPaolo Bonzini g_context->inject_page_fault = kvm_inject_page_fault; 5652c50d8ae3SPaolo Bonzini 5653c50d8ae3SPaolo Bonzini /* 56545efac074SPaolo Bonzini * L2 page tables are never shadowed, so there is no need to sync 56555efac074SPaolo Bonzini * SPTEs. 56565efac074SPaolo Bonzini */ 56579fd4a4e3SLai Jiangshan g_context->sync_spte = NULL; 56585efac074SPaolo Bonzini 56595efac074SPaolo Bonzini /* 5660c50d8ae3SPaolo Bonzini * Note that arch.mmu->gva_to_gpa translates l2_gpa to l1_gpa using 5661c50d8ae3SPaolo Bonzini * L1's nested page tables (e.g. EPT12). The nested translation 5662c50d8ae3SPaolo Bonzini * of l2_gva to l1_gpa is done by arch.nested_mmu.gva_to_gpa using 5663c50d8ae3SPaolo Bonzini * L2's page tables as the first level of translation and L1's 5664c50d8ae3SPaolo Bonzini * nested page tables as the second level of translation. Basically 5665c50d8ae3SPaolo Bonzini * the gva_to_gpa functions between mmu and nested_mmu are swapped. 5666c50d8ae3SPaolo Bonzini */ 5667fa4b5588SSean Christopherson if (!is_paging(vcpu)) 56681f5a21eeSLai Jiangshan g_context->gva_to_gpa = nonpaging_gva_to_gpa; 5669fa4b5588SSean Christopherson else if (is_long_mode(vcpu)) 56701f5a21eeSLai Jiangshan g_context->gva_to_gpa = paging64_gva_to_gpa; 5671fa4b5588SSean Christopherson else if (is_pae(vcpu)) 56721f5a21eeSLai Jiangshan g_context->gva_to_gpa = paging64_gva_to_gpa; 5673fa4b5588SSean Christopherson else 56741f5a21eeSLai Jiangshan g_context->gva_to_gpa = paging32_gva_to_gpa; 5675fa4b5588SSean Christopherson 5676533f9a4bSSean Christopherson reset_guest_paging_metadata(vcpu, g_context); 5677c50d8ae3SPaolo Bonzini } 5678c50d8ae3SPaolo Bonzini 5679c9060662SSean Christopherson void kvm_init_mmu(struct kvm_vcpu *vcpu) 5680c50d8ae3SPaolo Bonzini { 568139e7e2bfSPaolo Bonzini struct kvm_mmu_role_regs regs = vcpu_to_role_regs(vcpu); 5682a7f1de9bSPaolo Bonzini union kvm_cpu_role cpu_role = kvm_calc_cpu_role(vcpu, ®s); 568339e7e2bfSPaolo Bonzini 5684c50d8ae3SPaolo Bonzini if (mmu_is_nested(vcpu)) 5685a7f1de9bSPaolo Bonzini init_kvm_nested_mmu(vcpu, cpu_role); 5686c50d8ae3SPaolo Bonzini else if (tdp_enabled) 5687a7f1de9bSPaolo Bonzini init_kvm_tdp_mmu(vcpu, cpu_role); 5688c50d8ae3SPaolo Bonzini else 5689a7f1de9bSPaolo Bonzini init_kvm_softmmu(vcpu, cpu_role); 5690c50d8ae3SPaolo Bonzini } 5691c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_init_mmu); 5692c50d8ae3SPaolo Bonzini 569349c6f875SSean Christopherson void kvm_mmu_after_set_cpuid(struct kvm_vcpu *vcpu) 569449c6f875SSean Christopherson { 569549c6f875SSean Christopherson /* 569649c6f875SSean Christopherson * Invalidate all MMU roles to force them to reinitialize as CPUID 569749c6f875SSean Christopherson * information is factored into reserved bit calculations. 5698feb627e8SVitaly Kuznetsov * 5699feb627e8SVitaly Kuznetsov * Correctly handling multiple vCPU models with respect to paging and 5700feb627e8SVitaly Kuznetsov * physical address properties) in a single VM would require tracking 5701feb627e8SVitaly Kuznetsov * all relevant CPUID information in kvm_mmu_page_role. That is very 5702feb627e8SVitaly Kuznetsov * undesirable as it would increase the memory requirements for 5703338068b5SSean Christopherson * gfn_write_track (see struct kvm_mmu_page_role comments). For now 5704338068b5SSean Christopherson * that problem is swept under the rug; KVM's CPUID API is horrific and 5705feb627e8SVitaly Kuznetsov * it's all but impossible to solve it without introducing a new API. 570649c6f875SSean Christopherson */ 57071bc26cb9SSean Christopherson vcpu->arch.root_mmu.root_role.invalid = 1; 57081bc26cb9SSean Christopherson vcpu->arch.guest_mmu.root_role.invalid = 1; 57091bc26cb9SSean Christopherson vcpu->arch.nested_mmu.root_role.invalid = 1; 5710e5ed0fb0SPaolo Bonzini vcpu->arch.root_mmu.cpu_role.ext.valid = 0; 5711e5ed0fb0SPaolo Bonzini vcpu->arch.guest_mmu.cpu_role.ext.valid = 0; 5712e5ed0fb0SPaolo Bonzini vcpu->arch.nested_mmu.cpu_role.ext.valid = 0; 571349c6f875SSean Christopherson kvm_mmu_reset_context(vcpu); 571463f5a190SSean Christopherson 571563f5a190SSean Christopherson /* 5716feb627e8SVitaly Kuznetsov * Changing guest CPUID after KVM_RUN is forbidden, see the comment in 5717feb627e8SVitaly Kuznetsov * kvm_arch_vcpu_ioctl(). 571863f5a190SSean Christopherson */ 5719fb3146b4SSean Christopherson KVM_BUG_ON(kvm_vcpu_has_run(vcpu), vcpu->kvm); 572049c6f875SSean Christopherson } 572149c6f875SSean Christopherson 5722c50d8ae3SPaolo Bonzini void kvm_mmu_reset_context(struct kvm_vcpu *vcpu) 5723c50d8ae3SPaolo Bonzini { 5724c50d8ae3SPaolo Bonzini kvm_mmu_unload(vcpu); 5725c9060662SSean Christopherson kvm_init_mmu(vcpu); 5726c50d8ae3SPaolo Bonzini } 5727c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_mmu_reset_context); 5728c50d8ae3SPaolo Bonzini 5729c50d8ae3SPaolo Bonzini int kvm_mmu_load(struct kvm_vcpu *vcpu) 5730c50d8ae3SPaolo Bonzini { 5731c50d8ae3SPaolo Bonzini int r; 5732c50d8ae3SPaolo Bonzini 5733347a0d0dSPaolo Bonzini r = mmu_topup_memory_caches(vcpu, !vcpu->arch.mmu->root_role.direct); 5734c50d8ae3SPaolo Bonzini if (r) 5735c50d8ae3SPaolo Bonzini goto out; 5736748e52b9SSean Christopherson r = mmu_alloc_special_roots(vcpu); 5737c50d8ae3SPaolo Bonzini if (r) 5738c50d8ae3SPaolo Bonzini goto out; 5739347a0d0dSPaolo Bonzini if (vcpu->arch.mmu->root_role.direct) 57406e6ec584SSean Christopherson r = mmu_alloc_direct_roots(vcpu); 57416e6ec584SSean Christopherson else 57426e6ec584SSean Christopherson r = mmu_alloc_shadow_roots(vcpu); 5743c50d8ae3SPaolo Bonzini if (r) 5744c50d8ae3SPaolo Bonzini goto out; 5745a91f387bSSean Christopherson 5746a91f387bSSean Christopherson kvm_mmu_sync_roots(vcpu); 5747a91f387bSSean Christopherson 5748727a7e27SPaolo Bonzini kvm_mmu_load_pgd(vcpu); 5749db01416bSSean Christopherson 5750db01416bSSean Christopherson /* 5751db01416bSSean Christopherson * Flush any TLB entries for the new root, the provenance of the root 5752db01416bSSean Christopherson * is unknown. Even if KVM ensures there are no stale TLB entries 5753db01416bSSean Christopherson * for a freed root, in theory another hypervisor could have left 5754db01416bSSean Christopherson * stale entries. Flushing on alloc also allows KVM to skip the TLB 5755db01416bSSean Christopherson * flush when freeing a root (see kvm_tdp_mmu_put_root()). 5756db01416bSSean Christopherson */ 5757e27bc044SSean Christopherson static_call(kvm_x86_flush_tlb_current)(vcpu); 5758c50d8ae3SPaolo Bonzini out: 5759c50d8ae3SPaolo Bonzini return r; 5760c50d8ae3SPaolo Bonzini } 5761c50d8ae3SPaolo Bonzini 5762c50d8ae3SPaolo Bonzini void kvm_mmu_unload(struct kvm_vcpu *vcpu) 5763c50d8ae3SPaolo Bonzini { 57640c1c92f1SPaolo Bonzini struct kvm *kvm = vcpu->kvm; 57650c1c92f1SPaolo Bonzini 57660c1c92f1SPaolo Bonzini kvm_mmu_free_roots(kvm, &vcpu->arch.root_mmu, KVM_MMU_ROOTS_ALL); 576720ba462dSSean Christopherson WARN_ON_ONCE(VALID_PAGE(vcpu->arch.root_mmu.root.hpa)); 57680c1c92f1SPaolo Bonzini kvm_mmu_free_roots(kvm, &vcpu->arch.guest_mmu, KVM_MMU_ROOTS_ALL); 576920ba462dSSean Christopherson WARN_ON_ONCE(VALID_PAGE(vcpu->arch.guest_mmu.root.hpa)); 57706d58f275SPaolo Bonzini vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY); 5771c50d8ae3SPaolo Bonzini } 5772c50d8ae3SPaolo Bonzini 5773527d5cd7SSean Christopherson static bool is_obsolete_root(struct kvm *kvm, hpa_t root_hpa) 5774527d5cd7SSean Christopherson { 5775527d5cd7SSean Christopherson struct kvm_mmu_page *sp; 5776527d5cd7SSean Christopherson 5777527d5cd7SSean Christopherson if (!VALID_PAGE(root_hpa)) 5778527d5cd7SSean Christopherson return false; 5779527d5cd7SSean Christopherson 5780527d5cd7SSean Christopherson /* 5781527d5cd7SSean Christopherson * When freeing obsolete roots, treat roots as obsolete if they don't 57820e3223d8SSean Christopherson * have an associated shadow page, as it's impossible to determine if 57830e3223d8SSean Christopherson * such roots are fresh or stale. This does mean KVM will get false 5784527d5cd7SSean Christopherson * positives and free roots that don't strictly need to be freed, but 5785527d5cd7SSean Christopherson * such false positives are relatively rare: 5786527d5cd7SSean Christopherson * 57870e3223d8SSean Christopherson * (a) only PAE paging and nested NPT have roots without shadow pages 57880e3223d8SSean Christopherson * (or any shadow paging flavor with a dummy root, see note below) 5789527d5cd7SSean Christopherson * (b) remote reloads due to a memslot update obsoletes _all_ roots 5790527d5cd7SSean Christopherson * (c) KVM doesn't track previous roots for PAE paging, and the guest 5791527d5cd7SSean Christopherson * is unlikely to zap an in-use PGD. 57920e3223d8SSean Christopherson * 57930e3223d8SSean Christopherson * Note! Dummy roots are unique in that they are obsoleted by memslot 57940e3223d8SSean Christopherson * _creation_! See also FNAME(fetch). 5795527d5cd7SSean Christopherson */ 5796c5f2d564SSean Christopherson sp = root_to_sp(root_hpa); 5797527d5cd7SSean Christopherson return !sp || is_obsolete_sp(kvm, sp); 5798527d5cd7SSean Christopherson } 5799527d5cd7SSean Christopherson 5800527d5cd7SSean Christopherson static void __kvm_mmu_free_obsolete_roots(struct kvm *kvm, struct kvm_mmu *mmu) 5801527d5cd7SSean Christopherson { 5802527d5cd7SSean Christopherson unsigned long roots_to_free = 0; 5803527d5cd7SSean Christopherson int i; 5804527d5cd7SSean Christopherson 5805527d5cd7SSean Christopherson if (is_obsolete_root(kvm, mmu->root.hpa)) 5806527d5cd7SSean Christopherson roots_to_free |= KVM_MMU_ROOT_CURRENT; 5807527d5cd7SSean Christopherson 5808527d5cd7SSean Christopherson for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) { 5809cf4a8693SShaoqin Huang if (is_obsolete_root(kvm, mmu->prev_roots[i].hpa)) 5810527d5cd7SSean Christopherson roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i); 5811527d5cd7SSean Christopherson } 5812527d5cd7SSean Christopherson 5813527d5cd7SSean Christopherson if (roots_to_free) 5814527d5cd7SSean Christopherson kvm_mmu_free_roots(kvm, mmu, roots_to_free); 5815527d5cd7SSean Christopherson } 5816527d5cd7SSean Christopherson 5817527d5cd7SSean Christopherson void kvm_mmu_free_obsolete_roots(struct kvm_vcpu *vcpu) 5818527d5cd7SSean Christopherson { 5819527d5cd7SSean Christopherson __kvm_mmu_free_obsolete_roots(vcpu->kvm, &vcpu->arch.root_mmu); 5820527d5cd7SSean Christopherson __kvm_mmu_free_obsolete_roots(vcpu->kvm, &vcpu->arch.guest_mmu); 5821527d5cd7SSean Christopherson } 5822527d5cd7SSean Christopherson 5823c50d8ae3SPaolo Bonzini static u64 mmu_pte_write_fetch_gpte(struct kvm_vcpu *vcpu, gpa_t *gpa, 5824c50d8ae3SPaolo Bonzini int *bytes) 5825c50d8ae3SPaolo Bonzini { 5826c50d8ae3SPaolo Bonzini u64 gentry = 0; 5827c50d8ae3SPaolo Bonzini int r; 5828c50d8ae3SPaolo Bonzini 5829c50d8ae3SPaolo Bonzini /* 5830c50d8ae3SPaolo Bonzini * Assume that the pte write on a page table of the same type 5831c50d8ae3SPaolo Bonzini * as the current vcpu paging mode since we update the sptes only 5832c50d8ae3SPaolo Bonzini * when they have the same mode. 5833c50d8ae3SPaolo Bonzini */ 5834c50d8ae3SPaolo Bonzini if (is_pae(vcpu) && *bytes == 4) { 5835c50d8ae3SPaolo Bonzini /* Handle a 32-bit guest writing two halves of a 64-bit gpte */ 5836c50d8ae3SPaolo Bonzini *gpa &= ~(gpa_t)7; 5837c50d8ae3SPaolo Bonzini *bytes = 8; 5838c50d8ae3SPaolo Bonzini } 5839c50d8ae3SPaolo Bonzini 5840c50d8ae3SPaolo Bonzini if (*bytes == 4 || *bytes == 8) { 5841c50d8ae3SPaolo Bonzini r = kvm_vcpu_read_guest_atomic(vcpu, *gpa, &gentry, *bytes); 5842c50d8ae3SPaolo Bonzini if (r) 5843c50d8ae3SPaolo Bonzini gentry = 0; 5844c50d8ae3SPaolo Bonzini } 5845c50d8ae3SPaolo Bonzini 5846c50d8ae3SPaolo Bonzini return gentry; 5847c50d8ae3SPaolo Bonzini } 5848c50d8ae3SPaolo Bonzini 5849c50d8ae3SPaolo Bonzini /* 5850c50d8ae3SPaolo Bonzini * If we're seeing too many writes to a page, it may no longer be a page table, 5851c50d8ae3SPaolo Bonzini * or we may be forking, in which case it is better to unmap the page. 5852c50d8ae3SPaolo Bonzini */ 5853c50d8ae3SPaolo Bonzini static bool detect_write_flooding(struct kvm_mmu_page *sp) 5854c50d8ae3SPaolo Bonzini { 5855c50d8ae3SPaolo Bonzini /* 5856c50d8ae3SPaolo Bonzini * Skip write-flooding detected for the sp whose level is 1, because 5857c50d8ae3SPaolo Bonzini * it can become unsync, then the guest page is not write-protected. 5858c50d8ae3SPaolo Bonzini */ 58593bae0459SSean Christopherson if (sp->role.level == PG_LEVEL_4K) 5860c50d8ae3SPaolo Bonzini return false; 5861c50d8ae3SPaolo Bonzini 5862c50d8ae3SPaolo Bonzini atomic_inc(&sp->write_flooding_count); 5863c50d8ae3SPaolo Bonzini return atomic_read(&sp->write_flooding_count) >= 3; 5864c50d8ae3SPaolo Bonzini } 5865c50d8ae3SPaolo Bonzini 5866c50d8ae3SPaolo Bonzini /* 5867c50d8ae3SPaolo Bonzini * Misaligned accesses are too much trouble to fix up; also, they usually 5868c50d8ae3SPaolo Bonzini * indicate a page is not used as a page table. 5869c50d8ae3SPaolo Bonzini */ 5870c50d8ae3SPaolo Bonzini static bool detect_write_misaligned(struct kvm_mmu_page *sp, gpa_t gpa, 5871c50d8ae3SPaolo Bonzini int bytes) 5872c50d8ae3SPaolo Bonzini { 5873c50d8ae3SPaolo Bonzini unsigned offset, pte_size, misaligned; 5874c50d8ae3SPaolo Bonzini 5875c50d8ae3SPaolo Bonzini offset = offset_in_page(gpa); 5876bb3b394dSLai Jiangshan pte_size = sp->role.has_4_byte_gpte ? 4 : 8; 5877c50d8ae3SPaolo Bonzini 5878c50d8ae3SPaolo Bonzini /* 5879c50d8ae3SPaolo Bonzini * Sometimes, the OS only writes the last one bytes to update status 5880c50d8ae3SPaolo Bonzini * bits, for example, in linux, andb instruction is used in clear_bit(). 5881c50d8ae3SPaolo Bonzini */ 5882c50d8ae3SPaolo Bonzini if (!(offset & (pte_size - 1)) && bytes == 1) 5883c50d8ae3SPaolo Bonzini return false; 5884c50d8ae3SPaolo Bonzini 5885c50d8ae3SPaolo Bonzini misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1); 5886c50d8ae3SPaolo Bonzini misaligned |= bytes < 4; 5887c50d8ae3SPaolo Bonzini 5888c50d8ae3SPaolo Bonzini return misaligned; 5889c50d8ae3SPaolo Bonzini } 5890c50d8ae3SPaolo Bonzini 5891c50d8ae3SPaolo Bonzini static u64 *get_written_sptes(struct kvm_mmu_page *sp, gpa_t gpa, int *nspte) 5892c50d8ae3SPaolo Bonzini { 5893c50d8ae3SPaolo Bonzini unsigned page_offset, quadrant; 5894c50d8ae3SPaolo Bonzini u64 *spte; 5895c50d8ae3SPaolo Bonzini int level; 5896c50d8ae3SPaolo Bonzini 5897c50d8ae3SPaolo Bonzini page_offset = offset_in_page(gpa); 5898c50d8ae3SPaolo Bonzini level = sp->role.level; 5899c50d8ae3SPaolo Bonzini *nspte = 1; 5900bb3b394dSLai Jiangshan if (sp->role.has_4_byte_gpte) { 5901c50d8ae3SPaolo Bonzini page_offset <<= 1; /* 32->64 */ 5902c50d8ae3SPaolo Bonzini /* 5903c50d8ae3SPaolo Bonzini * A 32-bit pde maps 4MB while the shadow pdes map 5904c50d8ae3SPaolo Bonzini * only 2MB. So we need to double the offset again 5905c50d8ae3SPaolo Bonzini * and zap two pdes instead of one. 5906c50d8ae3SPaolo Bonzini */ 5907c50d8ae3SPaolo Bonzini if (level == PT32_ROOT_LEVEL) { 5908c50d8ae3SPaolo Bonzini page_offset &= ~7; /* kill rounding error */ 5909c50d8ae3SPaolo Bonzini page_offset <<= 1; 5910c50d8ae3SPaolo Bonzini *nspte = 2; 5911c50d8ae3SPaolo Bonzini } 5912c50d8ae3SPaolo Bonzini quadrant = page_offset >> PAGE_SHIFT; 5913c50d8ae3SPaolo Bonzini page_offset &= ~PAGE_MASK; 5914c50d8ae3SPaolo Bonzini if (quadrant != sp->role.quadrant) 5915c50d8ae3SPaolo Bonzini return NULL; 5916c50d8ae3SPaolo Bonzini } 5917c50d8ae3SPaolo Bonzini 5918c50d8ae3SPaolo Bonzini spte = &sp->spt[page_offset / sizeof(*spte)]; 5919c50d8ae3SPaolo Bonzini return spte; 5920c50d8ae3SPaolo Bonzini } 5921c50d8ae3SPaolo Bonzini 592293284446SSean Christopherson void kvm_mmu_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new, 592393284446SSean Christopherson int bytes) 5924c50d8ae3SPaolo Bonzini { 5925c50d8ae3SPaolo Bonzini gfn_t gfn = gpa >> PAGE_SHIFT; 5926c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 5927c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 5928c50d8ae3SPaolo Bonzini u64 entry, gentry, *spte; 5929c50d8ae3SPaolo Bonzini int npte; 593006152b2dSLai Jiangshan bool flush = false; 5931c50d8ae3SPaolo Bonzini 5932c50d8ae3SPaolo Bonzini /* 5933226d9b8fSSean Christopherson * When emulating guest writes, ensure the written value is visible to 5934226d9b8fSSean Christopherson * any task that is handling page faults before checking whether or not 5935226d9b8fSSean Christopherson * KVM is shadowing a guest PTE. This ensures either KVM will create 5936226d9b8fSSean Christopherson * the correct SPTE in the page fault handler, or this task will see 5937226d9b8fSSean Christopherson * a non-zero indirect_shadow_pages. Pairs with the smp_mb() in 5938226d9b8fSSean Christopherson * account_shadowed(). 5939c50d8ae3SPaolo Bonzini */ 5940226d9b8fSSean Christopherson smp_mb(); 5941226d9b8fSSean Christopherson if (!vcpu->kvm->arch.indirect_shadow_pages) 5942c50d8ae3SPaolo Bonzini return; 5943c50d8ae3SPaolo Bonzini 5944531810caSBen Gardon write_lock(&vcpu->kvm->mmu_lock); 5945c50d8ae3SPaolo Bonzini 5946c50d8ae3SPaolo Bonzini gentry = mmu_pte_write_fetch_gpte(vcpu, &gpa, &bytes); 5947c50d8ae3SPaolo Bonzini 5948c50d8ae3SPaolo Bonzini ++vcpu->kvm->stat.mmu_pte_write; 5949c50d8ae3SPaolo Bonzini 5950767d8d8dSLai Jiangshan for_each_gfn_valid_sp_with_gptes(vcpu->kvm, sp, gfn) { 5951c50d8ae3SPaolo Bonzini if (detect_write_misaligned(sp, gpa, bytes) || 5952c50d8ae3SPaolo Bonzini detect_write_flooding(sp)) { 5953c50d8ae3SPaolo Bonzini kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list); 5954c50d8ae3SPaolo Bonzini ++vcpu->kvm->stat.mmu_flooded; 5955c50d8ae3SPaolo Bonzini continue; 5956c50d8ae3SPaolo Bonzini } 5957c50d8ae3SPaolo Bonzini 5958c50d8ae3SPaolo Bonzini spte = get_written_sptes(sp, gpa, &npte); 5959c50d8ae3SPaolo Bonzini if (!spte) 5960c50d8ae3SPaolo Bonzini continue; 5961c50d8ae3SPaolo Bonzini 5962c50d8ae3SPaolo Bonzini while (npte--) { 5963c50d8ae3SPaolo Bonzini entry = *spte; 59642de4085cSBen Gardon mmu_page_zap_pte(vcpu->kvm, sp, spte, NULL); 5965c5e2184dSSean Christopherson if (gentry && sp->role.level != PG_LEVEL_4K) 5966c5e2184dSSean Christopherson ++vcpu->kvm->stat.mmu_pde_zapped; 59671441ca14SJunaid Shahid if (is_shadow_present_pte(entry)) 596806152b2dSLai Jiangshan flush = true; 5969c50d8ae3SPaolo Bonzini ++spte; 5970c50d8ae3SPaolo Bonzini } 5971c50d8ae3SPaolo Bonzini } 597206152b2dSLai Jiangshan kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, flush); 5973531810caSBen Gardon write_unlock(&vcpu->kvm->mmu_lock); 5974c50d8ae3SPaolo Bonzini } 5975c50d8ae3SPaolo Bonzini 59761075d41eSSean Christopherson int noinline kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, u64 error_code, 5977c50d8ae3SPaolo Bonzini void *insn, int insn_len) 5978c50d8ae3SPaolo Bonzini { 597992daa48bSSean Christopherson int r, emulation_type = EMULTYPE_PF; 5980347a0d0dSPaolo Bonzini bool direct = vcpu->arch.mmu->root_role.direct; 5981c50d8ae3SPaolo Bonzini 598220ba462dSSean Christopherson if (WARN_ON_ONCE(!VALID_PAGE(vcpu->arch.mmu->root.hpa))) 5983ddce6208SSean Christopherson return RET_PF_RETRY; 5984ddce6208SSean Christopherson 5985b3d5dc62SSean Christopherson /* 5986b3d5dc62SSean Christopherson * Except for reserved faults (emulated MMIO is shared-only), set the 5987b3d5dc62SSean Christopherson * PFERR_PRIVATE_ACCESS flag for software-protected VMs based on the gfn's 5988b3d5dc62SSean Christopherson * current attributes, which are the source of truth for such VMs. Note, 5989b3d5dc62SSean Christopherson * this wrong for nested MMUs as the GPA is an L2 GPA, but KVM doesn't 5990b3d5dc62SSean Christopherson * currently supported nested virtualization (among many other things) 5991b3d5dc62SSean Christopherson * for software-protected VMs. 5992b3d5dc62SSean Christopherson */ 5993b3d5dc62SSean Christopherson if (IS_ENABLED(CONFIG_KVM_SW_PROTECTED_VM) && 5994b3d5dc62SSean Christopherson !(error_code & PFERR_RSVD_MASK) && 5995b3d5dc62SSean Christopherson vcpu->kvm->arch.vm_type == KVM_X86_SW_PROTECTED_VM && 5996b3d5dc62SSean Christopherson kvm_mem_is_private(vcpu->kvm, gpa_to_gfn(cr2_or_gpa))) 5997b3d5dc62SSean Christopherson error_code |= PFERR_PRIVATE_ACCESS; 5998b3d5dc62SSean Christopherson 5999c50d8ae3SPaolo Bonzini r = RET_PF_INVALID; 6000c50d8ae3SPaolo Bonzini if (unlikely(error_code & PFERR_RSVD_MASK)) { 600107702e5aSSean Christopherson if (WARN_ON_ONCE(error_code & PFERR_PRIVATE_ACCESS)) 600207702e5aSSean Christopherson return -EFAULT; 600307702e5aSSean Christopherson 6004736c291cSSean Christopherson r = handle_mmio_page_fault(vcpu, cr2_or_gpa, direct); 6005c50d8ae3SPaolo Bonzini if (r == RET_PF_EMULATE) 6006c50d8ae3SPaolo Bonzini goto emulate; 6007c50d8ae3SPaolo Bonzini } 6008c50d8ae3SPaolo Bonzini 6009c50d8ae3SPaolo Bonzini if (r == RET_PF_INVALID) { 60105186ec22SSean Christopherson vcpu->stat.pf_taken++; 60115186ec22SSean Christopherson 6012c9710130SIsaku Yamahata r = kvm_mmu_do_page_fault(vcpu, cr2_or_gpa, error_code, false, 601358ef2469SPaolo Bonzini &emulation_type, NULL); 601419025e7bSSean Christopherson if (KVM_BUG_ON(r == RET_PF_INVALID, vcpu->kvm)) 60157b367bc9SSean Christopherson return -EIO; 6016c50d8ae3SPaolo Bonzini } 6017c50d8ae3SPaolo Bonzini 6018c50d8ae3SPaolo Bonzini if (r < 0) 6019c50d8ae3SPaolo Bonzini return r; 6020f5e7f00cSSean Christopherson 6021f5e7f00cSSean Christopherson if (r == RET_PF_FIXED) 6022f5e7f00cSSean Christopherson vcpu->stat.pf_fixed++; 6023f5e7f00cSSean Christopherson else if (r == RET_PF_EMULATE) 6024f5e7f00cSSean Christopherson vcpu->stat.pf_emulate++; 6025f5e7f00cSSean Christopherson else if (r == RET_PF_SPURIOUS) 6026f5e7f00cSSean Christopherson vcpu->stat.pf_spurious++; 6027f5e7f00cSSean Christopherson 602883a2ba4cSSean Christopherson if (r != RET_PF_EMULATE) 602983a2ba4cSSean Christopherson return 1; 6030c50d8ae3SPaolo Bonzini 6031c50d8ae3SPaolo Bonzini /* 6032c50d8ae3SPaolo Bonzini * Before emulating the instruction, check if the error code 6033c50d8ae3SPaolo Bonzini * was due to a RO violation while translating the guest page. 6034c50d8ae3SPaolo Bonzini * This can occur when using nested virtualization with nested 6035c50d8ae3SPaolo Bonzini * paging in both guests. If true, we simply unprotect the page 6036c50d8ae3SPaolo Bonzini * and resume the guest. 6037c50d8ae3SPaolo Bonzini */ 6038347a0d0dSPaolo Bonzini if (vcpu->arch.mmu->root_role.direct && 6039c50d8ae3SPaolo Bonzini (error_code & PFERR_NESTED_GUEST_PAGE) == PFERR_NESTED_GUEST_PAGE) { 6040736c291cSSean Christopherson kvm_mmu_unprotect_page(vcpu->kvm, gpa_to_gfn(cr2_or_gpa)); 6041c50d8ae3SPaolo Bonzini return 1; 6042c50d8ae3SPaolo Bonzini } 6043c50d8ae3SPaolo Bonzini 6044c50d8ae3SPaolo Bonzini /* 6045c50d8ae3SPaolo Bonzini * vcpu->arch.mmu.page_fault returned RET_PF_EMULATE, but we can still 6046c50d8ae3SPaolo Bonzini * optimistically try to just unprotect the page and let the processor 6047c50d8ae3SPaolo Bonzini * re-execute the instruction that caused the page fault. Do not allow 6048c50d8ae3SPaolo Bonzini * retrying MMIO emulation, as it's not only pointless but could also 6049c50d8ae3SPaolo Bonzini * cause us to enter an infinite loop because the processor will keep 6050c50d8ae3SPaolo Bonzini * faulting on the non-existent MMIO address. Retrying an instruction 6051c50d8ae3SPaolo Bonzini * from a nested guest is also pointless and dangerous as we are only 6052c50d8ae3SPaolo Bonzini * explicitly shadowing L1's page tables, i.e. unprotecting something 6053c50d8ae3SPaolo Bonzini * for L1 isn't going to magically fix whatever issue cause L2 to fail. 6054c50d8ae3SPaolo Bonzini */ 6055736c291cSSean Christopherson if (!mmio_info_in_cache(vcpu, cr2_or_gpa, direct) && !is_guest_mode(vcpu)) 605692daa48bSSean Christopherson emulation_type |= EMULTYPE_ALLOW_RETRY_PF; 6057c50d8ae3SPaolo Bonzini emulate: 6058736c291cSSean Christopherson return x86_emulate_instruction(vcpu, cr2_or_gpa, emulation_type, insn, 6059c50d8ae3SPaolo Bonzini insn_len); 6060c50d8ae3SPaolo Bonzini } 6061c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_mmu_page_fault); 6062c50d8ae3SPaolo Bonzini 6063bca99c03SSean Christopherson void kvm_mmu_print_sptes(struct kvm_vcpu *vcpu, gpa_t gpa, const char *msg) 6064bca99c03SSean Christopherson { 6065bca99c03SSean Christopherson u64 sptes[PT64_ROOT_MAX_LEVEL + 1]; 6066bca99c03SSean Christopherson int root_level, leaf, level; 6067bca99c03SSean Christopherson 6068bca99c03SSean Christopherson leaf = get_sptes_lockless(vcpu, gpa, sptes, &root_level); 6069bca99c03SSean Christopherson if (unlikely(leaf < 0)) 6070bca99c03SSean Christopherson return; 6071bca99c03SSean Christopherson 6072bca99c03SSean Christopherson pr_err("%s %llx", msg, gpa); 6073bca99c03SSean Christopherson for (level = root_level; level >= leaf; level--) 6074bca99c03SSean Christopherson pr_cont(", spte[%d] = 0x%llx", level, sptes[level]); 6075bca99c03SSean Christopherson pr_cont("\n"); 6076bca99c03SSean Christopherson } 6077bca99c03SSean Christopherson EXPORT_SYMBOL_GPL(kvm_mmu_print_sptes); 6078bca99c03SSean Christopherson 60799fd4a4e3SLai Jiangshan static void __kvm_mmu_invalidate_addr(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu, 60809fd4a4e3SLai Jiangshan u64 addr, hpa_t root_hpa) 60819fd4a4e3SLai Jiangshan { 60829fd4a4e3SLai Jiangshan struct kvm_shadow_walk_iterator iterator; 60839fd4a4e3SLai Jiangshan 60849fd4a4e3SLai Jiangshan vcpu_clear_mmio_info(vcpu, addr); 60859fd4a4e3SLai Jiangshan 6086762b33ebSLike Xu /* 6087762b33ebSLike Xu * Walking and synchronizing SPTEs both assume they are operating in 6088762b33ebSLike Xu * the context of the current MMU, and would need to be reworked if 6089762b33ebSLike Xu * this is ever used to sync the guest_mmu, e.g. to emulate INVEPT. 6090762b33ebSLike Xu */ 6091762b33ebSLike Xu if (WARN_ON_ONCE(mmu != vcpu->arch.mmu)) 6092762b33ebSLike Xu return; 6093762b33ebSLike Xu 60949fd4a4e3SLai Jiangshan if (!VALID_PAGE(root_hpa)) 60959fd4a4e3SLai Jiangshan return; 60969fd4a4e3SLai Jiangshan 60979fd4a4e3SLai Jiangshan write_lock(&vcpu->kvm->mmu_lock); 60989fd4a4e3SLai Jiangshan for_each_shadow_entry_using_root(vcpu, root_hpa, addr, iterator) { 60999fd4a4e3SLai Jiangshan struct kvm_mmu_page *sp = sptep_to_sp(iterator.sptep); 61009fd4a4e3SLai Jiangshan 61019fd4a4e3SLai Jiangshan if (sp->unsync) { 610219ace7d6SLai Jiangshan int ret = kvm_sync_spte(vcpu, sp, iterator.index); 61039fd4a4e3SLai Jiangshan 61049fd4a4e3SLai Jiangshan if (ret < 0) 61059fd4a4e3SLai Jiangshan mmu_page_zap_pte(vcpu->kvm, sp, iterator.sptep, NULL); 61069fd4a4e3SLai Jiangshan if (ret) 61079fd4a4e3SLai Jiangshan kvm_flush_remote_tlbs_sptep(vcpu->kvm, iterator.sptep); 61089fd4a4e3SLai Jiangshan } 61099fd4a4e3SLai Jiangshan 61109fd4a4e3SLai Jiangshan if (!sp->unsync_children) 61119fd4a4e3SLai Jiangshan break; 61129fd4a4e3SLai Jiangshan } 61139fd4a4e3SLai Jiangshan write_unlock(&vcpu->kvm->mmu_lock); 61149fd4a4e3SLai Jiangshan } 61159fd4a4e3SLai Jiangshan 6116753b43c9SLai Jiangshan void kvm_mmu_invalidate_addr(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu, 6117cd42853eSLai Jiangshan u64 addr, unsigned long roots) 6118c50d8ae3SPaolo Bonzini { 6119c50d8ae3SPaolo Bonzini int i; 6120c50d8ae3SPaolo Bonzini 6121cd42853eSLai Jiangshan WARN_ON_ONCE(roots & ~KVM_MMU_ROOTS_ALL); 6122cd42853eSLai Jiangshan 61235efac074SPaolo Bonzini /* It's actually a GPA for vcpu->arch.guest_mmu. */ 61245efac074SPaolo Bonzini if (mmu != &vcpu->arch.guest_mmu) { 61255efac074SPaolo Bonzini /* INVLPG on a non-canonical address is a NOP according to the SDM. */ 6126753b43c9SLai Jiangshan if (is_noncanonical_address(addr, vcpu)) 6127c50d8ae3SPaolo Bonzini return; 6128c50d8ae3SPaolo Bonzini 6129753b43c9SLai Jiangshan static_call(kvm_x86_flush_tlb_gva)(vcpu, addr); 61305efac074SPaolo Bonzini } 61315efac074SPaolo Bonzini 61329fd4a4e3SLai Jiangshan if (!mmu->sync_spte) 61335efac074SPaolo Bonzini return; 61345efac074SPaolo Bonzini 6135cd42853eSLai Jiangshan if (roots & KVM_MMU_ROOT_CURRENT) 61369fd4a4e3SLai Jiangshan __kvm_mmu_invalidate_addr(vcpu, mmu, addr, mmu->root.hpa); 6137c50d8ae3SPaolo Bonzini 6138cd42853eSLai Jiangshan for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) { 6139ed335278SLai Jiangshan if (roots & KVM_MMU_ROOT_PREVIOUS(i)) 61409fd4a4e3SLai Jiangshan __kvm_mmu_invalidate_addr(vcpu, mmu, addr, mmu->prev_roots[i].hpa); 61415efac074SPaolo Bonzini } 61425efac074SPaolo Bonzini } 61432c86c444SLai Jiangshan EXPORT_SYMBOL_GPL(kvm_mmu_invalidate_addr); 6144c50d8ae3SPaolo Bonzini 61455efac074SPaolo Bonzini void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva) 61465efac074SPaolo Bonzini { 6147cd42853eSLai Jiangshan /* 6148cd42853eSLai Jiangshan * INVLPG is required to invalidate any global mappings for the VA, 6149cd42853eSLai Jiangshan * irrespective of PCID. Blindly sync all roots as it would take 6150cd42853eSLai Jiangshan * roughly the same amount of work/time to determine whether any of the 6151cd42853eSLai Jiangshan * previous roots have a global mapping. 6152cd42853eSLai Jiangshan * 6153cd42853eSLai Jiangshan * Mappings not reachable via the current or previous cached roots will 6154cd42853eSLai Jiangshan * be synced when switching to that new cr3, so nothing needs to be 6155cd42853eSLai Jiangshan * done here for them. 6156cd42853eSLai Jiangshan */ 6157cd42853eSLai Jiangshan kvm_mmu_invalidate_addr(vcpu, vcpu->arch.walk_mmu, gva, KVM_MMU_ROOTS_ALL); 6158c50d8ae3SPaolo Bonzini ++vcpu->stat.invlpg; 6159c50d8ae3SPaolo Bonzini } 6160c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_mmu_invlpg); 6161c50d8ae3SPaolo Bonzini 61625efac074SPaolo Bonzini 6163c50d8ae3SPaolo Bonzini void kvm_mmu_invpcid_gva(struct kvm_vcpu *vcpu, gva_t gva, unsigned long pcid) 6164c50d8ae3SPaolo Bonzini { 6165c50d8ae3SPaolo Bonzini struct kvm_mmu *mmu = vcpu->arch.mmu; 61669ebc3f51SLai Jiangshan unsigned long roots = 0; 6167c50d8ae3SPaolo Bonzini uint i; 6168c50d8ae3SPaolo Bonzini 61699ebc3f51SLai Jiangshan if (pcid == kvm_get_active_pcid(vcpu)) 61709ebc3f51SLai Jiangshan roots |= KVM_MMU_ROOT_CURRENT; 6171c50d8ae3SPaolo Bonzini 6172c50d8ae3SPaolo Bonzini for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) { 6173c50d8ae3SPaolo Bonzini if (VALID_PAGE(mmu->prev_roots[i].hpa) && 61749ebc3f51SLai Jiangshan pcid == kvm_get_pcid(vcpu, mmu->prev_roots[i].pgd)) 61759ebc3f51SLai Jiangshan roots |= KVM_MMU_ROOT_PREVIOUS(i); 6176c50d8ae3SPaolo Bonzini } 6177c50d8ae3SPaolo Bonzini 61789ebc3f51SLai Jiangshan if (roots) 61799ebc3f51SLai Jiangshan kvm_mmu_invalidate_addr(vcpu, mmu, gva, roots); 6180c50d8ae3SPaolo Bonzini ++vcpu->stat.invlpg; 6181c50d8ae3SPaolo Bonzini 6182c50d8ae3SPaolo Bonzini /* 6183c50d8ae3SPaolo Bonzini * Mappings not reachable via the current cr3 or the prev_roots will be 6184c50d8ae3SPaolo Bonzini * synced when switching to that cr3, so nothing needs to be done here 6185c50d8ae3SPaolo Bonzini * for them. 6186c50d8ae3SPaolo Bonzini */ 6187c50d8ae3SPaolo Bonzini } 6188c50d8ae3SPaolo Bonzini 6189746700d2SWei Huang void kvm_configure_mmu(bool enable_tdp, int tdp_forced_root_level, 6190746700d2SWei Huang int tdp_max_root_level, int tdp_huge_page_level) 6191c50d8ae3SPaolo Bonzini { 6192bde77235SSean Christopherson tdp_enabled = enable_tdp; 6193746700d2SWei Huang tdp_root_level = tdp_forced_root_level; 619483013059SSean Christopherson max_tdp_level = tdp_max_root_level; 6195703c335dSSean Christopherson 61961f98f2bdSDavid Matlack #ifdef CONFIG_X86_64 61971f98f2bdSDavid Matlack tdp_mmu_enabled = tdp_mmu_allowed && tdp_enabled; 61981f98f2bdSDavid Matlack #endif 6199703c335dSSean Christopherson /* 62001d92d2e8SSean Christopherson * max_huge_page_level reflects KVM's MMU capabilities irrespective 6201703c335dSSean Christopherson * of kernel support, e.g. KVM may be capable of using 1GB pages when 6202703c335dSSean Christopherson * the kernel is not. But, KVM never creates a page size greater than 6203703c335dSSean Christopherson * what is used by the kernel for any given HVA, i.e. the kernel's 6204703c335dSSean Christopherson * capabilities are ultimately consulted by kvm_mmu_hugepage_adjust(). 6205703c335dSSean Christopherson */ 6206703c335dSSean Christopherson if (tdp_enabled) 62071d92d2e8SSean Christopherson max_huge_page_level = tdp_huge_page_level; 6208703c335dSSean Christopherson else if (boot_cpu_has(X86_FEATURE_GBPAGES)) 62091d92d2e8SSean Christopherson max_huge_page_level = PG_LEVEL_1G; 6210703c335dSSean Christopherson else 62111d92d2e8SSean Christopherson max_huge_page_level = PG_LEVEL_2M; 6212c50d8ae3SPaolo Bonzini } 6213bde77235SSean Christopherson EXPORT_SYMBOL_GPL(kvm_configure_mmu); 6214c50d8ae3SPaolo Bonzini 6215c50d8ae3SPaolo Bonzini /* The return value indicates if tlb flush on all vcpus is needed. */ 6216727ae377SSean Christopherson typedef bool (*slot_rmaps_handler) (struct kvm *kvm, 6217269e9552SHamza Mahfooz struct kvm_rmap_head *rmap_head, 6218269e9552SHamza Mahfooz const struct kvm_memory_slot *slot); 6219c50d8ae3SPaolo Bonzini 6220727ae377SSean Christopherson static __always_inline bool __walk_slot_rmaps(struct kvm *kvm, 6221727ae377SSean Christopherson const struct kvm_memory_slot *slot, 6222727ae377SSean Christopherson slot_rmaps_handler fn, 6223727ae377SSean Christopherson int start_level, int end_level, 6224727ae377SSean Christopherson gfn_t start_gfn, gfn_t end_gfn, 6225727ae377SSean Christopherson bool flush_on_yield, bool flush) 6226c50d8ae3SPaolo Bonzini { 6227c50d8ae3SPaolo Bonzini struct slot_rmap_walk_iterator iterator; 6228c50d8ae3SPaolo Bonzini 6229eddd9e83SSean Christopherson lockdep_assert_held_write(&kvm->mmu_lock); 6230eddd9e83SSean Christopherson 6231727ae377SSean Christopherson for_each_slot_rmap_range(slot, start_level, end_level, start_gfn, 6232c50d8ae3SPaolo Bonzini end_gfn, &iterator) { 6233c50d8ae3SPaolo Bonzini if (iterator.rmap) 6234727ae377SSean Christopherson flush |= fn(kvm, iterator.rmap, slot); 6235c50d8ae3SPaolo Bonzini 6236531810caSBen Gardon if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) { 6237302695a5SSean Christopherson if (flush && flush_on_yield) { 62388c63e8c2SDavid Matlack kvm_flush_remote_tlbs_range(kvm, start_gfn, 6239c50d8ae3SPaolo Bonzini iterator.gfn - start_gfn + 1); 6240c50d8ae3SPaolo Bonzini flush = false; 6241c50d8ae3SPaolo Bonzini } 6242531810caSBen Gardon cond_resched_rwlock_write(&kvm->mmu_lock); 6243c50d8ae3SPaolo Bonzini } 6244c50d8ae3SPaolo Bonzini } 6245c50d8ae3SPaolo Bonzini 6246c50d8ae3SPaolo Bonzini return flush; 6247c50d8ae3SPaolo Bonzini } 6248c50d8ae3SPaolo Bonzini 6249727ae377SSean Christopherson static __always_inline bool walk_slot_rmaps(struct kvm *kvm, 6250727ae377SSean Christopherson const struct kvm_memory_slot *slot, 6251727ae377SSean Christopherson slot_rmaps_handler fn, 6252727ae377SSean Christopherson int start_level, int end_level, 6253302695a5SSean Christopherson bool flush_on_yield) 6254c50d8ae3SPaolo Bonzini { 6255727ae377SSean Christopherson return __walk_slot_rmaps(kvm, slot, fn, start_level, end_level, 6256727ae377SSean Christopherson slot->base_gfn, slot->base_gfn + slot->npages - 1, 62571a61b7dbSSean Christopherson flush_on_yield, false); 6258c50d8ae3SPaolo Bonzini } 6259c50d8ae3SPaolo Bonzini 6260727ae377SSean Christopherson static __always_inline bool walk_slot_rmaps_4k(struct kvm *kvm, 6261727ae377SSean Christopherson const struct kvm_memory_slot *slot, 6262727ae377SSean Christopherson slot_rmaps_handler fn, 6263727ae377SSean Christopherson bool flush_on_yield) 6264c50d8ae3SPaolo Bonzini { 6265727ae377SSean Christopherson return walk_slot_rmaps(kvm, slot, fn, PG_LEVEL_4K, PG_LEVEL_4K, flush_on_yield); 6266c50d8ae3SPaolo Bonzini } 6267c50d8ae3SPaolo Bonzini 6268c50d8ae3SPaolo Bonzini static void free_mmu_pages(struct kvm_mmu *mmu) 6269c50d8ae3SPaolo Bonzini { 62704a98623dSSean Christopherson if (!tdp_enabled && mmu->pae_root) 62714a98623dSSean Christopherson set_memory_encrypted((unsigned long)mmu->pae_root, 1); 6272c50d8ae3SPaolo Bonzini free_page((unsigned long)mmu->pae_root); 627303ca4589SSean Christopherson free_page((unsigned long)mmu->pml4_root); 6274cb0f722aSWei Huang free_page((unsigned long)mmu->pml5_root); 6275c50d8ae3SPaolo Bonzini } 6276c50d8ae3SPaolo Bonzini 627704d28e37SSean Christopherson static int __kvm_mmu_create(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu) 6278c50d8ae3SPaolo Bonzini { 6279c50d8ae3SPaolo Bonzini struct page *page; 6280c50d8ae3SPaolo Bonzini int i; 6281c50d8ae3SPaolo Bonzini 6282b9e5603cSPaolo Bonzini mmu->root.hpa = INVALID_PAGE; 6283b9e5603cSPaolo Bonzini mmu->root.pgd = 0; 628404d28e37SSean Christopherson for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) 628504d28e37SSean Christopherson mmu->prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID; 628604d28e37SSean Christopherson 628727f4fca2SLai Jiangshan /* vcpu->arch.guest_mmu isn't used when !tdp_enabled. */ 628827f4fca2SLai Jiangshan if (!tdp_enabled && mmu == &vcpu->arch.guest_mmu) 628927f4fca2SLai Jiangshan return 0; 629027f4fca2SLai Jiangshan 6291c50d8ae3SPaolo Bonzini /* 6292c50d8ae3SPaolo Bonzini * When using PAE paging, the four PDPTEs are treated as 'root' pages, 6293c50d8ae3SPaolo Bonzini * while the PDP table is a per-vCPU construct that's allocated at MMU 6294c50d8ae3SPaolo Bonzini * creation. When emulating 32-bit mode, cr3 is only 32 bits even on 6295c50d8ae3SPaolo Bonzini * x86_64. Therefore we need to allocate the PDP table in the first 629604d45551SSean Christopherson * 4GB of memory, which happens to fit the DMA32 zone. TDP paging 629704d45551SSean Christopherson * generally doesn't use PAE paging and can skip allocating the PDP 629804d45551SSean Christopherson * table. The main exception, handled here, is SVM's 32-bit NPT. The 629904d45551SSean Christopherson * other exception is for shadowing L1's 32-bit or PAE NPT on 64-bit 630084432316SLai Jiangshan * KVM; that horror is handled on-demand by mmu_alloc_special_roots(). 6301c50d8ae3SPaolo Bonzini */ 6302d468d94bSSean Christopherson if (tdp_enabled && kvm_mmu_get_tdp_level(vcpu) > PT32E_ROOT_LEVEL) 6303c50d8ae3SPaolo Bonzini return 0; 6304c50d8ae3SPaolo Bonzini 6305c50d8ae3SPaolo Bonzini page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_DMA32); 6306c50d8ae3SPaolo Bonzini if (!page) 6307c50d8ae3SPaolo Bonzini return -ENOMEM; 6308c50d8ae3SPaolo Bonzini 6309c50d8ae3SPaolo Bonzini mmu->pae_root = page_address(page); 63104a98623dSSean Christopherson 63114a98623dSSean Christopherson /* 63124a98623dSSean Christopherson * CR3 is only 32 bits when PAE paging is used, thus it's impossible to 63134a98623dSSean Christopherson * get the CPU to treat the PDPTEs as encrypted. Decrypt the page so 63144a98623dSSean Christopherson * that KVM's writes and the CPU's reads get along. Note, this is 63154a98623dSSean Christopherson * only necessary when using shadow paging, as 64-bit NPT can get at 63164a98623dSSean Christopherson * the C-bit even when shadowing 32-bit NPT, and SME isn't supported 63174a98623dSSean Christopherson * by 32-bit kernels (when KVM itself uses 32-bit NPT). 63184a98623dSSean Christopherson */ 63194a98623dSSean Christopherson if (!tdp_enabled) 63204a98623dSSean Christopherson set_memory_decrypted((unsigned long)mmu->pae_root, 1); 63214a98623dSSean Christopherson else 6322e54f1ff2SKai Huang WARN_ON_ONCE(shadow_me_value); 63234a98623dSSean Christopherson 6324c50d8ae3SPaolo Bonzini for (i = 0; i < 4; ++i) 6325c834e5e4SSean Christopherson mmu->pae_root[i] = INVALID_PAE_ROOT; 6326c50d8ae3SPaolo Bonzini 6327c50d8ae3SPaolo Bonzini return 0; 6328c50d8ae3SPaolo Bonzini } 6329c50d8ae3SPaolo Bonzini 6330c50d8ae3SPaolo Bonzini int kvm_mmu_create(struct kvm_vcpu *vcpu) 6331c50d8ae3SPaolo Bonzini { 6332c50d8ae3SPaolo Bonzini int ret; 6333c50d8ae3SPaolo Bonzini 63345962bfb7SSean Christopherson vcpu->arch.mmu_pte_list_desc_cache.kmem_cache = pte_list_desc_cache; 63355f6078f9SSean Christopherson vcpu->arch.mmu_pte_list_desc_cache.gfp_zero = __GFP_ZERO; 63365f6078f9SSean Christopherson 63375962bfb7SSean Christopherson vcpu->arch.mmu_page_header_cache.kmem_cache = mmu_page_header_cache; 63385f6078f9SSean Christopherson vcpu->arch.mmu_page_header_cache.gfp_zero = __GFP_ZERO; 63395962bfb7SSean Christopherson 6340d8fa2031SSean Christopherson vcpu->arch.mmu_shadow_page_cache.init_value = 6341d8fa2031SSean Christopherson SHADOW_NONPRESENT_VALUE; 6342d8fa2031SSean Christopherson if (!vcpu->arch.mmu_shadow_page_cache.init_value) 634396880883SSean Christopherson vcpu->arch.mmu_shadow_page_cache.gfp_zero = __GFP_ZERO; 634496880883SSean Christopherson 6345c50d8ae3SPaolo Bonzini vcpu->arch.mmu = &vcpu->arch.root_mmu; 6346c50d8ae3SPaolo Bonzini vcpu->arch.walk_mmu = &vcpu->arch.root_mmu; 6347c50d8ae3SPaolo Bonzini 634804d28e37SSean Christopherson ret = __kvm_mmu_create(vcpu, &vcpu->arch.guest_mmu); 6349c50d8ae3SPaolo Bonzini if (ret) 6350c50d8ae3SPaolo Bonzini return ret; 6351c50d8ae3SPaolo Bonzini 635204d28e37SSean Christopherson ret = __kvm_mmu_create(vcpu, &vcpu->arch.root_mmu); 6353c50d8ae3SPaolo Bonzini if (ret) 6354c50d8ae3SPaolo Bonzini goto fail_allocate_root; 6355c50d8ae3SPaolo Bonzini 6356c50d8ae3SPaolo Bonzini return ret; 6357c50d8ae3SPaolo Bonzini fail_allocate_root: 6358c50d8ae3SPaolo Bonzini free_mmu_pages(&vcpu->arch.guest_mmu); 6359c50d8ae3SPaolo Bonzini return ret; 6360c50d8ae3SPaolo Bonzini } 6361c50d8ae3SPaolo Bonzini 6362c50d8ae3SPaolo Bonzini #define BATCH_ZAP_PAGES 10 6363c50d8ae3SPaolo Bonzini static void kvm_zap_obsolete_pages(struct kvm *kvm) 6364c50d8ae3SPaolo Bonzini { 6365c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, *node; 6366c50d8ae3SPaolo Bonzini int nr_zapped, batch = 0; 6367b28cb0cdSSean Christopherson bool unstable; 6368c50d8ae3SPaolo Bonzini 6369c50d8ae3SPaolo Bonzini restart: 6370c50d8ae3SPaolo Bonzini list_for_each_entry_safe_reverse(sp, node, 6371c50d8ae3SPaolo Bonzini &kvm->arch.active_mmu_pages, link) { 6372c50d8ae3SPaolo Bonzini /* 6373c50d8ae3SPaolo Bonzini * No obsolete valid page exists before a newly created page 6374c50d8ae3SPaolo Bonzini * since active_mmu_pages is a FIFO list. 6375c50d8ae3SPaolo Bonzini */ 6376c50d8ae3SPaolo Bonzini if (!is_obsolete_sp(kvm, sp)) 6377c50d8ae3SPaolo Bonzini break; 6378c50d8ae3SPaolo Bonzini 6379c50d8ae3SPaolo Bonzini /* 6380f95eec9bSSean Christopherson * Invalid pages should never land back on the list of active 6381f95eec9bSSean Christopherson * pages. Skip the bogus page, otherwise we'll get stuck in an 6382f95eec9bSSean Christopherson * infinite loop if the page gets put back on the list (again). 6383c50d8ae3SPaolo Bonzini */ 638420ba462dSSean Christopherson if (WARN_ON_ONCE(sp->role.invalid)) 6385c50d8ae3SPaolo Bonzini continue; 6386c50d8ae3SPaolo Bonzini 6387c50d8ae3SPaolo Bonzini /* 6388c50d8ae3SPaolo Bonzini * No need to flush the TLB since we're only zapping shadow 6389c50d8ae3SPaolo Bonzini * pages with an obsolete generation number and all vCPUS have 6390c50d8ae3SPaolo Bonzini * loaded a new root, i.e. the shadow pages being zapped cannot 6391c50d8ae3SPaolo Bonzini * be in active use by the guest. 6392c50d8ae3SPaolo Bonzini */ 6393c50d8ae3SPaolo Bonzini if (batch >= BATCH_ZAP_PAGES && 6394531810caSBen Gardon cond_resched_rwlock_write(&kvm->mmu_lock)) { 6395c50d8ae3SPaolo Bonzini batch = 0; 6396c50d8ae3SPaolo Bonzini goto restart; 6397c50d8ae3SPaolo Bonzini } 6398c50d8ae3SPaolo Bonzini 6399b28cb0cdSSean Christopherson unstable = __kvm_mmu_prepare_zap_page(kvm, sp, 6400b28cb0cdSSean Christopherson &kvm->arch.zapped_obsolete_pages, &nr_zapped); 6401c50d8ae3SPaolo Bonzini batch += nr_zapped; 6402b28cb0cdSSean Christopherson 6403b28cb0cdSSean Christopherson if (unstable) 6404c50d8ae3SPaolo Bonzini goto restart; 6405c50d8ae3SPaolo Bonzini } 6406c50d8ae3SPaolo Bonzini 6407c50d8ae3SPaolo Bonzini /* 64087ae5840eSSean Christopherson * Kick all vCPUs (via remote TLB flush) before freeing the page tables 64097ae5840eSSean Christopherson * to ensure KVM is not in the middle of a lockless shadow page table 64107ae5840eSSean Christopherson * walk, which may reference the pages. The remote TLB flush itself is 64117ae5840eSSean Christopherson * not required and is simply a convenient way to kick vCPUs as needed. 64127ae5840eSSean Christopherson * KVM performs a local TLB flush when allocating a new root (see 64137ae5840eSSean Christopherson * kvm_mmu_load()), and the reload in the caller ensure no vCPUs are 64147ae5840eSSean Christopherson * running with an obsolete MMU. 6415c50d8ae3SPaolo Bonzini */ 6416c50d8ae3SPaolo Bonzini kvm_mmu_commit_zap_page(kvm, &kvm->arch.zapped_obsolete_pages); 6417c50d8ae3SPaolo Bonzini } 6418c50d8ae3SPaolo Bonzini 6419c50d8ae3SPaolo Bonzini /* 6420c50d8ae3SPaolo Bonzini * Fast invalidate all shadow pages and use lock-break technique 6421c50d8ae3SPaolo Bonzini * to zap obsolete pages. 6422c50d8ae3SPaolo Bonzini * 6423c50d8ae3SPaolo Bonzini * It's required when memslot is being deleted or VM is being 6424c50d8ae3SPaolo Bonzini * destroyed, in these cases, we should ensure that KVM MMU does 6425c50d8ae3SPaolo Bonzini * not use any resource of the being-deleted slot or all slots 6426c50d8ae3SPaolo Bonzini * after calling the function. 6427c50d8ae3SPaolo Bonzini */ 6428c50d8ae3SPaolo Bonzini static void kvm_mmu_zap_all_fast(struct kvm *kvm) 6429c50d8ae3SPaolo Bonzini { 6430c50d8ae3SPaolo Bonzini lockdep_assert_held(&kvm->slots_lock); 6431c50d8ae3SPaolo Bonzini 6432531810caSBen Gardon write_lock(&kvm->mmu_lock); 6433c50d8ae3SPaolo Bonzini trace_kvm_mmu_zap_all_fast(kvm); 6434c50d8ae3SPaolo Bonzini 6435c50d8ae3SPaolo Bonzini /* 6436c50d8ae3SPaolo Bonzini * Toggle mmu_valid_gen between '0' and '1'. Because slots_lock is 6437c50d8ae3SPaolo Bonzini * held for the entire duration of zapping obsolete pages, it's 6438c50d8ae3SPaolo Bonzini * impossible for there to be multiple invalid generations associated 6439c50d8ae3SPaolo Bonzini * with *valid* shadow pages at any given time, i.e. there is exactly 6440c50d8ae3SPaolo Bonzini * one valid generation and (at most) one invalid generation. 6441c50d8ae3SPaolo Bonzini */ 6442c50d8ae3SPaolo Bonzini kvm->arch.mmu_valid_gen = kvm->arch.mmu_valid_gen ? 0 : 1; 6443c50d8ae3SPaolo Bonzini 64442f6f66ccSSean Christopherson /* 64452f6f66ccSSean Christopherson * In order to ensure all vCPUs drop their soon-to-be invalid roots, 64462f6f66ccSSean Christopherson * invalidating TDP MMU roots must be done while holding mmu_lock for 64472f6f66ccSSean Christopherson * write and in the same critical section as making the reload request, 64482f6f66ccSSean Christopherson * e.g. before kvm_zap_obsolete_pages() could drop mmu_lock and yield. 6449b7cccd39SBen Gardon */ 64501f98f2bdSDavid Matlack if (tdp_mmu_enabled) 6451b7cccd39SBen Gardon kvm_tdp_mmu_invalidate_all_roots(kvm); 6452b7cccd39SBen Gardon 6453c50d8ae3SPaolo Bonzini /* 6454c50d8ae3SPaolo Bonzini * Notify all vcpus to reload its shadow page table and flush TLB. 6455c50d8ae3SPaolo Bonzini * Then all vcpus will switch to new shadow page table with the new 6456c50d8ae3SPaolo Bonzini * mmu_valid_gen. 6457c50d8ae3SPaolo Bonzini * 6458c50d8ae3SPaolo Bonzini * Note: we need to do this under the protection of mmu_lock, 6459c50d8ae3SPaolo Bonzini * otherwise, vcpu would purge shadow page but miss tlb flush. 6460c50d8ae3SPaolo Bonzini */ 6461527d5cd7SSean Christopherson kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_FREE_OBSOLETE_ROOTS); 6462c50d8ae3SPaolo Bonzini 6463c50d8ae3SPaolo Bonzini kvm_zap_obsolete_pages(kvm); 6464faaf05b0SBen Gardon 6465531810caSBen Gardon write_unlock(&kvm->mmu_lock); 64664c6654bdSBen Gardon 6467f28e9c7fSSean Christopherson /* 6468f28e9c7fSSean Christopherson * Zap the invalidated TDP MMU roots, all SPTEs must be dropped before 6469f28e9c7fSSean Christopherson * returning to the caller, e.g. if the zap is in response to a memslot 6470f28e9c7fSSean Christopherson * deletion, mmu_notifier callbacks will be unable to reach the SPTEs 6471f28e9c7fSSean Christopherson * associated with the deleted memslot once the update completes, and 6472f28e9c7fSSean Christopherson * Deferring the zap until the final reference to the root is put would 6473f28e9c7fSSean Christopherson * lead to use-after-free. 6474f28e9c7fSSean Christopherson */ 64751f98f2bdSDavid Matlack if (tdp_mmu_enabled) 64764c6654bdSBen Gardon kvm_tdp_mmu_zap_invalidated_roots(kvm); 6477c50d8ae3SPaolo Bonzini } 6478c50d8ae3SPaolo Bonzini 6479c50d8ae3SPaolo Bonzini static bool kvm_has_zapped_obsolete_pages(struct kvm *kvm) 6480c50d8ae3SPaolo Bonzini { 6481c50d8ae3SPaolo Bonzini return unlikely(!list_empty_careful(&kvm->arch.zapped_obsolete_pages)); 6482c50d8ae3SPaolo Bonzini } 6483c50d8ae3SPaolo Bonzini 64840df9dab8SSean Christopherson void kvm_mmu_init_vm(struct kvm *kvm) 6485c50d8ae3SPaolo Bonzini { 6486949019b9SSean Christopherson kvm->arch.shadow_mmio_value = shadow_mmio_value; 6487a1a39128SPaolo Bonzini INIT_LIST_HEAD(&kvm->arch.active_mmu_pages); 6488a1a39128SPaolo Bonzini INIT_LIST_HEAD(&kvm->arch.zapped_obsolete_pages); 648955c510e2SSean Christopherson INIT_LIST_HEAD(&kvm->arch.possible_nx_huge_pages); 6490ce25681dSSean Christopherson spin_lock_init(&kvm->arch.mmu_unsync_pages_lock); 6491ce25681dSSean Christopherson 64920df9dab8SSean Christopherson if (tdp_mmu_enabled) 64930df9dab8SSean Christopherson kvm_mmu_init_tdp_mmu(kvm); 6494fe5db27dSBen Gardon 6495ada51a9dSDavid Matlack kvm->arch.split_page_header_cache.kmem_cache = mmu_page_header_cache; 6496ada51a9dSDavid Matlack kvm->arch.split_page_header_cache.gfp_zero = __GFP_ZERO; 6497ada51a9dSDavid Matlack 6498ada51a9dSDavid Matlack kvm->arch.split_shadow_page_cache.gfp_zero = __GFP_ZERO; 6499ada51a9dSDavid Matlack 6500ada51a9dSDavid Matlack kvm->arch.split_desc_cache.kmem_cache = pte_list_desc_cache; 6501ada51a9dSDavid Matlack kvm->arch.split_desc_cache.gfp_zero = __GFP_ZERO; 6502c50d8ae3SPaolo Bonzini } 6503c50d8ae3SPaolo Bonzini 6504ada51a9dSDavid Matlack static void mmu_free_vm_memory_caches(struct kvm *kvm) 6505ada51a9dSDavid Matlack { 6506ada51a9dSDavid Matlack kvm_mmu_free_memory_cache(&kvm->arch.split_desc_cache); 6507ada51a9dSDavid Matlack kvm_mmu_free_memory_cache(&kvm->arch.split_page_header_cache); 6508ada51a9dSDavid Matlack kvm_mmu_free_memory_cache(&kvm->arch.split_shadow_page_cache); 6509ada51a9dSDavid Matlack } 6510ada51a9dSDavid Matlack 6511c50d8ae3SPaolo Bonzini void kvm_mmu_uninit_vm(struct kvm *kvm) 6512c50d8ae3SPaolo Bonzini { 651309732d2bSDavid Matlack if (tdp_mmu_enabled) 6514fe5db27dSBen Gardon kvm_mmu_uninit_tdp_mmu(kvm); 6515ada51a9dSDavid Matlack 6516ada51a9dSDavid Matlack mmu_free_vm_memory_caches(kvm); 6517c50d8ae3SPaolo Bonzini } 6518c50d8ae3SPaolo Bonzini 65192833eda0SSean Christopherson static bool kvm_rmap_zap_gfn_range(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end) 652021fa3246SSean Christopherson { 652121fa3246SSean Christopherson const struct kvm_memory_slot *memslot; 652221fa3246SSean Christopherson struct kvm_memslots *slots; 6523f4209439SMaciej S. Szmigiero struct kvm_memslot_iter iter; 652421fa3246SSean Christopherson bool flush = false; 652521fa3246SSean Christopherson gfn_t start, end; 6526f4209439SMaciej S. Szmigiero int i; 652721fa3246SSean Christopherson 652821fa3246SSean Christopherson if (!kvm_memslots_have_rmaps(kvm)) 652921fa3246SSean Christopherson return flush; 653021fa3246SSean Christopherson 6531eed52e43SSean Christopherson for (i = 0; i < kvm_arch_nr_memslot_as_ids(kvm); i++) { 653221fa3246SSean Christopherson slots = __kvm_memslots(kvm, i); 6533f4209439SMaciej S. Szmigiero 6534f4209439SMaciej S. Szmigiero kvm_for_each_memslot_in_gfn_range(&iter, slots, gfn_start, gfn_end) { 6535f4209439SMaciej S. Szmigiero memslot = iter.slot; 653621fa3246SSean Christopherson start = max(gfn_start, memslot->base_gfn); 653721fa3246SSean Christopherson end = min(gfn_end, memslot->base_gfn + memslot->npages); 6538f4209439SMaciej S. Szmigiero if (WARN_ON_ONCE(start >= end)) 653921fa3246SSean Christopherson continue; 654021fa3246SSean Christopherson 6541727ae377SSean Christopherson flush = __walk_slot_rmaps(kvm, memslot, __kvm_zap_rmap, 654221fa3246SSean Christopherson PG_LEVEL_4K, KVM_MAX_HUGEPAGE_LEVEL, 654321fa3246SSean Christopherson start, end - 1, true, flush); 654421fa3246SSean Christopherson } 654521fa3246SSean Christopherson } 654621fa3246SSean Christopherson 654721fa3246SSean Christopherson return flush; 654821fa3246SSean Christopherson } 654921fa3246SSean Christopherson 655088f58535SMaxim Levitsky /* 655188f58535SMaxim Levitsky * Invalidate (zap) SPTEs that cover GFNs from gfn_start and up to gfn_end 655288f58535SMaxim Levitsky * (not including it) 655388f58535SMaxim Levitsky */ 6554c50d8ae3SPaolo Bonzini void kvm_zap_gfn_range(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end) 6555c50d8ae3SPaolo Bonzini { 655621fa3246SSean Christopherson bool flush; 6557c50d8ae3SPaolo Bonzini 6558f4209439SMaciej S. Szmigiero if (WARN_ON_ONCE(gfn_end <= gfn_start)) 6559f4209439SMaciej S. Szmigiero return; 6560f4209439SMaciej S. Szmigiero 6561531810caSBen Gardon write_lock(&kvm->mmu_lock); 65625a324c24SSean Christopherson 65638569992dSChao Peng kvm_mmu_invalidate_begin(kvm); 65648569992dSChao Peng 65658569992dSChao Peng kvm_mmu_invalidate_range_add(kvm, gfn_start, gfn_end); 6566edb298c6SMaxim Levitsky 65672833eda0SSean Christopherson flush = kvm_rmap_zap_gfn_range(kvm, gfn_start, gfn_end); 65686103bc07SBen Gardon 6569441a5dfcSPaolo Bonzini if (tdp_mmu_enabled) 6570441a5dfcSPaolo Bonzini flush = kvm_tdp_mmu_zap_leafs(kvm, gfn_start, gfn_end, flush); 65715a324c24SSean Christopherson 65725a324c24SSean Christopherson if (flush) 65738c63e8c2SDavid Matlack kvm_flush_remote_tlbs_range(kvm, gfn_start, gfn_end - gfn_start); 65745a324c24SSean Christopherson 65758569992dSChao Peng kvm_mmu_invalidate_end(kvm); 6576edb298c6SMaxim Levitsky 65775a324c24SSean Christopherson write_unlock(&kvm->mmu_lock); 6578c50d8ae3SPaolo Bonzini } 6579c50d8ae3SPaolo Bonzini 6580c50d8ae3SPaolo Bonzini static bool slot_rmap_write_protect(struct kvm *kvm, 65810a234f5dSSean Christopherson struct kvm_rmap_head *rmap_head, 6582269e9552SHamza Mahfooz const struct kvm_memory_slot *slot) 6583c50d8ae3SPaolo Bonzini { 65841346bbb6SDavid Matlack return rmap_write_protect(rmap_head, false); 6585c50d8ae3SPaolo Bonzini } 6586c50d8ae3SPaolo Bonzini 6587c50d8ae3SPaolo Bonzini void kvm_mmu_slot_remove_write_access(struct kvm *kvm, 6588269e9552SHamza Mahfooz const struct kvm_memory_slot *memslot, 65893c9bd400SJay Zhou int start_level) 6590c50d8ae3SPaolo Bonzini { 6591e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) { 6592531810caSBen Gardon write_lock(&kvm->mmu_lock); 6593727ae377SSean Christopherson walk_slot_rmaps(kvm, memslot, slot_rmap_write_protect, 6594b64d740eSJunaid Shahid start_level, KVM_MAX_HUGEPAGE_LEVEL, false); 6595531810caSBen Gardon write_unlock(&kvm->mmu_lock); 6596e2209710SBen Gardon } 6597c50d8ae3SPaolo Bonzini 65981f98f2bdSDavid Matlack if (tdp_mmu_enabled) { 659924ae4cfaSBen Gardon read_lock(&kvm->mmu_lock); 6600b64d740eSJunaid Shahid kvm_tdp_mmu_wrprot_slot(kvm, memslot, start_level); 660124ae4cfaSBen Gardon read_unlock(&kvm->mmu_lock); 660224ae4cfaSBen Gardon } 6603c50d8ae3SPaolo Bonzini } 6604c50d8ae3SPaolo Bonzini 6605ada51a9dSDavid Matlack static inline bool need_topup(struct kvm_mmu_memory_cache *cache, int min) 6606ada51a9dSDavid Matlack { 6607ada51a9dSDavid Matlack return kvm_mmu_memory_cache_nr_free_objects(cache) < min; 6608ada51a9dSDavid Matlack } 6609ada51a9dSDavid Matlack 6610ada51a9dSDavid Matlack static bool need_topup_split_caches_or_resched(struct kvm *kvm) 6611ada51a9dSDavid Matlack { 6612ada51a9dSDavid Matlack if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) 6613ada51a9dSDavid Matlack return true; 6614ada51a9dSDavid Matlack 6615ada51a9dSDavid Matlack /* 6616ada51a9dSDavid Matlack * In the worst case, SPLIT_DESC_CACHE_MIN_NR_OBJECTS descriptors are needed 6617ada51a9dSDavid Matlack * to split a single huge page. Calculating how many are actually needed 6618ada51a9dSDavid Matlack * is possible but not worth the complexity. 6619ada51a9dSDavid Matlack */ 6620ada51a9dSDavid Matlack return need_topup(&kvm->arch.split_desc_cache, SPLIT_DESC_CACHE_MIN_NR_OBJECTS) || 6621ada51a9dSDavid Matlack need_topup(&kvm->arch.split_page_header_cache, 1) || 6622ada51a9dSDavid Matlack need_topup(&kvm->arch.split_shadow_page_cache, 1); 6623ada51a9dSDavid Matlack } 6624ada51a9dSDavid Matlack 6625ada51a9dSDavid Matlack static int topup_split_caches(struct kvm *kvm) 6626ada51a9dSDavid Matlack { 6627b9b71f43SSean Christopherson /* 6628b9b71f43SSean Christopherson * Allocating rmap list entries when splitting huge pages for nested 6629dfd4eb44SSean Christopherson * MMUs is uncommon as KVM needs to use a list if and only if there is 6630b9b71f43SSean Christopherson * more than one rmap entry for a gfn, i.e. requires an L1 gfn to be 6631dfd4eb44SSean Christopherson * aliased by multiple L2 gfns and/or from multiple nested roots with 6632dfd4eb44SSean Christopherson * different roles. Aliasing gfns when using TDP is atypical for VMMs; 6633dfd4eb44SSean Christopherson * a few gfns are often aliased during boot, e.g. when remapping BIOS, 6634dfd4eb44SSean Christopherson * but aliasing rarely occurs post-boot or for many gfns. If there is 6635dfd4eb44SSean Christopherson * only one rmap entry, rmap->val points directly at that one entry and 6636dfd4eb44SSean Christopherson * doesn't need to allocate a list. Buffer the cache by the default 6637dfd4eb44SSean Christopherson * capacity so that KVM doesn't have to drop mmu_lock to topup if KVM 6638b9b71f43SSean Christopherson * encounters an aliased gfn or two. 6639b9b71f43SSean Christopherson */ 6640b9b71f43SSean Christopherson const int capacity = SPLIT_DESC_CACHE_MIN_NR_OBJECTS + 6641b9b71f43SSean Christopherson KVM_ARCH_NR_OBJS_PER_MEMORY_CACHE; 6642ada51a9dSDavid Matlack int r; 6643ada51a9dSDavid Matlack 6644ada51a9dSDavid Matlack lockdep_assert_held(&kvm->slots_lock); 6645ada51a9dSDavid Matlack 6646b9b71f43SSean Christopherson r = __kvm_mmu_topup_memory_cache(&kvm->arch.split_desc_cache, capacity, 6647ada51a9dSDavid Matlack SPLIT_DESC_CACHE_MIN_NR_OBJECTS); 6648ada51a9dSDavid Matlack if (r) 6649ada51a9dSDavid Matlack return r; 6650ada51a9dSDavid Matlack 6651ada51a9dSDavid Matlack r = kvm_mmu_topup_memory_cache(&kvm->arch.split_page_header_cache, 1); 6652ada51a9dSDavid Matlack if (r) 6653ada51a9dSDavid Matlack return r; 6654ada51a9dSDavid Matlack 6655ada51a9dSDavid Matlack return kvm_mmu_topup_memory_cache(&kvm->arch.split_shadow_page_cache, 1); 6656ada51a9dSDavid Matlack } 6657ada51a9dSDavid Matlack 6658ada51a9dSDavid Matlack static struct kvm_mmu_page *shadow_mmu_get_sp_for_split(struct kvm *kvm, u64 *huge_sptep) 6659ada51a9dSDavid Matlack { 6660ada51a9dSDavid Matlack struct kvm_mmu_page *huge_sp = sptep_to_sp(huge_sptep); 6661ada51a9dSDavid Matlack struct shadow_page_caches caches = {}; 6662ada51a9dSDavid Matlack union kvm_mmu_page_role role; 6663ada51a9dSDavid Matlack unsigned int access; 6664ada51a9dSDavid Matlack gfn_t gfn; 6665ada51a9dSDavid Matlack 666679e48cecSSean Christopherson gfn = kvm_mmu_page_get_gfn(huge_sp, spte_index(huge_sptep)); 666779e48cecSSean Christopherson access = kvm_mmu_page_get_access(huge_sp, spte_index(huge_sptep)); 6668ada51a9dSDavid Matlack 6669ada51a9dSDavid Matlack /* 6670ada51a9dSDavid Matlack * Note, huge page splitting always uses direct shadow pages, regardless 6671ada51a9dSDavid Matlack * of whether the huge page itself is mapped by a direct or indirect 6672ada51a9dSDavid Matlack * shadow page, since the huge page region itself is being directly 6673ada51a9dSDavid Matlack * mapped with smaller pages. 6674ada51a9dSDavid Matlack */ 6675ada51a9dSDavid Matlack role = kvm_mmu_child_role(huge_sptep, /*direct=*/true, access); 6676ada51a9dSDavid Matlack 6677ada51a9dSDavid Matlack /* Direct SPs do not require a shadowed_info_cache. */ 6678ada51a9dSDavid Matlack caches.page_header_cache = &kvm->arch.split_page_header_cache; 6679ada51a9dSDavid Matlack caches.shadow_page_cache = &kvm->arch.split_shadow_page_cache; 6680ada51a9dSDavid Matlack 6681ada51a9dSDavid Matlack /* Safe to pass NULL for vCPU since requesting a direct SP. */ 6682ada51a9dSDavid Matlack return __kvm_mmu_get_shadow_page(kvm, NULL, &caches, gfn, role); 6683ada51a9dSDavid Matlack } 6684ada51a9dSDavid Matlack 6685ada51a9dSDavid Matlack static void shadow_mmu_split_huge_page(struct kvm *kvm, 6686ada51a9dSDavid Matlack const struct kvm_memory_slot *slot, 6687ada51a9dSDavid Matlack u64 *huge_sptep) 6688ada51a9dSDavid Matlack 6689ada51a9dSDavid Matlack { 6690ada51a9dSDavid Matlack struct kvm_mmu_memory_cache *cache = &kvm->arch.split_desc_cache; 6691ada51a9dSDavid Matlack u64 huge_spte = READ_ONCE(*huge_sptep); 6692ada51a9dSDavid Matlack struct kvm_mmu_page *sp; 669303787394SPaolo Bonzini bool flush = false; 6694ada51a9dSDavid Matlack u64 *sptep, spte; 6695ada51a9dSDavid Matlack gfn_t gfn; 6696ada51a9dSDavid Matlack int index; 6697ada51a9dSDavid Matlack 6698ada51a9dSDavid Matlack sp = shadow_mmu_get_sp_for_split(kvm, huge_sptep); 6699ada51a9dSDavid Matlack 6700ada51a9dSDavid Matlack for (index = 0; index < SPTE_ENT_PER_PAGE; index++) { 6701ada51a9dSDavid Matlack sptep = &sp->spt[index]; 6702ada51a9dSDavid Matlack gfn = kvm_mmu_page_get_gfn(sp, index); 6703ada51a9dSDavid Matlack 6704ada51a9dSDavid Matlack /* 6705ada51a9dSDavid Matlack * The SP may already have populated SPTEs, e.g. if this huge 6706ada51a9dSDavid Matlack * page is aliased by multiple sptes with the same access 6707ada51a9dSDavid Matlack * permissions. These entries are guaranteed to map the same 6708ada51a9dSDavid Matlack * gfn-to-pfn translation since the SP is direct, so no need to 6709ada51a9dSDavid Matlack * modify them. 6710ada51a9dSDavid Matlack * 671103787394SPaolo Bonzini * However, if a given SPTE points to a lower level page table, 671203787394SPaolo Bonzini * that lower level page table may only be partially populated. 671303787394SPaolo Bonzini * Installing such SPTEs would effectively unmap a potion of the 671403787394SPaolo Bonzini * huge page. Unmapping guest memory always requires a TLB flush 671503787394SPaolo Bonzini * since a subsequent operation on the unmapped regions would 671603787394SPaolo Bonzini * fail to detect the need to flush. 6717ada51a9dSDavid Matlack */ 671803787394SPaolo Bonzini if (is_shadow_present_pte(*sptep)) { 671903787394SPaolo Bonzini flush |= !is_last_spte(*sptep, sp->role.level); 6720ada51a9dSDavid Matlack continue; 672103787394SPaolo Bonzini } 6722ada51a9dSDavid Matlack 6723ada51a9dSDavid Matlack spte = make_huge_page_split_spte(kvm, huge_spte, sp->role, index); 6724ada51a9dSDavid Matlack mmu_spte_set(sptep, spte); 6725ada51a9dSDavid Matlack __rmap_add(kvm, cache, slot, sptep, gfn, sp->role.access); 6726ada51a9dSDavid Matlack } 6727ada51a9dSDavid Matlack 672803787394SPaolo Bonzini __link_shadow_page(kvm, cache, huge_sptep, sp, flush); 6729ada51a9dSDavid Matlack } 6730ada51a9dSDavid Matlack 6731ada51a9dSDavid Matlack static int shadow_mmu_try_split_huge_page(struct kvm *kvm, 6732ada51a9dSDavid Matlack const struct kvm_memory_slot *slot, 6733ada51a9dSDavid Matlack u64 *huge_sptep) 6734ada51a9dSDavid Matlack { 6735ada51a9dSDavid Matlack struct kvm_mmu_page *huge_sp = sptep_to_sp(huge_sptep); 6736ada51a9dSDavid Matlack int level, r = 0; 6737ada51a9dSDavid Matlack gfn_t gfn; 6738ada51a9dSDavid Matlack u64 spte; 6739ada51a9dSDavid Matlack 6740ada51a9dSDavid Matlack /* Grab information for the tracepoint before dropping the MMU lock. */ 674179e48cecSSean Christopherson gfn = kvm_mmu_page_get_gfn(huge_sp, spte_index(huge_sptep)); 6742ada51a9dSDavid Matlack level = huge_sp->role.level; 6743ada51a9dSDavid Matlack spte = *huge_sptep; 6744ada51a9dSDavid Matlack 6745ada51a9dSDavid Matlack if (kvm_mmu_available_pages(kvm) <= KVM_MIN_FREE_MMU_PAGES) { 6746ada51a9dSDavid Matlack r = -ENOSPC; 6747ada51a9dSDavid Matlack goto out; 6748ada51a9dSDavid Matlack } 6749ada51a9dSDavid Matlack 6750ada51a9dSDavid Matlack if (need_topup_split_caches_or_resched(kvm)) { 6751ada51a9dSDavid Matlack write_unlock(&kvm->mmu_lock); 6752ada51a9dSDavid Matlack cond_resched(); 6753ada51a9dSDavid Matlack /* 6754ada51a9dSDavid Matlack * If the topup succeeds, return -EAGAIN to indicate that the 6755ada51a9dSDavid Matlack * rmap iterator should be restarted because the MMU lock was 6756ada51a9dSDavid Matlack * dropped. 6757ada51a9dSDavid Matlack */ 6758ada51a9dSDavid Matlack r = topup_split_caches(kvm) ?: -EAGAIN; 6759ada51a9dSDavid Matlack write_lock(&kvm->mmu_lock); 6760ada51a9dSDavid Matlack goto out; 6761ada51a9dSDavid Matlack } 6762ada51a9dSDavid Matlack 6763ada51a9dSDavid Matlack shadow_mmu_split_huge_page(kvm, slot, huge_sptep); 6764ada51a9dSDavid Matlack 6765ada51a9dSDavid Matlack out: 6766ada51a9dSDavid Matlack trace_kvm_mmu_split_huge_page(gfn, spte, level, r); 6767ada51a9dSDavid Matlack return r; 6768ada51a9dSDavid Matlack } 6769ada51a9dSDavid Matlack 6770ada51a9dSDavid Matlack static bool shadow_mmu_try_split_huge_pages(struct kvm *kvm, 6771ada51a9dSDavid Matlack struct kvm_rmap_head *rmap_head, 6772ada51a9dSDavid Matlack const struct kvm_memory_slot *slot) 6773ada51a9dSDavid Matlack { 6774ada51a9dSDavid Matlack struct rmap_iterator iter; 6775ada51a9dSDavid Matlack struct kvm_mmu_page *sp; 6776ada51a9dSDavid Matlack u64 *huge_sptep; 6777ada51a9dSDavid Matlack int r; 6778ada51a9dSDavid Matlack 6779ada51a9dSDavid Matlack restart: 6780ada51a9dSDavid Matlack for_each_rmap_spte(rmap_head, &iter, huge_sptep) { 6781ada51a9dSDavid Matlack sp = sptep_to_sp(huge_sptep); 6782ada51a9dSDavid Matlack 6783ada51a9dSDavid Matlack /* TDP MMU is enabled, so rmap only contains nested MMU SPs. */ 6784ada51a9dSDavid Matlack if (WARN_ON_ONCE(!sp->role.guest_mode)) 6785ada51a9dSDavid Matlack continue; 6786ada51a9dSDavid Matlack 6787ada51a9dSDavid Matlack /* The rmaps should never contain non-leaf SPTEs. */ 6788ada51a9dSDavid Matlack if (WARN_ON_ONCE(!is_large_pte(*huge_sptep))) 6789ada51a9dSDavid Matlack continue; 6790ada51a9dSDavid Matlack 6791ada51a9dSDavid Matlack /* SPs with level >PG_LEVEL_4K should never by unsync. */ 6792ada51a9dSDavid Matlack if (WARN_ON_ONCE(sp->unsync)) 6793ada51a9dSDavid Matlack continue; 6794ada51a9dSDavid Matlack 6795ada51a9dSDavid Matlack /* Don't bother splitting huge pages on invalid SPs. */ 6796ada51a9dSDavid Matlack if (sp->role.invalid) 6797ada51a9dSDavid Matlack continue; 6798ada51a9dSDavid Matlack 6799ada51a9dSDavid Matlack r = shadow_mmu_try_split_huge_page(kvm, slot, huge_sptep); 6800ada51a9dSDavid Matlack 6801ada51a9dSDavid Matlack /* 6802ada51a9dSDavid Matlack * The split succeeded or needs to be retried because the MMU 6803ada51a9dSDavid Matlack * lock was dropped. Either way, restart the iterator to get it 6804ada51a9dSDavid Matlack * back into a consistent state. 6805ada51a9dSDavid Matlack */ 6806ada51a9dSDavid Matlack if (!r || r == -EAGAIN) 6807ada51a9dSDavid Matlack goto restart; 6808ada51a9dSDavid Matlack 6809ada51a9dSDavid Matlack /* The split failed and shouldn't be retried (e.g. -ENOMEM). */ 6810ada51a9dSDavid Matlack break; 6811ada51a9dSDavid Matlack } 6812ada51a9dSDavid Matlack 6813ada51a9dSDavid Matlack return false; 6814ada51a9dSDavid Matlack } 6815ada51a9dSDavid Matlack 6816ada51a9dSDavid Matlack static void kvm_shadow_mmu_try_split_huge_pages(struct kvm *kvm, 6817ada51a9dSDavid Matlack const struct kvm_memory_slot *slot, 6818ada51a9dSDavid Matlack gfn_t start, gfn_t end, 6819ada51a9dSDavid Matlack int target_level) 6820ada51a9dSDavid Matlack { 6821ada51a9dSDavid Matlack int level; 6822ada51a9dSDavid Matlack 6823ada51a9dSDavid Matlack /* 6824ada51a9dSDavid Matlack * Split huge pages starting with KVM_MAX_HUGEPAGE_LEVEL and working 6825ada51a9dSDavid Matlack * down to the target level. This ensures pages are recursively split 6826ada51a9dSDavid Matlack * all the way to the target level. There's no need to split pages 6827ada51a9dSDavid Matlack * already at the target level. 6828ada51a9dSDavid Matlack */ 6829727ae377SSean Christopherson for (level = KVM_MAX_HUGEPAGE_LEVEL; level > target_level; level--) 6830727ae377SSean Christopherson __walk_slot_rmaps(kvm, slot, shadow_mmu_try_split_huge_pages, 6831ada51a9dSDavid Matlack level, level, start, end - 1, true, false); 6832ada51a9dSDavid Matlack } 6833ada51a9dSDavid Matlack 6834cb00a70bSDavid Matlack /* Must be called with the mmu_lock held in write-mode. */ 6835cb00a70bSDavid Matlack void kvm_mmu_try_split_huge_pages(struct kvm *kvm, 6836cb00a70bSDavid Matlack const struct kvm_memory_slot *memslot, 6837cb00a70bSDavid Matlack u64 start, u64 end, 6838cb00a70bSDavid Matlack int target_level) 6839cb00a70bSDavid Matlack { 68401f98f2bdSDavid Matlack if (!tdp_mmu_enabled) 6841ada51a9dSDavid Matlack return; 6842ada51a9dSDavid Matlack 6843ada51a9dSDavid Matlack if (kvm_memslots_have_rmaps(kvm)) 6844ada51a9dSDavid Matlack kvm_shadow_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level); 6845ada51a9dSDavid Matlack 6846ada51a9dSDavid Matlack kvm_tdp_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level, false); 6847cb00a70bSDavid Matlack 6848cb00a70bSDavid Matlack /* 684954aa699eSBjorn Helgaas * A TLB flush is unnecessary at this point for the same reasons as in 6850cb00a70bSDavid Matlack * kvm_mmu_slot_try_split_huge_pages(). 6851cb00a70bSDavid Matlack */ 6852cb00a70bSDavid Matlack } 6853cb00a70bSDavid Matlack 6854a3fe5dbdSDavid Matlack void kvm_mmu_slot_try_split_huge_pages(struct kvm *kvm, 6855a3fe5dbdSDavid Matlack const struct kvm_memory_slot *memslot, 6856a3fe5dbdSDavid Matlack int target_level) 6857a3fe5dbdSDavid Matlack { 6858a3fe5dbdSDavid Matlack u64 start = memslot->base_gfn; 6859a3fe5dbdSDavid Matlack u64 end = start + memslot->npages; 6860a3fe5dbdSDavid Matlack 68611f98f2bdSDavid Matlack if (!tdp_mmu_enabled) 6862ada51a9dSDavid Matlack return; 6863ada51a9dSDavid Matlack 6864ada51a9dSDavid Matlack if (kvm_memslots_have_rmaps(kvm)) { 6865ada51a9dSDavid Matlack write_lock(&kvm->mmu_lock); 6866ada51a9dSDavid Matlack kvm_shadow_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level); 6867ada51a9dSDavid Matlack write_unlock(&kvm->mmu_lock); 6868ada51a9dSDavid Matlack } 6869ada51a9dSDavid Matlack 6870a3fe5dbdSDavid Matlack read_lock(&kvm->mmu_lock); 6871cb00a70bSDavid Matlack kvm_tdp_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level, true); 6872a3fe5dbdSDavid Matlack read_unlock(&kvm->mmu_lock); 6873a3fe5dbdSDavid Matlack 6874a3fe5dbdSDavid Matlack /* 6875a3fe5dbdSDavid Matlack * No TLB flush is necessary here. KVM will flush TLBs after 6876a3fe5dbdSDavid Matlack * write-protecting and/or clearing dirty on the newly split SPTEs to 6877a3fe5dbdSDavid Matlack * ensure that guest writes are reflected in the dirty log before the 6878a3fe5dbdSDavid Matlack * ioctl to enable dirty logging on this memslot completes. Since the 6879a3fe5dbdSDavid Matlack * split SPTEs retain the write and dirty bits of the huge SPTE, it is 6880a3fe5dbdSDavid Matlack * safe for KVM to decide if a TLB flush is necessary based on the split 6881a3fe5dbdSDavid Matlack * SPTEs. 6882a3fe5dbdSDavid Matlack */ 6883a3fe5dbdSDavid Matlack } 6884a3fe5dbdSDavid Matlack 6885c50d8ae3SPaolo Bonzini static bool kvm_mmu_zap_collapsible_spte(struct kvm *kvm, 68860a234f5dSSean Christopherson struct kvm_rmap_head *rmap_head, 6887269e9552SHamza Mahfooz const struct kvm_memory_slot *slot) 6888c50d8ae3SPaolo Bonzini { 6889c50d8ae3SPaolo Bonzini u64 *sptep; 6890c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 6891c50d8ae3SPaolo Bonzini int need_tlb_flush = 0; 6892c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 6893c50d8ae3SPaolo Bonzini 6894c50d8ae3SPaolo Bonzini restart: 6895c50d8ae3SPaolo Bonzini for_each_rmap_spte(rmap_head, &iter, sptep) { 689657354682SSean Christopherson sp = sptep_to_sp(sptep); 6897c50d8ae3SPaolo Bonzini 6898c50d8ae3SPaolo Bonzini /* 6899c50d8ae3SPaolo Bonzini * We cannot do huge page mapping for indirect shadow pages, 6900c50d8ae3SPaolo Bonzini * which are found on the last rmap (level = 1) when not using 6901c50d8ae3SPaolo Bonzini * tdp; such shadow pages are synced with the page table in 6902c50d8ae3SPaolo Bonzini * the guest, and the guest page table is using 4K page size 6903c50d8ae3SPaolo Bonzini * mapping if the indirect sp has level = 1. 6904c50d8ae3SPaolo Bonzini */ 69055d49f08cSSean Christopherson if (sp->role.direct && 69069eba50f8SSean Christopherson sp->role.level < kvm_mmu_max_mapping_level(kvm, slot, sp->gfn, 6907a8ac499bSSean Christopherson PG_LEVEL_NUM)) { 69089202aee8SSean Christopherson kvm_zap_one_rmap_spte(kvm, rmap_head, sptep); 6909c50d8ae3SPaolo Bonzini 69108a1300ffSSean Christopherson if (kvm_available_flush_remote_tlbs_range()) 69111b2dc736SHou Wenlong kvm_flush_remote_tlbs_sptep(kvm, sptep); 6912c50d8ae3SPaolo Bonzini else 6913c50d8ae3SPaolo Bonzini need_tlb_flush = 1; 6914c50d8ae3SPaolo Bonzini 6915c50d8ae3SPaolo Bonzini goto restart; 6916c50d8ae3SPaolo Bonzini } 6917c50d8ae3SPaolo Bonzini } 6918c50d8ae3SPaolo Bonzini 6919c50d8ae3SPaolo Bonzini return need_tlb_flush; 6920c50d8ae3SPaolo Bonzini } 6921c63cf135SBrijesh Singh EXPORT_SYMBOL_GPL(kvm_zap_gfn_range); 6922c50d8ae3SPaolo Bonzini 692320d49186SDavid Matlack static void kvm_rmap_zap_collapsible_sptes(struct kvm *kvm, 692420d49186SDavid Matlack const struct kvm_memory_slot *slot) 692520d49186SDavid Matlack { 692620d49186SDavid Matlack /* 692720d49186SDavid Matlack * Note, use KVM_MAX_HUGEPAGE_LEVEL - 1 since there's no need to zap 692820d49186SDavid Matlack * pages that are already mapped at the maximum hugepage level. 692920d49186SDavid Matlack */ 6930727ae377SSean Christopherson if (walk_slot_rmaps(kvm, slot, kvm_mmu_zap_collapsible_spte, 693120d49186SDavid Matlack PG_LEVEL_4K, KVM_MAX_HUGEPAGE_LEVEL - 1, true)) 6932619b5072SDavid Matlack kvm_flush_remote_tlbs_memslot(kvm, slot); 693320d49186SDavid Matlack } 693420d49186SDavid Matlack 6935c50d8ae3SPaolo Bonzini void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm, 6936269e9552SHamza Mahfooz const struct kvm_memory_slot *slot) 6937c50d8ae3SPaolo Bonzini { 6938e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) { 6939531810caSBen Gardon write_lock(&kvm->mmu_lock); 694020d49186SDavid Matlack kvm_rmap_zap_collapsible_sptes(kvm, slot); 6941531810caSBen Gardon write_unlock(&kvm->mmu_lock); 6942e2209710SBen Gardon } 69432db6f772SBen Gardon 69441f98f2bdSDavid Matlack if (tdp_mmu_enabled) { 69452db6f772SBen Gardon read_lock(&kvm->mmu_lock); 69464b85c921SSean Christopherson kvm_tdp_mmu_zap_collapsible_sptes(kvm, slot); 69472db6f772SBen Gardon read_unlock(&kvm->mmu_lock); 69482db6f772SBen Gardon } 6949c50d8ae3SPaolo Bonzini } 6950c50d8ae3SPaolo Bonzini 6951c50d8ae3SPaolo Bonzini void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm, 6952269e9552SHamza Mahfooz const struct kvm_memory_slot *memslot) 6953c50d8ae3SPaolo Bonzini { 6954e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) { 6955531810caSBen Gardon write_lock(&kvm->mmu_lock); 6956610265eaSDavid Matlack /* 6957610265eaSDavid Matlack * Clear dirty bits only on 4k SPTEs since the legacy MMU only 6958610265eaSDavid Matlack * support dirty logging at a 4k granularity. 6959610265eaSDavid Matlack */ 6960727ae377SSean Christopherson walk_slot_rmaps_4k(kvm, memslot, __rmap_clear_dirty, false); 6961531810caSBen Gardon write_unlock(&kvm->mmu_lock); 6962e2209710SBen Gardon } 6963c50d8ae3SPaolo Bonzini 69641f98f2bdSDavid Matlack if (tdp_mmu_enabled) { 696524ae4cfaSBen Gardon read_lock(&kvm->mmu_lock); 6966b64d740eSJunaid Shahid kvm_tdp_mmu_clear_dirty_slot(kvm, memslot); 696724ae4cfaSBen Gardon read_unlock(&kvm->mmu_lock); 696824ae4cfaSBen Gardon } 696924ae4cfaSBen Gardon 6970c50d8ae3SPaolo Bonzini /* 6971b64d740eSJunaid Shahid * The caller will flush the TLBs after this function returns. 6972b64d740eSJunaid Shahid * 6973c50d8ae3SPaolo Bonzini * It's also safe to flush TLBs out of mmu lock here as currently this 6974c50d8ae3SPaolo Bonzini * function is only used for dirty logging, in which case flushing TLB 6975c50d8ae3SPaolo Bonzini * out of mmu lock also guarantees no dirty pages will be lost in 6976c50d8ae3SPaolo Bonzini * dirty_bitmap. 6977c50d8ae3SPaolo Bonzini */ 6978c50d8ae3SPaolo Bonzini } 6979c50d8ae3SPaolo Bonzini 6980db0d70e6SSean Christopherson static void kvm_mmu_zap_all(struct kvm *kvm) 6981c50d8ae3SPaolo Bonzini { 6982c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, *node; 6983c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 6984c50d8ae3SPaolo Bonzini int ign; 6985c50d8ae3SPaolo Bonzini 6986531810caSBen Gardon write_lock(&kvm->mmu_lock); 6987c50d8ae3SPaolo Bonzini restart: 6988c50d8ae3SPaolo Bonzini list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link) { 698920ba462dSSean Christopherson if (WARN_ON_ONCE(sp->role.invalid)) 6990c50d8ae3SPaolo Bonzini continue; 6991c50d8ae3SPaolo Bonzini if (__kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list, &ign)) 6992c50d8ae3SPaolo Bonzini goto restart; 6993531810caSBen Gardon if (cond_resched_rwlock_write(&kvm->mmu_lock)) 6994c50d8ae3SPaolo Bonzini goto restart; 6995c50d8ae3SPaolo Bonzini } 6996c50d8ae3SPaolo Bonzini 6997c50d8ae3SPaolo Bonzini kvm_mmu_commit_zap_page(kvm, &invalid_list); 6998faaf05b0SBen Gardon 69991f98f2bdSDavid Matlack if (tdp_mmu_enabled) 7000faaf05b0SBen Gardon kvm_tdp_mmu_zap_all(kvm); 7001faaf05b0SBen Gardon 7002531810caSBen Gardon write_unlock(&kvm->mmu_lock); 7003c50d8ae3SPaolo Bonzini } 7004c50d8ae3SPaolo Bonzini 7005db0d70e6SSean Christopherson void kvm_arch_flush_shadow_all(struct kvm *kvm) 7006db0d70e6SSean Christopherson { 7007db0d70e6SSean Christopherson kvm_mmu_zap_all(kvm); 7008db0d70e6SSean Christopherson } 7009db0d70e6SSean Christopherson 7010db0d70e6SSean Christopherson void kvm_arch_flush_shadow_memslot(struct kvm *kvm, 7011db0d70e6SSean Christopherson struct kvm_memory_slot *slot) 7012db0d70e6SSean Christopherson { 7013eeb87272SSean Christopherson kvm_mmu_zap_all_fast(kvm); 7014db0d70e6SSean Christopherson } 7015db0d70e6SSean Christopherson 7016c50d8ae3SPaolo Bonzini void kvm_mmu_invalidate_mmio_sptes(struct kvm *kvm, u64 gen) 7017c50d8ae3SPaolo Bonzini { 701820ba462dSSean Christopherson WARN_ON_ONCE(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS); 7019c50d8ae3SPaolo Bonzini 7020c50d8ae3SPaolo Bonzini gen &= MMIO_SPTE_GEN_MASK; 7021c50d8ae3SPaolo Bonzini 7022c50d8ae3SPaolo Bonzini /* 7023c50d8ae3SPaolo Bonzini * Generation numbers are incremented in multiples of the number of 7024c50d8ae3SPaolo Bonzini * address spaces in order to provide unique generations across all 7025c50d8ae3SPaolo Bonzini * address spaces. Strip what is effectively the address space 7026c50d8ae3SPaolo Bonzini * modifier prior to checking for a wrap of the MMIO generation so 7027c50d8ae3SPaolo Bonzini * that a wrap in any address space is detected. 7028c50d8ae3SPaolo Bonzini */ 7029eed52e43SSean Christopherson gen &= ~((u64)kvm_arch_nr_memslot_as_ids(kvm) - 1); 7030c50d8ae3SPaolo Bonzini 7031c50d8ae3SPaolo Bonzini /* 7032c50d8ae3SPaolo Bonzini * The very rare case: if the MMIO generation number has wrapped, 7033c50d8ae3SPaolo Bonzini * zap all shadow pages. 7034c50d8ae3SPaolo Bonzini */ 7035c50d8ae3SPaolo Bonzini if (unlikely(gen == 0)) { 70368d20bd63SSean Christopherson kvm_debug_ratelimited("zapping shadow pages for mmio generation wraparound\n"); 7037c50d8ae3SPaolo Bonzini kvm_mmu_zap_all_fast(kvm); 7038c50d8ae3SPaolo Bonzini } 7039c50d8ae3SPaolo Bonzini } 7040c50d8ae3SPaolo Bonzini 7041f3d90f90SSean Christopherson static unsigned long mmu_shrink_scan(struct shrinker *shrink, 7042f3d90f90SSean Christopherson struct shrink_control *sc) 7043c50d8ae3SPaolo Bonzini { 7044c50d8ae3SPaolo Bonzini struct kvm *kvm; 7045c50d8ae3SPaolo Bonzini int nr_to_scan = sc->nr_to_scan; 7046c50d8ae3SPaolo Bonzini unsigned long freed = 0; 7047c50d8ae3SPaolo Bonzini 7048c50d8ae3SPaolo Bonzini mutex_lock(&kvm_lock); 7049c50d8ae3SPaolo Bonzini 7050c50d8ae3SPaolo Bonzini list_for_each_entry(kvm, &vm_list, vm_list) { 7051c50d8ae3SPaolo Bonzini int idx; 7052c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 7053c50d8ae3SPaolo Bonzini 7054c50d8ae3SPaolo Bonzini /* 7055c50d8ae3SPaolo Bonzini * Never scan more than sc->nr_to_scan VM instances. 7056c50d8ae3SPaolo Bonzini * Will not hit this condition practically since we do not try 7057c50d8ae3SPaolo Bonzini * to shrink more than one VM and it is very unlikely to see 7058c50d8ae3SPaolo Bonzini * !n_used_mmu_pages so many times. 7059c50d8ae3SPaolo Bonzini */ 7060c50d8ae3SPaolo Bonzini if (!nr_to_scan--) 7061c50d8ae3SPaolo Bonzini break; 7062c50d8ae3SPaolo Bonzini /* 7063c50d8ae3SPaolo Bonzini * n_used_mmu_pages is accessed without holding kvm->mmu_lock 7064c50d8ae3SPaolo Bonzini * here. We may skip a VM instance errorneosly, but we do not 7065c50d8ae3SPaolo Bonzini * want to shrink a VM that only started to populate its MMU 7066c50d8ae3SPaolo Bonzini * anyway. 7067c50d8ae3SPaolo Bonzini */ 7068c50d8ae3SPaolo Bonzini if (!kvm->arch.n_used_mmu_pages && 7069c50d8ae3SPaolo Bonzini !kvm_has_zapped_obsolete_pages(kvm)) 7070c50d8ae3SPaolo Bonzini continue; 7071c50d8ae3SPaolo Bonzini 7072c50d8ae3SPaolo Bonzini idx = srcu_read_lock(&kvm->srcu); 7073531810caSBen Gardon write_lock(&kvm->mmu_lock); 7074c50d8ae3SPaolo Bonzini 7075c50d8ae3SPaolo Bonzini if (kvm_has_zapped_obsolete_pages(kvm)) { 7076c50d8ae3SPaolo Bonzini kvm_mmu_commit_zap_page(kvm, 7077c50d8ae3SPaolo Bonzini &kvm->arch.zapped_obsolete_pages); 7078c50d8ae3SPaolo Bonzini goto unlock; 7079c50d8ae3SPaolo Bonzini } 7080c50d8ae3SPaolo Bonzini 7081ebdb292dSSean Christopherson freed = kvm_mmu_zap_oldest_mmu_pages(kvm, sc->nr_to_scan); 7082c50d8ae3SPaolo Bonzini 7083c50d8ae3SPaolo Bonzini unlock: 7084531810caSBen Gardon write_unlock(&kvm->mmu_lock); 7085c50d8ae3SPaolo Bonzini srcu_read_unlock(&kvm->srcu, idx); 7086c50d8ae3SPaolo Bonzini 7087c50d8ae3SPaolo Bonzini /* 7088c50d8ae3SPaolo Bonzini * unfair on small ones 7089c50d8ae3SPaolo Bonzini * per-vm shrinkers cry out 7090c50d8ae3SPaolo Bonzini * sadness comes quickly 7091c50d8ae3SPaolo Bonzini */ 7092c50d8ae3SPaolo Bonzini list_move_tail(&kvm->vm_list, &vm_list); 7093c50d8ae3SPaolo Bonzini break; 7094c50d8ae3SPaolo Bonzini } 7095c50d8ae3SPaolo Bonzini 7096c50d8ae3SPaolo Bonzini mutex_unlock(&kvm_lock); 7097c50d8ae3SPaolo Bonzini return freed; 7098c50d8ae3SPaolo Bonzini } 7099c50d8ae3SPaolo Bonzini 7100f3d90f90SSean Christopherson static unsigned long mmu_shrink_count(struct shrinker *shrink, 7101f3d90f90SSean Christopherson struct shrink_control *sc) 7102c50d8ae3SPaolo Bonzini { 7103c50d8ae3SPaolo Bonzini return percpu_counter_read_positive(&kvm_total_used_mmu_pages); 7104c50d8ae3SPaolo Bonzini } 7105c50d8ae3SPaolo Bonzini 7106e5985c40SQi Zheng static struct shrinker *mmu_shrinker; 7107c50d8ae3SPaolo Bonzini 7108c50d8ae3SPaolo Bonzini static void mmu_destroy_caches(void) 7109c50d8ae3SPaolo Bonzini { 7110c50d8ae3SPaolo Bonzini kmem_cache_destroy(pte_list_desc_cache); 7111c50d8ae3SPaolo Bonzini kmem_cache_destroy(mmu_page_header_cache); 7112c50d8ae3SPaolo Bonzini } 7113c50d8ae3SPaolo Bonzini 71140b210fafSSean Christopherson static int get_nx_huge_pages(char *buffer, const struct kernel_param *kp) 71150b210fafSSean Christopherson { 71160b210fafSSean Christopherson if (nx_hugepage_mitigation_hard_disabled) 71171d6664faSLike Xu return sysfs_emit(buffer, "never\n"); 71180b210fafSSean Christopherson 71190b210fafSSean Christopherson return param_get_bool(buffer, kp); 71200b210fafSSean Christopherson } 71210b210fafSSean Christopherson 7122c50d8ae3SPaolo Bonzini static bool get_nx_auto_mode(void) 7123c50d8ae3SPaolo Bonzini { 7124c50d8ae3SPaolo Bonzini /* Return true when CPU has the bug, and mitigations are ON */ 7125c50d8ae3SPaolo Bonzini return boot_cpu_has_bug(X86_BUG_ITLB_MULTIHIT) && !cpu_mitigations_off(); 7126c50d8ae3SPaolo Bonzini } 7127c50d8ae3SPaolo Bonzini 7128c50d8ae3SPaolo Bonzini static void __set_nx_huge_pages(bool val) 7129c50d8ae3SPaolo Bonzini { 7130c50d8ae3SPaolo Bonzini nx_huge_pages = itlb_multihit_kvm_mitigation = val; 7131c50d8ae3SPaolo Bonzini } 7132c50d8ae3SPaolo Bonzini 7133c50d8ae3SPaolo Bonzini static int set_nx_huge_pages(const char *val, const struct kernel_param *kp) 7134c50d8ae3SPaolo Bonzini { 7135c50d8ae3SPaolo Bonzini bool old_val = nx_huge_pages; 7136c50d8ae3SPaolo Bonzini bool new_val; 7137c50d8ae3SPaolo Bonzini 71380b210fafSSean Christopherson if (nx_hugepage_mitigation_hard_disabled) 71390b210fafSSean Christopherson return -EPERM; 71400b210fafSSean Christopherson 7141c50d8ae3SPaolo Bonzini /* In "auto" mode deploy workaround only if CPU has the bug. */ 71420b210fafSSean Christopherson if (sysfs_streq(val, "off")) { 7143c50d8ae3SPaolo Bonzini new_val = 0; 71440b210fafSSean Christopherson } else if (sysfs_streq(val, "force")) { 7145c50d8ae3SPaolo Bonzini new_val = 1; 71460b210fafSSean Christopherson } else if (sysfs_streq(val, "auto")) { 7147c50d8ae3SPaolo Bonzini new_val = get_nx_auto_mode(); 71480b210fafSSean Christopherson } else if (sysfs_streq(val, "never")) { 71490b210fafSSean Christopherson new_val = 0; 71500b210fafSSean Christopherson 71510b210fafSSean Christopherson mutex_lock(&kvm_lock); 71520b210fafSSean Christopherson if (!list_empty(&vm_list)) { 71530b210fafSSean Christopherson mutex_unlock(&kvm_lock); 71540b210fafSSean Christopherson return -EBUSY; 71550b210fafSSean Christopherson } 71560b210fafSSean Christopherson nx_hugepage_mitigation_hard_disabled = true; 71570b210fafSSean Christopherson mutex_unlock(&kvm_lock); 71580b210fafSSean Christopherson } else if (kstrtobool(val, &new_val) < 0) { 7159c50d8ae3SPaolo Bonzini return -EINVAL; 71600b210fafSSean Christopherson } 7161c50d8ae3SPaolo Bonzini 7162c50d8ae3SPaolo Bonzini __set_nx_huge_pages(new_val); 7163c50d8ae3SPaolo Bonzini 7164c50d8ae3SPaolo Bonzini if (new_val != old_val) { 7165c50d8ae3SPaolo Bonzini struct kvm *kvm; 7166c50d8ae3SPaolo Bonzini 7167c50d8ae3SPaolo Bonzini mutex_lock(&kvm_lock); 7168c50d8ae3SPaolo Bonzini 7169c50d8ae3SPaolo Bonzini list_for_each_entry(kvm, &vm_list, vm_list) { 7170c50d8ae3SPaolo Bonzini mutex_lock(&kvm->slots_lock); 7171c50d8ae3SPaolo Bonzini kvm_mmu_zap_all_fast(kvm); 7172c50d8ae3SPaolo Bonzini mutex_unlock(&kvm->slots_lock); 7173c50d8ae3SPaolo Bonzini 717455c510e2SSean Christopherson wake_up_process(kvm->arch.nx_huge_page_recovery_thread); 7175c50d8ae3SPaolo Bonzini } 7176c50d8ae3SPaolo Bonzini mutex_unlock(&kvm_lock); 7177c50d8ae3SPaolo Bonzini } 7178c50d8ae3SPaolo Bonzini 7179c50d8ae3SPaolo Bonzini return 0; 7180c50d8ae3SPaolo Bonzini } 7181c50d8ae3SPaolo Bonzini 71821d0e8480SSean Christopherson /* 71831d0e8480SSean Christopherson * nx_huge_pages needs to be resolved to true/false when kvm.ko is loaded, as 71841d0e8480SSean Christopherson * its default value of -1 is technically undefined behavior for a boolean. 7185c3e0c8c2SSean Christopherson * Forward the module init call to SPTE code so that it too can handle module 7186c3e0c8c2SSean Christopherson * params that need to be resolved/snapshot. 71871d0e8480SSean Christopherson */ 7188982bae43SSean Christopherson void __init kvm_mmu_x86_module_init(void) 7189c50d8ae3SPaolo Bonzini { 7190c50d8ae3SPaolo Bonzini if (nx_huge_pages == -1) 7191c50d8ae3SPaolo Bonzini __set_nx_huge_pages(get_nx_auto_mode()); 7192c3e0c8c2SSean Christopherson 71931f98f2bdSDavid Matlack /* 71941f98f2bdSDavid Matlack * Snapshot userspace's desire to enable the TDP MMU. Whether or not the 71951f98f2bdSDavid Matlack * TDP MMU is actually enabled is determined in kvm_configure_mmu() 71961f98f2bdSDavid Matlack * when the vendor module is loaded. 71971f98f2bdSDavid Matlack */ 71981f98f2bdSDavid Matlack tdp_mmu_allowed = tdp_mmu_enabled; 71991f98f2bdSDavid Matlack 7200c3e0c8c2SSean Christopherson kvm_mmu_spte_module_init(); 72011d0e8480SSean Christopherson } 72021d0e8480SSean Christopherson 72031d0e8480SSean Christopherson /* 72041d0e8480SSean Christopherson * The bulk of the MMU initialization is deferred until the vendor module is 72051d0e8480SSean Christopherson * loaded as many of the masks/values may be modified by VMX or SVM, i.e. need 72061d0e8480SSean Christopherson * to be reset when a potentially different vendor module is loaded. 72071d0e8480SSean Christopherson */ 72081d0e8480SSean Christopherson int kvm_mmu_vendor_module_init(void) 72091d0e8480SSean Christopherson { 72101d0e8480SSean Christopherson int ret = -ENOMEM; 7211c50d8ae3SPaolo Bonzini 7212c50d8ae3SPaolo Bonzini /* 7213c50d8ae3SPaolo Bonzini * MMU roles use union aliasing which is, generally speaking, an 7214c50d8ae3SPaolo Bonzini * undefined behavior. However, we supposedly know how compilers behave 7215c50d8ae3SPaolo Bonzini * and the current status quo is unlikely to change. Guardians below are 7216c50d8ae3SPaolo Bonzini * supposed to let us know if the assumption becomes false. 7217c50d8ae3SPaolo Bonzini */ 7218c50d8ae3SPaolo Bonzini BUILD_BUG_ON(sizeof(union kvm_mmu_page_role) != sizeof(u32)); 7219c50d8ae3SPaolo Bonzini BUILD_BUG_ON(sizeof(union kvm_mmu_extended_role) != sizeof(u32)); 72207a7ae829SPaolo Bonzini BUILD_BUG_ON(sizeof(union kvm_cpu_role) != sizeof(u64)); 7221c50d8ae3SPaolo Bonzini 7222c50d8ae3SPaolo Bonzini kvm_mmu_reset_all_pte_masks(); 7223c50d8ae3SPaolo Bonzini 72240dbd0546SKunwu Chan pte_list_desc_cache = KMEM_CACHE(pte_list_desc, SLAB_ACCOUNT); 7225c50d8ae3SPaolo Bonzini if (!pte_list_desc_cache) 7226c50d8ae3SPaolo Bonzini goto out; 7227c50d8ae3SPaolo Bonzini 7228c50d8ae3SPaolo Bonzini mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header", 7229c50d8ae3SPaolo Bonzini sizeof(struct kvm_mmu_page), 7230c50d8ae3SPaolo Bonzini 0, SLAB_ACCOUNT, NULL); 7231c50d8ae3SPaolo Bonzini if (!mmu_page_header_cache) 7232c50d8ae3SPaolo Bonzini goto out; 7233c50d8ae3SPaolo Bonzini 7234c50d8ae3SPaolo Bonzini if (percpu_counter_init(&kvm_total_used_mmu_pages, 0, GFP_KERNEL)) 7235c50d8ae3SPaolo Bonzini goto out; 7236c50d8ae3SPaolo Bonzini 7237e5985c40SQi Zheng mmu_shrinker = shrinker_alloc(0, "x86-mmu"); 7238e5985c40SQi Zheng if (!mmu_shrinker) 7239d7c9bfb9SMiaohe Lin goto out_shrinker; 7240c50d8ae3SPaolo Bonzini 7241e5985c40SQi Zheng mmu_shrinker->count_objects = mmu_shrink_count; 7242e5985c40SQi Zheng mmu_shrinker->scan_objects = mmu_shrink_scan; 7243e5985c40SQi Zheng mmu_shrinker->seeks = DEFAULT_SEEKS * 10; 7244e5985c40SQi Zheng 7245e5985c40SQi Zheng shrinker_register(mmu_shrinker); 7246e5985c40SQi Zheng 7247c50d8ae3SPaolo Bonzini return 0; 7248c50d8ae3SPaolo Bonzini 7249d7c9bfb9SMiaohe Lin out_shrinker: 7250d7c9bfb9SMiaohe Lin percpu_counter_destroy(&kvm_total_used_mmu_pages); 7251c50d8ae3SPaolo Bonzini out: 7252c50d8ae3SPaolo Bonzini mmu_destroy_caches(); 7253c50d8ae3SPaolo Bonzini return ret; 7254c50d8ae3SPaolo Bonzini } 7255c50d8ae3SPaolo Bonzini 7256c50d8ae3SPaolo Bonzini void kvm_mmu_destroy(struct kvm_vcpu *vcpu) 7257c50d8ae3SPaolo Bonzini { 7258c50d8ae3SPaolo Bonzini kvm_mmu_unload(vcpu); 7259c50d8ae3SPaolo Bonzini free_mmu_pages(&vcpu->arch.root_mmu); 7260c50d8ae3SPaolo Bonzini free_mmu_pages(&vcpu->arch.guest_mmu); 7261c50d8ae3SPaolo Bonzini mmu_free_memory_caches(vcpu); 7262c50d8ae3SPaolo Bonzini } 7263c50d8ae3SPaolo Bonzini 72641d0e8480SSean Christopherson void kvm_mmu_vendor_module_exit(void) 7265c50d8ae3SPaolo Bonzini { 7266c50d8ae3SPaolo Bonzini mmu_destroy_caches(); 7267c50d8ae3SPaolo Bonzini percpu_counter_destroy(&kvm_total_used_mmu_pages); 7268e5985c40SQi Zheng shrinker_free(mmu_shrinker); 7269c50d8ae3SPaolo Bonzini } 7270c50d8ae3SPaolo Bonzini 7271f47491d7SSean Christopherson /* 7272f47491d7SSean Christopherson * Calculate the effective recovery period, accounting for '0' meaning "let KVM 7273f47491d7SSean Christopherson * select a halving time of 1 hour". Returns true if recovery is enabled. 7274f47491d7SSean Christopherson */ 7275f47491d7SSean Christopherson static bool calc_nx_huge_pages_recovery_period(uint *period) 7276f47491d7SSean Christopherson { 7277f47491d7SSean Christopherson /* 7278f47491d7SSean Christopherson * Use READ_ONCE to get the params, this may be called outside of the 7279f47491d7SSean Christopherson * param setters, e.g. by the kthread to compute its next timeout. 7280f47491d7SSean Christopherson */ 7281f47491d7SSean Christopherson bool enabled = READ_ONCE(nx_huge_pages); 7282f47491d7SSean Christopherson uint ratio = READ_ONCE(nx_huge_pages_recovery_ratio); 7283f47491d7SSean Christopherson 7284f47491d7SSean Christopherson if (!enabled || !ratio) 7285f47491d7SSean Christopherson return false; 7286f47491d7SSean Christopherson 7287f47491d7SSean Christopherson *period = READ_ONCE(nx_huge_pages_recovery_period_ms); 7288f47491d7SSean Christopherson if (!*period) { 7289f47491d7SSean Christopherson /* Make sure the period is not less than one second. */ 7290f47491d7SSean Christopherson ratio = min(ratio, 3600u); 7291f47491d7SSean Christopherson *period = 60 * 60 * 1000 / ratio; 7292f47491d7SSean Christopherson } 7293f47491d7SSean Christopherson return true; 7294f47491d7SSean Christopherson } 7295f47491d7SSean Christopherson 72964dfe4f40SJunaid Shahid static int set_nx_huge_pages_recovery_param(const char *val, const struct kernel_param *kp) 7297c50d8ae3SPaolo Bonzini { 72984dfe4f40SJunaid Shahid bool was_recovery_enabled, is_recovery_enabled; 72994dfe4f40SJunaid Shahid uint old_period, new_period; 7300c50d8ae3SPaolo Bonzini int err; 7301c50d8ae3SPaolo Bonzini 73020b210fafSSean Christopherson if (nx_hugepage_mitigation_hard_disabled) 73030b210fafSSean Christopherson return -EPERM; 73040b210fafSSean Christopherson 7305f47491d7SSean Christopherson was_recovery_enabled = calc_nx_huge_pages_recovery_period(&old_period); 73064dfe4f40SJunaid Shahid 7307c50d8ae3SPaolo Bonzini err = param_set_uint(val, kp); 7308c50d8ae3SPaolo Bonzini if (err) 7309c50d8ae3SPaolo Bonzini return err; 7310c50d8ae3SPaolo Bonzini 7311f47491d7SSean Christopherson is_recovery_enabled = calc_nx_huge_pages_recovery_period(&new_period); 73124dfe4f40SJunaid Shahid 7313f47491d7SSean Christopherson if (is_recovery_enabled && 73144dfe4f40SJunaid Shahid (!was_recovery_enabled || old_period > new_period)) { 7315c50d8ae3SPaolo Bonzini struct kvm *kvm; 7316c50d8ae3SPaolo Bonzini 7317c50d8ae3SPaolo Bonzini mutex_lock(&kvm_lock); 7318c50d8ae3SPaolo Bonzini 7319c50d8ae3SPaolo Bonzini list_for_each_entry(kvm, &vm_list, vm_list) 732055c510e2SSean Christopherson wake_up_process(kvm->arch.nx_huge_page_recovery_thread); 7321c50d8ae3SPaolo Bonzini 7322c50d8ae3SPaolo Bonzini mutex_unlock(&kvm_lock); 7323c50d8ae3SPaolo Bonzini } 7324c50d8ae3SPaolo Bonzini 7325c50d8ae3SPaolo Bonzini return err; 7326c50d8ae3SPaolo Bonzini } 7327c50d8ae3SPaolo Bonzini 732855c510e2SSean Christopherson static void kvm_recover_nx_huge_pages(struct kvm *kvm) 7329c50d8ae3SPaolo Bonzini { 7330ade74e14SSean Christopherson unsigned long nx_lpage_splits = kvm->stat.nx_lpage_splits; 7331eb298605SDavid Matlack struct kvm_memory_slot *slot; 7332c50d8ae3SPaolo Bonzini int rcu_idx; 7333c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 7334c50d8ae3SPaolo Bonzini unsigned int ratio; 7335c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 7336048f4980SSean Christopherson bool flush = false; 7337c50d8ae3SPaolo Bonzini ulong to_zap; 7338c50d8ae3SPaolo Bonzini 7339c50d8ae3SPaolo Bonzini rcu_idx = srcu_read_lock(&kvm->srcu); 7340531810caSBen Gardon write_lock(&kvm->mmu_lock); 7341c50d8ae3SPaolo Bonzini 7342bb95dfb9SSean Christopherson /* 7343bb95dfb9SSean Christopherson * Zapping TDP MMU shadow pages, including the remote TLB flush, must 7344bb95dfb9SSean Christopherson * be done under RCU protection, because the pages are freed via RCU 7345bb95dfb9SSean Christopherson * callback. 7346bb95dfb9SSean Christopherson */ 7347bb95dfb9SSean Christopherson rcu_read_lock(); 7348bb95dfb9SSean Christopherson 7349c50d8ae3SPaolo Bonzini ratio = READ_ONCE(nx_huge_pages_recovery_ratio); 7350ade74e14SSean Christopherson to_zap = ratio ? DIV_ROUND_UP(nx_lpage_splits, ratio) : 0; 73517d919c7aSSean Christopherson for ( ; to_zap; --to_zap) { 735255c510e2SSean Christopherson if (list_empty(&kvm->arch.possible_nx_huge_pages)) 73537d919c7aSSean Christopherson break; 73547d919c7aSSean Christopherson 7355c50d8ae3SPaolo Bonzini /* 7356c50d8ae3SPaolo Bonzini * We use a separate list instead of just using active_mmu_pages 735755c510e2SSean Christopherson * because the number of shadow pages that be replaced with an 735855c510e2SSean Christopherson * NX huge page is expected to be relatively small compared to 735955c510e2SSean Christopherson * the total number of shadow pages. And because the TDP MMU 736055c510e2SSean Christopherson * doesn't use active_mmu_pages. 7361c50d8ae3SPaolo Bonzini */ 736255c510e2SSean Christopherson sp = list_first_entry(&kvm->arch.possible_nx_huge_pages, 7363c50d8ae3SPaolo Bonzini struct kvm_mmu_page, 736455c510e2SSean Christopherson possible_nx_huge_page_link); 736555c510e2SSean Christopherson WARN_ON_ONCE(!sp->nx_huge_page_disallowed); 7366eb298605SDavid Matlack WARN_ON_ONCE(!sp->role.direct); 7367eb298605SDavid Matlack 7368eb298605SDavid Matlack /* 7369eb298605SDavid Matlack * Unaccount and do not attempt to recover any NX Huge Pages 7370eb298605SDavid Matlack * that are being dirty tracked, as they would just be faulted 7371eb298605SDavid Matlack * back in as 4KiB pages. The NX Huge Pages in this slot will be 7372eb298605SDavid Matlack * recovered, along with all the other huge pages in the slot, 7373eb298605SDavid Matlack * when dirty logging is disabled. 73746c7b2202SPaolo Bonzini * 73756c7b2202SPaolo Bonzini * Since gfn_to_memslot() is relatively expensive, it helps to 73766c7b2202SPaolo Bonzini * skip it if it the test cannot possibly return true. On the 73776c7b2202SPaolo Bonzini * other hand, if any memslot has logging enabled, chances are 73786c7b2202SPaolo Bonzini * good that all of them do, in which case unaccount_nx_huge_page() 73796c7b2202SPaolo Bonzini * is much cheaper than zapping the page. 73806c7b2202SPaolo Bonzini * 73816c7b2202SPaolo Bonzini * If a memslot update is in progress, reading an incorrect value 73826c7b2202SPaolo Bonzini * of kvm->nr_memslots_dirty_logging is not a problem: if it is 73836c7b2202SPaolo Bonzini * becoming zero, gfn_to_memslot() will be done unnecessarily; if 73846c7b2202SPaolo Bonzini * it is becoming nonzero, the page will be zapped unnecessarily. 73856c7b2202SPaolo Bonzini * Either way, this only affects efficiency in racy situations, 73866c7b2202SPaolo Bonzini * and not correctness. 7387eb298605SDavid Matlack */ 73886c7b2202SPaolo Bonzini slot = NULL; 73896c7b2202SPaolo Bonzini if (atomic_read(&kvm->nr_memslots_dirty_logging)) { 7390817fa998SSean Christopherson struct kvm_memslots *slots; 7391817fa998SSean Christopherson 7392817fa998SSean Christopherson slots = kvm_memslots_for_spte_role(kvm, sp->role); 7393817fa998SSean Christopherson slot = __gfn_to_memslot(slots, sp->gfn); 73946c7b2202SPaolo Bonzini WARN_ON_ONCE(!slot); 739529cf0f50SBen Gardon } 7396c50d8ae3SPaolo Bonzini 7397eb298605SDavid Matlack if (slot && kvm_slot_dirty_track_enabled(slot)) 7398eb298605SDavid Matlack unaccount_nx_huge_page(kvm, sp); 7399eb298605SDavid Matlack else if (is_tdp_mmu_page(sp)) 7400c50d8ae3SPaolo Bonzini flush |= kvm_tdp_mmu_zap_sp(kvm, sp); 74013a056757SSean Christopherson else 7402c50d8ae3SPaolo Bonzini kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list); 740355c510e2SSean Christopherson WARN_ON_ONCE(sp->nx_huge_page_disallowed); 7404c50d8ae3SPaolo Bonzini 7405531810caSBen Gardon if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) { 7406048f4980SSean Christopherson kvm_mmu_remote_flush_or_zap(kvm, &invalid_list, flush); 7407bb95dfb9SSean Christopherson rcu_read_unlock(); 7408bb95dfb9SSean Christopherson 7409531810caSBen Gardon cond_resched_rwlock_write(&kvm->mmu_lock); 7410048f4980SSean Christopherson flush = false; 7411bb95dfb9SSean Christopherson 7412bb95dfb9SSean Christopherson rcu_read_lock(); 7413c50d8ae3SPaolo Bonzini } 7414c50d8ae3SPaolo Bonzini } 7415048f4980SSean Christopherson kvm_mmu_remote_flush_or_zap(kvm, &invalid_list, flush); 7416c50d8ae3SPaolo Bonzini 7417bb95dfb9SSean Christopherson rcu_read_unlock(); 7418bb95dfb9SSean Christopherson 7419531810caSBen Gardon write_unlock(&kvm->mmu_lock); 7420c50d8ae3SPaolo Bonzini srcu_read_unlock(&kvm->srcu, rcu_idx); 7421c50d8ae3SPaolo Bonzini } 7422c50d8ae3SPaolo Bonzini 742355c510e2SSean Christopherson static long get_nx_huge_page_recovery_timeout(u64 start_time) 7424c50d8ae3SPaolo Bonzini { 7425f47491d7SSean Christopherson bool enabled; 7426f47491d7SSean Christopherson uint period; 74274dfe4f40SJunaid Shahid 7428f47491d7SSean Christopherson enabled = calc_nx_huge_pages_recovery_period(&period); 74294dfe4f40SJunaid Shahid 7430f47491d7SSean Christopherson return enabled ? start_time + msecs_to_jiffies(period) - get_jiffies_64() 7431c50d8ae3SPaolo Bonzini : MAX_SCHEDULE_TIMEOUT; 7432c50d8ae3SPaolo Bonzini } 7433c50d8ae3SPaolo Bonzini 743455c510e2SSean Christopherson static int kvm_nx_huge_page_recovery_worker(struct kvm *kvm, uintptr_t data) 7435c50d8ae3SPaolo Bonzini { 7436c50d8ae3SPaolo Bonzini u64 start_time; 7437c50d8ae3SPaolo Bonzini long remaining_time; 7438c50d8ae3SPaolo Bonzini 7439c50d8ae3SPaolo Bonzini while (true) { 7440c50d8ae3SPaolo Bonzini start_time = get_jiffies_64(); 744155c510e2SSean Christopherson remaining_time = get_nx_huge_page_recovery_timeout(start_time); 7442c50d8ae3SPaolo Bonzini 7443c50d8ae3SPaolo Bonzini set_current_state(TASK_INTERRUPTIBLE); 7444c50d8ae3SPaolo Bonzini while (!kthread_should_stop() && remaining_time > 0) { 7445c50d8ae3SPaolo Bonzini schedule_timeout(remaining_time); 744655c510e2SSean Christopherson remaining_time = get_nx_huge_page_recovery_timeout(start_time); 7447c50d8ae3SPaolo Bonzini set_current_state(TASK_INTERRUPTIBLE); 7448c50d8ae3SPaolo Bonzini } 7449c50d8ae3SPaolo Bonzini 7450c50d8ae3SPaolo Bonzini set_current_state(TASK_RUNNING); 7451c50d8ae3SPaolo Bonzini 7452c50d8ae3SPaolo Bonzini if (kthread_should_stop()) 7453c50d8ae3SPaolo Bonzini return 0; 7454c50d8ae3SPaolo Bonzini 745555c510e2SSean Christopherson kvm_recover_nx_huge_pages(kvm); 7456c50d8ae3SPaolo Bonzini } 7457c50d8ae3SPaolo Bonzini } 7458c50d8ae3SPaolo Bonzini 7459c50d8ae3SPaolo Bonzini int kvm_mmu_post_init_vm(struct kvm *kvm) 7460c50d8ae3SPaolo Bonzini { 7461c50d8ae3SPaolo Bonzini int err; 7462c50d8ae3SPaolo Bonzini 74630b210fafSSean Christopherson if (nx_hugepage_mitigation_hard_disabled) 74640b210fafSSean Christopherson return 0; 74650b210fafSSean Christopherson 746655c510e2SSean Christopherson err = kvm_vm_create_worker_thread(kvm, kvm_nx_huge_page_recovery_worker, 0, 7467c50d8ae3SPaolo Bonzini "kvm-nx-lpage-recovery", 746855c510e2SSean Christopherson &kvm->arch.nx_huge_page_recovery_thread); 7469c50d8ae3SPaolo Bonzini if (!err) 747055c510e2SSean Christopherson kthread_unpark(kvm->arch.nx_huge_page_recovery_thread); 7471c50d8ae3SPaolo Bonzini 7472c50d8ae3SPaolo Bonzini return err; 7473c50d8ae3SPaolo Bonzini } 7474c50d8ae3SPaolo Bonzini 7475c50d8ae3SPaolo Bonzini void kvm_mmu_pre_destroy_vm(struct kvm *kvm) 7476c50d8ae3SPaolo Bonzini { 747755c510e2SSean Christopherson if (kvm->arch.nx_huge_page_recovery_thread) 747855c510e2SSean Christopherson kthread_stop(kvm->arch.nx_huge_page_recovery_thread); 7479c50d8ae3SPaolo Bonzini } 748090b4fe17SChao Peng 748190b4fe17SChao Peng #ifdef CONFIG_KVM_GENERIC_MEMORY_ATTRIBUTES 74828dd2eee9SChao Peng bool kvm_arch_pre_set_memory_attributes(struct kvm *kvm, 74838dd2eee9SChao Peng struct kvm_gfn_range *range) 74848dd2eee9SChao Peng { 74858dd2eee9SChao Peng /* 74868dd2eee9SChao Peng * Zap SPTEs even if the slot can't be mapped PRIVATE. KVM x86 only 74878dd2eee9SChao Peng * supports KVM_MEMORY_ATTRIBUTE_PRIVATE, and so it *seems* like KVM 74888dd2eee9SChao Peng * can simply ignore such slots. But if userspace is making memory 74898dd2eee9SChao Peng * PRIVATE, then KVM must prevent the guest from accessing the memory 74908dd2eee9SChao Peng * as shared. And if userspace is making memory SHARED and this point 74918dd2eee9SChao Peng * is reached, then at least one page within the range was previously 74928dd2eee9SChao Peng * PRIVATE, i.e. the slot's possible hugepage ranges are changing. 74938dd2eee9SChao Peng * Zapping SPTEs in this case ensures KVM will reassess whether or not 74948dd2eee9SChao Peng * a hugepage can be used for affected ranges. 74958dd2eee9SChao Peng */ 74968dd2eee9SChao Peng if (WARN_ON_ONCE(!kvm_arch_has_private_mem(kvm))) 74978dd2eee9SChao Peng return false; 74988dd2eee9SChao Peng 74998dd2eee9SChao Peng return kvm_unmap_gfn_range(kvm, range); 75008dd2eee9SChao Peng } 75018dd2eee9SChao Peng 750290b4fe17SChao Peng static bool hugepage_test_mixed(struct kvm_memory_slot *slot, gfn_t gfn, 750390b4fe17SChao Peng int level) 750490b4fe17SChao Peng { 750590b4fe17SChao Peng return lpage_info_slot(gfn, slot, level)->disallow_lpage & KVM_LPAGE_MIXED_FLAG; 750690b4fe17SChao Peng } 750790b4fe17SChao Peng 750890b4fe17SChao Peng static void hugepage_clear_mixed(struct kvm_memory_slot *slot, gfn_t gfn, 750990b4fe17SChao Peng int level) 751090b4fe17SChao Peng { 751190b4fe17SChao Peng lpage_info_slot(gfn, slot, level)->disallow_lpage &= ~KVM_LPAGE_MIXED_FLAG; 751290b4fe17SChao Peng } 751390b4fe17SChao Peng 751490b4fe17SChao Peng static void hugepage_set_mixed(struct kvm_memory_slot *slot, gfn_t gfn, 751590b4fe17SChao Peng int level) 751690b4fe17SChao Peng { 751790b4fe17SChao Peng lpage_info_slot(gfn, slot, level)->disallow_lpage |= KVM_LPAGE_MIXED_FLAG; 751890b4fe17SChao Peng } 751990b4fe17SChao Peng 752090b4fe17SChao Peng static bool hugepage_has_attrs(struct kvm *kvm, struct kvm_memory_slot *slot, 752190b4fe17SChao Peng gfn_t gfn, int level, unsigned long attrs) 752290b4fe17SChao Peng { 752390b4fe17SChao Peng const unsigned long start = gfn; 752490b4fe17SChao Peng const unsigned long end = start + KVM_PAGES_PER_HPAGE(level); 752590b4fe17SChao Peng 752690b4fe17SChao Peng if (level == PG_LEVEL_2M) 752790b4fe17SChao Peng return kvm_range_has_memory_attributes(kvm, start, end, attrs); 752890b4fe17SChao Peng 752990b4fe17SChao Peng for (gfn = start; gfn < end; gfn += KVM_PAGES_PER_HPAGE(level - 1)) { 753090b4fe17SChao Peng if (hugepage_test_mixed(slot, gfn, level - 1) || 753190b4fe17SChao Peng attrs != kvm_get_memory_attributes(kvm, gfn)) 753290b4fe17SChao Peng return false; 753390b4fe17SChao Peng } 753490b4fe17SChao Peng return true; 753590b4fe17SChao Peng } 753690b4fe17SChao Peng 753790b4fe17SChao Peng bool kvm_arch_post_set_memory_attributes(struct kvm *kvm, 753890b4fe17SChao Peng struct kvm_gfn_range *range) 753990b4fe17SChao Peng { 754090b4fe17SChao Peng unsigned long attrs = range->arg.attributes; 754190b4fe17SChao Peng struct kvm_memory_slot *slot = range->slot; 754290b4fe17SChao Peng int level; 754390b4fe17SChao Peng 754490b4fe17SChao Peng lockdep_assert_held_write(&kvm->mmu_lock); 754590b4fe17SChao Peng lockdep_assert_held(&kvm->slots_lock); 754690b4fe17SChao Peng 754790b4fe17SChao Peng /* 754890b4fe17SChao Peng * Calculate which ranges can be mapped with hugepages even if the slot 754990b4fe17SChao Peng * can't map memory PRIVATE. KVM mustn't create a SHARED hugepage over 755090b4fe17SChao Peng * a range that has PRIVATE GFNs, and conversely converting a range to 755190b4fe17SChao Peng * SHARED may now allow hugepages. 755290b4fe17SChao Peng */ 755390b4fe17SChao Peng if (WARN_ON_ONCE(!kvm_arch_has_private_mem(kvm))) 755490b4fe17SChao Peng return false; 755590b4fe17SChao Peng 755690b4fe17SChao Peng /* 755790b4fe17SChao Peng * The sequence matters here: upper levels consume the result of lower 755890b4fe17SChao Peng * level's scanning. 755990b4fe17SChao Peng */ 756090b4fe17SChao Peng for (level = PG_LEVEL_2M; level <= KVM_MAX_HUGEPAGE_LEVEL; level++) { 756190b4fe17SChao Peng gfn_t nr_pages = KVM_PAGES_PER_HPAGE(level); 756290b4fe17SChao Peng gfn_t gfn = gfn_round_for_level(range->start, level); 756390b4fe17SChao Peng 756490b4fe17SChao Peng /* Process the head page if it straddles the range. */ 756590b4fe17SChao Peng if (gfn != range->start || gfn + nr_pages > range->end) { 756690b4fe17SChao Peng /* 756790b4fe17SChao Peng * Skip mixed tracking if the aligned gfn isn't covered 756890b4fe17SChao Peng * by the memslot, KVM can't use a hugepage due to the 756990b4fe17SChao Peng * misaligned address regardless of memory attributes. 757090b4fe17SChao Peng */ 7571992b54bdSRick Edgecombe if (gfn >= slot->base_gfn && 7572992b54bdSRick Edgecombe gfn + nr_pages <= slot->base_gfn + slot->npages) { 757390b4fe17SChao Peng if (hugepage_has_attrs(kvm, slot, gfn, level, attrs)) 757490b4fe17SChao Peng hugepage_clear_mixed(slot, gfn, level); 757590b4fe17SChao Peng else 757690b4fe17SChao Peng hugepage_set_mixed(slot, gfn, level); 757790b4fe17SChao Peng } 757890b4fe17SChao Peng gfn += nr_pages; 757990b4fe17SChao Peng } 758090b4fe17SChao Peng 758190b4fe17SChao Peng /* 758290b4fe17SChao Peng * Pages entirely covered by the range are guaranteed to have 758390b4fe17SChao Peng * only the attributes which were just set. 758490b4fe17SChao Peng */ 758590b4fe17SChao Peng for ( ; gfn + nr_pages <= range->end; gfn += nr_pages) 758690b4fe17SChao Peng hugepage_clear_mixed(slot, gfn, level); 758790b4fe17SChao Peng 758890b4fe17SChao Peng /* 758990b4fe17SChao Peng * Process the last tail page if it straddles the range and is 759090b4fe17SChao Peng * contained by the memslot. Like the head page, KVM can't 759190b4fe17SChao Peng * create a hugepage if the slot size is misaligned. 759290b4fe17SChao Peng */ 759390b4fe17SChao Peng if (gfn < range->end && 759490b4fe17SChao Peng (gfn + nr_pages) <= (slot->base_gfn + slot->npages)) { 759590b4fe17SChao Peng if (hugepage_has_attrs(kvm, slot, gfn, level, attrs)) 759690b4fe17SChao Peng hugepage_clear_mixed(slot, gfn, level); 759790b4fe17SChao Peng else 759890b4fe17SChao Peng hugepage_set_mixed(slot, gfn, level); 759990b4fe17SChao Peng } 760090b4fe17SChao Peng } 760190b4fe17SChao Peng return false; 760290b4fe17SChao Peng } 760390b4fe17SChao Peng 760490b4fe17SChao Peng void kvm_mmu_init_memslot_memory_attributes(struct kvm *kvm, 760590b4fe17SChao Peng struct kvm_memory_slot *slot) 760690b4fe17SChao Peng { 760790b4fe17SChao Peng int level; 760890b4fe17SChao Peng 760990b4fe17SChao Peng if (!kvm_arch_has_private_mem(kvm)) 761090b4fe17SChao Peng return; 761190b4fe17SChao Peng 761290b4fe17SChao Peng for (level = PG_LEVEL_2M; level <= KVM_MAX_HUGEPAGE_LEVEL; level++) { 761390b4fe17SChao Peng /* 761490b4fe17SChao Peng * Don't bother tracking mixed attributes for pages that can't 761590b4fe17SChao Peng * be huge due to alignment, i.e. process only pages that are 761690b4fe17SChao Peng * entirely contained by the memslot. 761790b4fe17SChao Peng */ 761890b4fe17SChao Peng gfn_t end = gfn_round_for_level(slot->base_gfn + slot->npages, level); 761990b4fe17SChao Peng gfn_t start = gfn_round_for_level(slot->base_gfn, level); 762090b4fe17SChao Peng gfn_t nr_pages = KVM_PAGES_PER_HPAGE(level); 762190b4fe17SChao Peng gfn_t gfn; 762290b4fe17SChao Peng 762390b4fe17SChao Peng if (start < slot->base_gfn) 762490b4fe17SChao Peng start += nr_pages; 762590b4fe17SChao Peng 762690b4fe17SChao Peng /* 762790b4fe17SChao Peng * Unlike setting attributes, every potential hugepage needs to 762890b4fe17SChao Peng * be manually checked as the attributes may already be mixed. 762990b4fe17SChao Peng */ 763090b4fe17SChao Peng for (gfn = start; gfn < end; gfn += nr_pages) { 763190b4fe17SChao Peng unsigned long attrs = kvm_get_memory_attributes(kvm, gfn); 763290b4fe17SChao Peng 763390b4fe17SChao Peng if (hugepage_has_attrs(kvm, slot, gfn, level, attrs)) 763490b4fe17SChao Peng hugepage_clear_mixed(slot, gfn, level); 763590b4fe17SChao Peng else 763690b4fe17SChao Peng hugepage_set_mixed(slot, gfn, level); 763790b4fe17SChao Peng } 763890b4fe17SChao Peng } 763990b4fe17SChao Peng } 764090b4fe17SChao Peng #endif 7641