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" 28c50d8ae3SPaolo Bonzini #include "cpuid.h" 295a9624afSPaolo Bonzini #include "spte.h" 30c50d8ae3SPaolo Bonzini 31c50d8ae3SPaolo Bonzini #include <linux/kvm_host.h> 32c50d8ae3SPaolo Bonzini #include <linux/types.h> 33c50d8ae3SPaolo Bonzini #include <linux/string.h> 34c50d8ae3SPaolo Bonzini #include <linux/mm.h> 35c50d8ae3SPaolo Bonzini #include <linux/highmem.h> 36c50d8ae3SPaolo Bonzini #include <linux/moduleparam.h> 37c50d8ae3SPaolo Bonzini #include <linux/export.h> 38c50d8ae3SPaolo Bonzini #include <linux/swap.h> 39c50d8ae3SPaolo Bonzini #include <linux/hugetlb.h> 40c50d8ae3SPaolo Bonzini #include <linux/compiler.h> 41c50d8ae3SPaolo Bonzini #include <linux/srcu.h> 42c50d8ae3SPaolo Bonzini #include <linux/slab.h> 43c50d8ae3SPaolo Bonzini #include <linux/sched/signal.h> 44c50d8ae3SPaolo Bonzini #include <linux/uaccess.h> 45c50d8ae3SPaolo Bonzini #include <linux/hash.h> 46c50d8ae3SPaolo Bonzini #include <linux/kern_levels.h> 4711b36fe7SChristophe JAILLET #include <linux/kstrtox.h> 48c50d8ae3SPaolo Bonzini #include <linux/kthread.h> 49c50d8ae3SPaolo Bonzini 50c50d8ae3SPaolo Bonzini #include <asm/page.h> 51eb243d1dSIngo Molnar #include <asm/memtype.h> 52c50d8ae3SPaolo Bonzini #include <asm/cmpxchg.h> 53c50d8ae3SPaolo Bonzini #include <asm/io.h> 544a98623dSSean Christopherson #include <asm/set_memory.h> 55c50d8ae3SPaolo Bonzini #include <asm/vmx.h> 56c50d8ae3SPaolo Bonzini #include <asm/kvm_page_track.h> 57c50d8ae3SPaolo Bonzini #include "trace.h" 58c50d8ae3SPaolo Bonzini 59c50d8ae3SPaolo Bonzini extern bool itlb_multihit_kvm_mitigation; 60c50d8ae3SPaolo Bonzini 61a9d6496dSShaokun Zhang int __read_mostly nx_huge_pages = -1; 624dfe4f40SJunaid Shahid static uint __read_mostly nx_huge_pages_recovery_period_ms; 63c50d8ae3SPaolo Bonzini #ifdef CONFIG_PREEMPT_RT 64c50d8ae3SPaolo Bonzini /* Recovery can cause latency spikes, disable it for PREEMPT_RT. */ 65c50d8ae3SPaolo Bonzini static uint __read_mostly nx_huge_pages_recovery_ratio = 0; 66c50d8ae3SPaolo Bonzini #else 67c50d8ae3SPaolo Bonzini static uint __read_mostly nx_huge_pages_recovery_ratio = 60; 68c50d8ae3SPaolo Bonzini #endif 69c50d8ae3SPaolo Bonzini 70c50d8ae3SPaolo Bonzini static int set_nx_huge_pages(const char *val, const struct kernel_param *kp); 714dfe4f40SJunaid Shahid static int set_nx_huge_pages_recovery_param(const char *val, const struct kernel_param *kp); 72c50d8ae3SPaolo Bonzini 73d5d6c18dSJoe Perches static const struct kernel_param_ops nx_huge_pages_ops = { 74c50d8ae3SPaolo Bonzini .set = set_nx_huge_pages, 75c50d8ae3SPaolo Bonzini .get = param_get_bool, 76c50d8ae3SPaolo Bonzini }; 77c50d8ae3SPaolo Bonzini 784dfe4f40SJunaid Shahid static const struct kernel_param_ops nx_huge_pages_recovery_param_ops = { 794dfe4f40SJunaid Shahid .set = set_nx_huge_pages_recovery_param, 80c50d8ae3SPaolo Bonzini .get = param_get_uint, 81c50d8ae3SPaolo Bonzini }; 82c50d8ae3SPaolo Bonzini 83c50d8ae3SPaolo Bonzini module_param_cb(nx_huge_pages, &nx_huge_pages_ops, &nx_huge_pages, 0644); 84c50d8ae3SPaolo Bonzini __MODULE_PARM_TYPE(nx_huge_pages, "bool"); 854dfe4f40SJunaid Shahid module_param_cb(nx_huge_pages_recovery_ratio, &nx_huge_pages_recovery_param_ops, 86c50d8ae3SPaolo Bonzini &nx_huge_pages_recovery_ratio, 0644); 87c50d8ae3SPaolo Bonzini __MODULE_PARM_TYPE(nx_huge_pages_recovery_ratio, "uint"); 884dfe4f40SJunaid Shahid module_param_cb(nx_huge_pages_recovery_period_ms, &nx_huge_pages_recovery_param_ops, 894dfe4f40SJunaid Shahid &nx_huge_pages_recovery_period_ms, 0644); 904dfe4f40SJunaid Shahid __MODULE_PARM_TYPE(nx_huge_pages_recovery_period_ms, "uint"); 91c50d8ae3SPaolo Bonzini 9271fe7013SSean Christopherson static bool __read_mostly force_flush_and_sync_on_reuse; 9371fe7013SSean Christopherson module_param_named(flush_on_reuse, force_flush_and_sync_on_reuse, bool, 0644); 9471fe7013SSean Christopherson 95c50d8ae3SPaolo Bonzini /* 96c50d8ae3SPaolo Bonzini * When setting this variable to true it enables Two-Dimensional-Paging 97c50d8ae3SPaolo Bonzini * where the hardware walks 2 page tables: 98c50d8ae3SPaolo Bonzini * 1. the guest-virtual to guest-physical 99c50d8ae3SPaolo Bonzini * 2. while doing 1. it walks guest-physical to host-physical 100c50d8ae3SPaolo Bonzini * If the hardware supports that we don't need to do shadow paging. 101c50d8ae3SPaolo Bonzini */ 102c50d8ae3SPaolo Bonzini bool tdp_enabled = false; 103c50d8ae3SPaolo Bonzini 1047f604e92SDavid Matlack static bool __ro_after_init tdp_mmu_allowed; 1051f98f2bdSDavid Matlack 1061f98f2bdSDavid Matlack #ifdef CONFIG_X86_64 1071f98f2bdSDavid Matlack bool __read_mostly tdp_mmu_enabled = true; 1081f98f2bdSDavid Matlack module_param_named(tdp_mmu, tdp_mmu_enabled, bool, 0444); 1091f98f2bdSDavid Matlack #endif 1101f98f2bdSDavid Matlack 1111d92d2e8SSean Christopherson static int max_huge_page_level __read_mostly; 112746700d2SWei Huang static int tdp_root_level __read_mostly; 11383013059SSean Christopherson static int max_tdp_level __read_mostly; 114703c335dSSean Christopherson 115c50d8ae3SPaolo Bonzini #ifdef MMU_DEBUG 1165a9624afSPaolo Bonzini bool dbg = 0; 117c50d8ae3SPaolo Bonzini module_param(dbg, bool, 0644); 118c50d8ae3SPaolo Bonzini #endif 119c50d8ae3SPaolo Bonzini 120c50d8ae3SPaolo Bonzini #define PTE_PREFETCH_NUM 8 121c50d8ae3SPaolo Bonzini 122c50d8ae3SPaolo Bonzini #include <trace/events/kvm.h> 123c50d8ae3SPaolo Bonzini 124dc1cff96SPeter Xu /* make pte_list_desc fit well in cache lines */ 12513236e25SPeter Xu #define PTE_LIST_EXT 14 126c50d8ae3SPaolo Bonzini 12713236e25SPeter Xu /* 128141705b7SLai Jiangshan * struct pte_list_desc is the core data structure used to implement a custom 129141705b7SLai Jiangshan * list for tracking a set of related SPTEs, e.g. all the SPTEs that map a 130141705b7SLai Jiangshan * given GFN when used in the context of rmaps. Using a custom list allows KVM 131141705b7SLai Jiangshan * to optimize for the common case where many GFNs will have at most a handful 132141705b7SLai Jiangshan * of SPTEs pointing at them, i.e. allows packing multiple SPTEs into a small 133141705b7SLai Jiangshan * memory footprint, which in turn improves runtime performance by exploiting 134141705b7SLai Jiangshan * cache locality. 135141705b7SLai Jiangshan * 136141705b7SLai Jiangshan * A list is comprised of one or more pte_list_desc objects (descriptors). 137141705b7SLai Jiangshan * Each individual descriptor stores up to PTE_LIST_EXT SPTEs. If a descriptor 138141705b7SLai Jiangshan * is full and a new SPTEs needs to be added, a new descriptor is allocated and 139141705b7SLai Jiangshan * becomes the head of the list. This means that by definitions, all tail 140141705b7SLai Jiangshan * descriptors are full. 141141705b7SLai Jiangshan * 142141705b7SLai Jiangshan * Note, the meta data fields are deliberately placed at the start of the 143141705b7SLai Jiangshan * structure to optimize the cacheline layout; accessing the descriptor will 144141705b7SLai Jiangshan * touch only a single cacheline so long as @spte_count<=6 (or if only the 145141705b7SLai Jiangshan * descriptors metadata is accessed). 14613236e25SPeter Xu */ 147c50d8ae3SPaolo Bonzini struct pte_list_desc { 148c50d8ae3SPaolo Bonzini struct pte_list_desc *more; 149141705b7SLai Jiangshan /* The number of PTEs stored in _this_ descriptor. */ 150141705b7SLai Jiangshan u32 spte_count; 151141705b7SLai Jiangshan /* The number of PTEs stored in all tails of this descriptor. */ 152141705b7SLai Jiangshan u32 tail_count; 15313236e25SPeter Xu u64 *sptes[PTE_LIST_EXT]; 154c50d8ae3SPaolo Bonzini }; 155c50d8ae3SPaolo Bonzini 156c50d8ae3SPaolo Bonzini struct kvm_shadow_walk_iterator { 157c50d8ae3SPaolo Bonzini u64 addr; 158c50d8ae3SPaolo Bonzini hpa_t shadow_addr; 159c50d8ae3SPaolo Bonzini u64 *sptep; 160c50d8ae3SPaolo Bonzini int level; 161c50d8ae3SPaolo Bonzini unsigned index; 162c50d8ae3SPaolo Bonzini }; 163c50d8ae3SPaolo Bonzini 164c50d8ae3SPaolo Bonzini #define for_each_shadow_entry_using_root(_vcpu, _root, _addr, _walker) \ 165c50d8ae3SPaolo Bonzini for (shadow_walk_init_using_root(&(_walker), (_vcpu), \ 166c50d8ae3SPaolo Bonzini (_root), (_addr)); \ 167c50d8ae3SPaolo Bonzini shadow_walk_okay(&(_walker)); \ 168c50d8ae3SPaolo Bonzini shadow_walk_next(&(_walker))) 169c50d8ae3SPaolo Bonzini 170c50d8ae3SPaolo Bonzini #define for_each_shadow_entry(_vcpu, _addr, _walker) \ 171c50d8ae3SPaolo Bonzini for (shadow_walk_init(&(_walker), _vcpu, _addr); \ 172c50d8ae3SPaolo Bonzini shadow_walk_okay(&(_walker)); \ 173c50d8ae3SPaolo Bonzini shadow_walk_next(&(_walker))) 174c50d8ae3SPaolo Bonzini 175c50d8ae3SPaolo Bonzini #define for_each_shadow_entry_lockless(_vcpu, _addr, _walker, spte) \ 176c50d8ae3SPaolo Bonzini for (shadow_walk_init(&(_walker), _vcpu, _addr); \ 177c50d8ae3SPaolo Bonzini shadow_walk_okay(&(_walker)) && \ 178c50d8ae3SPaolo Bonzini ({ spte = mmu_spte_get_lockless(_walker.sptep); 1; }); \ 179c50d8ae3SPaolo Bonzini __shadow_walk_next(&(_walker), spte)) 180c50d8ae3SPaolo Bonzini 181c50d8ae3SPaolo Bonzini static struct kmem_cache *pte_list_desc_cache; 18202c00b3aSBen Gardon struct kmem_cache *mmu_page_header_cache; 183c50d8ae3SPaolo Bonzini static struct percpu_counter kvm_total_used_mmu_pages; 184c50d8ae3SPaolo Bonzini 185c50d8ae3SPaolo Bonzini static void mmu_spte_set(u64 *sptep, u64 spte); 186c50d8ae3SPaolo Bonzini 187594e91a1SSean Christopherson struct kvm_mmu_role_regs { 188594e91a1SSean Christopherson const unsigned long cr0; 189594e91a1SSean Christopherson const unsigned long cr4; 190594e91a1SSean Christopherson const u64 efer; 191594e91a1SSean Christopherson }; 192594e91a1SSean Christopherson 193c50d8ae3SPaolo Bonzini #define CREATE_TRACE_POINTS 194c50d8ae3SPaolo Bonzini #include "mmutrace.h" 195c50d8ae3SPaolo Bonzini 196594e91a1SSean Christopherson /* 197594e91a1SSean Christopherson * Yes, lot's of underscores. They're a hint that you probably shouldn't be 1987a458f0eSPaolo Bonzini * reading from the role_regs. Once the root_role is constructed, it becomes 199594e91a1SSean Christopherson * the single source of truth for the MMU's state. 200594e91a1SSean Christopherson */ 201594e91a1SSean Christopherson #define BUILD_MMU_ROLE_REGS_ACCESSOR(reg, name, flag) \ 20282ffa13fSPaolo Bonzini static inline bool __maybe_unused \ 20382ffa13fSPaolo Bonzini ____is_##reg##_##name(const struct kvm_mmu_role_regs *regs) \ 204594e91a1SSean Christopherson { \ 205594e91a1SSean Christopherson return !!(regs->reg & flag); \ 206594e91a1SSean Christopherson } 207594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr0, pg, X86_CR0_PG); 208594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr0, wp, X86_CR0_WP); 209594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pse, X86_CR4_PSE); 210594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pae, X86_CR4_PAE); 211594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, smep, X86_CR4_SMEP); 212594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, smap, X86_CR4_SMAP); 213594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pke, X86_CR4_PKE); 214594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, la57, X86_CR4_LA57); 215594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(efer, nx, EFER_NX); 216594e91a1SSean Christopherson BUILD_MMU_ROLE_REGS_ACCESSOR(efer, lma, EFER_LMA); 217594e91a1SSean Christopherson 21860667724SSean Christopherson /* 21960667724SSean Christopherson * The MMU itself (with a valid role) is the single source of truth for the 22060667724SSean Christopherson * MMU. Do not use the regs used to build the MMU/role, nor the vCPU. The 22160667724SSean Christopherson * regs don't account for dependencies, e.g. clearing CR4 bits if CR0.PG=1, 22260667724SSean Christopherson * and the vCPU may be incorrect/irrelevant. 22360667724SSean Christopherson */ 22460667724SSean Christopherson #define BUILD_MMU_ROLE_ACCESSOR(base_or_ext, reg, name) \ 2254ac21457SPaolo Bonzini static inline bool __maybe_unused is_##reg##_##name(struct kvm_mmu *mmu) \ 22660667724SSean Christopherson { \ 227e5ed0fb0SPaolo Bonzini return !!(mmu->cpu_role. base_or_ext . reg##_##name); \ 22860667724SSean Christopherson } 22960667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(base, cr0, wp); 23060667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(ext, cr4, pse); 23160667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(ext, cr4, smep); 23260667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(ext, cr4, smap); 23360667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(ext, cr4, pke); 23460667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(ext, cr4, la57); 23560667724SSean Christopherson BUILD_MMU_ROLE_ACCESSOR(base, efer, nx); 23656b321f9SPaolo Bonzini BUILD_MMU_ROLE_ACCESSOR(ext, efer, lma); 23760667724SSean Christopherson 238faf72962SPaolo Bonzini static inline bool is_cr0_pg(struct kvm_mmu *mmu) 239faf72962SPaolo Bonzini { 240faf72962SPaolo Bonzini return mmu->cpu_role.base.level > 0; 241faf72962SPaolo Bonzini } 242faf72962SPaolo Bonzini 243faf72962SPaolo Bonzini static inline bool is_cr4_pae(struct kvm_mmu *mmu) 244faf72962SPaolo Bonzini { 245faf72962SPaolo Bonzini return !mmu->cpu_role.base.has_4_byte_gpte; 246faf72962SPaolo Bonzini } 247faf72962SPaolo Bonzini 248594e91a1SSean Christopherson static struct kvm_mmu_role_regs vcpu_to_role_regs(struct kvm_vcpu *vcpu) 249594e91a1SSean Christopherson { 250594e91a1SSean Christopherson struct kvm_mmu_role_regs regs = { 251594e91a1SSean Christopherson .cr0 = kvm_read_cr0_bits(vcpu, KVM_MMU_CR0_ROLE_BITS), 252594e91a1SSean Christopherson .cr4 = kvm_read_cr4_bits(vcpu, KVM_MMU_CR4_ROLE_BITS), 253594e91a1SSean Christopherson .efer = vcpu->arch.efer, 254594e91a1SSean Christopherson }; 255594e91a1SSean Christopherson 256594e91a1SSean Christopherson return regs; 257594e91a1SSean Christopherson } 258c50d8ae3SPaolo Bonzini 2592fdcc1b3SPaolo Bonzini static unsigned long get_guest_cr3(struct kvm_vcpu *vcpu) 260c50d8ae3SPaolo Bonzini { 2612fdcc1b3SPaolo Bonzini return kvm_read_cr3(vcpu); 262c50d8ae3SPaolo Bonzini } 263c50d8ae3SPaolo Bonzini 2642fdcc1b3SPaolo Bonzini static inline unsigned long kvm_mmu_get_guest_pgd(struct kvm_vcpu *vcpu, 2652fdcc1b3SPaolo Bonzini struct kvm_mmu *mmu) 266c50d8ae3SPaolo Bonzini { 2672fdcc1b3SPaolo Bonzini if (IS_ENABLED(CONFIG_RETPOLINE) && mmu->get_guest_pgd == get_guest_cr3) 2682fdcc1b3SPaolo Bonzini return kvm_read_cr3(vcpu); 269c50d8ae3SPaolo Bonzini 2702fdcc1b3SPaolo Bonzini return mmu->get_guest_pgd(vcpu); 2712fdcc1b3SPaolo Bonzini } 272c50d8ae3SPaolo Bonzini 2738a1300ffSSean Christopherson static inline bool kvm_available_flush_remote_tlbs_range(void) 274c50d8ae3SPaolo Bonzini { 2758a1300ffSSean Christopherson return kvm_x86_ops.flush_remote_tlbs_range; 276c50d8ae3SPaolo Bonzini } 277c50d8ae3SPaolo Bonzini 2789d4655daSDavid Matlack void kvm_flush_remote_tlbs_range(struct kvm *kvm, gfn_t start_gfn, 2799d4655daSDavid Matlack gfn_t nr_pages) 280c50d8ae3SPaolo Bonzini { 28128e4b459SDavid Matlack int ret = -EOPNOTSUPP; 282c50d8ae3SPaolo Bonzini 2838a1300ffSSean Christopherson if (kvm_x86_ops.flush_remote_tlbs_range) 2849ed3bf41SSean Christopherson ret = static_call(kvm_x86_flush_remote_tlbs_range)(kvm, start_gfn, 2859ed3bf41SSean Christopherson nr_pages); 286c50d8ae3SPaolo Bonzini if (ret) 287c50d8ae3SPaolo Bonzini kvm_flush_remote_tlbs(kvm); 288c50d8ae3SPaolo Bonzini } 289c50d8ae3SPaolo Bonzini 2901b2dc736SHou Wenlong static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index); 2911b2dc736SHou Wenlong 2921b2dc736SHou Wenlong /* Flush the range of guest memory mapped by the given SPTE. */ 2931b2dc736SHou Wenlong static void kvm_flush_remote_tlbs_sptep(struct kvm *kvm, u64 *sptep) 2941b2dc736SHou Wenlong { 2951b2dc736SHou Wenlong struct kvm_mmu_page *sp = sptep_to_sp(sptep); 2961b2dc736SHou Wenlong gfn_t gfn = kvm_mmu_page_get_gfn(sp, spte_index(sptep)); 2971b2dc736SHou Wenlong 2981b2dc736SHou Wenlong kvm_flush_remote_tlbs_gfn(kvm, gfn, sp->role.level); 2991b2dc736SHou Wenlong } 3001b2dc736SHou Wenlong 3018f79b064SBen Gardon static void mark_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, u64 gfn, 3028f79b064SBen Gardon unsigned int access) 3038f79b064SBen Gardon { 304c236d962SSean Christopherson u64 spte = make_mmio_spte(vcpu, gfn, access); 3058f79b064SBen Gardon 306c236d962SSean Christopherson trace_mark_mmio_spte(sptep, gfn, spte); 307c236d962SSean Christopherson mmu_spte_set(sptep, spte); 308c50d8ae3SPaolo Bonzini } 309c50d8ae3SPaolo Bonzini 310c50d8ae3SPaolo Bonzini static gfn_t get_mmio_spte_gfn(u64 spte) 311c50d8ae3SPaolo Bonzini { 312c50d8ae3SPaolo Bonzini u64 gpa = spte & shadow_nonpresent_or_rsvd_lower_gfn_mask; 313c50d8ae3SPaolo Bonzini 3148a967d65SPaolo Bonzini gpa |= (spte >> SHADOW_NONPRESENT_OR_RSVD_MASK_LEN) 315c50d8ae3SPaolo Bonzini & shadow_nonpresent_or_rsvd_mask; 316c50d8ae3SPaolo Bonzini 317c50d8ae3SPaolo Bonzini return gpa >> PAGE_SHIFT; 318c50d8ae3SPaolo Bonzini } 319c50d8ae3SPaolo Bonzini 320c50d8ae3SPaolo Bonzini static unsigned get_mmio_spte_access(u64 spte) 321c50d8ae3SPaolo Bonzini { 322c50d8ae3SPaolo Bonzini return spte & shadow_mmio_access_mask; 323c50d8ae3SPaolo Bonzini } 324c50d8ae3SPaolo Bonzini 325c50d8ae3SPaolo Bonzini static bool check_mmio_spte(struct kvm_vcpu *vcpu, u64 spte) 326c50d8ae3SPaolo Bonzini { 327c50d8ae3SPaolo Bonzini u64 kvm_gen, spte_gen, gen; 328c50d8ae3SPaolo Bonzini 329c50d8ae3SPaolo Bonzini gen = kvm_vcpu_memslots(vcpu)->generation; 330c50d8ae3SPaolo Bonzini if (unlikely(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS)) 331c50d8ae3SPaolo Bonzini return false; 332c50d8ae3SPaolo Bonzini 333c50d8ae3SPaolo Bonzini kvm_gen = gen & MMIO_SPTE_GEN_MASK; 334c50d8ae3SPaolo Bonzini spte_gen = get_mmio_spte_generation(spte); 335c50d8ae3SPaolo Bonzini 336c50d8ae3SPaolo Bonzini trace_check_mmio_spte(spte, kvm_gen, spte_gen); 337c50d8ae3SPaolo Bonzini return likely(kvm_gen == spte_gen); 338c50d8ae3SPaolo Bonzini } 339c50d8ae3SPaolo Bonzini 340c50d8ae3SPaolo Bonzini static int is_cpuid_PSE36(void) 341c50d8ae3SPaolo Bonzini { 342c50d8ae3SPaolo Bonzini return 1; 343c50d8ae3SPaolo Bonzini } 344c50d8ae3SPaolo Bonzini 345c50d8ae3SPaolo Bonzini #ifdef CONFIG_X86_64 346c50d8ae3SPaolo Bonzini static void __set_spte(u64 *sptep, u64 spte) 347c50d8ae3SPaolo Bonzini { 348c50d8ae3SPaolo Bonzini WRITE_ONCE(*sptep, spte); 349c50d8ae3SPaolo Bonzini } 350c50d8ae3SPaolo Bonzini 351c50d8ae3SPaolo Bonzini static void __update_clear_spte_fast(u64 *sptep, u64 spte) 352c50d8ae3SPaolo Bonzini { 353c50d8ae3SPaolo Bonzini WRITE_ONCE(*sptep, spte); 354c50d8ae3SPaolo Bonzini } 355c50d8ae3SPaolo Bonzini 356c50d8ae3SPaolo Bonzini static u64 __update_clear_spte_slow(u64 *sptep, u64 spte) 357c50d8ae3SPaolo Bonzini { 358c50d8ae3SPaolo Bonzini return xchg(sptep, spte); 359c50d8ae3SPaolo Bonzini } 360c50d8ae3SPaolo Bonzini 361c50d8ae3SPaolo Bonzini static u64 __get_spte_lockless(u64 *sptep) 362c50d8ae3SPaolo Bonzini { 363c50d8ae3SPaolo Bonzini return READ_ONCE(*sptep); 364c50d8ae3SPaolo Bonzini } 365c50d8ae3SPaolo Bonzini #else 366c50d8ae3SPaolo Bonzini union split_spte { 367c50d8ae3SPaolo Bonzini struct { 368c50d8ae3SPaolo Bonzini u32 spte_low; 369c50d8ae3SPaolo Bonzini u32 spte_high; 370c50d8ae3SPaolo Bonzini }; 371c50d8ae3SPaolo Bonzini u64 spte; 372c50d8ae3SPaolo Bonzini }; 373c50d8ae3SPaolo Bonzini 374c50d8ae3SPaolo Bonzini static void count_spte_clear(u64 *sptep, u64 spte) 375c50d8ae3SPaolo Bonzini { 37657354682SSean Christopherson struct kvm_mmu_page *sp = sptep_to_sp(sptep); 377c50d8ae3SPaolo Bonzini 378c50d8ae3SPaolo Bonzini if (is_shadow_present_pte(spte)) 379c50d8ae3SPaolo Bonzini return; 380c50d8ae3SPaolo Bonzini 381c50d8ae3SPaolo Bonzini /* Ensure the spte is completely set before we increase the count */ 382c50d8ae3SPaolo Bonzini smp_wmb(); 383c50d8ae3SPaolo Bonzini sp->clear_spte_count++; 384c50d8ae3SPaolo Bonzini } 385c50d8ae3SPaolo Bonzini 386c50d8ae3SPaolo Bonzini static void __set_spte(u64 *sptep, u64 spte) 387c50d8ae3SPaolo Bonzini { 388c50d8ae3SPaolo Bonzini union split_spte *ssptep, sspte; 389c50d8ae3SPaolo Bonzini 390c50d8ae3SPaolo Bonzini ssptep = (union split_spte *)sptep; 391c50d8ae3SPaolo Bonzini sspte = (union split_spte)spte; 392c50d8ae3SPaolo Bonzini 393c50d8ae3SPaolo Bonzini ssptep->spte_high = sspte.spte_high; 394c50d8ae3SPaolo Bonzini 395c50d8ae3SPaolo Bonzini /* 396c50d8ae3SPaolo Bonzini * If we map the spte from nonpresent to present, We should store 397c50d8ae3SPaolo Bonzini * the high bits firstly, then set present bit, so cpu can not 398c50d8ae3SPaolo Bonzini * fetch this spte while we are setting the spte. 399c50d8ae3SPaolo Bonzini */ 400c50d8ae3SPaolo Bonzini smp_wmb(); 401c50d8ae3SPaolo Bonzini 402c50d8ae3SPaolo Bonzini WRITE_ONCE(ssptep->spte_low, sspte.spte_low); 403c50d8ae3SPaolo Bonzini } 404c50d8ae3SPaolo Bonzini 405c50d8ae3SPaolo Bonzini static void __update_clear_spte_fast(u64 *sptep, u64 spte) 406c50d8ae3SPaolo Bonzini { 407c50d8ae3SPaolo Bonzini union split_spte *ssptep, sspte; 408c50d8ae3SPaolo Bonzini 409c50d8ae3SPaolo Bonzini ssptep = (union split_spte *)sptep; 410c50d8ae3SPaolo Bonzini sspte = (union split_spte)spte; 411c50d8ae3SPaolo Bonzini 412c50d8ae3SPaolo Bonzini WRITE_ONCE(ssptep->spte_low, sspte.spte_low); 413c50d8ae3SPaolo Bonzini 414c50d8ae3SPaolo Bonzini /* 415c50d8ae3SPaolo Bonzini * If we map the spte from present to nonpresent, we should clear 416c50d8ae3SPaolo Bonzini * present bit firstly to avoid vcpu fetch the old high bits. 417c50d8ae3SPaolo Bonzini */ 418c50d8ae3SPaolo Bonzini smp_wmb(); 419c50d8ae3SPaolo Bonzini 420c50d8ae3SPaolo Bonzini ssptep->spte_high = sspte.spte_high; 421c50d8ae3SPaolo Bonzini count_spte_clear(sptep, spte); 422c50d8ae3SPaolo Bonzini } 423c50d8ae3SPaolo Bonzini 424c50d8ae3SPaolo Bonzini static u64 __update_clear_spte_slow(u64 *sptep, u64 spte) 425c50d8ae3SPaolo Bonzini { 426c50d8ae3SPaolo Bonzini union split_spte *ssptep, sspte, orig; 427c50d8ae3SPaolo Bonzini 428c50d8ae3SPaolo Bonzini ssptep = (union split_spte *)sptep; 429c50d8ae3SPaolo Bonzini sspte = (union split_spte)spte; 430c50d8ae3SPaolo Bonzini 431c50d8ae3SPaolo Bonzini /* xchg acts as a barrier before the setting of the high bits */ 432c50d8ae3SPaolo Bonzini orig.spte_low = xchg(&ssptep->spte_low, sspte.spte_low); 433c50d8ae3SPaolo Bonzini orig.spte_high = ssptep->spte_high; 434c50d8ae3SPaolo Bonzini ssptep->spte_high = sspte.spte_high; 435c50d8ae3SPaolo Bonzini count_spte_clear(sptep, spte); 436c50d8ae3SPaolo Bonzini 437c50d8ae3SPaolo Bonzini return orig.spte; 438c50d8ae3SPaolo Bonzini } 439c50d8ae3SPaolo Bonzini 440c50d8ae3SPaolo Bonzini /* 441c50d8ae3SPaolo Bonzini * The idea using the light way get the spte on x86_32 guest is from 442c50d8ae3SPaolo Bonzini * gup_get_pte (mm/gup.c). 443c50d8ae3SPaolo Bonzini * 444aed02fe3SSean Christopherson * An spte tlb flush may be pending, because kvm_set_pte_rmap 445c50d8ae3SPaolo Bonzini * coalesces them and we are running out of the MMU lock. Therefore 446c50d8ae3SPaolo Bonzini * we need to protect against in-progress updates of the spte. 447c50d8ae3SPaolo Bonzini * 448c50d8ae3SPaolo Bonzini * Reading the spte while an update is in progress may get the old value 449c50d8ae3SPaolo Bonzini * for the high part of the spte. The race is fine for a present->non-present 450c50d8ae3SPaolo Bonzini * change (because the high part of the spte is ignored for non-present spte), 451c50d8ae3SPaolo Bonzini * but for a present->present change we must reread the spte. 452c50d8ae3SPaolo Bonzini * 453c50d8ae3SPaolo Bonzini * All such changes are done in two steps (present->non-present and 454c50d8ae3SPaolo Bonzini * non-present->present), hence it is enough to count the number of 455c50d8ae3SPaolo Bonzini * present->non-present updates: if it changed while reading the spte, 456c50d8ae3SPaolo Bonzini * we might have hit the race. This is done using clear_spte_count. 457c50d8ae3SPaolo Bonzini */ 458c50d8ae3SPaolo Bonzini static u64 __get_spte_lockless(u64 *sptep) 459c50d8ae3SPaolo Bonzini { 46057354682SSean Christopherson struct kvm_mmu_page *sp = sptep_to_sp(sptep); 461c50d8ae3SPaolo Bonzini union split_spte spte, *orig = (union split_spte *)sptep; 462c50d8ae3SPaolo Bonzini int count; 463c50d8ae3SPaolo Bonzini 464c50d8ae3SPaolo Bonzini retry: 465c50d8ae3SPaolo Bonzini count = sp->clear_spte_count; 466c50d8ae3SPaolo Bonzini smp_rmb(); 467c50d8ae3SPaolo Bonzini 468c50d8ae3SPaolo Bonzini spte.spte_low = orig->spte_low; 469c50d8ae3SPaolo Bonzini smp_rmb(); 470c50d8ae3SPaolo Bonzini 471c50d8ae3SPaolo Bonzini spte.spte_high = orig->spte_high; 472c50d8ae3SPaolo Bonzini smp_rmb(); 473c50d8ae3SPaolo Bonzini 474c50d8ae3SPaolo Bonzini if (unlikely(spte.spte_low != orig->spte_low || 475c50d8ae3SPaolo Bonzini count != sp->clear_spte_count)) 476c50d8ae3SPaolo Bonzini goto retry; 477c50d8ae3SPaolo Bonzini 478c50d8ae3SPaolo Bonzini return spte.spte; 479c50d8ae3SPaolo Bonzini } 480c50d8ae3SPaolo Bonzini #endif 481c50d8ae3SPaolo Bonzini 482c50d8ae3SPaolo Bonzini /* Rules for using mmu_spte_set: 483c50d8ae3SPaolo Bonzini * Set the sptep from nonpresent to present. 484c50d8ae3SPaolo Bonzini * Note: the sptep being assigned *must* be either not present 485c50d8ae3SPaolo Bonzini * or in a state where the hardware will not attempt to update 486c50d8ae3SPaolo Bonzini * the spte. 487c50d8ae3SPaolo Bonzini */ 488c50d8ae3SPaolo Bonzini static void mmu_spte_set(u64 *sptep, u64 new_spte) 489c50d8ae3SPaolo Bonzini { 490c50d8ae3SPaolo Bonzini WARN_ON(is_shadow_present_pte(*sptep)); 491c50d8ae3SPaolo Bonzini __set_spte(sptep, new_spte); 492c50d8ae3SPaolo Bonzini } 493c50d8ae3SPaolo Bonzini 494c50d8ae3SPaolo Bonzini /* 495c50d8ae3SPaolo Bonzini * Update the SPTE (excluding the PFN), but do not track changes in its 496c50d8ae3SPaolo Bonzini * accessed/dirty status. 497c50d8ae3SPaolo Bonzini */ 498c50d8ae3SPaolo Bonzini static u64 mmu_spte_update_no_track(u64 *sptep, u64 new_spte) 499c50d8ae3SPaolo Bonzini { 500c50d8ae3SPaolo Bonzini u64 old_spte = *sptep; 501c50d8ae3SPaolo Bonzini 502c50d8ae3SPaolo Bonzini WARN_ON(!is_shadow_present_pte(new_spte)); 503115111efSDavid Matlack check_spte_writable_invariants(new_spte); 504c50d8ae3SPaolo Bonzini 505c50d8ae3SPaolo Bonzini if (!is_shadow_present_pte(old_spte)) { 506c50d8ae3SPaolo Bonzini mmu_spte_set(sptep, new_spte); 507c50d8ae3SPaolo Bonzini return old_spte; 508c50d8ae3SPaolo Bonzini } 509c50d8ae3SPaolo Bonzini 510c50d8ae3SPaolo Bonzini if (!spte_has_volatile_bits(old_spte)) 511c50d8ae3SPaolo Bonzini __update_clear_spte_fast(sptep, new_spte); 512c50d8ae3SPaolo Bonzini else 513c50d8ae3SPaolo Bonzini old_spte = __update_clear_spte_slow(sptep, new_spte); 514c50d8ae3SPaolo Bonzini 515c50d8ae3SPaolo Bonzini WARN_ON(spte_to_pfn(old_spte) != spte_to_pfn(new_spte)); 516c50d8ae3SPaolo Bonzini 517c50d8ae3SPaolo Bonzini return old_spte; 518c50d8ae3SPaolo Bonzini } 519c50d8ae3SPaolo Bonzini 520c50d8ae3SPaolo Bonzini /* Rules for using mmu_spte_update: 521c50d8ae3SPaolo Bonzini * Update the state bits, it means the mapped pfn is not changed. 522c50d8ae3SPaolo Bonzini * 52302844ac1SDavid Matlack * Whenever an MMU-writable SPTE is overwritten with a read-only SPTE, remote 52402844ac1SDavid Matlack * TLBs must be flushed. Otherwise rmap_write_protect will find a read-only 52502844ac1SDavid Matlack * spte, even though the writable spte might be cached on a CPU's TLB. 526c50d8ae3SPaolo Bonzini * 527c50d8ae3SPaolo Bonzini * Returns true if the TLB needs to be flushed 528c50d8ae3SPaolo Bonzini */ 529c50d8ae3SPaolo Bonzini static bool mmu_spte_update(u64 *sptep, u64 new_spte) 530c50d8ae3SPaolo Bonzini { 531c50d8ae3SPaolo Bonzini bool flush = false; 532c50d8ae3SPaolo Bonzini u64 old_spte = mmu_spte_update_no_track(sptep, new_spte); 533c50d8ae3SPaolo Bonzini 534c50d8ae3SPaolo Bonzini if (!is_shadow_present_pte(old_spte)) 535c50d8ae3SPaolo Bonzini return false; 536c50d8ae3SPaolo Bonzini 537c50d8ae3SPaolo Bonzini /* 538c50d8ae3SPaolo Bonzini * For the spte updated out of mmu-lock is safe, since 539c50d8ae3SPaolo Bonzini * we always atomically update it, see the comments in 540c50d8ae3SPaolo Bonzini * spte_has_volatile_bits(). 541c50d8ae3SPaolo Bonzini */ 542706c9c55SSean Christopherson if (is_mmu_writable_spte(old_spte) && 543c50d8ae3SPaolo Bonzini !is_writable_pte(new_spte)) 544c50d8ae3SPaolo Bonzini flush = true; 545c50d8ae3SPaolo Bonzini 546c50d8ae3SPaolo Bonzini /* 547c50d8ae3SPaolo Bonzini * Flush TLB when accessed/dirty states are changed in the page tables, 548c50d8ae3SPaolo Bonzini * to guarantee consistency between TLB and page tables. 549c50d8ae3SPaolo Bonzini */ 550c50d8ae3SPaolo Bonzini 551c50d8ae3SPaolo Bonzini if (is_accessed_spte(old_spte) && !is_accessed_spte(new_spte)) { 552c50d8ae3SPaolo Bonzini flush = true; 553c50d8ae3SPaolo Bonzini kvm_set_pfn_accessed(spte_to_pfn(old_spte)); 554c50d8ae3SPaolo Bonzini } 555c50d8ae3SPaolo Bonzini 556c50d8ae3SPaolo Bonzini if (is_dirty_spte(old_spte) && !is_dirty_spte(new_spte)) { 557c50d8ae3SPaolo Bonzini flush = true; 558c50d8ae3SPaolo Bonzini kvm_set_pfn_dirty(spte_to_pfn(old_spte)); 559c50d8ae3SPaolo Bonzini } 560c50d8ae3SPaolo Bonzini 561c50d8ae3SPaolo Bonzini return flush; 562c50d8ae3SPaolo Bonzini } 563c50d8ae3SPaolo Bonzini 564c50d8ae3SPaolo Bonzini /* 565c50d8ae3SPaolo Bonzini * Rules for using mmu_spte_clear_track_bits: 566c50d8ae3SPaolo Bonzini * It sets the sptep from present to nonpresent, and track the 567c50d8ae3SPaolo Bonzini * state bits, it is used to clear the last level sptep. 5687fa2a347SSean Christopherson * Returns the old PTE. 569c50d8ae3SPaolo Bonzini */ 57035d539c3SSean Christopherson static u64 mmu_spte_clear_track_bits(struct kvm *kvm, u64 *sptep) 571c50d8ae3SPaolo Bonzini { 572c50d8ae3SPaolo Bonzini kvm_pfn_t pfn; 573c50d8ae3SPaolo Bonzini u64 old_spte = *sptep; 57471f51d2cSMingwei Zhang int level = sptep_to_sp(sptep)->role.level; 575b14b2690SSean Christopherson struct page *page; 576c50d8ae3SPaolo Bonzini 57754eb3ef5SSean Christopherson if (!is_shadow_present_pte(old_spte) || 57854eb3ef5SSean Christopherson !spte_has_volatile_bits(old_spte)) 579c50d8ae3SPaolo Bonzini __update_clear_spte_fast(sptep, 0ull); 580c50d8ae3SPaolo Bonzini else 581c50d8ae3SPaolo Bonzini old_spte = __update_clear_spte_slow(sptep, 0ull); 582c50d8ae3SPaolo Bonzini 583c50d8ae3SPaolo Bonzini if (!is_shadow_present_pte(old_spte)) 5847fa2a347SSean Christopherson return old_spte; 585c50d8ae3SPaolo Bonzini 58671f51d2cSMingwei Zhang kvm_update_page_stats(kvm, level, -1); 58771f51d2cSMingwei Zhang 588c50d8ae3SPaolo Bonzini pfn = spte_to_pfn(old_spte); 589c50d8ae3SPaolo Bonzini 590c50d8ae3SPaolo Bonzini /* 591b14b2690SSean Christopherson * KVM doesn't hold a reference to any pages mapped into the guest, and 592b14b2690SSean Christopherson * instead uses the mmu_notifier to ensure that KVM unmaps any pages 593b14b2690SSean Christopherson * before they are reclaimed. Sanity check that, if the pfn is backed 594b14b2690SSean Christopherson * by a refcounted page, the refcount is elevated. 595c50d8ae3SPaolo Bonzini */ 596b14b2690SSean Christopherson page = kvm_pfn_to_refcounted_page(pfn); 597b14b2690SSean Christopherson WARN_ON(page && !page_count(page)); 598c50d8ae3SPaolo Bonzini 599c50d8ae3SPaolo Bonzini if (is_accessed_spte(old_spte)) 600c50d8ae3SPaolo Bonzini kvm_set_pfn_accessed(pfn); 601c50d8ae3SPaolo Bonzini 602c50d8ae3SPaolo Bonzini if (is_dirty_spte(old_spte)) 603c50d8ae3SPaolo Bonzini kvm_set_pfn_dirty(pfn); 604c50d8ae3SPaolo Bonzini 6057fa2a347SSean Christopherson return old_spte; 606c50d8ae3SPaolo Bonzini } 607c50d8ae3SPaolo Bonzini 608c50d8ae3SPaolo Bonzini /* 609c50d8ae3SPaolo Bonzini * Rules for using mmu_spte_clear_no_track: 610c50d8ae3SPaolo Bonzini * Directly clear spte without caring the state bits of sptep, 611c50d8ae3SPaolo Bonzini * it is used to set the upper level spte. 612c50d8ae3SPaolo Bonzini */ 613c50d8ae3SPaolo Bonzini static void mmu_spte_clear_no_track(u64 *sptep) 614c50d8ae3SPaolo Bonzini { 615c50d8ae3SPaolo Bonzini __update_clear_spte_fast(sptep, 0ull); 616c50d8ae3SPaolo Bonzini } 617c50d8ae3SPaolo Bonzini 618c50d8ae3SPaolo Bonzini static u64 mmu_spte_get_lockless(u64 *sptep) 619c50d8ae3SPaolo Bonzini { 620c50d8ae3SPaolo Bonzini return __get_spte_lockless(sptep); 621c50d8ae3SPaolo Bonzini } 622c50d8ae3SPaolo Bonzini 623c50d8ae3SPaolo Bonzini /* Returns the Accessed status of the PTE and resets it at the same time. */ 624c50d8ae3SPaolo Bonzini static bool mmu_spte_age(u64 *sptep) 625c50d8ae3SPaolo Bonzini { 626c50d8ae3SPaolo Bonzini u64 spte = mmu_spte_get_lockless(sptep); 627c50d8ae3SPaolo Bonzini 628c50d8ae3SPaolo Bonzini if (!is_accessed_spte(spte)) 629c50d8ae3SPaolo Bonzini return false; 630c50d8ae3SPaolo Bonzini 631c50d8ae3SPaolo Bonzini if (spte_ad_enabled(spte)) { 632c50d8ae3SPaolo Bonzini clear_bit((ffs(shadow_accessed_mask) - 1), 633c50d8ae3SPaolo Bonzini (unsigned long *)sptep); 634c50d8ae3SPaolo Bonzini } else { 635c50d8ae3SPaolo Bonzini /* 636c50d8ae3SPaolo Bonzini * Capture the dirty status of the page, so that it doesn't get 637c50d8ae3SPaolo Bonzini * lost when the SPTE is marked for access tracking. 638c50d8ae3SPaolo Bonzini */ 639c50d8ae3SPaolo Bonzini if (is_writable_pte(spte)) 640c50d8ae3SPaolo Bonzini kvm_set_pfn_dirty(spte_to_pfn(spte)); 641c50d8ae3SPaolo Bonzini 642c50d8ae3SPaolo Bonzini spte = mark_spte_for_access_track(spte); 643c50d8ae3SPaolo Bonzini mmu_spte_update_no_track(sptep, spte); 644c50d8ae3SPaolo Bonzini } 645c50d8ae3SPaolo Bonzini 646c50d8ae3SPaolo Bonzini return true; 647c50d8ae3SPaolo Bonzini } 648c50d8ae3SPaolo Bonzini 64978fdd2f0SSean Christopherson static inline bool is_tdp_mmu_active(struct kvm_vcpu *vcpu) 65078fdd2f0SSean Christopherson { 65178fdd2f0SSean Christopherson return tdp_mmu_enabled && vcpu->arch.mmu->root_role.direct; 65278fdd2f0SSean Christopherson } 65378fdd2f0SSean Christopherson 654c50d8ae3SPaolo Bonzini static void walk_shadow_page_lockless_begin(struct kvm_vcpu *vcpu) 655c50d8ae3SPaolo Bonzini { 65678fdd2f0SSean Christopherson if (is_tdp_mmu_active(vcpu)) { 657c5c8c7c5SDavid Matlack kvm_tdp_mmu_walk_lockless_begin(); 658c5c8c7c5SDavid Matlack } else { 659c50d8ae3SPaolo Bonzini /* 660c50d8ae3SPaolo Bonzini * Prevent page table teardown by making any free-er wait during 661c50d8ae3SPaolo Bonzini * kvm_flush_remote_tlbs() IPI to all active vcpus. 662c50d8ae3SPaolo Bonzini */ 663c50d8ae3SPaolo Bonzini local_irq_disable(); 664c50d8ae3SPaolo Bonzini 665c50d8ae3SPaolo Bonzini /* 666c50d8ae3SPaolo Bonzini * Make sure a following spte read is not reordered ahead of the write 667c50d8ae3SPaolo Bonzini * to vcpu->mode. 668c50d8ae3SPaolo Bonzini */ 669c50d8ae3SPaolo Bonzini smp_store_mb(vcpu->mode, READING_SHADOW_PAGE_TABLES); 670c50d8ae3SPaolo Bonzini } 671c5c8c7c5SDavid Matlack } 672c50d8ae3SPaolo Bonzini 673c50d8ae3SPaolo Bonzini static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu) 674c50d8ae3SPaolo Bonzini { 67578fdd2f0SSean Christopherson if (is_tdp_mmu_active(vcpu)) { 676c5c8c7c5SDavid Matlack kvm_tdp_mmu_walk_lockless_end(); 677c5c8c7c5SDavid Matlack } else { 678c50d8ae3SPaolo Bonzini /* 679c50d8ae3SPaolo Bonzini * Make sure the write to vcpu->mode is not reordered in front of 680c50d8ae3SPaolo Bonzini * reads to sptes. If it does, kvm_mmu_commit_zap_page() can see us 681c50d8ae3SPaolo Bonzini * OUTSIDE_GUEST_MODE and proceed to free the shadow page table. 682c50d8ae3SPaolo Bonzini */ 683c50d8ae3SPaolo Bonzini smp_store_release(&vcpu->mode, OUTSIDE_GUEST_MODE); 684c50d8ae3SPaolo Bonzini local_irq_enable(); 685c50d8ae3SPaolo Bonzini } 686c5c8c7c5SDavid Matlack } 687c50d8ae3SPaolo Bonzini 688378f5cd6SSean Christopherson static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu, bool maybe_indirect) 689c50d8ae3SPaolo Bonzini { 690c50d8ae3SPaolo Bonzini int r; 691c50d8ae3SPaolo Bonzini 692531281adSSean Christopherson /* 1 rmap, 1 parent PTE per level, and the prefetched rmaps. */ 69394ce87efSSean Christopherson r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache, 694531281adSSean Christopherson 1 + PT64_ROOT_MAX_LEVEL + PTE_PREFETCH_NUM); 695c50d8ae3SPaolo Bonzini if (r) 696c50d8ae3SPaolo Bonzini return r; 69794ce87efSSean Christopherson r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_shadow_page_cache, 698171a90d7SSean Christopherson PT64_ROOT_MAX_LEVEL); 699171a90d7SSean Christopherson if (r) 700171a90d7SSean Christopherson return r; 701378f5cd6SSean Christopherson if (maybe_indirect) { 7026a97575dSDavid Matlack r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_shadowed_info_cache, 703171a90d7SSean Christopherson PT64_ROOT_MAX_LEVEL); 704c50d8ae3SPaolo Bonzini if (r) 705c50d8ae3SPaolo Bonzini return r; 706378f5cd6SSean Christopherson } 70794ce87efSSean Christopherson return kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache, 708531281adSSean Christopherson PT64_ROOT_MAX_LEVEL); 709c50d8ae3SPaolo Bonzini } 710c50d8ae3SPaolo Bonzini 711c50d8ae3SPaolo Bonzini static void mmu_free_memory_caches(struct kvm_vcpu *vcpu) 712c50d8ae3SPaolo Bonzini { 71394ce87efSSean Christopherson kvm_mmu_free_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache); 71494ce87efSSean Christopherson kvm_mmu_free_memory_cache(&vcpu->arch.mmu_shadow_page_cache); 7156a97575dSDavid Matlack kvm_mmu_free_memory_cache(&vcpu->arch.mmu_shadowed_info_cache); 71694ce87efSSean Christopherson kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache); 717c50d8ae3SPaolo Bonzini } 718c50d8ae3SPaolo Bonzini 719c50d8ae3SPaolo Bonzini static void mmu_free_pte_list_desc(struct pte_list_desc *pte_list_desc) 720c50d8ae3SPaolo Bonzini { 721c50d8ae3SPaolo Bonzini kmem_cache_free(pte_list_desc_cache, pte_list_desc); 722c50d8ae3SPaolo Bonzini } 723c50d8ae3SPaolo Bonzini 7246a97575dSDavid Matlack static bool sp_has_gptes(struct kvm_mmu_page *sp); 7256a97575dSDavid Matlack 726c50d8ae3SPaolo Bonzini static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index) 727c50d8ae3SPaolo Bonzini { 72884e5ffd0SLai Jiangshan if (sp->role.passthrough) 72984e5ffd0SLai Jiangshan return sp->gfn; 73084e5ffd0SLai Jiangshan 731c50d8ae3SPaolo Bonzini if (!sp->role.direct) 7326a97575dSDavid Matlack return sp->shadowed_translation[index] >> PAGE_SHIFT; 733c50d8ae3SPaolo Bonzini 7342ca3129eSSean Christopherson return sp->gfn + (index << ((sp->role.level - 1) * SPTE_LEVEL_BITS)); 735c50d8ae3SPaolo Bonzini } 736c50d8ae3SPaolo Bonzini 7376a97575dSDavid Matlack /* 7386a97575dSDavid Matlack * For leaf SPTEs, fetch the *guest* access permissions being shadowed. Note 7396a97575dSDavid Matlack * that the SPTE itself may have a more constrained access permissions that 7406a97575dSDavid Matlack * what the guest enforces. For example, a guest may create an executable 7416a97575dSDavid Matlack * huge PTE but KVM may disallow execution to mitigate iTLB multihit. 7426a97575dSDavid Matlack */ 7436a97575dSDavid Matlack static u32 kvm_mmu_page_get_access(struct kvm_mmu_page *sp, int index) 744c50d8ae3SPaolo Bonzini { 7456a97575dSDavid Matlack if (sp_has_gptes(sp)) 7466a97575dSDavid Matlack return sp->shadowed_translation[index] & ACC_ALL; 7476a97575dSDavid Matlack 7486a97575dSDavid Matlack /* 7496a97575dSDavid Matlack * For direct MMUs (e.g. TDP or non-paging guests) or passthrough SPs, 7506a97575dSDavid Matlack * KVM is not shadowing any guest page tables, so the "guest access 7516a97575dSDavid Matlack * permissions" are just ACC_ALL. 7526a97575dSDavid Matlack * 7536a97575dSDavid Matlack * For direct SPs in indirect MMUs (shadow paging), i.e. when KVM 7546a97575dSDavid Matlack * is shadowing a guest huge page with small pages, the guest access 7556a97575dSDavid Matlack * permissions being shadowed are the access permissions of the huge 7566a97575dSDavid Matlack * page. 7576a97575dSDavid Matlack * 7586a97575dSDavid Matlack * In both cases, sp->role.access contains the correct access bits. 7596a97575dSDavid Matlack */ 7606a97575dSDavid Matlack return sp->role.access; 7616a97575dSDavid Matlack } 7626a97575dSDavid Matlack 76372ae5822SSean Christopherson static void kvm_mmu_page_set_translation(struct kvm_mmu_page *sp, int index, 76472ae5822SSean Christopherson gfn_t gfn, unsigned int access) 7656a97575dSDavid Matlack { 7666a97575dSDavid Matlack if (sp_has_gptes(sp)) { 7676a97575dSDavid Matlack sp->shadowed_translation[index] = (gfn << PAGE_SHIFT) | access; 76884e5ffd0SLai Jiangshan return; 76984e5ffd0SLai Jiangshan } 77084e5ffd0SLai Jiangshan 7716a97575dSDavid Matlack WARN_ONCE(access != kvm_mmu_page_get_access(sp, index), 7726a97575dSDavid Matlack "access mismatch under %s page %llx (expected %u, got %u)\n", 7736a97575dSDavid Matlack sp->role.passthrough ? "passthrough" : "direct", 7746a97575dSDavid Matlack sp->gfn, kvm_mmu_page_get_access(sp, index), access); 7756a97575dSDavid Matlack 7766a97575dSDavid Matlack WARN_ONCE(gfn != kvm_mmu_page_get_gfn(sp, index), 7776a97575dSDavid Matlack "gfn mismatch under %s page %llx (expected %llx, got %llx)\n", 7786a97575dSDavid Matlack sp->role.passthrough ? "passthrough" : "direct", 7796a97575dSDavid Matlack sp->gfn, kvm_mmu_page_get_gfn(sp, index), gfn); 780c50d8ae3SPaolo Bonzini } 781c50d8ae3SPaolo Bonzini 78272ae5822SSean Christopherson static void kvm_mmu_page_set_access(struct kvm_mmu_page *sp, int index, 78372ae5822SSean Christopherson unsigned int access) 7846a97575dSDavid Matlack { 7856a97575dSDavid Matlack gfn_t gfn = kvm_mmu_page_get_gfn(sp, index); 7866a97575dSDavid Matlack 7876a97575dSDavid Matlack kvm_mmu_page_set_translation(sp, index, gfn, access); 788c50d8ae3SPaolo Bonzini } 789c50d8ae3SPaolo Bonzini 790c50d8ae3SPaolo Bonzini /* 791c50d8ae3SPaolo Bonzini * Return the pointer to the large page information for a given gfn, 792c50d8ae3SPaolo Bonzini * handling slots that are not large page aligned. 793c50d8ae3SPaolo Bonzini */ 794c50d8ae3SPaolo Bonzini static struct kvm_lpage_info *lpage_info_slot(gfn_t gfn, 7958ca6f063SBen Gardon const struct kvm_memory_slot *slot, int level) 796c50d8ae3SPaolo Bonzini { 797c50d8ae3SPaolo Bonzini unsigned long idx; 798c50d8ae3SPaolo Bonzini 799c50d8ae3SPaolo Bonzini idx = gfn_to_index(gfn, slot->base_gfn, level); 800c50d8ae3SPaolo Bonzini return &slot->arch.lpage_info[level - 2][idx]; 801c50d8ae3SPaolo Bonzini } 802c50d8ae3SPaolo Bonzini 803269e9552SHamza Mahfooz static void update_gfn_disallow_lpage_count(const struct kvm_memory_slot *slot, 804c50d8ae3SPaolo Bonzini gfn_t gfn, int count) 805c50d8ae3SPaolo Bonzini { 806c50d8ae3SPaolo Bonzini struct kvm_lpage_info *linfo; 807c50d8ae3SPaolo Bonzini int i; 808c50d8ae3SPaolo Bonzini 8093bae0459SSean Christopherson for (i = PG_LEVEL_2M; i <= KVM_MAX_HUGEPAGE_LEVEL; ++i) { 810c50d8ae3SPaolo Bonzini linfo = lpage_info_slot(gfn, slot, i); 811c50d8ae3SPaolo Bonzini linfo->disallow_lpage += count; 812c50d8ae3SPaolo Bonzini WARN_ON(linfo->disallow_lpage < 0); 813c50d8ae3SPaolo Bonzini } 814c50d8ae3SPaolo Bonzini } 815c50d8ae3SPaolo Bonzini 816269e9552SHamza Mahfooz void kvm_mmu_gfn_disallow_lpage(const struct kvm_memory_slot *slot, gfn_t gfn) 817c50d8ae3SPaolo Bonzini { 818c50d8ae3SPaolo Bonzini update_gfn_disallow_lpage_count(slot, gfn, 1); 819c50d8ae3SPaolo Bonzini } 820c50d8ae3SPaolo Bonzini 821269e9552SHamza Mahfooz void kvm_mmu_gfn_allow_lpage(const struct kvm_memory_slot *slot, gfn_t gfn) 822c50d8ae3SPaolo Bonzini { 823c50d8ae3SPaolo Bonzini update_gfn_disallow_lpage_count(slot, gfn, -1); 824c50d8ae3SPaolo Bonzini } 825c50d8ae3SPaolo Bonzini 826c50d8ae3SPaolo Bonzini static void account_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp) 827c50d8ae3SPaolo Bonzini { 828c50d8ae3SPaolo Bonzini struct kvm_memslots *slots; 829c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot; 830c50d8ae3SPaolo Bonzini gfn_t gfn; 831c50d8ae3SPaolo Bonzini 832c50d8ae3SPaolo Bonzini kvm->arch.indirect_shadow_pages++; 833c50d8ae3SPaolo Bonzini gfn = sp->gfn; 834c50d8ae3SPaolo Bonzini slots = kvm_memslots_for_spte_role(kvm, sp->role); 835c50d8ae3SPaolo Bonzini slot = __gfn_to_memslot(slots, gfn); 836c50d8ae3SPaolo Bonzini 837c50d8ae3SPaolo Bonzini /* the non-leaf shadow pages are keeping readonly. */ 8383bae0459SSean Christopherson if (sp->role.level > PG_LEVEL_4K) 839c50d8ae3SPaolo Bonzini return kvm_slot_page_track_add_page(kvm, slot, gfn, 840c50d8ae3SPaolo Bonzini KVM_PAGE_TRACK_WRITE); 841c50d8ae3SPaolo Bonzini 842c50d8ae3SPaolo Bonzini kvm_mmu_gfn_disallow_lpage(slot, gfn); 843be911771SDavid Matlack 844be911771SDavid Matlack if (kvm_mmu_slot_gfn_write_protect(kvm, slot, gfn, PG_LEVEL_4K)) 8454ad980aeSHou Wenlong kvm_flush_remote_tlbs_gfn(kvm, gfn, PG_LEVEL_4K); 846c50d8ae3SPaolo Bonzini } 847c50d8ae3SPaolo Bonzini 84861f94478SSean Christopherson void track_possible_nx_huge_page(struct kvm *kvm, struct kvm_mmu_page *sp) 849c50d8ae3SPaolo Bonzini { 850428e9216SSean Christopherson /* 851428e9216SSean Christopherson * If it's possible to replace the shadow page with an NX huge page, 852428e9216SSean Christopherson * i.e. if the shadow page is the only thing currently preventing KVM 853428e9216SSean Christopherson * from using a huge page, add the shadow page to the list of "to be 854428e9216SSean Christopherson * zapped for NX recovery" pages. Note, the shadow page can already be 855428e9216SSean Christopherson * on the list if KVM is reusing an existing shadow page, i.e. if KVM 856428e9216SSean Christopherson * links a shadow page at multiple points. 857428e9216SSean Christopherson */ 85861f94478SSean Christopherson if (!list_empty(&sp->possible_nx_huge_page_link)) 859c50d8ae3SPaolo Bonzini return; 860c50d8ae3SPaolo Bonzini 861c50d8ae3SPaolo Bonzini ++kvm->stat.nx_lpage_splits; 86255c510e2SSean Christopherson list_add_tail(&sp->possible_nx_huge_page_link, 86355c510e2SSean Christopherson &kvm->arch.possible_nx_huge_pages); 864c50d8ae3SPaolo Bonzini } 865c50d8ae3SPaolo Bonzini 86661f94478SSean Christopherson static void account_nx_huge_page(struct kvm *kvm, struct kvm_mmu_page *sp, 86761f94478SSean Christopherson bool nx_huge_page_possible) 86861f94478SSean Christopherson { 86961f94478SSean Christopherson sp->nx_huge_page_disallowed = true; 87061f94478SSean Christopherson 87161f94478SSean Christopherson if (nx_huge_page_possible) 87261f94478SSean Christopherson track_possible_nx_huge_page(kvm, sp); 873c50d8ae3SPaolo Bonzini } 874c50d8ae3SPaolo Bonzini 875c50d8ae3SPaolo Bonzini static void unaccount_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp) 876c50d8ae3SPaolo Bonzini { 877c50d8ae3SPaolo Bonzini struct kvm_memslots *slots; 878c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot; 879c50d8ae3SPaolo Bonzini gfn_t gfn; 880c50d8ae3SPaolo Bonzini 881c50d8ae3SPaolo Bonzini kvm->arch.indirect_shadow_pages--; 882c50d8ae3SPaolo Bonzini gfn = sp->gfn; 883c50d8ae3SPaolo Bonzini slots = kvm_memslots_for_spte_role(kvm, sp->role); 884c50d8ae3SPaolo Bonzini slot = __gfn_to_memslot(slots, gfn); 8853bae0459SSean Christopherson if (sp->role.level > PG_LEVEL_4K) 886c50d8ae3SPaolo Bonzini return kvm_slot_page_track_remove_page(kvm, slot, gfn, 887c50d8ae3SPaolo Bonzini KVM_PAGE_TRACK_WRITE); 888c50d8ae3SPaolo Bonzini 889c50d8ae3SPaolo Bonzini kvm_mmu_gfn_allow_lpage(slot, gfn); 890c50d8ae3SPaolo Bonzini } 891c50d8ae3SPaolo Bonzini 89261f94478SSean Christopherson void untrack_possible_nx_huge_page(struct kvm *kvm, struct kvm_mmu_page *sp) 893c50d8ae3SPaolo Bonzini { 89455c510e2SSean Christopherson if (list_empty(&sp->possible_nx_huge_page_link)) 895428e9216SSean Christopherson return; 896428e9216SSean Christopherson 897c50d8ae3SPaolo Bonzini --kvm->stat.nx_lpage_splits; 89855c510e2SSean Christopherson list_del_init(&sp->possible_nx_huge_page_link); 899c50d8ae3SPaolo Bonzini } 900c50d8ae3SPaolo Bonzini 90161f94478SSean Christopherson static void unaccount_nx_huge_page(struct kvm *kvm, struct kvm_mmu_page *sp) 90261f94478SSean Christopherson { 90361f94478SSean Christopherson sp->nx_huge_page_disallowed = false; 90461f94478SSean Christopherson 90561f94478SSean Christopherson untrack_possible_nx_huge_page(kvm, sp); 906c50d8ae3SPaolo Bonzini } 907c50d8ae3SPaolo Bonzini 908f3d90f90SSean Christopherson static struct kvm_memory_slot *gfn_to_memslot_dirty_bitmap(struct kvm_vcpu *vcpu, 909f3d90f90SSean Christopherson gfn_t gfn, 910c50d8ae3SPaolo Bonzini bool no_dirty_log) 911c50d8ae3SPaolo Bonzini { 912c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot; 913c50d8ae3SPaolo Bonzini 914c50d8ae3SPaolo Bonzini slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn); 91591b0d268SPaolo Bonzini if (!slot || slot->flags & KVM_MEMSLOT_INVALID) 91691b0d268SPaolo Bonzini return NULL; 917044c59c4SPeter Xu if (no_dirty_log && kvm_slot_dirty_track_enabled(slot)) 91891b0d268SPaolo Bonzini return NULL; 919c50d8ae3SPaolo Bonzini 920c50d8ae3SPaolo Bonzini return slot; 921c50d8ae3SPaolo Bonzini } 922c50d8ae3SPaolo Bonzini 923c50d8ae3SPaolo Bonzini /* 924c50d8ae3SPaolo Bonzini * About rmap_head encoding: 925c50d8ae3SPaolo Bonzini * 926c50d8ae3SPaolo Bonzini * If the bit zero of rmap_head->val is clear, then it points to the only spte 927c50d8ae3SPaolo Bonzini * in this rmap chain. Otherwise, (rmap_head->val & ~1) points to a struct 928c50d8ae3SPaolo Bonzini * pte_list_desc containing more mappings. 929c50d8ae3SPaolo Bonzini */ 930c50d8ae3SPaolo Bonzini 931c50d8ae3SPaolo Bonzini /* 932c50d8ae3SPaolo Bonzini * Returns the number of pointers in the rmap chain, not counting the new one. 933c50d8ae3SPaolo Bonzini */ 9342ff9039aSDavid Matlack static int pte_list_add(struct kvm_mmu_memory_cache *cache, u64 *spte, 935c50d8ae3SPaolo Bonzini struct kvm_rmap_head *rmap_head) 936c50d8ae3SPaolo Bonzini { 937c50d8ae3SPaolo Bonzini struct pte_list_desc *desc; 93813236e25SPeter Xu int count = 0; 939c50d8ae3SPaolo Bonzini 940c50d8ae3SPaolo Bonzini if (!rmap_head->val) { 941805a0f83SStephen Zhang rmap_printk("%p %llx 0->1\n", spte, *spte); 942c50d8ae3SPaolo Bonzini rmap_head->val = (unsigned long)spte; 943c50d8ae3SPaolo Bonzini } else if (!(rmap_head->val & 1)) { 944805a0f83SStephen Zhang rmap_printk("%p %llx 1->many\n", spte, *spte); 9452ff9039aSDavid Matlack desc = kvm_mmu_memory_cache_alloc(cache); 946c50d8ae3SPaolo Bonzini desc->sptes[0] = (u64 *)rmap_head->val; 947c50d8ae3SPaolo Bonzini desc->sptes[1] = spte; 94813236e25SPeter Xu desc->spte_count = 2; 949141705b7SLai Jiangshan desc->tail_count = 0; 950c50d8ae3SPaolo Bonzini rmap_head->val = (unsigned long)desc | 1; 951c50d8ae3SPaolo Bonzini ++count; 952c50d8ae3SPaolo Bonzini } else { 953805a0f83SStephen Zhang rmap_printk("%p %llx many->many\n", spte, *spte); 954c50d8ae3SPaolo Bonzini desc = (struct pte_list_desc *)(rmap_head->val & ~1ul); 955141705b7SLai Jiangshan count = desc->tail_count + desc->spte_count; 956141705b7SLai Jiangshan 957141705b7SLai Jiangshan /* 958141705b7SLai Jiangshan * If the previous head is full, allocate a new head descriptor 959141705b7SLai Jiangshan * as tail descriptors are always kept full. 960141705b7SLai Jiangshan */ 961141705b7SLai Jiangshan if (desc->spte_count == PTE_LIST_EXT) { 962141705b7SLai Jiangshan desc = kvm_mmu_memory_cache_alloc(cache); 963141705b7SLai Jiangshan desc->more = (struct pte_list_desc *)(rmap_head->val & ~1ul); 96413236e25SPeter Xu desc->spte_count = 0; 965141705b7SLai Jiangshan desc->tail_count = count; 966141705b7SLai Jiangshan rmap_head->val = (unsigned long)desc | 1; 967c6c4f961SLi RongQing } 96813236e25SPeter Xu desc->sptes[desc->spte_count++] = spte; 969c50d8ae3SPaolo Bonzini } 970c50d8ae3SPaolo Bonzini return count; 971c50d8ae3SPaolo Bonzini } 972c50d8ae3SPaolo Bonzini 973f3d90f90SSean Christopherson static void pte_list_desc_remove_entry(struct kvm_rmap_head *rmap_head, 974141705b7SLai Jiangshan struct pte_list_desc *desc, int i) 975c50d8ae3SPaolo Bonzini { 976141705b7SLai Jiangshan struct pte_list_desc *head_desc = (struct pte_list_desc *)(rmap_head->val & ~1ul); 977141705b7SLai Jiangshan int j = head_desc->spte_count - 1; 978c50d8ae3SPaolo Bonzini 979141705b7SLai Jiangshan /* 980141705b7SLai Jiangshan * The head descriptor should never be empty. A new head is added only 981141705b7SLai Jiangshan * when adding an entry and the previous head is full, and heads are 982141705b7SLai Jiangshan * removed (this flow) when they become empty. 983141705b7SLai Jiangshan */ 984141705b7SLai Jiangshan BUG_ON(j < 0); 985141705b7SLai Jiangshan 986141705b7SLai Jiangshan /* 987141705b7SLai Jiangshan * Replace the to-be-freed SPTE with the last valid entry from the head 988141705b7SLai Jiangshan * descriptor to ensure that tail descriptors are full at all times. 989141705b7SLai Jiangshan * Note, this also means that tail_count is stable for each descriptor. 990141705b7SLai Jiangshan */ 991141705b7SLai Jiangshan desc->sptes[i] = head_desc->sptes[j]; 992141705b7SLai Jiangshan head_desc->sptes[j] = NULL; 993141705b7SLai Jiangshan head_desc->spte_count--; 994141705b7SLai Jiangshan if (head_desc->spte_count) 995c50d8ae3SPaolo Bonzini return; 996141705b7SLai Jiangshan 997141705b7SLai Jiangshan /* 998141705b7SLai Jiangshan * The head descriptor is empty. If there are no tail descriptors, 999141705b7SLai Jiangshan * nullify the rmap head to mark the list as emtpy, else point the rmap 1000141705b7SLai Jiangshan * head at the next descriptor, i.e. the new head. 1001141705b7SLai Jiangshan */ 1002141705b7SLai Jiangshan if (!head_desc->more) 1003fe3c2b4cSMiaohe Lin rmap_head->val = 0; 1004c50d8ae3SPaolo Bonzini else 1005141705b7SLai Jiangshan rmap_head->val = (unsigned long)head_desc->more | 1; 1006141705b7SLai Jiangshan mmu_free_pte_list_desc(head_desc); 1007c50d8ae3SPaolo Bonzini } 1008c50d8ae3SPaolo Bonzini 10093c2e1037SSean Christopherson static void pte_list_remove(u64 *spte, struct kvm_rmap_head *rmap_head) 1010c50d8ae3SPaolo Bonzini { 1011c50d8ae3SPaolo Bonzini struct pte_list_desc *desc; 1012c50d8ae3SPaolo Bonzini int i; 1013c50d8ae3SPaolo Bonzini 1014c50d8ae3SPaolo Bonzini if (!rmap_head->val) { 1015c50d8ae3SPaolo Bonzini pr_err("%s: %p 0->BUG\n", __func__, spte); 1016c50d8ae3SPaolo Bonzini BUG(); 1017c50d8ae3SPaolo Bonzini } else if (!(rmap_head->val & 1)) { 1018805a0f83SStephen Zhang rmap_printk("%p 1->0\n", spte); 1019c50d8ae3SPaolo Bonzini if ((u64 *)rmap_head->val != spte) { 1020c50d8ae3SPaolo Bonzini pr_err("%s: %p 1->BUG\n", __func__, spte); 1021c50d8ae3SPaolo Bonzini BUG(); 1022c50d8ae3SPaolo Bonzini } 1023c50d8ae3SPaolo Bonzini rmap_head->val = 0; 1024c50d8ae3SPaolo Bonzini } else { 1025805a0f83SStephen Zhang rmap_printk("%p many->many\n", spte); 1026c50d8ae3SPaolo Bonzini desc = (struct pte_list_desc *)(rmap_head->val & ~1ul); 1027c50d8ae3SPaolo Bonzini while (desc) { 102813236e25SPeter Xu for (i = 0; i < desc->spte_count; ++i) { 1029c50d8ae3SPaolo Bonzini if (desc->sptes[i] == spte) { 1030141705b7SLai Jiangshan pte_list_desc_remove_entry(rmap_head, desc, i); 1031c50d8ae3SPaolo Bonzini return; 1032c50d8ae3SPaolo Bonzini } 1033c50d8ae3SPaolo Bonzini } 1034c50d8ae3SPaolo Bonzini desc = desc->more; 1035c50d8ae3SPaolo Bonzini } 1036c50d8ae3SPaolo Bonzini pr_err("%s: %p many->many\n", __func__, spte); 1037c50d8ae3SPaolo Bonzini BUG(); 1038c50d8ae3SPaolo Bonzini } 1039c50d8ae3SPaolo Bonzini } 1040c50d8ae3SPaolo Bonzini 10419202aee8SSean Christopherson static void kvm_zap_one_rmap_spte(struct kvm *kvm, 10429202aee8SSean Christopherson struct kvm_rmap_head *rmap_head, u64 *sptep) 1043c50d8ae3SPaolo Bonzini { 104471f51d2cSMingwei Zhang mmu_spte_clear_track_bits(kvm, sptep); 10453c2e1037SSean Christopherson pte_list_remove(sptep, rmap_head); 1046c50d8ae3SPaolo Bonzini } 1047c50d8ae3SPaolo Bonzini 10489202aee8SSean Christopherson /* Return true if at least one SPTE was zapped, false otherwise */ 10499202aee8SSean Christopherson static bool kvm_zap_all_rmap_sptes(struct kvm *kvm, 10509202aee8SSean Christopherson struct kvm_rmap_head *rmap_head) 1051a75b5404SPeter Xu { 1052a75b5404SPeter Xu struct pte_list_desc *desc, *next; 1053a75b5404SPeter Xu int i; 1054a75b5404SPeter Xu 1055a75b5404SPeter Xu if (!rmap_head->val) 1056a75b5404SPeter Xu return false; 1057a75b5404SPeter Xu 1058a75b5404SPeter Xu if (!(rmap_head->val & 1)) { 105971f51d2cSMingwei Zhang mmu_spte_clear_track_bits(kvm, (u64 *)rmap_head->val); 1060a75b5404SPeter Xu goto out; 1061a75b5404SPeter Xu } 1062a75b5404SPeter Xu 1063a75b5404SPeter Xu desc = (struct pte_list_desc *)(rmap_head->val & ~1ul); 1064a75b5404SPeter Xu 1065a75b5404SPeter Xu for (; desc; desc = next) { 1066a75b5404SPeter Xu for (i = 0; i < desc->spte_count; i++) 106771f51d2cSMingwei Zhang mmu_spte_clear_track_bits(kvm, desc->sptes[i]); 1068a75b5404SPeter Xu next = desc->more; 1069a75b5404SPeter Xu mmu_free_pte_list_desc(desc); 1070a75b5404SPeter Xu } 1071a75b5404SPeter Xu out: 1072a75b5404SPeter Xu /* rmap_head is meaningless now, remember to reset it */ 1073a75b5404SPeter Xu rmap_head->val = 0; 1074a75b5404SPeter Xu return true; 1075a75b5404SPeter Xu } 1076a75b5404SPeter Xu 10773bcd0662SPeter Xu unsigned int pte_list_count(struct kvm_rmap_head *rmap_head) 10783bcd0662SPeter Xu { 10793bcd0662SPeter Xu struct pte_list_desc *desc; 10803bcd0662SPeter Xu 10813bcd0662SPeter Xu if (!rmap_head->val) 10823bcd0662SPeter Xu return 0; 10833bcd0662SPeter Xu else if (!(rmap_head->val & 1)) 10843bcd0662SPeter Xu return 1; 10853bcd0662SPeter Xu 10863bcd0662SPeter Xu desc = (struct pte_list_desc *)(rmap_head->val & ~1ul); 1087141705b7SLai Jiangshan return desc->tail_count + desc->spte_count; 10883bcd0662SPeter Xu } 10893bcd0662SPeter Xu 109093e083d4SDavid Matlack static struct kvm_rmap_head *gfn_to_rmap(gfn_t gfn, int level, 1091269e9552SHamza Mahfooz const struct kvm_memory_slot *slot) 1092c50d8ae3SPaolo Bonzini { 1093c50d8ae3SPaolo Bonzini unsigned long idx; 1094c50d8ae3SPaolo Bonzini 1095c50d8ae3SPaolo Bonzini idx = gfn_to_index(gfn, slot->base_gfn, level); 10963bae0459SSean Christopherson return &slot->arch.rmap[level - PG_LEVEL_4K][idx]; 1097c50d8ae3SPaolo Bonzini } 1098c50d8ae3SPaolo Bonzini 1099c50d8ae3SPaolo Bonzini static void rmap_remove(struct kvm *kvm, u64 *spte) 1100c50d8ae3SPaolo Bonzini { 1101601f8af0SDavid Matlack struct kvm_memslots *slots; 1102601f8af0SDavid Matlack struct kvm_memory_slot *slot; 1103c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 1104c50d8ae3SPaolo Bonzini gfn_t gfn; 1105c50d8ae3SPaolo Bonzini struct kvm_rmap_head *rmap_head; 1106c50d8ae3SPaolo Bonzini 110757354682SSean Christopherson sp = sptep_to_sp(spte); 110879e48cecSSean Christopherson gfn = kvm_mmu_page_get_gfn(sp, spte_index(spte)); 1109601f8af0SDavid Matlack 1110601f8af0SDavid Matlack /* 111168be1306SDavid Matlack * Unlike rmap_add, rmap_remove does not run in the context of a vCPU 111268be1306SDavid Matlack * so we have to determine which memslots to use based on context 111368be1306SDavid Matlack * information in sp->role. 1114601f8af0SDavid Matlack */ 1115601f8af0SDavid Matlack slots = kvm_memslots_for_spte_role(kvm, sp->role); 1116601f8af0SDavid Matlack 1117601f8af0SDavid Matlack slot = __gfn_to_memslot(slots, gfn); 111893e083d4SDavid Matlack rmap_head = gfn_to_rmap(gfn, sp->role.level, slot); 1119601f8af0SDavid Matlack 11203c2e1037SSean Christopherson pte_list_remove(spte, rmap_head); 1121c50d8ae3SPaolo Bonzini } 1122c50d8ae3SPaolo Bonzini 1123c50d8ae3SPaolo Bonzini /* 1124c50d8ae3SPaolo Bonzini * Used by the following functions to iterate through the sptes linked by a 1125c50d8ae3SPaolo Bonzini * rmap. All fields are private and not assumed to be used outside. 1126c50d8ae3SPaolo Bonzini */ 1127c50d8ae3SPaolo Bonzini struct rmap_iterator { 1128c50d8ae3SPaolo Bonzini /* private fields */ 1129c50d8ae3SPaolo Bonzini struct pte_list_desc *desc; /* holds the sptep if not NULL */ 1130c50d8ae3SPaolo Bonzini int pos; /* index of the sptep */ 1131c50d8ae3SPaolo Bonzini }; 1132c50d8ae3SPaolo Bonzini 1133c50d8ae3SPaolo Bonzini /* 1134c50d8ae3SPaolo Bonzini * Iteration must be started by this function. This should also be used after 1135c50d8ae3SPaolo Bonzini * removing/dropping sptes from the rmap link because in such cases the 11360a03cbdaSMiaohe Lin * information in the iterator may not be valid. 1137c50d8ae3SPaolo Bonzini * 1138c50d8ae3SPaolo Bonzini * Returns sptep if found, NULL otherwise. 1139c50d8ae3SPaolo Bonzini */ 1140c50d8ae3SPaolo Bonzini static u64 *rmap_get_first(struct kvm_rmap_head *rmap_head, 1141c50d8ae3SPaolo Bonzini struct rmap_iterator *iter) 1142c50d8ae3SPaolo Bonzini { 1143c50d8ae3SPaolo Bonzini u64 *sptep; 1144c50d8ae3SPaolo Bonzini 1145c50d8ae3SPaolo Bonzini if (!rmap_head->val) 1146c50d8ae3SPaolo Bonzini return NULL; 1147c50d8ae3SPaolo Bonzini 1148c50d8ae3SPaolo Bonzini if (!(rmap_head->val & 1)) { 1149c50d8ae3SPaolo Bonzini iter->desc = NULL; 1150c50d8ae3SPaolo Bonzini sptep = (u64 *)rmap_head->val; 1151c50d8ae3SPaolo Bonzini goto out; 1152c50d8ae3SPaolo Bonzini } 1153c50d8ae3SPaolo Bonzini 1154c50d8ae3SPaolo Bonzini iter->desc = (struct pte_list_desc *)(rmap_head->val & ~1ul); 1155c50d8ae3SPaolo Bonzini iter->pos = 0; 1156c50d8ae3SPaolo Bonzini sptep = iter->desc->sptes[iter->pos]; 1157c50d8ae3SPaolo Bonzini out: 1158c50d8ae3SPaolo Bonzini BUG_ON(!is_shadow_present_pte(*sptep)); 1159c50d8ae3SPaolo Bonzini return sptep; 1160c50d8ae3SPaolo Bonzini } 1161c50d8ae3SPaolo Bonzini 1162c50d8ae3SPaolo Bonzini /* 1163c50d8ae3SPaolo Bonzini * Must be used with a valid iterator: e.g. after rmap_get_first(). 1164c50d8ae3SPaolo Bonzini * 1165c50d8ae3SPaolo Bonzini * Returns sptep if found, NULL otherwise. 1166c50d8ae3SPaolo Bonzini */ 1167c50d8ae3SPaolo Bonzini static u64 *rmap_get_next(struct rmap_iterator *iter) 1168c50d8ae3SPaolo Bonzini { 1169c50d8ae3SPaolo Bonzini u64 *sptep; 1170c50d8ae3SPaolo Bonzini 1171c50d8ae3SPaolo Bonzini if (iter->desc) { 1172c50d8ae3SPaolo Bonzini if (iter->pos < PTE_LIST_EXT - 1) { 1173c50d8ae3SPaolo Bonzini ++iter->pos; 1174c50d8ae3SPaolo Bonzini sptep = iter->desc->sptes[iter->pos]; 1175c50d8ae3SPaolo Bonzini if (sptep) 1176c50d8ae3SPaolo Bonzini goto out; 1177c50d8ae3SPaolo Bonzini } 1178c50d8ae3SPaolo Bonzini 1179c50d8ae3SPaolo Bonzini iter->desc = iter->desc->more; 1180c50d8ae3SPaolo Bonzini 1181c50d8ae3SPaolo Bonzini if (iter->desc) { 1182c50d8ae3SPaolo Bonzini iter->pos = 0; 1183c50d8ae3SPaolo Bonzini /* desc->sptes[0] cannot be NULL */ 1184c50d8ae3SPaolo Bonzini sptep = iter->desc->sptes[iter->pos]; 1185c50d8ae3SPaolo Bonzini goto out; 1186c50d8ae3SPaolo Bonzini } 1187c50d8ae3SPaolo Bonzini } 1188c50d8ae3SPaolo Bonzini 1189c50d8ae3SPaolo Bonzini return NULL; 1190c50d8ae3SPaolo Bonzini out: 1191c50d8ae3SPaolo Bonzini BUG_ON(!is_shadow_present_pte(*sptep)); 1192c50d8ae3SPaolo Bonzini return sptep; 1193c50d8ae3SPaolo Bonzini } 1194c50d8ae3SPaolo Bonzini 1195c50d8ae3SPaolo Bonzini #define for_each_rmap_spte(_rmap_head_, _iter_, _spte_) \ 1196c50d8ae3SPaolo Bonzini for (_spte_ = rmap_get_first(_rmap_head_, _iter_); \ 1197c50d8ae3SPaolo Bonzini _spte_; _spte_ = rmap_get_next(_iter_)) 1198c50d8ae3SPaolo Bonzini 1199c50d8ae3SPaolo Bonzini static void drop_spte(struct kvm *kvm, u64 *sptep) 1200c50d8ae3SPaolo Bonzini { 120171f51d2cSMingwei Zhang u64 old_spte = mmu_spte_clear_track_bits(kvm, sptep); 12027fa2a347SSean Christopherson 12037fa2a347SSean Christopherson if (is_shadow_present_pte(old_spte)) 1204c50d8ae3SPaolo Bonzini rmap_remove(kvm, sptep); 1205c50d8ae3SPaolo Bonzini } 1206c50d8ae3SPaolo Bonzini 120703787394SPaolo Bonzini static void drop_large_spte(struct kvm *kvm, u64 *sptep, bool flush) 1208c50d8ae3SPaolo Bonzini { 12090cd8dc73SPaolo Bonzini struct kvm_mmu_page *sp; 12100cd8dc73SPaolo Bonzini 12110cd8dc73SPaolo Bonzini sp = sptep_to_sp(sptep); 12120cd8dc73SPaolo Bonzini WARN_ON(sp->role.level == PG_LEVEL_4K); 12130cd8dc73SPaolo Bonzini 1214c50d8ae3SPaolo Bonzini drop_spte(kvm, sptep); 121503787394SPaolo Bonzini 121603787394SPaolo Bonzini if (flush) 12171b2dc736SHou Wenlong kvm_flush_remote_tlbs_sptep(kvm, sptep); 1218c50d8ae3SPaolo Bonzini } 1219c50d8ae3SPaolo Bonzini 1220c50d8ae3SPaolo Bonzini /* 1221c50d8ae3SPaolo Bonzini * Write-protect on the specified @sptep, @pt_protect indicates whether 1222c50d8ae3SPaolo Bonzini * spte write-protection is caused by protecting shadow page table. 1223c50d8ae3SPaolo Bonzini * 1224c50d8ae3SPaolo Bonzini * Note: write protection is difference between dirty logging and spte 1225c50d8ae3SPaolo Bonzini * protection: 1226c50d8ae3SPaolo Bonzini * - for dirty logging, the spte can be set to writable at anytime if 1227c50d8ae3SPaolo Bonzini * its dirty bitmap is properly set. 1228c50d8ae3SPaolo Bonzini * - for spte protection, the spte can be writable only after unsync-ing 1229c50d8ae3SPaolo Bonzini * shadow page. 1230c50d8ae3SPaolo Bonzini * 1231c50d8ae3SPaolo Bonzini * Return true if tlb need be flushed. 1232c50d8ae3SPaolo Bonzini */ 1233c50d8ae3SPaolo Bonzini static bool spte_write_protect(u64 *sptep, bool pt_protect) 1234c50d8ae3SPaolo Bonzini { 1235c50d8ae3SPaolo Bonzini u64 spte = *sptep; 1236c50d8ae3SPaolo Bonzini 1237c50d8ae3SPaolo Bonzini if (!is_writable_pte(spte) && 1238706c9c55SSean Christopherson !(pt_protect && is_mmu_writable_spte(spte))) 1239c50d8ae3SPaolo Bonzini return false; 1240c50d8ae3SPaolo Bonzini 1241805a0f83SStephen Zhang rmap_printk("spte %p %llx\n", sptep, *sptep); 1242c50d8ae3SPaolo Bonzini 1243c50d8ae3SPaolo Bonzini if (pt_protect) 12445fc3424fSSean Christopherson spte &= ~shadow_mmu_writable_mask; 1245c50d8ae3SPaolo Bonzini spte = spte & ~PT_WRITABLE_MASK; 1246c50d8ae3SPaolo Bonzini 1247c50d8ae3SPaolo Bonzini return mmu_spte_update(sptep, spte); 1248c50d8ae3SPaolo Bonzini } 1249c50d8ae3SPaolo Bonzini 12501346bbb6SDavid Matlack static bool rmap_write_protect(struct kvm_rmap_head *rmap_head, 1251c50d8ae3SPaolo Bonzini bool pt_protect) 1252c50d8ae3SPaolo Bonzini { 1253c50d8ae3SPaolo Bonzini u64 *sptep; 1254c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 1255c50d8ae3SPaolo Bonzini bool flush = false; 1256c50d8ae3SPaolo Bonzini 1257c50d8ae3SPaolo Bonzini for_each_rmap_spte(rmap_head, &iter, sptep) 1258c50d8ae3SPaolo Bonzini flush |= spte_write_protect(sptep, pt_protect); 1259c50d8ae3SPaolo Bonzini 1260c50d8ae3SPaolo Bonzini return flush; 1261c50d8ae3SPaolo Bonzini } 1262c50d8ae3SPaolo Bonzini 1263c50d8ae3SPaolo Bonzini static bool spte_clear_dirty(u64 *sptep) 1264c50d8ae3SPaolo Bonzini { 1265c50d8ae3SPaolo Bonzini u64 spte = *sptep; 1266c50d8ae3SPaolo Bonzini 1267805a0f83SStephen Zhang rmap_printk("spte %p %llx\n", sptep, *sptep); 1268c50d8ae3SPaolo Bonzini 1269c50d8ae3SPaolo Bonzini MMU_WARN_ON(!spte_ad_enabled(spte)); 1270c50d8ae3SPaolo Bonzini spte &= ~shadow_dirty_mask; 1271c50d8ae3SPaolo Bonzini return mmu_spte_update(sptep, spte); 1272c50d8ae3SPaolo Bonzini } 1273c50d8ae3SPaolo Bonzini 1274c50d8ae3SPaolo Bonzini static bool spte_wrprot_for_clear_dirty(u64 *sptep) 1275c50d8ae3SPaolo Bonzini { 1276c50d8ae3SPaolo Bonzini bool was_writable = test_and_clear_bit(PT_WRITABLE_SHIFT, 1277c50d8ae3SPaolo Bonzini (unsigned long *)sptep); 1278c50d8ae3SPaolo Bonzini if (was_writable && !spte_ad_enabled(*sptep)) 1279c50d8ae3SPaolo Bonzini kvm_set_pfn_dirty(spte_to_pfn(*sptep)); 1280c50d8ae3SPaolo Bonzini 1281c50d8ae3SPaolo Bonzini return was_writable; 1282c50d8ae3SPaolo Bonzini } 1283c50d8ae3SPaolo Bonzini 1284c50d8ae3SPaolo Bonzini /* 1285c50d8ae3SPaolo Bonzini * Gets the GFN ready for another round of dirty logging by clearing the 1286c50d8ae3SPaolo Bonzini * - D bit on ad-enabled SPTEs, and 1287c50d8ae3SPaolo Bonzini * - W bit on ad-disabled SPTEs. 1288c50d8ae3SPaolo Bonzini * Returns true iff any D or W bits were cleared. 1289c50d8ae3SPaolo Bonzini */ 12900a234f5dSSean Christopherson static bool __rmap_clear_dirty(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1291269e9552SHamza Mahfooz const struct kvm_memory_slot *slot) 1292c50d8ae3SPaolo Bonzini { 1293c50d8ae3SPaolo Bonzini u64 *sptep; 1294c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 1295c50d8ae3SPaolo Bonzini bool flush = false; 1296c50d8ae3SPaolo Bonzini 1297c50d8ae3SPaolo Bonzini for_each_rmap_spte(rmap_head, &iter, sptep) 1298c50d8ae3SPaolo Bonzini if (spte_ad_need_write_protect(*sptep)) 1299c50d8ae3SPaolo Bonzini flush |= spte_wrprot_for_clear_dirty(sptep); 1300c50d8ae3SPaolo Bonzini else 1301c50d8ae3SPaolo Bonzini flush |= spte_clear_dirty(sptep); 1302c50d8ae3SPaolo Bonzini 1303c50d8ae3SPaolo Bonzini return flush; 1304c50d8ae3SPaolo Bonzini } 1305c50d8ae3SPaolo Bonzini 1306c50d8ae3SPaolo Bonzini /** 1307c50d8ae3SPaolo Bonzini * kvm_mmu_write_protect_pt_masked - write protect selected PT level pages 1308c50d8ae3SPaolo Bonzini * @kvm: kvm instance 1309c50d8ae3SPaolo Bonzini * @slot: slot to protect 1310c50d8ae3SPaolo Bonzini * @gfn_offset: start of the BITS_PER_LONG pages we care about 1311c50d8ae3SPaolo Bonzini * @mask: indicates which pages we should protect 1312c50d8ae3SPaolo Bonzini * 131389212919SKeqian Zhu * Used when we do not need to care about huge page mappings. 1314c50d8ae3SPaolo Bonzini */ 1315c50d8ae3SPaolo Bonzini static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm, 1316c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot, 1317c50d8ae3SPaolo Bonzini gfn_t gfn_offset, unsigned long mask) 1318c50d8ae3SPaolo Bonzini { 1319c50d8ae3SPaolo Bonzini struct kvm_rmap_head *rmap_head; 1320c50d8ae3SPaolo Bonzini 13211f98f2bdSDavid Matlack if (tdp_mmu_enabled) 1322a6a0b05dSBen Gardon kvm_tdp_mmu_clear_dirty_pt_masked(kvm, slot, 1323a6a0b05dSBen Gardon slot->base_gfn + gfn_offset, mask, true); 1324e2209710SBen Gardon 1325e2209710SBen Gardon if (!kvm_memslots_have_rmaps(kvm)) 1326e2209710SBen Gardon return; 1327e2209710SBen Gardon 1328c50d8ae3SPaolo Bonzini while (mask) { 132993e083d4SDavid Matlack rmap_head = gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask), 13303bae0459SSean Christopherson PG_LEVEL_4K, slot); 13311346bbb6SDavid Matlack rmap_write_protect(rmap_head, false); 1332c50d8ae3SPaolo Bonzini 1333c50d8ae3SPaolo Bonzini /* clear the first set bit */ 1334c50d8ae3SPaolo Bonzini mask &= mask - 1; 1335c50d8ae3SPaolo Bonzini } 1336c50d8ae3SPaolo Bonzini } 1337c50d8ae3SPaolo Bonzini 1338c50d8ae3SPaolo Bonzini /** 1339c50d8ae3SPaolo Bonzini * kvm_mmu_clear_dirty_pt_masked - clear MMU D-bit for PT level pages, or write 1340c50d8ae3SPaolo Bonzini * protect the page if the D-bit isn't supported. 1341c50d8ae3SPaolo Bonzini * @kvm: kvm instance 1342c50d8ae3SPaolo Bonzini * @slot: slot to clear D-bit 1343c50d8ae3SPaolo Bonzini * @gfn_offset: start of the BITS_PER_LONG pages we care about 1344c50d8ae3SPaolo Bonzini * @mask: indicates which pages we should clear D-bit 1345c50d8ae3SPaolo Bonzini * 1346c50d8ae3SPaolo Bonzini * Used for PML to re-log the dirty GPAs after userspace querying dirty_bitmap. 1347c50d8ae3SPaolo Bonzini */ 1348a018eba5SSean Christopherson static void kvm_mmu_clear_dirty_pt_masked(struct kvm *kvm, 1349c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot, 1350c50d8ae3SPaolo Bonzini gfn_t gfn_offset, unsigned long mask) 1351c50d8ae3SPaolo Bonzini { 1352c50d8ae3SPaolo Bonzini struct kvm_rmap_head *rmap_head; 1353c50d8ae3SPaolo Bonzini 13541f98f2bdSDavid Matlack if (tdp_mmu_enabled) 1355a6a0b05dSBen Gardon kvm_tdp_mmu_clear_dirty_pt_masked(kvm, slot, 1356a6a0b05dSBen Gardon slot->base_gfn + gfn_offset, mask, false); 1357e2209710SBen Gardon 1358e2209710SBen Gardon if (!kvm_memslots_have_rmaps(kvm)) 1359e2209710SBen Gardon return; 1360e2209710SBen Gardon 1361c50d8ae3SPaolo Bonzini while (mask) { 136293e083d4SDavid Matlack rmap_head = gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask), 13633bae0459SSean Christopherson PG_LEVEL_4K, slot); 13640a234f5dSSean Christopherson __rmap_clear_dirty(kvm, rmap_head, slot); 1365c50d8ae3SPaolo Bonzini 1366c50d8ae3SPaolo Bonzini /* clear the first set bit */ 1367c50d8ae3SPaolo Bonzini mask &= mask - 1; 1368c50d8ae3SPaolo Bonzini } 1369c50d8ae3SPaolo Bonzini } 1370c50d8ae3SPaolo Bonzini 1371c50d8ae3SPaolo Bonzini /** 1372c50d8ae3SPaolo Bonzini * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected 1373c50d8ae3SPaolo Bonzini * PT level pages. 1374c50d8ae3SPaolo Bonzini * 1375c50d8ae3SPaolo Bonzini * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to 1376c50d8ae3SPaolo Bonzini * enable dirty logging for them. 1377c50d8ae3SPaolo Bonzini * 137889212919SKeqian Zhu * We need to care about huge page mappings: e.g. during dirty logging we may 137989212919SKeqian Zhu * have such mappings. 1380c50d8ae3SPaolo Bonzini */ 1381c50d8ae3SPaolo Bonzini void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm, 1382c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot, 1383c50d8ae3SPaolo Bonzini gfn_t gfn_offset, unsigned long mask) 1384c50d8ae3SPaolo Bonzini { 138589212919SKeqian Zhu /* 138689212919SKeqian Zhu * Huge pages are NOT write protected when we start dirty logging in 138789212919SKeqian Zhu * initially-all-set mode; must write protect them here so that they 138889212919SKeqian Zhu * are split to 4K on the first write. 138989212919SKeqian Zhu * 139089212919SKeqian Zhu * The gfn_offset is guaranteed to be aligned to 64, but the base_gfn 139189212919SKeqian Zhu * of memslot has no such restriction, so the range can cross two large 139289212919SKeqian Zhu * pages. 139389212919SKeqian Zhu */ 139489212919SKeqian Zhu if (kvm_dirty_log_manual_protect_and_init_set(kvm)) { 139589212919SKeqian Zhu gfn_t start = slot->base_gfn + gfn_offset + __ffs(mask); 139689212919SKeqian Zhu gfn_t end = slot->base_gfn + gfn_offset + __fls(mask); 139789212919SKeqian Zhu 1398cb00a70bSDavid Matlack if (READ_ONCE(eager_page_split)) 1399cb00a70bSDavid Matlack kvm_mmu_try_split_huge_pages(kvm, slot, start, end, PG_LEVEL_4K); 1400cb00a70bSDavid Matlack 140189212919SKeqian Zhu kvm_mmu_slot_gfn_write_protect(kvm, slot, start, PG_LEVEL_2M); 140289212919SKeqian Zhu 140389212919SKeqian Zhu /* Cross two large pages? */ 140489212919SKeqian Zhu if (ALIGN(start << PAGE_SHIFT, PMD_SIZE) != 140589212919SKeqian Zhu ALIGN(end << PAGE_SHIFT, PMD_SIZE)) 140689212919SKeqian Zhu kvm_mmu_slot_gfn_write_protect(kvm, slot, end, 140789212919SKeqian Zhu PG_LEVEL_2M); 140889212919SKeqian Zhu } 140989212919SKeqian Zhu 141089212919SKeqian Zhu /* Now handle 4K PTEs. */ 1411a018eba5SSean Christopherson if (kvm_x86_ops.cpu_dirty_log_size) 1412a018eba5SSean Christopherson kvm_mmu_clear_dirty_pt_masked(kvm, slot, gfn_offset, mask); 1413c50d8ae3SPaolo Bonzini else 1414c50d8ae3SPaolo Bonzini kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask); 1415c50d8ae3SPaolo Bonzini } 1416c50d8ae3SPaolo Bonzini 1417fb04a1edSPeter Xu int kvm_cpu_dirty_log_size(void) 1418fb04a1edSPeter Xu { 14196dd03800SSean Christopherson return kvm_x86_ops.cpu_dirty_log_size; 1420fb04a1edSPeter Xu } 1421fb04a1edSPeter Xu 1422c50d8ae3SPaolo Bonzini bool kvm_mmu_slot_gfn_write_protect(struct kvm *kvm, 14233ad93562SKeqian Zhu struct kvm_memory_slot *slot, u64 gfn, 14243ad93562SKeqian Zhu int min_level) 1425c50d8ae3SPaolo Bonzini { 1426c50d8ae3SPaolo Bonzini struct kvm_rmap_head *rmap_head; 1427c50d8ae3SPaolo Bonzini int i; 1428c50d8ae3SPaolo Bonzini bool write_protected = false; 1429c50d8ae3SPaolo Bonzini 1430e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) { 14313ad93562SKeqian Zhu for (i = min_level; i <= KVM_MAX_HUGEPAGE_LEVEL; ++i) { 143293e083d4SDavid Matlack rmap_head = gfn_to_rmap(gfn, i, slot); 14331346bbb6SDavid Matlack write_protected |= rmap_write_protect(rmap_head, true); 1434c50d8ae3SPaolo Bonzini } 1435e2209710SBen Gardon } 1436c50d8ae3SPaolo Bonzini 14371f98f2bdSDavid Matlack if (tdp_mmu_enabled) 143846044f72SBen Gardon write_protected |= 14393ad93562SKeqian Zhu kvm_tdp_mmu_write_protect_gfn(kvm, slot, gfn, min_level); 144046044f72SBen Gardon 1441c50d8ae3SPaolo Bonzini return write_protected; 1442c50d8ae3SPaolo Bonzini } 1443c50d8ae3SPaolo Bonzini 1444cf48f9e2SDavid Matlack static bool kvm_vcpu_write_protect_gfn(struct kvm_vcpu *vcpu, u64 gfn) 1445c50d8ae3SPaolo Bonzini { 1446c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot; 1447c50d8ae3SPaolo Bonzini 1448c50d8ae3SPaolo Bonzini slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn); 14493ad93562SKeqian Zhu return kvm_mmu_slot_gfn_write_protect(vcpu->kvm, slot, gfn, PG_LEVEL_4K); 1450c50d8ae3SPaolo Bonzini } 1451c50d8ae3SPaolo Bonzini 1452f8480721SSean Christopherson static bool __kvm_zap_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1453269e9552SHamza Mahfooz const struct kvm_memory_slot *slot) 1454c50d8ae3SPaolo Bonzini { 14559202aee8SSean Christopherson return kvm_zap_all_rmap_sptes(kvm, rmap_head); 1456c50d8ae3SPaolo Bonzini } 1457c50d8ae3SPaolo Bonzini 1458f8480721SSean Christopherson static bool kvm_zap_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1459c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot, gfn_t gfn, int level, 14603039bcc7SSean Christopherson pte_t unused) 1461c50d8ae3SPaolo Bonzini { 1462f8480721SSean Christopherson return __kvm_zap_rmap(kvm, rmap_head, slot); 1463c50d8ae3SPaolo Bonzini } 1464c50d8ae3SPaolo Bonzini 1465aed02fe3SSean Christopherson static bool kvm_set_pte_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1466c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot, gfn_t gfn, int level, 14673039bcc7SSean Christopherson pte_t pte) 1468c50d8ae3SPaolo Bonzini { 1469c50d8ae3SPaolo Bonzini u64 *sptep; 1470c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 147198a26b69SVihas Mak bool need_flush = false; 1472c50d8ae3SPaolo Bonzini u64 new_spte; 1473c50d8ae3SPaolo Bonzini kvm_pfn_t new_pfn; 1474c50d8ae3SPaolo Bonzini 14753039bcc7SSean Christopherson WARN_ON(pte_huge(pte)); 14763039bcc7SSean Christopherson new_pfn = pte_pfn(pte); 1477c50d8ae3SPaolo Bonzini 1478c50d8ae3SPaolo Bonzini restart: 1479c50d8ae3SPaolo Bonzini for_each_rmap_spte(rmap_head, &iter, sptep) { 1480805a0f83SStephen Zhang rmap_printk("spte %p %llx gfn %llx (%d)\n", 1481c50d8ae3SPaolo Bonzini sptep, *sptep, gfn, level); 1482c50d8ae3SPaolo Bonzini 148398a26b69SVihas Mak need_flush = true; 1484c50d8ae3SPaolo Bonzini 14853039bcc7SSean Christopherson if (pte_write(pte)) { 14869202aee8SSean Christopherson kvm_zap_one_rmap_spte(kvm, rmap_head, sptep); 1487c50d8ae3SPaolo Bonzini goto restart; 1488c50d8ae3SPaolo Bonzini } else { 1489cb3eedabSPaolo Bonzini new_spte = kvm_mmu_changed_pte_notifier_make_spte( 1490cb3eedabSPaolo Bonzini *sptep, new_pfn); 1491c50d8ae3SPaolo Bonzini 149271f51d2cSMingwei Zhang mmu_spte_clear_track_bits(kvm, sptep); 1493c50d8ae3SPaolo Bonzini mmu_spte_set(sptep, new_spte); 1494c50d8ae3SPaolo Bonzini } 1495c50d8ae3SPaolo Bonzini } 1496c50d8ae3SPaolo Bonzini 14978a1300ffSSean Christopherson if (need_flush && kvm_available_flush_remote_tlbs_range()) { 14989ffe9265SHou Wenlong kvm_flush_remote_tlbs_gfn(kvm, gfn, level); 149998a26b69SVihas Mak return false; 1500c50d8ae3SPaolo Bonzini } 1501c50d8ae3SPaolo Bonzini 1502c50d8ae3SPaolo Bonzini return need_flush; 1503c50d8ae3SPaolo Bonzini } 1504c50d8ae3SPaolo Bonzini 1505c50d8ae3SPaolo Bonzini struct slot_rmap_walk_iterator { 1506c50d8ae3SPaolo Bonzini /* input fields. */ 1507269e9552SHamza Mahfooz const struct kvm_memory_slot *slot; 1508c50d8ae3SPaolo Bonzini gfn_t start_gfn; 1509c50d8ae3SPaolo Bonzini gfn_t end_gfn; 1510c50d8ae3SPaolo Bonzini int start_level; 1511c50d8ae3SPaolo Bonzini int end_level; 1512c50d8ae3SPaolo Bonzini 1513c50d8ae3SPaolo Bonzini /* output fields. */ 1514c50d8ae3SPaolo Bonzini gfn_t gfn; 1515c50d8ae3SPaolo Bonzini struct kvm_rmap_head *rmap; 1516c50d8ae3SPaolo Bonzini int level; 1517c50d8ae3SPaolo Bonzini 1518c50d8ae3SPaolo Bonzini /* private field. */ 1519c50d8ae3SPaolo Bonzini struct kvm_rmap_head *end_rmap; 1520c50d8ae3SPaolo Bonzini }; 1521c50d8ae3SPaolo Bonzini 1522f3d90f90SSean Christopherson static void rmap_walk_init_level(struct slot_rmap_walk_iterator *iterator, 1523f3d90f90SSean Christopherson int level) 1524c50d8ae3SPaolo Bonzini { 1525c50d8ae3SPaolo Bonzini iterator->level = level; 1526c50d8ae3SPaolo Bonzini iterator->gfn = iterator->start_gfn; 152793e083d4SDavid Matlack iterator->rmap = gfn_to_rmap(iterator->gfn, level, iterator->slot); 152893e083d4SDavid Matlack iterator->end_rmap = gfn_to_rmap(iterator->end_gfn, level, iterator->slot); 1529c50d8ae3SPaolo Bonzini } 1530c50d8ae3SPaolo Bonzini 1531f3d90f90SSean Christopherson static void slot_rmap_walk_init(struct slot_rmap_walk_iterator *iterator, 1532f3d90f90SSean Christopherson const struct kvm_memory_slot *slot, 1533f3d90f90SSean Christopherson int start_level, int end_level, 1534f3d90f90SSean Christopherson gfn_t start_gfn, gfn_t end_gfn) 1535c50d8ae3SPaolo Bonzini { 1536c50d8ae3SPaolo Bonzini iterator->slot = slot; 1537c50d8ae3SPaolo Bonzini iterator->start_level = start_level; 1538c50d8ae3SPaolo Bonzini iterator->end_level = end_level; 1539c50d8ae3SPaolo Bonzini iterator->start_gfn = start_gfn; 1540c50d8ae3SPaolo Bonzini iterator->end_gfn = end_gfn; 1541c50d8ae3SPaolo Bonzini 1542c50d8ae3SPaolo Bonzini rmap_walk_init_level(iterator, iterator->start_level); 1543c50d8ae3SPaolo Bonzini } 1544c50d8ae3SPaolo Bonzini 1545c50d8ae3SPaolo Bonzini static bool slot_rmap_walk_okay(struct slot_rmap_walk_iterator *iterator) 1546c50d8ae3SPaolo Bonzini { 1547c50d8ae3SPaolo Bonzini return !!iterator->rmap; 1548c50d8ae3SPaolo Bonzini } 1549c50d8ae3SPaolo Bonzini 1550c50d8ae3SPaolo Bonzini static void slot_rmap_walk_next(struct slot_rmap_walk_iterator *iterator) 1551c50d8ae3SPaolo Bonzini { 15526ba1e04fSVipin Sharma while (++iterator->rmap <= iterator->end_rmap) { 1553c50d8ae3SPaolo Bonzini iterator->gfn += (1UL << KVM_HPAGE_GFN_SHIFT(iterator->level)); 15546ba1e04fSVipin Sharma 15556ba1e04fSVipin Sharma if (iterator->rmap->val) 1556c50d8ae3SPaolo Bonzini return; 1557c50d8ae3SPaolo Bonzini } 1558c50d8ae3SPaolo Bonzini 1559c50d8ae3SPaolo Bonzini if (++iterator->level > iterator->end_level) { 1560c50d8ae3SPaolo Bonzini iterator->rmap = NULL; 1561c50d8ae3SPaolo Bonzini return; 1562c50d8ae3SPaolo Bonzini } 1563c50d8ae3SPaolo Bonzini 1564c50d8ae3SPaolo Bonzini rmap_walk_init_level(iterator, iterator->level); 1565c50d8ae3SPaolo Bonzini } 1566c50d8ae3SPaolo Bonzini 1567c50d8ae3SPaolo Bonzini #define for_each_slot_rmap_range(_slot_, _start_level_, _end_level_, \ 1568c50d8ae3SPaolo Bonzini _start_gfn, _end_gfn, _iter_) \ 1569c50d8ae3SPaolo Bonzini for (slot_rmap_walk_init(_iter_, _slot_, _start_level_, \ 1570c50d8ae3SPaolo Bonzini _end_level_, _start_gfn, _end_gfn); \ 1571c50d8ae3SPaolo Bonzini slot_rmap_walk_okay(_iter_); \ 1572c50d8ae3SPaolo Bonzini slot_rmap_walk_next(_iter_)) 1573c50d8ae3SPaolo Bonzini 15743039bcc7SSean Christopherson typedef bool (*rmap_handler_t)(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1575c1b91493SSean Christopherson struct kvm_memory_slot *slot, gfn_t gfn, 15763039bcc7SSean Christopherson int level, pte_t pte); 1577c1b91493SSean Christopherson 15783039bcc7SSean Christopherson static __always_inline bool kvm_handle_gfn_range(struct kvm *kvm, 15793039bcc7SSean Christopherson struct kvm_gfn_range *range, 1580c1b91493SSean Christopherson rmap_handler_t handler) 1581c50d8ae3SPaolo Bonzini { 1582c50d8ae3SPaolo Bonzini struct slot_rmap_walk_iterator iterator; 15833039bcc7SSean Christopherson bool ret = false; 1584c50d8ae3SPaolo Bonzini 15853039bcc7SSean Christopherson for_each_slot_rmap_range(range->slot, PG_LEVEL_4K, KVM_MAX_HUGEPAGE_LEVEL, 15863039bcc7SSean Christopherson range->start, range->end - 1, &iterator) 15873039bcc7SSean Christopherson ret |= handler(kvm, iterator.rmap, range->slot, iterator.gfn, 15883039bcc7SSean Christopherson iterator.level, range->pte); 1589c50d8ae3SPaolo Bonzini 1590c50d8ae3SPaolo Bonzini return ret; 1591c50d8ae3SPaolo Bonzini } 1592c50d8ae3SPaolo Bonzini 15933039bcc7SSean Christopherson bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range) 1594c50d8ae3SPaolo Bonzini { 1595e2209710SBen Gardon bool flush = false; 1596c50d8ae3SPaolo Bonzini 1597e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) 1598f8480721SSean Christopherson flush = kvm_handle_gfn_range(kvm, range, kvm_zap_rmap); 1599063afacdSBen Gardon 16001f98f2bdSDavid Matlack if (tdp_mmu_enabled) 1601c7785d85SHou Wenlong flush = kvm_tdp_mmu_unmap_gfn_range(kvm, range, flush); 1602063afacdSBen Gardon 1603*0a8a5f2cSSean Christopherson if (range->slot->id == APIC_ACCESS_PAGE_PRIVATE_MEMSLOT) 1604*0a8a5f2cSSean Christopherson kvm_make_all_cpus_request(kvm, KVM_REQ_APIC_PAGE_RELOAD); 1605*0a8a5f2cSSean Christopherson 16063039bcc7SSean Christopherson return flush; 1607c50d8ae3SPaolo Bonzini } 1608c50d8ae3SPaolo Bonzini 16093039bcc7SSean Christopherson bool kvm_set_spte_gfn(struct kvm *kvm, struct kvm_gfn_range *range) 1610c50d8ae3SPaolo Bonzini { 1611e2209710SBen Gardon bool flush = false; 16121d8dd6b3SBen Gardon 1613e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) 1614aed02fe3SSean Christopherson flush = kvm_handle_gfn_range(kvm, range, kvm_set_pte_rmap); 16151d8dd6b3SBen Gardon 16161f98f2bdSDavid Matlack if (tdp_mmu_enabled) 16173039bcc7SSean Christopherson flush |= kvm_tdp_mmu_set_spte_gfn(kvm, range); 16181d8dd6b3SBen Gardon 16193039bcc7SSean Christopherson return flush; 1620c50d8ae3SPaolo Bonzini } 1621c50d8ae3SPaolo Bonzini 1622aed02fe3SSean Christopherson static bool kvm_age_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1623c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot, gfn_t gfn, int level, 16243039bcc7SSean Christopherson pte_t unused) 1625c50d8ae3SPaolo Bonzini { 1626c50d8ae3SPaolo Bonzini u64 *sptep; 16273f649ab7SKees Cook struct rmap_iterator iter; 1628c50d8ae3SPaolo Bonzini int young = 0; 1629c50d8ae3SPaolo Bonzini 1630c50d8ae3SPaolo Bonzini for_each_rmap_spte(rmap_head, &iter, sptep) 1631c50d8ae3SPaolo Bonzini young |= mmu_spte_age(sptep); 1632c50d8ae3SPaolo Bonzini 1633c50d8ae3SPaolo Bonzini return young; 1634c50d8ae3SPaolo Bonzini } 1635c50d8ae3SPaolo Bonzini 1636aed02fe3SSean Christopherson static bool kvm_test_age_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head, 1637c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot, gfn_t gfn, 16383039bcc7SSean Christopherson int level, pte_t unused) 1639c50d8ae3SPaolo Bonzini { 1640c50d8ae3SPaolo Bonzini u64 *sptep; 1641c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 1642c50d8ae3SPaolo Bonzini 1643c50d8ae3SPaolo Bonzini for_each_rmap_spte(rmap_head, &iter, sptep) 1644c50d8ae3SPaolo Bonzini if (is_accessed_spte(*sptep)) 164598a26b69SVihas Mak return true; 164698a26b69SVihas Mak return false; 1647c50d8ae3SPaolo Bonzini } 1648c50d8ae3SPaolo Bonzini 1649c50d8ae3SPaolo Bonzini #define RMAP_RECYCLE_THRESHOLD 1000 1650c50d8ae3SPaolo Bonzini 16512ff9039aSDavid Matlack static void __rmap_add(struct kvm *kvm, 16522ff9039aSDavid Matlack struct kvm_mmu_memory_cache *cache, 16532ff9039aSDavid Matlack const struct kvm_memory_slot *slot, 165472ae5822SSean Christopherson u64 *spte, gfn_t gfn, unsigned int access) 1655c50d8ae3SPaolo Bonzini { 1656c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 165768be1306SDavid Matlack struct kvm_rmap_head *rmap_head; 165868be1306SDavid Matlack int rmap_count; 1659c50d8ae3SPaolo Bonzini 166057354682SSean Christopherson sp = sptep_to_sp(spte); 166179e48cecSSean Christopherson kvm_mmu_page_set_translation(sp, spte_index(spte), gfn, access); 166281cb4657SDavid Matlack kvm_update_page_stats(kvm, sp->role.level, 1); 166381cb4657SDavid Matlack 166493e083d4SDavid Matlack rmap_head = gfn_to_rmap(gfn, sp->role.level, slot); 16652ff9039aSDavid Matlack rmap_count = pte_list_add(cache, spte, rmap_head); 1666c50d8ae3SPaolo Bonzini 1667604f5332SMiaohe Lin if (rmap_count > kvm->stat.max_mmu_rmap_size) 1668604f5332SMiaohe Lin kvm->stat.max_mmu_rmap_size = rmap_count; 166968be1306SDavid Matlack if (rmap_count > RMAP_RECYCLE_THRESHOLD) { 16709202aee8SSean Christopherson kvm_zap_all_rmap_sptes(kvm, rmap_head); 16711b2dc736SHou Wenlong kvm_flush_remote_tlbs_gfn(kvm, gfn, sp->role.level); 167268be1306SDavid Matlack } 1673c50d8ae3SPaolo Bonzini } 1674c50d8ae3SPaolo Bonzini 16752ff9039aSDavid Matlack static void rmap_add(struct kvm_vcpu *vcpu, const struct kvm_memory_slot *slot, 167672ae5822SSean Christopherson u64 *spte, gfn_t gfn, unsigned int access) 16772ff9039aSDavid Matlack { 16782ff9039aSDavid Matlack struct kvm_mmu_memory_cache *cache = &vcpu->arch.mmu_pte_list_desc_cache; 16792ff9039aSDavid Matlack 16806a97575dSDavid Matlack __rmap_add(vcpu->kvm, cache, slot, spte, gfn, access); 16812ff9039aSDavid Matlack } 16822ff9039aSDavid Matlack 16833039bcc7SSean Christopherson bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range) 1684c50d8ae3SPaolo Bonzini { 1685e2209710SBen Gardon bool young = false; 1686f8e14497SBen Gardon 1687e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) 1688aed02fe3SSean Christopherson young = kvm_handle_gfn_range(kvm, range, kvm_age_rmap); 16893039bcc7SSean Christopherson 16901f98f2bdSDavid Matlack if (tdp_mmu_enabled) 16913039bcc7SSean Christopherson young |= kvm_tdp_mmu_age_gfn_range(kvm, range); 1692f8e14497SBen Gardon 1693f8e14497SBen Gardon return young; 1694c50d8ae3SPaolo Bonzini } 1695c50d8ae3SPaolo Bonzini 16963039bcc7SSean Christopherson bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range) 1697c50d8ae3SPaolo Bonzini { 1698e2209710SBen Gardon bool young = false; 1699f8e14497SBen Gardon 1700e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) 1701aed02fe3SSean Christopherson young = kvm_handle_gfn_range(kvm, range, kvm_test_age_rmap); 17023039bcc7SSean Christopherson 17031f98f2bdSDavid Matlack if (tdp_mmu_enabled) 17043039bcc7SSean Christopherson young |= kvm_tdp_mmu_test_age_gfn(kvm, range); 1705f8e14497SBen Gardon 1706f8e14497SBen Gardon return young; 1707c50d8ae3SPaolo Bonzini } 1708c50d8ae3SPaolo Bonzini 1709c50d8ae3SPaolo Bonzini #ifdef MMU_DEBUG 1710c50d8ae3SPaolo Bonzini static int is_empty_shadow_page(u64 *spt) 1711c50d8ae3SPaolo Bonzini { 1712c50d8ae3SPaolo Bonzini u64 *pos; 1713c50d8ae3SPaolo Bonzini u64 *end; 1714c50d8ae3SPaolo Bonzini 17153adbdf81SMiaohe Lin for (pos = spt, end = pos + SPTE_ENT_PER_PAGE; pos != end; pos++) 1716c50d8ae3SPaolo Bonzini if (is_shadow_present_pte(*pos)) { 1717c50d8ae3SPaolo Bonzini printk(KERN_ERR "%s: %p %llx\n", __func__, 1718c50d8ae3SPaolo Bonzini pos, *pos); 1719c50d8ae3SPaolo Bonzini return 0; 1720c50d8ae3SPaolo Bonzini } 1721c50d8ae3SPaolo Bonzini return 1; 1722c50d8ae3SPaolo Bonzini } 1723c50d8ae3SPaolo Bonzini #endif 1724c50d8ae3SPaolo Bonzini 1725c50d8ae3SPaolo Bonzini /* 1726c50d8ae3SPaolo Bonzini * This value is the sum of all of the kvm instances's 1727c50d8ae3SPaolo Bonzini * kvm->arch.n_used_mmu_pages values. We need a global, 1728c50d8ae3SPaolo Bonzini * aggregate version in order to make the slab shrinker 1729c50d8ae3SPaolo Bonzini * faster 1730c50d8ae3SPaolo Bonzini */ 1731d5aaad6fSSean Christopherson static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, long nr) 1732c50d8ae3SPaolo Bonzini { 1733c50d8ae3SPaolo Bonzini kvm->arch.n_used_mmu_pages += nr; 1734c50d8ae3SPaolo Bonzini percpu_counter_add(&kvm_total_used_mmu_pages, nr); 1735c50d8ae3SPaolo Bonzini } 1736c50d8ae3SPaolo Bonzini 173743a063caSYosry Ahmed static void kvm_account_mmu_page(struct kvm *kvm, struct kvm_mmu_page *sp) 173843a063caSYosry Ahmed { 173943a063caSYosry Ahmed kvm_mod_used_mmu_pages(kvm, +1); 174043a063caSYosry Ahmed kvm_account_pgtable_pages((void *)sp->spt, +1); 174143a063caSYosry Ahmed } 174243a063caSYosry Ahmed 174343a063caSYosry Ahmed static void kvm_unaccount_mmu_page(struct kvm *kvm, struct kvm_mmu_page *sp) 174443a063caSYosry Ahmed { 174543a063caSYosry Ahmed kvm_mod_used_mmu_pages(kvm, -1); 174643a063caSYosry Ahmed kvm_account_pgtable_pages((void *)sp->spt, -1); 174743a063caSYosry Ahmed } 174843a063caSYosry Ahmed 174987654643SDavid Matlack static void kvm_mmu_free_shadow_page(struct kvm_mmu_page *sp) 1750c50d8ae3SPaolo Bonzini { 1751c50d8ae3SPaolo Bonzini MMU_WARN_ON(!is_empty_shadow_page(sp->spt)); 1752c50d8ae3SPaolo Bonzini hlist_del(&sp->hash_link); 1753c50d8ae3SPaolo Bonzini list_del(&sp->link); 1754c50d8ae3SPaolo Bonzini free_page((unsigned long)sp->spt); 1755c50d8ae3SPaolo Bonzini if (!sp->role.direct) 17566a97575dSDavid Matlack free_page((unsigned long)sp->shadowed_translation); 1757c50d8ae3SPaolo Bonzini kmem_cache_free(mmu_page_header_cache, sp); 1758c50d8ae3SPaolo Bonzini } 1759c50d8ae3SPaolo Bonzini 1760c50d8ae3SPaolo Bonzini static unsigned kvm_page_table_hashfn(gfn_t gfn) 1761c50d8ae3SPaolo Bonzini { 1762c50d8ae3SPaolo Bonzini return hash_64(gfn, KVM_MMU_HASH_SHIFT); 1763c50d8ae3SPaolo Bonzini } 1764c50d8ae3SPaolo Bonzini 17652ff9039aSDavid Matlack static void mmu_page_add_parent_pte(struct kvm_mmu_memory_cache *cache, 1766c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, u64 *parent_pte) 1767c50d8ae3SPaolo Bonzini { 1768c50d8ae3SPaolo Bonzini if (!parent_pte) 1769c50d8ae3SPaolo Bonzini return; 1770c50d8ae3SPaolo Bonzini 17712ff9039aSDavid Matlack pte_list_add(cache, parent_pte, &sp->parent_ptes); 1772c50d8ae3SPaolo Bonzini } 1773c50d8ae3SPaolo Bonzini 1774c50d8ae3SPaolo Bonzini static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp, 1775c50d8ae3SPaolo Bonzini u64 *parent_pte) 1776c50d8ae3SPaolo Bonzini { 17773c2e1037SSean Christopherson pte_list_remove(parent_pte, &sp->parent_ptes); 1778c50d8ae3SPaolo Bonzini } 1779c50d8ae3SPaolo Bonzini 1780c50d8ae3SPaolo Bonzini static void drop_parent_pte(struct kvm_mmu_page *sp, 1781c50d8ae3SPaolo Bonzini u64 *parent_pte) 1782c50d8ae3SPaolo Bonzini { 1783c50d8ae3SPaolo Bonzini mmu_page_remove_parent_pte(sp, parent_pte); 1784c50d8ae3SPaolo Bonzini mmu_spte_clear_no_track(parent_pte); 1785c50d8ae3SPaolo Bonzini } 1786c50d8ae3SPaolo Bonzini 1787c50d8ae3SPaolo Bonzini static void mark_unsync(u64 *spte); 1788c50d8ae3SPaolo Bonzini static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp) 1789c50d8ae3SPaolo Bonzini { 1790c50d8ae3SPaolo Bonzini u64 *sptep; 1791c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 1792c50d8ae3SPaolo Bonzini 1793c50d8ae3SPaolo Bonzini for_each_rmap_spte(&sp->parent_ptes, &iter, sptep) { 1794c50d8ae3SPaolo Bonzini mark_unsync(sptep); 1795c50d8ae3SPaolo Bonzini } 1796c50d8ae3SPaolo Bonzini } 1797c50d8ae3SPaolo Bonzini 1798c50d8ae3SPaolo Bonzini static void mark_unsync(u64 *spte) 1799c50d8ae3SPaolo Bonzini { 1800c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 1801c50d8ae3SPaolo Bonzini 180257354682SSean Christopherson sp = sptep_to_sp(spte); 180379e48cecSSean Christopherson if (__test_and_set_bit(spte_index(spte), sp->unsync_child_bitmap)) 1804c50d8ae3SPaolo Bonzini return; 1805c50d8ae3SPaolo Bonzini if (sp->unsync_children++) 1806c50d8ae3SPaolo Bonzini return; 1807c50d8ae3SPaolo Bonzini kvm_mmu_mark_parents_unsync(sp); 1808c50d8ae3SPaolo Bonzini } 1809c50d8ae3SPaolo Bonzini 1810c50d8ae3SPaolo Bonzini #define KVM_PAGE_ARRAY_NR 16 1811c50d8ae3SPaolo Bonzini 1812c50d8ae3SPaolo Bonzini struct kvm_mmu_pages { 1813c50d8ae3SPaolo Bonzini struct mmu_page_and_offset { 1814c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 1815c50d8ae3SPaolo Bonzini unsigned int idx; 1816c50d8ae3SPaolo Bonzini } page[KVM_PAGE_ARRAY_NR]; 1817c50d8ae3SPaolo Bonzini unsigned int nr; 1818c50d8ae3SPaolo Bonzini }; 1819c50d8ae3SPaolo Bonzini 1820c50d8ae3SPaolo Bonzini static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp, 1821c50d8ae3SPaolo Bonzini int idx) 1822c50d8ae3SPaolo Bonzini { 1823c50d8ae3SPaolo Bonzini int i; 1824c50d8ae3SPaolo Bonzini 1825c50d8ae3SPaolo Bonzini if (sp->unsync) 1826c50d8ae3SPaolo Bonzini for (i=0; i < pvec->nr; i++) 1827c50d8ae3SPaolo Bonzini if (pvec->page[i].sp == sp) 1828c50d8ae3SPaolo Bonzini return 0; 1829c50d8ae3SPaolo Bonzini 1830c50d8ae3SPaolo Bonzini pvec->page[pvec->nr].sp = sp; 1831c50d8ae3SPaolo Bonzini pvec->page[pvec->nr].idx = idx; 1832c50d8ae3SPaolo Bonzini pvec->nr++; 1833c50d8ae3SPaolo Bonzini return (pvec->nr == KVM_PAGE_ARRAY_NR); 1834c50d8ae3SPaolo Bonzini } 1835c50d8ae3SPaolo Bonzini 1836c50d8ae3SPaolo Bonzini static inline void clear_unsync_child_bit(struct kvm_mmu_page *sp, int idx) 1837c50d8ae3SPaolo Bonzini { 1838c50d8ae3SPaolo Bonzini --sp->unsync_children; 1839c50d8ae3SPaolo Bonzini WARN_ON((int)sp->unsync_children < 0); 1840c50d8ae3SPaolo Bonzini __clear_bit(idx, sp->unsync_child_bitmap); 1841c50d8ae3SPaolo Bonzini } 1842c50d8ae3SPaolo Bonzini 1843c50d8ae3SPaolo Bonzini static int __mmu_unsync_walk(struct kvm_mmu_page *sp, 1844c50d8ae3SPaolo Bonzini struct kvm_mmu_pages *pvec) 1845c50d8ae3SPaolo Bonzini { 1846c50d8ae3SPaolo Bonzini int i, ret, nr_unsync_leaf = 0; 1847c50d8ae3SPaolo Bonzini 1848c50d8ae3SPaolo Bonzini for_each_set_bit(i, sp->unsync_child_bitmap, 512) { 1849c50d8ae3SPaolo Bonzini struct kvm_mmu_page *child; 1850c50d8ae3SPaolo Bonzini u64 ent = sp->spt[i]; 1851c50d8ae3SPaolo Bonzini 1852c50d8ae3SPaolo Bonzini if (!is_shadow_present_pte(ent) || is_large_pte(ent)) { 1853c50d8ae3SPaolo Bonzini clear_unsync_child_bit(sp, i); 1854c50d8ae3SPaolo Bonzini continue; 1855c50d8ae3SPaolo Bonzini } 1856c50d8ae3SPaolo Bonzini 18575e3edd7eSSean Christopherson child = spte_to_child_sp(ent); 1858c50d8ae3SPaolo Bonzini 1859c50d8ae3SPaolo Bonzini if (child->unsync_children) { 1860c50d8ae3SPaolo Bonzini if (mmu_pages_add(pvec, child, i)) 1861c50d8ae3SPaolo Bonzini return -ENOSPC; 1862c50d8ae3SPaolo Bonzini 1863c50d8ae3SPaolo Bonzini ret = __mmu_unsync_walk(child, pvec); 1864c50d8ae3SPaolo Bonzini if (!ret) { 1865c50d8ae3SPaolo Bonzini clear_unsync_child_bit(sp, i); 1866c50d8ae3SPaolo Bonzini continue; 1867c50d8ae3SPaolo Bonzini } else if (ret > 0) { 1868c50d8ae3SPaolo Bonzini nr_unsync_leaf += ret; 1869c50d8ae3SPaolo Bonzini } else 1870c50d8ae3SPaolo Bonzini return ret; 1871c50d8ae3SPaolo Bonzini } else if (child->unsync) { 1872c50d8ae3SPaolo Bonzini nr_unsync_leaf++; 1873c50d8ae3SPaolo Bonzini if (mmu_pages_add(pvec, child, i)) 1874c50d8ae3SPaolo Bonzini return -ENOSPC; 1875c50d8ae3SPaolo Bonzini } else 1876c50d8ae3SPaolo Bonzini clear_unsync_child_bit(sp, i); 1877c50d8ae3SPaolo Bonzini } 1878c50d8ae3SPaolo Bonzini 1879c50d8ae3SPaolo Bonzini return nr_unsync_leaf; 1880c50d8ae3SPaolo Bonzini } 1881c50d8ae3SPaolo Bonzini 1882c50d8ae3SPaolo Bonzini #define INVALID_INDEX (-1) 1883c50d8ae3SPaolo Bonzini 1884c50d8ae3SPaolo Bonzini static int mmu_unsync_walk(struct kvm_mmu_page *sp, 1885c50d8ae3SPaolo Bonzini struct kvm_mmu_pages *pvec) 1886c50d8ae3SPaolo Bonzini { 1887c50d8ae3SPaolo Bonzini pvec->nr = 0; 1888c50d8ae3SPaolo Bonzini if (!sp->unsync_children) 1889c50d8ae3SPaolo Bonzini return 0; 1890c50d8ae3SPaolo Bonzini 1891c50d8ae3SPaolo Bonzini mmu_pages_add(pvec, sp, INVALID_INDEX); 1892c50d8ae3SPaolo Bonzini return __mmu_unsync_walk(sp, pvec); 1893c50d8ae3SPaolo Bonzini } 1894c50d8ae3SPaolo Bonzini 1895c50d8ae3SPaolo Bonzini static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp) 1896c50d8ae3SPaolo Bonzini { 1897c50d8ae3SPaolo Bonzini WARN_ON(!sp->unsync); 1898c50d8ae3SPaolo Bonzini trace_kvm_mmu_sync_page(sp); 1899c50d8ae3SPaolo Bonzini sp->unsync = 0; 1900c50d8ae3SPaolo Bonzini --kvm->stat.mmu_unsync; 1901c50d8ae3SPaolo Bonzini } 1902c50d8ae3SPaolo Bonzini 1903c50d8ae3SPaolo Bonzini static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp, 1904c50d8ae3SPaolo Bonzini struct list_head *invalid_list); 1905c50d8ae3SPaolo Bonzini static void kvm_mmu_commit_zap_page(struct kvm *kvm, 1906c50d8ae3SPaolo Bonzini struct list_head *invalid_list); 1907c50d8ae3SPaolo Bonzini 1908767d8d8dSLai Jiangshan static bool sp_has_gptes(struct kvm_mmu_page *sp) 1909767d8d8dSLai Jiangshan { 1910767d8d8dSLai Jiangshan if (sp->role.direct) 1911767d8d8dSLai Jiangshan return false; 1912767d8d8dSLai Jiangshan 191384e5ffd0SLai Jiangshan if (sp->role.passthrough) 191484e5ffd0SLai Jiangshan return false; 191584e5ffd0SLai Jiangshan 1916767d8d8dSLai Jiangshan return true; 1917767d8d8dSLai Jiangshan } 1918767d8d8dSLai Jiangshan 1919ac101b7cSSean Christopherson #define for_each_valid_sp(_kvm, _sp, _list) \ 1920ac101b7cSSean Christopherson hlist_for_each_entry(_sp, _list, hash_link) \ 1921c50d8ae3SPaolo Bonzini if (is_obsolete_sp((_kvm), (_sp))) { \ 1922c50d8ae3SPaolo Bonzini } else 1923c50d8ae3SPaolo Bonzini 1924767d8d8dSLai Jiangshan #define for_each_gfn_valid_sp_with_gptes(_kvm, _sp, _gfn) \ 1925ac101b7cSSean Christopherson for_each_valid_sp(_kvm, _sp, \ 1926ac101b7cSSean Christopherson &(_kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(_gfn)]) \ 1927767d8d8dSLai Jiangshan if ((_sp)->gfn != (_gfn) || !sp_has_gptes(_sp)) {} else 1928c50d8ae3SPaolo Bonzini 192990e44470SLai Jiangshan static bool kvm_sync_page_check(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp) 193090e44470SLai Jiangshan { 193190e44470SLai Jiangshan union kvm_mmu_page_role root_role = vcpu->arch.mmu->root_role; 193290e44470SLai Jiangshan 193390e44470SLai Jiangshan /* 193490e44470SLai Jiangshan * Ignore various flags when verifying that it's safe to sync a shadow 193590e44470SLai Jiangshan * page using the current MMU context. 193690e44470SLai Jiangshan * 193790e44470SLai Jiangshan * - level: not part of the overall MMU role and will never match as the MMU's 193890e44470SLai Jiangshan * level tracks the root level 193990e44470SLai Jiangshan * - access: updated based on the new guest PTE 194090e44470SLai Jiangshan * - quadrant: not part of the overall MMU role (similar to level) 194190e44470SLai Jiangshan */ 194290e44470SLai Jiangshan const union kvm_mmu_page_role sync_role_ign = { 194390e44470SLai Jiangshan .level = 0xf, 194490e44470SLai Jiangshan .access = 0x7, 194590e44470SLai Jiangshan .quadrant = 0x3, 194690e44470SLai Jiangshan .passthrough = 0x1, 194790e44470SLai Jiangshan }; 194890e44470SLai Jiangshan 194990e44470SLai Jiangshan /* 195090e44470SLai Jiangshan * Direct pages can never be unsync, and KVM should never attempt to 195190e44470SLai Jiangshan * sync a shadow page for a different MMU context, e.g. if the role 195290e44470SLai Jiangshan * differs then the memslot lookup (SMM vs. non-SMM) will be bogus, the 195390e44470SLai Jiangshan * reserved bits checks will be wrong, etc... 195490e44470SLai Jiangshan */ 1955c3c6c9fcSLai Jiangshan if (WARN_ON_ONCE(sp->role.direct || !vcpu->arch.mmu->sync_spte || 195690e44470SLai Jiangshan (sp->role.word ^ root_role.word) & ~sync_role_ign.word)) 195790e44470SLai Jiangshan return false; 195890e44470SLai Jiangshan 195990e44470SLai Jiangshan return true; 196090e44470SLai Jiangshan } 196190e44470SLai Jiangshan 196219ace7d6SLai Jiangshan static int kvm_sync_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp, int i) 196319ace7d6SLai Jiangshan { 196419ace7d6SLai Jiangshan if (!sp->spt[i]) 196519ace7d6SLai Jiangshan return 0; 196619ace7d6SLai Jiangshan 196719ace7d6SLai Jiangshan return vcpu->arch.mmu->sync_spte(vcpu, sp, i); 196819ace7d6SLai Jiangshan } 196919ace7d6SLai Jiangshan 197090e44470SLai Jiangshan static int __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp) 197190e44470SLai Jiangshan { 1972c3c6c9fcSLai Jiangshan int flush = 0; 1973c3c6c9fcSLai Jiangshan int i; 1974c3c6c9fcSLai Jiangshan 197590e44470SLai Jiangshan if (!kvm_sync_page_check(vcpu, sp)) 197690e44470SLai Jiangshan return -1; 197790e44470SLai Jiangshan 1978c3c6c9fcSLai Jiangshan for (i = 0; i < SPTE_ENT_PER_PAGE; i++) { 197919ace7d6SLai Jiangshan int ret = kvm_sync_spte(vcpu, sp, i); 1980c3c6c9fcSLai Jiangshan 1981c3c6c9fcSLai Jiangshan if (ret < -1) 1982c3c6c9fcSLai Jiangshan return -1; 1983c3c6c9fcSLai Jiangshan flush |= ret; 1984c3c6c9fcSLai Jiangshan } 1985c3c6c9fcSLai Jiangshan 1986c3c6c9fcSLai Jiangshan /* 1987c3c6c9fcSLai Jiangshan * Note, any flush is purely for KVM's correctness, e.g. when dropping 1988c3c6c9fcSLai Jiangshan * an existing SPTE or clearing W/A/D bits to ensure an mmu_notifier 1989c3c6c9fcSLai Jiangshan * unmap or dirty logging event doesn't fail to flush. The guest is 1990c3c6c9fcSLai Jiangshan * responsible for flushing the TLB to ensure any changes in protection 1991c3c6c9fcSLai Jiangshan * bits are recognized, i.e. until the guest flushes or page faults on 1992c3c6c9fcSLai Jiangshan * a relevant address, KVM is architecturally allowed to let vCPUs use 1993c3c6c9fcSLai Jiangshan * cached translations with the old protection bits. 1994c3c6c9fcSLai Jiangshan */ 1995c3c6c9fcSLai Jiangshan return flush; 199690e44470SLai Jiangshan } 199790e44470SLai Jiangshan 19988d5678a7SHou Wenlong static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp, 1999c50d8ae3SPaolo Bonzini struct list_head *invalid_list) 2000c50d8ae3SPaolo Bonzini { 200190e44470SLai Jiangshan int ret = __kvm_sync_page(vcpu, sp); 2002c3e5e415SLai Jiangshan 20038d5678a7SHou Wenlong if (ret < 0) 2004c50d8ae3SPaolo Bonzini kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list); 20058d5678a7SHou Wenlong return ret; 2006c50d8ae3SPaolo Bonzini } 2007c50d8ae3SPaolo Bonzini 2008c50d8ae3SPaolo Bonzini static bool kvm_mmu_remote_flush_or_zap(struct kvm *kvm, 2009c50d8ae3SPaolo Bonzini struct list_head *invalid_list, 2010c50d8ae3SPaolo Bonzini bool remote_flush) 2011c50d8ae3SPaolo Bonzini { 2012c50d8ae3SPaolo Bonzini if (!remote_flush && list_empty(invalid_list)) 2013c50d8ae3SPaolo Bonzini return false; 2014c50d8ae3SPaolo Bonzini 2015c50d8ae3SPaolo Bonzini if (!list_empty(invalid_list)) 2016c50d8ae3SPaolo Bonzini kvm_mmu_commit_zap_page(kvm, invalid_list); 2017c50d8ae3SPaolo Bonzini else 2018c50d8ae3SPaolo Bonzini kvm_flush_remote_tlbs(kvm); 2019c50d8ae3SPaolo Bonzini return true; 2020c50d8ae3SPaolo Bonzini } 2021c50d8ae3SPaolo Bonzini 2022c50d8ae3SPaolo Bonzini static bool is_obsolete_sp(struct kvm *kvm, struct kvm_mmu_page *sp) 2023c50d8ae3SPaolo Bonzini { 2024a955cad8SSean Christopherson if (sp->role.invalid) 2025a955cad8SSean Christopherson return true; 2026a955cad8SSean Christopherson 2027fa3e4203SMiaohe Lin /* TDP MMU pages do not use the MMU generation. */ 2028de0322f5SSean Christopherson return !is_tdp_mmu_page(sp) && 2029c50d8ae3SPaolo Bonzini unlikely(sp->mmu_valid_gen != kvm->arch.mmu_valid_gen); 2030c50d8ae3SPaolo Bonzini } 2031c50d8ae3SPaolo Bonzini 2032c50d8ae3SPaolo Bonzini struct mmu_page_path { 2033c50d8ae3SPaolo Bonzini struct kvm_mmu_page *parent[PT64_ROOT_MAX_LEVEL]; 2034c50d8ae3SPaolo Bonzini unsigned int idx[PT64_ROOT_MAX_LEVEL]; 2035c50d8ae3SPaolo Bonzini }; 2036c50d8ae3SPaolo Bonzini 2037c50d8ae3SPaolo Bonzini #define for_each_sp(pvec, sp, parents, i) \ 2038c50d8ae3SPaolo Bonzini for (i = mmu_pages_first(&pvec, &parents); \ 2039c50d8ae3SPaolo Bonzini i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \ 2040c50d8ae3SPaolo Bonzini i = mmu_pages_next(&pvec, &parents, i)) 2041c50d8ae3SPaolo Bonzini 2042c50d8ae3SPaolo Bonzini static int mmu_pages_next(struct kvm_mmu_pages *pvec, 2043c50d8ae3SPaolo Bonzini struct mmu_page_path *parents, 2044c50d8ae3SPaolo Bonzini int i) 2045c50d8ae3SPaolo Bonzini { 2046c50d8ae3SPaolo Bonzini int n; 2047c50d8ae3SPaolo Bonzini 2048c50d8ae3SPaolo Bonzini for (n = i+1; n < pvec->nr; n++) { 2049c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp = pvec->page[n].sp; 2050c50d8ae3SPaolo Bonzini unsigned idx = pvec->page[n].idx; 2051c50d8ae3SPaolo Bonzini int level = sp->role.level; 2052c50d8ae3SPaolo Bonzini 2053c50d8ae3SPaolo Bonzini parents->idx[level-1] = idx; 20543bae0459SSean Christopherson if (level == PG_LEVEL_4K) 2055c50d8ae3SPaolo Bonzini break; 2056c50d8ae3SPaolo Bonzini 2057c50d8ae3SPaolo Bonzini parents->parent[level-2] = sp; 2058c50d8ae3SPaolo Bonzini } 2059c50d8ae3SPaolo Bonzini 2060c50d8ae3SPaolo Bonzini return n; 2061c50d8ae3SPaolo Bonzini } 2062c50d8ae3SPaolo Bonzini 2063c50d8ae3SPaolo Bonzini static int mmu_pages_first(struct kvm_mmu_pages *pvec, 2064c50d8ae3SPaolo Bonzini struct mmu_page_path *parents) 2065c50d8ae3SPaolo Bonzini { 2066c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2067c50d8ae3SPaolo Bonzini int level; 2068c50d8ae3SPaolo Bonzini 2069c50d8ae3SPaolo Bonzini if (pvec->nr == 0) 2070c50d8ae3SPaolo Bonzini return 0; 2071c50d8ae3SPaolo Bonzini 2072c50d8ae3SPaolo Bonzini WARN_ON(pvec->page[0].idx != INVALID_INDEX); 2073c50d8ae3SPaolo Bonzini 2074c50d8ae3SPaolo Bonzini sp = pvec->page[0].sp; 2075c50d8ae3SPaolo Bonzini level = sp->role.level; 20763bae0459SSean Christopherson WARN_ON(level == PG_LEVEL_4K); 2077c50d8ae3SPaolo Bonzini 2078c50d8ae3SPaolo Bonzini parents->parent[level-2] = sp; 2079c50d8ae3SPaolo Bonzini 2080c50d8ae3SPaolo Bonzini /* Also set up a sentinel. Further entries in pvec are all 2081c50d8ae3SPaolo Bonzini * children of sp, so this element is never overwritten. 2082c50d8ae3SPaolo Bonzini */ 2083c50d8ae3SPaolo Bonzini parents->parent[level-1] = NULL; 2084c50d8ae3SPaolo Bonzini return mmu_pages_next(pvec, parents, 0); 2085c50d8ae3SPaolo Bonzini } 2086c50d8ae3SPaolo Bonzini 2087c50d8ae3SPaolo Bonzini static void mmu_pages_clear_parents(struct mmu_page_path *parents) 2088c50d8ae3SPaolo Bonzini { 2089c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2090c50d8ae3SPaolo Bonzini unsigned int level = 0; 2091c50d8ae3SPaolo Bonzini 2092c50d8ae3SPaolo Bonzini do { 2093c50d8ae3SPaolo Bonzini unsigned int idx = parents->idx[level]; 2094c50d8ae3SPaolo Bonzini sp = parents->parent[level]; 2095c50d8ae3SPaolo Bonzini if (!sp) 2096c50d8ae3SPaolo Bonzini return; 2097c50d8ae3SPaolo Bonzini 2098c50d8ae3SPaolo Bonzini WARN_ON(idx == INVALID_INDEX); 2099c50d8ae3SPaolo Bonzini clear_unsync_child_bit(sp, idx); 2100c50d8ae3SPaolo Bonzini level++; 2101c50d8ae3SPaolo Bonzini } while (!sp->unsync_children); 2102c50d8ae3SPaolo Bonzini } 2103c50d8ae3SPaolo Bonzini 210465855ed8SLai Jiangshan static int mmu_sync_children(struct kvm_vcpu *vcpu, 210565855ed8SLai Jiangshan struct kvm_mmu_page *parent, bool can_yield) 2106c50d8ae3SPaolo Bonzini { 2107c50d8ae3SPaolo Bonzini int i; 2108c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2109c50d8ae3SPaolo Bonzini struct mmu_page_path parents; 2110c50d8ae3SPaolo Bonzini struct kvm_mmu_pages pages; 2111c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 2112c50d8ae3SPaolo Bonzini bool flush = false; 2113c50d8ae3SPaolo Bonzini 2114c50d8ae3SPaolo Bonzini while (mmu_unsync_walk(parent, &pages)) { 2115c50d8ae3SPaolo Bonzini bool protected = false; 2116c50d8ae3SPaolo Bonzini 2117c50d8ae3SPaolo Bonzini for_each_sp(pages, sp, parents, i) 2118cf48f9e2SDavid Matlack protected |= kvm_vcpu_write_protect_gfn(vcpu, sp->gfn); 2119c50d8ae3SPaolo Bonzini 2120c50d8ae3SPaolo Bonzini if (protected) { 21215591c069SLai Jiangshan kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, true); 2122c50d8ae3SPaolo Bonzini flush = false; 2123c50d8ae3SPaolo Bonzini } 2124c50d8ae3SPaolo Bonzini 2125c50d8ae3SPaolo Bonzini for_each_sp(pages, sp, parents, i) { 2126479a1efcSSean Christopherson kvm_unlink_unsync_page(vcpu->kvm, sp); 21278d5678a7SHou Wenlong flush |= kvm_sync_page(vcpu, sp, &invalid_list) > 0; 2128c50d8ae3SPaolo Bonzini mmu_pages_clear_parents(&parents); 2129c50d8ae3SPaolo Bonzini } 2130531810caSBen Gardon if (need_resched() || rwlock_needbreak(&vcpu->kvm->mmu_lock)) { 2131c3e5e415SLai Jiangshan kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, flush); 213265855ed8SLai Jiangshan if (!can_yield) { 213365855ed8SLai Jiangshan kvm_make_request(KVM_REQ_MMU_SYNC, vcpu); 213465855ed8SLai Jiangshan return -EINTR; 213565855ed8SLai Jiangshan } 213665855ed8SLai Jiangshan 2137531810caSBen Gardon cond_resched_rwlock_write(&vcpu->kvm->mmu_lock); 2138c50d8ae3SPaolo Bonzini flush = false; 2139c50d8ae3SPaolo Bonzini } 2140c50d8ae3SPaolo Bonzini } 2141c50d8ae3SPaolo Bonzini 2142c3e5e415SLai Jiangshan kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, flush); 214365855ed8SLai Jiangshan return 0; 2144c50d8ae3SPaolo Bonzini } 2145c50d8ae3SPaolo Bonzini 2146c50d8ae3SPaolo Bonzini static void __clear_sp_write_flooding_count(struct kvm_mmu_page *sp) 2147c50d8ae3SPaolo Bonzini { 2148c50d8ae3SPaolo Bonzini atomic_set(&sp->write_flooding_count, 0); 2149c50d8ae3SPaolo Bonzini } 2150c50d8ae3SPaolo Bonzini 2151c50d8ae3SPaolo Bonzini static void clear_sp_write_flooding_count(u64 *spte) 2152c50d8ae3SPaolo Bonzini { 215357354682SSean Christopherson __clear_sp_write_flooding_count(sptep_to_sp(spte)); 2154c50d8ae3SPaolo Bonzini } 2155c50d8ae3SPaolo Bonzini 2156cbd858b1SDavid Matlack /* 2157cbd858b1SDavid Matlack * The vCPU is required when finding indirect shadow pages; the shadow 2158cbd858b1SDavid Matlack * page may already exist and syncing it needs the vCPU pointer in 2159cbd858b1SDavid Matlack * order to read guest page tables. Direct shadow pages are never 2160cbd858b1SDavid Matlack * unsync, thus @vcpu can be NULL if @role.direct is true. 2161cbd858b1SDavid Matlack */ 21623cc736b3SDavid Matlack static struct kvm_mmu_page *kvm_mmu_find_shadow_page(struct kvm *kvm, 21633cc736b3SDavid Matlack struct kvm_vcpu *vcpu, 216494c81364SDavid Matlack gfn_t gfn, 216594c81364SDavid Matlack struct hlist_head *sp_list, 21662e65e842SDavid Matlack union kvm_mmu_page_role role) 2167c50d8ae3SPaolo Bonzini { 2168c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 21698d5678a7SHou Wenlong int ret; 2170c50d8ae3SPaolo Bonzini int collisions = 0; 2171c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 2172c50d8ae3SPaolo Bonzini 21733cc736b3SDavid Matlack for_each_valid_sp(kvm, sp, sp_list) { 2174c50d8ae3SPaolo Bonzini if (sp->gfn != gfn) { 2175c50d8ae3SPaolo Bonzini collisions++; 2176c50d8ae3SPaolo Bonzini continue; 2177c50d8ae3SPaolo Bonzini } 2178c50d8ae3SPaolo Bonzini 2179ddc16abbSSean Christopherson if (sp->role.word != role.word) { 2180ddc16abbSSean Christopherson /* 2181ddc16abbSSean Christopherson * If the guest is creating an upper-level page, zap 2182ddc16abbSSean Christopherson * unsync pages for the same gfn. While it's possible 2183ddc16abbSSean Christopherson * the guest is using recursive page tables, in all 2184ddc16abbSSean Christopherson * likelihood the guest has stopped using the unsync 2185ddc16abbSSean Christopherson * page and is installing a completely unrelated page. 2186ddc16abbSSean Christopherson * Unsync pages must not be left as is, because the new 2187ddc16abbSSean Christopherson * upper-level page will be write-protected. 2188ddc16abbSSean Christopherson */ 21892e65e842SDavid Matlack if (role.level > PG_LEVEL_4K && sp->unsync) 21903cc736b3SDavid Matlack kvm_mmu_prepare_zap_page(kvm, sp, 2191ddc16abbSSean Christopherson &invalid_list); 2192c50d8ae3SPaolo Bonzini continue; 2193ddc16abbSSean Christopherson } 2194c50d8ae3SPaolo Bonzini 2195bb924ca6SDavid Matlack /* unsync and write-flooding only apply to indirect SPs. */ 2196bb924ca6SDavid Matlack if (sp->role.direct) 219794c81364SDavid Matlack goto out; 2198fb58a9c3SSean Christopherson 2199c50d8ae3SPaolo Bonzini if (sp->unsync) { 2200cbd858b1SDavid Matlack if (KVM_BUG_ON(!vcpu, kvm)) 2201cbd858b1SDavid Matlack break; 2202cbd858b1SDavid Matlack 220307dc4f35SSean Christopherson /* 2204479a1efcSSean Christopherson * The page is good, but is stale. kvm_sync_page does 220507dc4f35SSean Christopherson * get the latest guest state, but (unlike mmu_unsync_children) 220607dc4f35SSean Christopherson * it doesn't write-protect the page or mark it synchronized! 220707dc4f35SSean Christopherson * This way the validity of the mapping is ensured, but the 220807dc4f35SSean Christopherson * overhead of write protection is not incurred until the 220907dc4f35SSean Christopherson * guest invalidates the TLB mapping. This allows multiple 221007dc4f35SSean Christopherson * SPs for a single gfn to be unsync. 221107dc4f35SSean Christopherson * 221207dc4f35SSean Christopherson * If the sync fails, the page is zapped. If so, break 221307dc4f35SSean Christopherson * in order to rebuild it. 2214c50d8ae3SPaolo Bonzini */ 22158d5678a7SHou Wenlong ret = kvm_sync_page(vcpu, sp, &invalid_list); 22168d5678a7SHou Wenlong if (ret < 0) 2217c50d8ae3SPaolo Bonzini break; 2218c50d8ae3SPaolo Bonzini 2219c50d8ae3SPaolo Bonzini WARN_ON(!list_empty(&invalid_list)); 22208d5678a7SHou Wenlong if (ret > 0) 22213cc736b3SDavid Matlack kvm_flush_remote_tlbs(kvm); 2222c50d8ae3SPaolo Bonzini } 2223c50d8ae3SPaolo Bonzini 2224c50d8ae3SPaolo Bonzini __clear_sp_write_flooding_count(sp); 2225fb58a9c3SSean Christopherson 2226c50d8ae3SPaolo Bonzini goto out; 2227c50d8ae3SPaolo Bonzini } 2228c50d8ae3SPaolo Bonzini 222994c81364SDavid Matlack sp = NULL; 22303cc736b3SDavid Matlack ++kvm->stat.mmu_cache_miss; 2231c50d8ae3SPaolo Bonzini 223294c81364SDavid Matlack out: 22333cc736b3SDavid Matlack kvm_mmu_commit_zap_page(kvm, &invalid_list); 223494c81364SDavid Matlack 22353cc736b3SDavid Matlack if (collisions > kvm->stat.max_mmu_page_hash_collisions) 22363cc736b3SDavid Matlack kvm->stat.max_mmu_page_hash_collisions = collisions; 223794c81364SDavid Matlack return sp; 223894c81364SDavid Matlack } 223994c81364SDavid Matlack 22402f8b1b53SDavid Matlack /* Caches used when allocating a new shadow page. */ 22412f8b1b53SDavid Matlack struct shadow_page_caches { 22422f8b1b53SDavid Matlack struct kvm_mmu_memory_cache *page_header_cache; 22432f8b1b53SDavid Matlack struct kvm_mmu_memory_cache *shadow_page_cache; 22446a97575dSDavid Matlack struct kvm_mmu_memory_cache *shadowed_info_cache; 22452f8b1b53SDavid Matlack }; 22462f8b1b53SDavid Matlack 2247336081fbSDavid Matlack static struct kvm_mmu_page *kvm_mmu_alloc_shadow_page(struct kvm *kvm, 22482f8b1b53SDavid Matlack struct shadow_page_caches *caches, 224994c81364SDavid Matlack gfn_t gfn, 225094c81364SDavid Matlack struct hlist_head *sp_list, 225194c81364SDavid Matlack union kvm_mmu_page_role role) 225294c81364SDavid Matlack { 2253c306aec8SDavid Matlack struct kvm_mmu_page *sp; 2254c306aec8SDavid Matlack 22552f8b1b53SDavid Matlack sp = kvm_mmu_memory_cache_alloc(caches->page_header_cache); 22562f8b1b53SDavid Matlack sp->spt = kvm_mmu_memory_cache_alloc(caches->shadow_page_cache); 2257c306aec8SDavid Matlack if (!role.direct) 22586a97575dSDavid Matlack sp->shadowed_translation = kvm_mmu_memory_cache_alloc(caches->shadowed_info_cache); 2259c306aec8SDavid Matlack 2260c306aec8SDavid Matlack set_page_private(virt_to_page(sp->spt), (unsigned long)sp); 2261c306aec8SDavid Matlack 226255c510e2SSean Christopherson INIT_LIST_HEAD(&sp->possible_nx_huge_page_link); 2263428e9216SSean Christopherson 2264c306aec8SDavid Matlack /* 2265c306aec8SDavid Matlack * active_mmu_pages must be a FIFO list, as kvm_zap_obsolete_pages() 2266c306aec8SDavid Matlack * depends on valid pages being added to the head of the list. See 2267c306aec8SDavid Matlack * comments in kvm_zap_obsolete_pages(). 2268c306aec8SDavid Matlack */ 2269336081fbSDavid Matlack sp->mmu_valid_gen = kvm->arch.mmu_valid_gen; 2270336081fbSDavid Matlack list_add(&sp->link, &kvm->arch.active_mmu_pages); 227143a063caSYosry Ahmed kvm_account_mmu_page(kvm, sp); 2272c50d8ae3SPaolo Bonzini 2273c50d8ae3SPaolo Bonzini sp->gfn = gfn; 2274c50d8ae3SPaolo Bonzini sp->role = role; 2275ac101b7cSSean Christopherson hlist_add_head(&sp->hash_link, sp_list); 2276be911771SDavid Matlack if (sp_has_gptes(sp)) 2277336081fbSDavid Matlack account_shadowed(kvm, sp); 2278ddc16abbSSean Christopherson 227994c81364SDavid Matlack return sp; 228094c81364SDavid Matlack } 228194c81364SDavid Matlack 2282cbd858b1SDavid Matlack /* Note, @vcpu may be NULL if @role.direct is true; see kvm_mmu_find_shadow_page. */ 22833cc736b3SDavid Matlack static struct kvm_mmu_page *__kvm_mmu_get_shadow_page(struct kvm *kvm, 22843cc736b3SDavid Matlack struct kvm_vcpu *vcpu, 22852f8b1b53SDavid Matlack struct shadow_page_caches *caches, 228687654643SDavid Matlack gfn_t gfn, 228794c81364SDavid Matlack union kvm_mmu_page_role role) 228894c81364SDavid Matlack { 228994c81364SDavid Matlack struct hlist_head *sp_list; 229094c81364SDavid Matlack struct kvm_mmu_page *sp; 229194c81364SDavid Matlack bool created = false; 229294c81364SDavid Matlack 22933cc736b3SDavid Matlack sp_list = &kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]; 229494c81364SDavid Matlack 22953cc736b3SDavid Matlack sp = kvm_mmu_find_shadow_page(kvm, vcpu, gfn, sp_list, role); 229694c81364SDavid Matlack if (!sp) { 229794c81364SDavid Matlack created = true; 22983cc736b3SDavid Matlack sp = kvm_mmu_alloc_shadow_page(kvm, caches, gfn, sp_list, role); 229994c81364SDavid Matlack } 230094c81364SDavid Matlack 230194c81364SDavid Matlack trace_kvm_mmu_get_page(sp, created); 2302c50d8ae3SPaolo Bonzini return sp; 2303c50d8ae3SPaolo Bonzini } 2304c50d8ae3SPaolo Bonzini 23052f8b1b53SDavid Matlack static struct kvm_mmu_page *kvm_mmu_get_shadow_page(struct kvm_vcpu *vcpu, 23062f8b1b53SDavid Matlack gfn_t gfn, 23072f8b1b53SDavid Matlack union kvm_mmu_page_role role) 23082f8b1b53SDavid Matlack { 23092f8b1b53SDavid Matlack struct shadow_page_caches caches = { 23102f8b1b53SDavid Matlack .page_header_cache = &vcpu->arch.mmu_page_header_cache, 23112f8b1b53SDavid Matlack .shadow_page_cache = &vcpu->arch.mmu_shadow_page_cache, 23126a97575dSDavid Matlack .shadowed_info_cache = &vcpu->arch.mmu_shadowed_info_cache, 23132f8b1b53SDavid Matlack }; 23142f8b1b53SDavid Matlack 23153cc736b3SDavid Matlack return __kvm_mmu_get_shadow_page(vcpu->kvm, vcpu, &caches, gfn, role); 23162f8b1b53SDavid Matlack } 23172f8b1b53SDavid Matlack 231839944ab9SSean Christopherson static union kvm_mmu_page_role kvm_mmu_child_role(u64 *sptep, bool direct, 231939944ab9SSean Christopherson unsigned int access) 23202e65e842SDavid Matlack { 23212e65e842SDavid Matlack struct kvm_mmu_page *parent_sp = sptep_to_sp(sptep); 23222e65e842SDavid Matlack union kvm_mmu_page_role role; 23232e65e842SDavid Matlack 23242e65e842SDavid Matlack role = parent_sp->role; 23252e65e842SDavid Matlack role.level--; 23262e65e842SDavid Matlack role.access = access; 23272e65e842SDavid Matlack role.direct = direct; 23282e65e842SDavid Matlack role.passthrough = 0; 23292e65e842SDavid Matlack 23302e65e842SDavid Matlack /* 23312e65e842SDavid Matlack * If the guest has 4-byte PTEs then that means it's using 32-bit, 23322e65e842SDavid Matlack * 2-level, non-PAE paging. KVM shadows such guests with PAE paging 23332e65e842SDavid Matlack * (i.e. 8-byte PTEs). The difference in PTE size means that KVM must 23342e65e842SDavid Matlack * shadow each guest page table with multiple shadow page tables, which 23352e65e842SDavid Matlack * requires extra bookkeeping in the role. 23362e65e842SDavid Matlack * 23372e65e842SDavid Matlack * Specifically, to shadow the guest's page directory (which covers a 23382e65e842SDavid Matlack * 4GiB address space), KVM uses 4 PAE page directories, each mapping 23392e65e842SDavid Matlack * 1GiB of the address space. @role.quadrant encodes which quarter of 23402e65e842SDavid Matlack * the address space each maps. 23412e65e842SDavid Matlack * 23422e65e842SDavid Matlack * To shadow the guest's page tables (which each map a 4MiB region), KVM 23432e65e842SDavid Matlack * uses 2 PAE page tables, each mapping a 2MiB region. For these, 23442e65e842SDavid Matlack * @role.quadrant encodes which half of the region they map. 23452e65e842SDavid Matlack * 234639944ab9SSean Christopherson * Concretely, a 4-byte PDE consumes bits 31:22, while an 8-byte PDE 234739944ab9SSean Christopherson * consumes bits 29:21. To consume bits 31:30, KVM's uses 4 shadow 234839944ab9SSean Christopherson * PDPTEs; those 4 PAE page directories are pre-allocated and their 234939944ab9SSean Christopherson * quadrant is assigned in mmu_alloc_root(). A 4-byte PTE consumes 235039944ab9SSean Christopherson * bits 21:12, while an 8-byte PTE consumes bits 20:12. To consume 235139944ab9SSean Christopherson * bit 21 in the PTE (the child here), KVM propagates that bit to the 235239944ab9SSean Christopherson * quadrant, i.e. sets quadrant to '0' or '1'. The parent 8-byte PDE 235339944ab9SSean Christopherson * covers bit 21 (see above), thus the quadrant is calculated from the 235439944ab9SSean Christopherson * _least_ significant bit of the PDE index. 23552e65e842SDavid Matlack */ 23562e65e842SDavid Matlack if (role.has_4_byte_gpte) { 23572e65e842SDavid Matlack WARN_ON_ONCE(role.level != PG_LEVEL_4K); 235879e48cecSSean Christopherson role.quadrant = spte_index(sptep) & 1; 23592e65e842SDavid Matlack } 23602e65e842SDavid Matlack 23612e65e842SDavid Matlack return role; 23622e65e842SDavid Matlack } 23632e65e842SDavid Matlack 23642e65e842SDavid Matlack static struct kvm_mmu_page *kvm_mmu_get_child_sp(struct kvm_vcpu *vcpu, 23652e65e842SDavid Matlack u64 *sptep, gfn_t gfn, 23662e65e842SDavid Matlack bool direct, unsigned int access) 23672e65e842SDavid Matlack { 23682e65e842SDavid Matlack union kvm_mmu_page_role role; 23692e65e842SDavid Matlack 23700cd8dc73SPaolo Bonzini if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) 23710cd8dc73SPaolo Bonzini return ERR_PTR(-EEXIST); 23720cd8dc73SPaolo Bonzini 23732e65e842SDavid Matlack role = kvm_mmu_child_role(sptep, direct, access); 237487654643SDavid Matlack return kvm_mmu_get_shadow_page(vcpu, gfn, role); 23752e65e842SDavid Matlack } 23762e65e842SDavid Matlack 2377c50d8ae3SPaolo Bonzini static void shadow_walk_init_using_root(struct kvm_shadow_walk_iterator *iterator, 2378c50d8ae3SPaolo Bonzini struct kvm_vcpu *vcpu, hpa_t root, 2379c50d8ae3SPaolo Bonzini u64 addr) 2380c50d8ae3SPaolo Bonzini { 2381c50d8ae3SPaolo Bonzini iterator->addr = addr; 2382c50d8ae3SPaolo Bonzini iterator->shadow_addr = root; 2383a972e29cSPaolo Bonzini iterator->level = vcpu->arch.mmu->root_role.level; 2384c50d8ae3SPaolo Bonzini 238512ec33a7SLai Jiangshan if (iterator->level >= PT64_ROOT_4LEVEL && 23864d25502aSPaolo Bonzini vcpu->arch.mmu->cpu_role.base.level < PT64_ROOT_4LEVEL && 2387347a0d0dSPaolo Bonzini !vcpu->arch.mmu->root_role.direct) 238812ec33a7SLai Jiangshan iterator->level = PT32E_ROOT_LEVEL; 2389c50d8ae3SPaolo Bonzini 2390c50d8ae3SPaolo Bonzini if (iterator->level == PT32E_ROOT_LEVEL) { 2391c50d8ae3SPaolo Bonzini /* 2392c50d8ae3SPaolo Bonzini * prev_root is currently only used for 64-bit hosts. So only 2393c50d8ae3SPaolo Bonzini * the active root_hpa is valid here. 2394c50d8ae3SPaolo Bonzini */ 2395b9e5603cSPaolo Bonzini BUG_ON(root != vcpu->arch.mmu->root.hpa); 2396c50d8ae3SPaolo Bonzini 2397c50d8ae3SPaolo Bonzini iterator->shadow_addr 2398c50d8ae3SPaolo Bonzini = vcpu->arch.mmu->pae_root[(addr >> 30) & 3]; 23992ca3129eSSean Christopherson iterator->shadow_addr &= SPTE_BASE_ADDR_MASK; 2400c50d8ae3SPaolo Bonzini --iterator->level; 2401c50d8ae3SPaolo Bonzini if (!iterator->shadow_addr) 2402c50d8ae3SPaolo Bonzini iterator->level = 0; 2403c50d8ae3SPaolo Bonzini } 2404c50d8ae3SPaolo Bonzini } 2405c50d8ae3SPaolo Bonzini 2406c50d8ae3SPaolo Bonzini static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator, 2407c50d8ae3SPaolo Bonzini struct kvm_vcpu *vcpu, u64 addr) 2408c50d8ae3SPaolo Bonzini { 2409b9e5603cSPaolo Bonzini shadow_walk_init_using_root(iterator, vcpu, vcpu->arch.mmu->root.hpa, 2410c50d8ae3SPaolo Bonzini addr); 2411c50d8ae3SPaolo Bonzini } 2412c50d8ae3SPaolo Bonzini 2413c50d8ae3SPaolo Bonzini static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator) 2414c50d8ae3SPaolo Bonzini { 24153bae0459SSean Christopherson if (iterator->level < PG_LEVEL_4K) 2416c50d8ae3SPaolo Bonzini return false; 2417c50d8ae3SPaolo Bonzini 24182ca3129eSSean Christopherson iterator->index = SPTE_INDEX(iterator->addr, iterator->level); 2419c50d8ae3SPaolo Bonzini iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index; 2420c50d8ae3SPaolo Bonzini return true; 2421c50d8ae3SPaolo Bonzini } 2422c50d8ae3SPaolo Bonzini 2423c50d8ae3SPaolo Bonzini static void __shadow_walk_next(struct kvm_shadow_walk_iterator *iterator, 2424c50d8ae3SPaolo Bonzini u64 spte) 2425c50d8ae3SPaolo Bonzini { 24263e44dce4SLai Jiangshan if (!is_shadow_present_pte(spte) || is_last_spte(spte, iterator->level)) { 2427c50d8ae3SPaolo Bonzini iterator->level = 0; 2428c50d8ae3SPaolo Bonzini return; 2429c50d8ae3SPaolo Bonzini } 2430c50d8ae3SPaolo Bonzini 24312ca3129eSSean Christopherson iterator->shadow_addr = spte & SPTE_BASE_ADDR_MASK; 2432c50d8ae3SPaolo Bonzini --iterator->level; 2433c50d8ae3SPaolo Bonzini } 2434c50d8ae3SPaolo Bonzini 2435c50d8ae3SPaolo Bonzini static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator) 2436c50d8ae3SPaolo Bonzini { 2437c50d8ae3SPaolo Bonzini __shadow_walk_next(iterator, *iterator->sptep); 2438c50d8ae3SPaolo Bonzini } 2439c50d8ae3SPaolo Bonzini 24400cd8dc73SPaolo Bonzini static void __link_shadow_page(struct kvm *kvm, 24410cd8dc73SPaolo Bonzini struct kvm_mmu_memory_cache *cache, u64 *sptep, 244203787394SPaolo Bonzini struct kvm_mmu_page *sp, bool flush) 2443c50d8ae3SPaolo Bonzini { 2444c50d8ae3SPaolo Bonzini u64 spte; 2445c50d8ae3SPaolo Bonzini 2446c50d8ae3SPaolo Bonzini BUILD_BUG_ON(VMX_EPT_WRITABLE_MASK != PT_WRITABLE_MASK); 2447c50d8ae3SPaolo Bonzini 24480cd8dc73SPaolo Bonzini /* 24490cd8dc73SPaolo Bonzini * If an SPTE is present already, it must be a leaf and therefore 245003787394SPaolo Bonzini * a large one. Drop it, and flush the TLB if needed, before 245103787394SPaolo Bonzini * installing sp. 24520cd8dc73SPaolo Bonzini */ 24530cd8dc73SPaolo Bonzini if (is_shadow_present_pte(*sptep)) 245403787394SPaolo Bonzini drop_large_spte(kvm, sptep, flush); 24550cd8dc73SPaolo Bonzini 2456cc4674d0SBen Gardon spte = make_nonleaf_spte(sp->spt, sp_ad_disabled(sp)); 2457c50d8ae3SPaolo Bonzini 2458c50d8ae3SPaolo Bonzini mmu_spte_set(sptep, spte); 2459c50d8ae3SPaolo Bonzini 24602ff9039aSDavid Matlack mmu_page_add_parent_pte(cache, sp, sptep); 2461c50d8ae3SPaolo Bonzini 2462c4a48868SLai Jiangshan /* 2463c4a48868SLai Jiangshan * The non-direct sub-pagetable must be updated before linking. For 2464c4a48868SLai Jiangshan * L1 sp, the pagetable is updated via kvm_sync_page() in 2465c4a48868SLai Jiangshan * kvm_mmu_find_shadow_page() without write-protecting the gfn, 2466c4a48868SLai Jiangshan * so sp->unsync can be true or false. For higher level non-direct 2467c4a48868SLai Jiangshan * sp, the pagetable is updated/synced via mmu_sync_children() in 2468c4a48868SLai Jiangshan * FNAME(fetch)(), so sp->unsync_children can only be false. 2469c4a48868SLai Jiangshan * WARN_ON_ONCE() if anything happens unexpectedly. 2470c4a48868SLai Jiangshan */ 2471c4a48868SLai Jiangshan if (WARN_ON_ONCE(sp->unsync_children) || sp->unsync) 2472c50d8ae3SPaolo Bonzini mark_unsync(sptep); 2473c50d8ae3SPaolo Bonzini } 2474c50d8ae3SPaolo Bonzini 24752ff9039aSDavid Matlack static void link_shadow_page(struct kvm_vcpu *vcpu, u64 *sptep, 24762ff9039aSDavid Matlack struct kvm_mmu_page *sp) 24772ff9039aSDavid Matlack { 247803787394SPaolo Bonzini __link_shadow_page(vcpu->kvm, &vcpu->arch.mmu_pte_list_desc_cache, sptep, sp, true); 24792ff9039aSDavid Matlack } 24802ff9039aSDavid Matlack 2481c50d8ae3SPaolo Bonzini static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep, 2482c50d8ae3SPaolo Bonzini unsigned direct_access) 2483c50d8ae3SPaolo Bonzini { 2484c50d8ae3SPaolo Bonzini if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) { 2485c50d8ae3SPaolo Bonzini struct kvm_mmu_page *child; 2486c50d8ae3SPaolo Bonzini 2487c50d8ae3SPaolo Bonzini /* 2488c50d8ae3SPaolo Bonzini * For the direct sp, if the guest pte's dirty bit 2489c50d8ae3SPaolo Bonzini * changed form clean to dirty, it will corrupt the 2490c50d8ae3SPaolo Bonzini * sp's access: allow writable in the read-only sp, 2491c50d8ae3SPaolo Bonzini * so we should update the spte at this point to get 2492c50d8ae3SPaolo Bonzini * a new sp with the correct access. 2493c50d8ae3SPaolo Bonzini */ 24945e3edd7eSSean Christopherson child = spte_to_child_sp(*sptep); 2495c50d8ae3SPaolo Bonzini if (child->role.access == direct_access) 2496c50d8ae3SPaolo Bonzini return; 2497c50d8ae3SPaolo Bonzini 2498c50d8ae3SPaolo Bonzini drop_parent_pte(child, sptep); 24993cdf9374SHou Wenlong kvm_flush_remote_tlbs_sptep(vcpu->kvm, sptep); 2500c50d8ae3SPaolo Bonzini } 2501c50d8ae3SPaolo Bonzini } 2502c50d8ae3SPaolo Bonzini 25032de4085cSBen Gardon /* Returns the number of zapped non-leaf child shadow pages. */ 25042de4085cSBen Gardon static int mmu_page_zap_pte(struct kvm *kvm, struct kvm_mmu_page *sp, 25052de4085cSBen Gardon u64 *spte, struct list_head *invalid_list) 2506c50d8ae3SPaolo Bonzini { 2507c50d8ae3SPaolo Bonzini u64 pte; 2508c50d8ae3SPaolo Bonzini struct kvm_mmu_page *child; 2509c50d8ae3SPaolo Bonzini 2510c50d8ae3SPaolo Bonzini pte = *spte; 2511c50d8ae3SPaolo Bonzini if (is_shadow_present_pte(pte)) { 2512c50d8ae3SPaolo Bonzini if (is_last_spte(pte, sp->role.level)) { 2513c50d8ae3SPaolo Bonzini drop_spte(kvm, spte); 2514c50d8ae3SPaolo Bonzini } else { 25155e3edd7eSSean Christopherson child = spte_to_child_sp(pte); 2516c50d8ae3SPaolo Bonzini drop_parent_pte(child, spte); 25172de4085cSBen Gardon 25182de4085cSBen Gardon /* 25192de4085cSBen Gardon * Recursively zap nested TDP SPs, parentless SPs are 25202de4085cSBen Gardon * unlikely to be used again in the near future. This 25212de4085cSBen Gardon * avoids retaining a large number of stale nested SPs. 25222de4085cSBen Gardon */ 25232de4085cSBen Gardon if (tdp_enabled && invalid_list && 25242de4085cSBen Gardon child->role.guest_mode && !child->parent_ptes.val) 25252de4085cSBen Gardon return kvm_mmu_prepare_zap_page(kvm, child, 25262de4085cSBen Gardon invalid_list); 2527c50d8ae3SPaolo Bonzini } 2528ace569e0SSean Christopherson } else if (is_mmio_spte(pte)) { 2529c50d8ae3SPaolo Bonzini mmu_spte_clear_no_track(spte); 2530ace569e0SSean Christopherson } 25312de4085cSBen Gardon return 0; 2532c50d8ae3SPaolo Bonzini } 2533c50d8ae3SPaolo Bonzini 25342de4085cSBen Gardon static int kvm_mmu_page_unlink_children(struct kvm *kvm, 25352de4085cSBen Gardon struct kvm_mmu_page *sp, 25362de4085cSBen Gardon struct list_head *invalid_list) 2537c50d8ae3SPaolo Bonzini { 25382de4085cSBen Gardon int zapped = 0; 2539c50d8ae3SPaolo Bonzini unsigned i; 2540c50d8ae3SPaolo Bonzini 25412ca3129eSSean Christopherson for (i = 0; i < SPTE_ENT_PER_PAGE; ++i) 25422de4085cSBen Gardon zapped += mmu_page_zap_pte(kvm, sp, sp->spt + i, invalid_list); 25432de4085cSBen Gardon 25442de4085cSBen Gardon return zapped; 2545c50d8ae3SPaolo Bonzini } 2546c50d8ae3SPaolo Bonzini 254761827671SJinrong Liang static void kvm_mmu_unlink_parents(struct kvm_mmu_page *sp) 2548c50d8ae3SPaolo Bonzini { 2549c50d8ae3SPaolo Bonzini u64 *sptep; 2550c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 2551c50d8ae3SPaolo Bonzini 2552c50d8ae3SPaolo Bonzini while ((sptep = rmap_get_first(&sp->parent_ptes, &iter))) 2553c50d8ae3SPaolo Bonzini drop_parent_pte(sp, sptep); 2554c50d8ae3SPaolo Bonzini } 2555c50d8ae3SPaolo Bonzini 2556c50d8ae3SPaolo Bonzini static int mmu_zap_unsync_children(struct kvm *kvm, 2557c50d8ae3SPaolo Bonzini struct kvm_mmu_page *parent, 2558c50d8ae3SPaolo Bonzini struct list_head *invalid_list) 2559c50d8ae3SPaolo Bonzini { 2560c50d8ae3SPaolo Bonzini int i, zapped = 0; 2561c50d8ae3SPaolo Bonzini struct mmu_page_path parents; 2562c50d8ae3SPaolo Bonzini struct kvm_mmu_pages pages; 2563c50d8ae3SPaolo Bonzini 25643bae0459SSean Christopherson if (parent->role.level == PG_LEVEL_4K) 2565c50d8ae3SPaolo Bonzini return 0; 2566c50d8ae3SPaolo Bonzini 2567c50d8ae3SPaolo Bonzini while (mmu_unsync_walk(parent, &pages)) { 2568c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2569c50d8ae3SPaolo Bonzini 2570c50d8ae3SPaolo Bonzini for_each_sp(pages, sp, parents, i) { 2571c50d8ae3SPaolo Bonzini kvm_mmu_prepare_zap_page(kvm, sp, invalid_list); 2572c50d8ae3SPaolo Bonzini mmu_pages_clear_parents(&parents); 2573c50d8ae3SPaolo Bonzini zapped++; 2574c50d8ae3SPaolo Bonzini } 2575c50d8ae3SPaolo Bonzini } 2576c50d8ae3SPaolo Bonzini 2577c50d8ae3SPaolo Bonzini return zapped; 2578c50d8ae3SPaolo Bonzini } 2579c50d8ae3SPaolo Bonzini 2580c50d8ae3SPaolo Bonzini static bool __kvm_mmu_prepare_zap_page(struct kvm *kvm, 2581c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, 2582c50d8ae3SPaolo Bonzini struct list_head *invalid_list, 2583c50d8ae3SPaolo Bonzini int *nr_zapped) 2584c50d8ae3SPaolo Bonzini { 2585527d5cd7SSean Christopherson bool list_unstable, zapped_root = false; 2586c50d8ae3SPaolo Bonzini 258747b0c2e4SKazuki Takiguchi lockdep_assert_held_write(&kvm->mmu_lock); 2588c50d8ae3SPaolo Bonzini trace_kvm_mmu_prepare_zap_page(sp); 2589c50d8ae3SPaolo Bonzini ++kvm->stat.mmu_shadow_zapped; 2590c50d8ae3SPaolo Bonzini *nr_zapped = mmu_zap_unsync_children(kvm, sp, invalid_list); 25912de4085cSBen Gardon *nr_zapped += kvm_mmu_page_unlink_children(kvm, sp, invalid_list); 259261827671SJinrong Liang kvm_mmu_unlink_parents(sp); 2593c50d8ae3SPaolo Bonzini 2594c50d8ae3SPaolo Bonzini /* Zapping children means active_mmu_pages has become unstable. */ 2595c50d8ae3SPaolo Bonzini list_unstable = *nr_zapped; 2596c50d8ae3SPaolo Bonzini 2597767d8d8dSLai Jiangshan if (!sp->role.invalid && sp_has_gptes(sp)) 2598c50d8ae3SPaolo Bonzini unaccount_shadowed(kvm, sp); 2599c50d8ae3SPaolo Bonzini 2600c50d8ae3SPaolo Bonzini if (sp->unsync) 2601c50d8ae3SPaolo Bonzini kvm_unlink_unsync_page(kvm, sp); 2602c50d8ae3SPaolo Bonzini if (!sp->root_count) { 2603c50d8ae3SPaolo Bonzini /* Count self */ 2604c50d8ae3SPaolo Bonzini (*nr_zapped)++; 2605f95eec9bSSean Christopherson 2606f95eec9bSSean Christopherson /* 2607f95eec9bSSean Christopherson * Already invalid pages (previously active roots) are not on 2608f95eec9bSSean Christopherson * the active page list. See list_del() in the "else" case of 2609f95eec9bSSean Christopherson * !sp->root_count. 2610f95eec9bSSean Christopherson */ 2611f95eec9bSSean Christopherson if (sp->role.invalid) 2612f95eec9bSSean Christopherson list_add(&sp->link, invalid_list); 2613f95eec9bSSean Christopherson else 2614c50d8ae3SPaolo Bonzini list_move(&sp->link, invalid_list); 261543a063caSYosry Ahmed kvm_unaccount_mmu_page(kvm, sp); 2616c50d8ae3SPaolo Bonzini } else { 2617f95eec9bSSean Christopherson /* 2618f95eec9bSSean Christopherson * Remove the active root from the active page list, the root 2619f95eec9bSSean Christopherson * will be explicitly freed when the root_count hits zero. 2620f95eec9bSSean Christopherson */ 2621f95eec9bSSean Christopherson list_del(&sp->link); 2622c50d8ae3SPaolo Bonzini 2623c50d8ae3SPaolo Bonzini /* 2624c50d8ae3SPaolo Bonzini * Obsolete pages cannot be used on any vCPUs, see the comment 2625c50d8ae3SPaolo Bonzini * in kvm_mmu_zap_all_fast(). Note, is_obsolete_sp() also 2626c50d8ae3SPaolo Bonzini * treats invalid shadow pages as being obsolete. 2627c50d8ae3SPaolo Bonzini */ 2628527d5cd7SSean Christopherson zapped_root = !is_obsolete_sp(kvm, sp); 2629c50d8ae3SPaolo Bonzini } 2630c50d8ae3SPaolo Bonzini 263155c510e2SSean Christopherson if (sp->nx_huge_page_disallowed) 263255c510e2SSean Christopherson unaccount_nx_huge_page(kvm, sp); 2633c50d8ae3SPaolo Bonzini 2634c50d8ae3SPaolo Bonzini sp->role.invalid = 1; 2635527d5cd7SSean Christopherson 2636527d5cd7SSean Christopherson /* 2637527d5cd7SSean Christopherson * Make the request to free obsolete roots after marking the root 2638527d5cd7SSean Christopherson * invalid, otherwise other vCPUs may not see it as invalid. 2639527d5cd7SSean Christopherson */ 2640527d5cd7SSean Christopherson if (zapped_root) 2641527d5cd7SSean Christopherson kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_FREE_OBSOLETE_ROOTS); 2642c50d8ae3SPaolo Bonzini return list_unstable; 2643c50d8ae3SPaolo Bonzini } 2644c50d8ae3SPaolo Bonzini 2645c50d8ae3SPaolo Bonzini static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp, 2646c50d8ae3SPaolo Bonzini struct list_head *invalid_list) 2647c50d8ae3SPaolo Bonzini { 2648c50d8ae3SPaolo Bonzini int nr_zapped; 2649c50d8ae3SPaolo Bonzini 2650c50d8ae3SPaolo Bonzini __kvm_mmu_prepare_zap_page(kvm, sp, invalid_list, &nr_zapped); 2651c50d8ae3SPaolo Bonzini return nr_zapped; 2652c50d8ae3SPaolo Bonzini } 2653c50d8ae3SPaolo Bonzini 2654c50d8ae3SPaolo Bonzini static void kvm_mmu_commit_zap_page(struct kvm *kvm, 2655c50d8ae3SPaolo Bonzini struct list_head *invalid_list) 2656c50d8ae3SPaolo Bonzini { 2657c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, *nsp; 2658c50d8ae3SPaolo Bonzini 2659c50d8ae3SPaolo Bonzini if (list_empty(invalid_list)) 2660c50d8ae3SPaolo Bonzini return; 2661c50d8ae3SPaolo Bonzini 2662c50d8ae3SPaolo Bonzini /* 2663c50d8ae3SPaolo Bonzini * We need to make sure everyone sees our modifications to 2664c50d8ae3SPaolo Bonzini * the page tables and see changes to vcpu->mode here. The barrier 2665c50d8ae3SPaolo Bonzini * in the kvm_flush_remote_tlbs() achieves this. This pairs 2666c50d8ae3SPaolo Bonzini * with vcpu_enter_guest and walk_shadow_page_lockless_begin/end. 2667c50d8ae3SPaolo Bonzini * 2668c50d8ae3SPaolo Bonzini * In addition, kvm_flush_remote_tlbs waits for all vcpus to exit 2669c50d8ae3SPaolo Bonzini * guest mode and/or lockless shadow page table walks. 2670c50d8ae3SPaolo Bonzini */ 2671c50d8ae3SPaolo Bonzini kvm_flush_remote_tlbs(kvm); 2672c50d8ae3SPaolo Bonzini 2673c50d8ae3SPaolo Bonzini list_for_each_entry_safe(sp, nsp, invalid_list, link) { 2674c50d8ae3SPaolo Bonzini WARN_ON(!sp->role.invalid || sp->root_count); 267587654643SDavid Matlack kvm_mmu_free_shadow_page(sp); 2676c50d8ae3SPaolo Bonzini } 2677c50d8ae3SPaolo Bonzini } 2678c50d8ae3SPaolo Bonzini 26796b82ef2cSSean Christopherson static unsigned long kvm_mmu_zap_oldest_mmu_pages(struct kvm *kvm, 26806b82ef2cSSean Christopherson unsigned long nr_to_zap) 2681c50d8ae3SPaolo Bonzini { 26826b82ef2cSSean Christopherson unsigned long total_zapped = 0; 26836b82ef2cSSean Christopherson struct kvm_mmu_page *sp, *tmp; 2684ba7888ddSSean Christopherson LIST_HEAD(invalid_list); 26856b82ef2cSSean Christopherson bool unstable; 26866b82ef2cSSean Christopherson int nr_zapped; 2687c50d8ae3SPaolo Bonzini 2688c50d8ae3SPaolo Bonzini if (list_empty(&kvm->arch.active_mmu_pages)) 2689ba7888ddSSean Christopherson return 0; 2690c50d8ae3SPaolo Bonzini 26916b82ef2cSSean Christopherson restart: 26928fc51726SSean Christopherson list_for_each_entry_safe_reverse(sp, tmp, &kvm->arch.active_mmu_pages, link) { 26936b82ef2cSSean Christopherson /* 26946b82ef2cSSean Christopherson * Don't zap active root pages, the page itself can't be freed 26956b82ef2cSSean Christopherson * and zapping it will just force vCPUs to realloc and reload. 26966b82ef2cSSean Christopherson */ 26976b82ef2cSSean Christopherson if (sp->root_count) 26986b82ef2cSSean Christopherson continue; 26996b82ef2cSSean Christopherson 27006b82ef2cSSean Christopherson unstable = __kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list, 27016b82ef2cSSean Christopherson &nr_zapped); 27026b82ef2cSSean Christopherson total_zapped += nr_zapped; 27036b82ef2cSSean Christopherson if (total_zapped >= nr_to_zap) 2704ba7888ddSSean Christopherson break; 2705ba7888ddSSean Christopherson 27066b82ef2cSSean Christopherson if (unstable) 27076b82ef2cSSean Christopherson goto restart; 2708ba7888ddSSean Christopherson } 27096b82ef2cSSean Christopherson 27106b82ef2cSSean Christopherson kvm_mmu_commit_zap_page(kvm, &invalid_list); 27116b82ef2cSSean Christopherson 27126b82ef2cSSean Christopherson kvm->stat.mmu_recycled += total_zapped; 27136b82ef2cSSean Christopherson return total_zapped; 27146b82ef2cSSean Christopherson } 27156b82ef2cSSean Christopherson 2716afe8d7e6SSean Christopherson static inline unsigned long kvm_mmu_available_pages(struct kvm *kvm) 2717afe8d7e6SSean Christopherson { 2718afe8d7e6SSean Christopherson if (kvm->arch.n_max_mmu_pages > kvm->arch.n_used_mmu_pages) 2719afe8d7e6SSean Christopherson return kvm->arch.n_max_mmu_pages - 2720afe8d7e6SSean Christopherson kvm->arch.n_used_mmu_pages; 2721afe8d7e6SSean Christopherson 2722afe8d7e6SSean Christopherson return 0; 2723c50d8ae3SPaolo Bonzini } 2724c50d8ae3SPaolo Bonzini 2725ba7888ddSSean Christopherson static int make_mmu_pages_available(struct kvm_vcpu *vcpu) 2726ba7888ddSSean Christopherson { 27276b82ef2cSSean Christopherson unsigned long avail = kvm_mmu_available_pages(vcpu->kvm); 2728ba7888ddSSean Christopherson 27296b82ef2cSSean Christopherson if (likely(avail >= KVM_MIN_FREE_MMU_PAGES)) 2730ba7888ddSSean Christopherson return 0; 2731ba7888ddSSean Christopherson 27326b82ef2cSSean Christopherson kvm_mmu_zap_oldest_mmu_pages(vcpu->kvm, KVM_REFILL_PAGES - avail); 2733ba7888ddSSean Christopherson 27346e6ec584SSean Christopherson /* 27356e6ec584SSean Christopherson * Note, this check is intentionally soft, it only guarantees that one 27366e6ec584SSean Christopherson * page is available, while the caller may end up allocating as many as 27376e6ec584SSean Christopherson * four pages, e.g. for PAE roots or for 5-level paging. Temporarily 27386e6ec584SSean Christopherson * exceeding the (arbitrary by default) limit will not harm the host, 2739c4342633SIngo Molnar * being too aggressive may unnecessarily kill the guest, and getting an 27406e6ec584SSean Christopherson * exact count is far more trouble than it's worth, especially in the 27416e6ec584SSean Christopherson * page fault paths. 27426e6ec584SSean Christopherson */ 2743ba7888ddSSean Christopherson if (!kvm_mmu_available_pages(vcpu->kvm)) 2744ba7888ddSSean Christopherson return -ENOSPC; 2745ba7888ddSSean Christopherson return 0; 2746ba7888ddSSean Christopherson } 2747ba7888ddSSean Christopherson 2748c50d8ae3SPaolo Bonzini /* 2749c50d8ae3SPaolo Bonzini * Changing the number of mmu pages allocated to the vm 2750c50d8ae3SPaolo Bonzini * Note: if goal_nr_mmu_pages is too small, you will get dead lock 2751c50d8ae3SPaolo Bonzini */ 2752c50d8ae3SPaolo Bonzini void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned long goal_nr_mmu_pages) 2753c50d8ae3SPaolo Bonzini { 2754531810caSBen Gardon write_lock(&kvm->mmu_lock); 2755c50d8ae3SPaolo Bonzini 2756c50d8ae3SPaolo Bonzini if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) { 27576b82ef2cSSean Christopherson kvm_mmu_zap_oldest_mmu_pages(kvm, kvm->arch.n_used_mmu_pages - 27586b82ef2cSSean Christopherson goal_nr_mmu_pages); 2759c50d8ae3SPaolo Bonzini 2760c50d8ae3SPaolo Bonzini goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages; 2761c50d8ae3SPaolo Bonzini } 2762c50d8ae3SPaolo Bonzini 2763c50d8ae3SPaolo Bonzini kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages; 2764c50d8ae3SPaolo Bonzini 2765531810caSBen Gardon write_unlock(&kvm->mmu_lock); 2766c50d8ae3SPaolo Bonzini } 2767c50d8ae3SPaolo Bonzini 2768c50d8ae3SPaolo Bonzini int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn) 2769c50d8ae3SPaolo Bonzini { 2770c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2771c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 2772c50d8ae3SPaolo Bonzini int r; 2773c50d8ae3SPaolo Bonzini 2774c50d8ae3SPaolo Bonzini pgprintk("%s: looking for gfn %llx\n", __func__, gfn); 2775c50d8ae3SPaolo Bonzini r = 0; 2776531810caSBen Gardon write_lock(&kvm->mmu_lock); 2777767d8d8dSLai Jiangshan for_each_gfn_valid_sp_with_gptes(kvm, sp, gfn) { 2778c50d8ae3SPaolo Bonzini pgprintk("%s: gfn %llx role %x\n", __func__, gfn, 2779c50d8ae3SPaolo Bonzini sp->role.word); 2780c50d8ae3SPaolo Bonzini r = 1; 2781c50d8ae3SPaolo Bonzini kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list); 2782c50d8ae3SPaolo Bonzini } 2783c50d8ae3SPaolo Bonzini kvm_mmu_commit_zap_page(kvm, &invalid_list); 2784531810caSBen Gardon write_unlock(&kvm->mmu_lock); 2785c50d8ae3SPaolo Bonzini 2786c50d8ae3SPaolo Bonzini return r; 2787c50d8ae3SPaolo Bonzini } 278896ad91aeSSean Christopherson 278996ad91aeSSean Christopherson static int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva) 279096ad91aeSSean Christopherson { 279196ad91aeSSean Christopherson gpa_t gpa; 279296ad91aeSSean Christopherson int r; 279396ad91aeSSean Christopherson 2794347a0d0dSPaolo Bonzini if (vcpu->arch.mmu->root_role.direct) 279596ad91aeSSean Christopherson return 0; 279696ad91aeSSean Christopherson 279796ad91aeSSean Christopherson gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL); 279896ad91aeSSean Christopherson 279996ad91aeSSean Christopherson r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT); 280096ad91aeSSean Christopherson 280196ad91aeSSean Christopherson return r; 280296ad91aeSSean Christopherson } 2803c50d8ae3SPaolo Bonzini 28044d78d0b3SBen Gardon static void kvm_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp) 2805c50d8ae3SPaolo Bonzini { 2806c50d8ae3SPaolo Bonzini trace_kvm_mmu_unsync_page(sp); 28074d78d0b3SBen Gardon ++kvm->stat.mmu_unsync; 2808c50d8ae3SPaolo Bonzini sp->unsync = 1; 2809c50d8ae3SPaolo Bonzini 2810c50d8ae3SPaolo Bonzini kvm_mmu_mark_parents_unsync(sp); 2811c50d8ae3SPaolo Bonzini } 2812c50d8ae3SPaolo Bonzini 28130337f585SSean Christopherson /* 28140337f585SSean Christopherson * Attempt to unsync any shadow pages that can be reached by the specified gfn, 28150337f585SSean Christopherson * KVM is creating a writable mapping for said gfn. Returns 0 if all pages 28160337f585SSean Christopherson * were marked unsync (or if there is no shadow page), -EPERM if the SPTE must 28170337f585SSean Christopherson * be write-protected. 28180337f585SSean Christopherson */ 28198283e36aSBen Gardon int mmu_try_to_unsync_pages(struct kvm *kvm, const struct kvm_memory_slot *slot, 28202839180cSPaolo Bonzini gfn_t gfn, bool can_unsync, bool prefetch) 2821c50d8ae3SPaolo Bonzini { 2822c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 2823ce25681dSSean Christopherson bool locked = false; 2824c50d8ae3SPaolo Bonzini 28250337f585SSean Christopherson /* 28260337f585SSean Christopherson * Force write-protection if the page is being tracked. Note, the page 28270337f585SSean Christopherson * track machinery is used to write-protect upper-level shadow pages, 28280337f585SSean Christopherson * i.e. this guards the role.level == 4K assertion below! 28290337f585SSean Christopherson */ 28304d78d0b3SBen Gardon if (kvm_slot_page_track_is_active(kvm, slot, gfn, KVM_PAGE_TRACK_WRITE)) 28310337f585SSean Christopherson return -EPERM; 2832c50d8ae3SPaolo Bonzini 28330337f585SSean Christopherson /* 28340337f585SSean Christopherson * The page is not write-tracked, mark existing shadow pages unsync 28350337f585SSean Christopherson * unless KVM is synchronizing an unsync SP (can_unsync = false). In 28360337f585SSean Christopherson * that case, KVM must complete emulation of the guest TLB flush before 28370337f585SSean Christopherson * allowing shadow pages to become unsync (writable by the guest). 28380337f585SSean Christopherson */ 2839767d8d8dSLai Jiangshan for_each_gfn_valid_sp_with_gptes(kvm, sp, gfn) { 2840c50d8ae3SPaolo Bonzini if (!can_unsync) 28410337f585SSean Christopherson return -EPERM; 2842c50d8ae3SPaolo Bonzini 2843c50d8ae3SPaolo Bonzini if (sp->unsync) 2844c50d8ae3SPaolo Bonzini continue; 2845c50d8ae3SPaolo Bonzini 28462839180cSPaolo Bonzini if (prefetch) 2847f1c4a88cSLai Jiangshan return -EEXIST; 2848f1c4a88cSLai Jiangshan 2849ce25681dSSean Christopherson /* 2850ce25681dSSean Christopherson * TDP MMU page faults require an additional spinlock as they 2851ce25681dSSean Christopherson * run with mmu_lock held for read, not write, and the unsync 2852ce25681dSSean Christopherson * logic is not thread safe. Take the spinklock regardless of 2853ce25681dSSean Christopherson * the MMU type to avoid extra conditionals/parameters, there's 2854ce25681dSSean Christopherson * no meaningful penalty if mmu_lock is held for write. 2855ce25681dSSean Christopherson */ 2856ce25681dSSean Christopherson if (!locked) { 2857ce25681dSSean Christopherson locked = true; 28584d78d0b3SBen Gardon spin_lock(&kvm->arch.mmu_unsync_pages_lock); 2859ce25681dSSean Christopherson 2860ce25681dSSean Christopherson /* 2861ce25681dSSean Christopherson * Recheck after taking the spinlock, a different vCPU 2862ce25681dSSean Christopherson * may have since marked the page unsync. A false 2863ce25681dSSean Christopherson * positive on the unprotected check above is not 2864ce25681dSSean Christopherson * possible as clearing sp->unsync _must_ hold mmu_lock 2865ce25681dSSean Christopherson * for write, i.e. unsync cannot transition from 0->1 2866ce25681dSSean Christopherson * while this CPU holds mmu_lock for read (or write). 2867ce25681dSSean Christopherson */ 2868ce25681dSSean Christopherson if (READ_ONCE(sp->unsync)) 2869ce25681dSSean Christopherson continue; 2870ce25681dSSean Christopherson } 2871ce25681dSSean Christopherson 28723bae0459SSean Christopherson WARN_ON(sp->role.level != PG_LEVEL_4K); 28734d78d0b3SBen Gardon kvm_unsync_page(kvm, sp); 2874c50d8ae3SPaolo Bonzini } 2875ce25681dSSean Christopherson if (locked) 28764d78d0b3SBen Gardon spin_unlock(&kvm->arch.mmu_unsync_pages_lock); 2877c50d8ae3SPaolo Bonzini 2878c50d8ae3SPaolo Bonzini /* 2879c50d8ae3SPaolo Bonzini * We need to ensure that the marking of unsync pages is visible 2880c50d8ae3SPaolo Bonzini * before the SPTE is updated to allow writes because 2881c50d8ae3SPaolo Bonzini * kvm_mmu_sync_roots() checks the unsync flags without holding 2882c50d8ae3SPaolo Bonzini * the MMU lock and so can race with this. If the SPTE was updated 2883c50d8ae3SPaolo Bonzini * before the page had been marked as unsync-ed, something like the 2884c50d8ae3SPaolo Bonzini * following could happen: 2885c50d8ae3SPaolo Bonzini * 2886c50d8ae3SPaolo Bonzini * CPU 1 CPU 2 2887c50d8ae3SPaolo Bonzini * --------------------------------------------------------------------- 2888c50d8ae3SPaolo Bonzini * 1.2 Host updates SPTE 2889c50d8ae3SPaolo Bonzini * to be writable 2890c50d8ae3SPaolo Bonzini * 2.1 Guest writes a GPTE for GVA X. 2891c50d8ae3SPaolo Bonzini * (GPTE being in the guest page table shadowed 2892c50d8ae3SPaolo Bonzini * by the SP from CPU 1.) 2893c50d8ae3SPaolo Bonzini * This reads SPTE during the page table walk. 2894c50d8ae3SPaolo Bonzini * Since SPTE.W is read as 1, there is no 2895c50d8ae3SPaolo Bonzini * fault. 2896c50d8ae3SPaolo Bonzini * 2897c50d8ae3SPaolo Bonzini * 2.2 Guest issues TLB flush. 2898c50d8ae3SPaolo Bonzini * That causes a VM Exit. 2899c50d8ae3SPaolo Bonzini * 29000337f585SSean Christopherson * 2.3 Walking of unsync pages sees sp->unsync is 29010337f585SSean Christopherson * false and skips the page. 2902c50d8ae3SPaolo Bonzini * 2903c50d8ae3SPaolo Bonzini * 2.4 Guest accesses GVA X. 2904c50d8ae3SPaolo Bonzini * Since the mapping in the SP was not updated, 2905c50d8ae3SPaolo Bonzini * so the old mapping for GVA X incorrectly 2906c50d8ae3SPaolo Bonzini * gets used. 2907c50d8ae3SPaolo Bonzini * 1.1 Host marks SP 2908c50d8ae3SPaolo Bonzini * as unsync 2909c50d8ae3SPaolo Bonzini * (sp->unsync = true) 2910c50d8ae3SPaolo Bonzini * 2911c50d8ae3SPaolo Bonzini * The write barrier below ensures that 1.1 happens before 1.2 and thus 2912264d3dc1SLai Jiangshan * the situation in 2.4 does not arise. It pairs with the read barrier 2913264d3dc1SLai Jiangshan * in is_unsync_root(), placed between 2.1's load of SPTE.W and 2.3. 2914c50d8ae3SPaolo Bonzini */ 2915c50d8ae3SPaolo Bonzini smp_wmb(); 2916c50d8ae3SPaolo Bonzini 29170337f585SSean Christopherson return 0; 2918c50d8ae3SPaolo Bonzini } 2919c50d8ae3SPaolo Bonzini 29208a9f566aSDavid Matlack static int mmu_set_spte(struct kvm_vcpu *vcpu, struct kvm_memory_slot *slot, 29218a9f566aSDavid Matlack u64 *sptep, unsigned int pte_access, gfn_t gfn, 2922a12f4381SPaolo Bonzini kvm_pfn_t pfn, struct kvm_page_fault *fault) 2923799a4190SBen Gardon { 2924d786c778SPaolo Bonzini struct kvm_mmu_page *sp = sptep_to_sp(sptep); 2925eb5cd7ffSPaolo Bonzini int level = sp->role.level; 2926c50d8ae3SPaolo Bonzini int was_rmapped = 0; 2927c4371c2aSSean Christopherson int ret = RET_PF_FIXED; 2928c50d8ae3SPaolo Bonzini bool flush = false; 2929ad67e480SPaolo Bonzini bool wrprot; 2930d786c778SPaolo Bonzini u64 spte; 2931c50d8ae3SPaolo Bonzini 2932a12f4381SPaolo Bonzini /* Prefetching always gets a writable pfn. */ 2933a12f4381SPaolo Bonzini bool host_writable = !fault || fault->map_writable; 29342839180cSPaolo Bonzini bool prefetch = !fault || fault->prefetch; 2935a12f4381SPaolo Bonzini bool write_fault = fault && fault->write; 2936c50d8ae3SPaolo Bonzini 2937c50d8ae3SPaolo Bonzini pgprintk("%s: spte %llx write_fault %d gfn %llx\n", __func__, 2938c50d8ae3SPaolo Bonzini *sptep, write_fault, gfn); 2939c50d8ae3SPaolo Bonzini 2940a54aa15cSSean Christopherson if (unlikely(is_noslot_pfn(pfn))) { 29411075d41eSSean Christopherson vcpu->stat.pf_mmio_spte_created++; 2942a54aa15cSSean Christopherson mark_mmio_spte(vcpu, sptep, gfn, pte_access); 2943a54aa15cSSean Christopherson return RET_PF_EMULATE; 2944a54aa15cSSean Christopherson } 2945a54aa15cSSean Christopherson 2946c50d8ae3SPaolo Bonzini if (is_shadow_present_pte(*sptep)) { 2947c50d8ae3SPaolo Bonzini /* 2948c50d8ae3SPaolo Bonzini * If we overwrite a PTE page pointer with a 2MB PMD, unlink 2949c50d8ae3SPaolo Bonzini * the parent of the now unreachable PTE. 2950c50d8ae3SPaolo Bonzini */ 29513bae0459SSean Christopherson if (level > PG_LEVEL_4K && !is_large_pte(*sptep)) { 2952c50d8ae3SPaolo Bonzini struct kvm_mmu_page *child; 2953c50d8ae3SPaolo Bonzini u64 pte = *sptep; 2954c50d8ae3SPaolo Bonzini 29555e3edd7eSSean Christopherson child = spte_to_child_sp(pte); 2956c50d8ae3SPaolo Bonzini drop_parent_pte(child, sptep); 2957c50d8ae3SPaolo Bonzini flush = true; 2958c50d8ae3SPaolo Bonzini } else if (pfn != spte_to_pfn(*sptep)) { 2959c50d8ae3SPaolo Bonzini pgprintk("hfn old %llx new %llx\n", 2960c50d8ae3SPaolo Bonzini spte_to_pfn(*sptep), pfn); 2961c50d8ae3SPaolo Bonzini drop_spte(vcpu->kvm, sptep); 2962c50d8ae3SPaolo Bonzini flush = true; 2963c50d8ae3SPaolo Bonzini } else 2964c50d8ae3SPaolo Bonzini was_rmapped = 1; 2965c50d8ae3SPaolo Bonzini } 2966c50d8ae3SPaolo Bonzini 29672839180cSPaolo Bonzini wrprot = make_spte(vcpu, sp, slot, pte_access, gfn, pfn, *sptep, prefetch, 29687158bee4SPaolo Bonzini true, host_writable, &spte); 2969d786c778SPaolo Bonzini 2970d786c778SPaolo Bonzini if (*sptep == spte) { 2971d786c778SPaolo Bonzini ret = RET_PF_SPURIOUS; 2972d786c778SPaolo Bonzini } else { 2973d786c778SPaolo Bonzini flush |= mmu_spte_update(sptep, spte); 29745959ff4aSMaxim Levitsky trace_kvm_mmu_set_spte(level, gfn, sptep); 2975c50d8ae3SPaolo Bonzini } 2976c50d8ae3SPaolo Bonzini 2977ad67e480SPaolo Bonzini if (wrprot) { 2978c50d8ae3SPaolo Bonzini if (write_fault) 2979c50d8ae3SPaolo Bonzini ret = RET_PF_EMULATE; 2980c50d8ae3SPaolo Bonzini } 2981c50d8ae3SPaolo Bonzini 2982d786c778SPaolo Bonzini if (flush) 29834ad980aeSHou Wenlong kvm_flush_remote_tlbs_gfn(vcpu->kvm, gfn, level); 2984c50d8ae3SPaolo Bonzini 2985c50d8ae3SPaolo Bonzini pgprintk("%s: setting spte %llx\n", __func__, *sptep); 2986c50d8ae3SPaolo Bonzini 2987c50d8ae3SPaolo Bonzini if (!was_rmapped) { 2988d786c778SPaolo Bonzini WARN_ON_ONCE(ret == RET_PF_SPURIOUS); 29896a97575dSDavid Matlack rmap_add(vcpu, slot, sptep, gfn, pte_access); 29906a97575dSDavid Matlack } else { 29916a97575dSDavid Matlack /* Already rmapped but the pte_access bits may have changed. */ 299279e48cecSSean Christopherson kvm_mmu_page_set_access(sp, spte_index(sptep), pte_access); 2993c50d8ae3SPaolo Bonzini } 2994c50d8ae3SPaolo Bonzini 2995c50d8ae3SPaolo Bonzini return ret; 2996c50d8ae3SPaolo Bonzini } 2997c50d8ae3SPaolo Bonzini 2998c50d8ae3SPaolo Bonzini static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu, 2999c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, 3000c50d8ae3SPaolo Bonzini u64 *start, u64 *end) 3001c50d8ae3SPaolo Bonzini { 3002c50d8ae3SPaolo Bonzini struct page *pages[PTE_PREFETCH_NUM]; 3003c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot; 30040a2b64c5SBen Gardon unsigned int access = sp->role.access; 3005c50d8ae3SPaolo Bonzini int i, ret; 3006c50d8ae3SPaolo Bonzini gfn_t gfn; 3007c50d8ae3SPaolo Bonzini 300879e48cecSSean Christopherson gfn = kvm_mmu_page_get_gfn(sp, spte_index(start)); 3009c50d8ae3SPaolo Bonzini slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, access & ACC_WRITE_MASK); 3010c50d8ae3SPaolo Bonzini if (!slot) 3011c50d8ae3SPaolo Bonzini return -1; 3012c50d8ae3SPaolo Bonzini 3013c50d8ae3SPaolo Bonzini ret = gfn_to_page_many_atomic(slot, gfn, pages, end - start); 3014c50d8ae3SPaolo Bonzini if (ret <= 0) 3015c50d8ae3SPaolo Bonzini return -1; 3016c50d8ae3SPaolo Bonzini 3017c50d8ae3SPaolo Bonzini for (i = 0; i < ret; i++, gfn++, start++) { 30188a9f566aSDavid Matlack mmu_set_spte(vcpu, slot, start, access, gfn, 3019a12f4381SPaolo Bonzini page_to_pfn(pages[i]), NULL); 3020c50d8ae3SPaolo Bonzini put_page(pages[i]); 3021c50d8ae3SPaolo Bonzini } 3022c50d8ae3SPaolo Bonzini 3023c50d8ae3SPaolo Bonzini return 0; 3024c50d8ae3SPaolo Bonzini } 3025c50d8ae3SPaolo Bonzini 3026c50d8ae3SPaolo Bonzini static void __direct_pte_prefetch(struct kvm_vcpu *vcpu, 3027c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, u64 *sptep) 3028c50d8ae3SPaolo Bonzini { 3029c50d8ae3SPaolo Bonzini u64 *spte, *start = NULL; 3030c50d8ae3SPaolo Bonzini int i; 3031c50d8ae3SPaolo Bonzini 3032c50d8ae3SPaolo Bonzini WARN_ON(!sp->role.direct); 3033c50d8ae3SPaolo Bonzini 303479e48cecSSean Christopherson i = spte_index(sptep) & ~(PTE_PREFETCH_NUM - 1); 3035c50d8ae3SPaolo Bonzini spte = sp->spt + i; 3036c50d8ae3SPaolo Bonzini 3037c50d8ae3SPaolo Bonzini for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) { 3038c50d8ae3SPaolo Bonzini if (is_shadow_present_pte(*spte) || spte == sptep) { 3039c50d8ae3SPaolo Bonzini if (!start) 3040c50d8ae3SPaolo Bonzini continue; 3041c50d8ae3SPaolo Bonzini if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0) 3042c6cecc4bSSean Christopherson return; 3043c50d8ae3SPaolo Bonzini start = NULL; 3044c50d8ae3SPaolo Bonzini } else if (!start) 3045c50d8ae3SPaolo Bonzini start = spte; 3046c50d8ae3SPaolo Bonzini } 3047c6cecc4bSSean Christopherson if (start) 3048c6cecc4bSSean Christopherson direct_pte_prefetch_many(vcpu, sp, start, spte); 3049c50d8ae3SPaolo Bonzini } 3050c50d8ae3SPaolo Bonzini 3051c50d8ae3SPaolo Bonzini static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep) 3052c50d8ae3SPaolo Bonzini { 3053c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 3054c50d8ae3SPaolo Bonzini 305557354682SSean Christopherson sp = sptep_to_sp(sptep); 3056c50d8ae3SPaolo Bonzini 3057c50d8ae3SPaolo Bonzini /* 3058c50d8ae3SPaolo Bonzini * Without accessed bits, there's no way to distinguish between 3059c50d8ae3SPaolo Bonzini * actually accessed translations and prefetched, so disable pte 3060c50d8ae3SPaolo Bonzini * prefetch if accessed bits aren't available. 3061c50d8ae3SPaolo Bonzini */ 3062c50d8ae3SPaolo Bonzini if (sp_ad_disabled(sp)) 3063c50d8ae3SPaolo Bonzini return; 3064c50d8ae3SPaolo Bonzini 30653bae0459SSean Christopherson if (sp->role.level > PG_LEVEL_4K) 3066c50d8ae3SPaolo Bonzini return; 3067c50d8ae3SPaolo Bonzini 30684a42d848SDavid Stevens /* 30694a42d848SDavid Stevens * If addresses are being invalidated, skip prefetching to avoid 30704a42d848SDavid Stevens * accidentally prefetching those addresses. 30714a42d848SDavid Stevens */ 307220ec3ebdSChao Peng if (unlikely(vcpu->kvm->mmu_invalidate_in_progress)) 30734a42d848SDavid Stevens return; 30744a42d848SDavid Stevens 3075c50d8ae3SPaolo Bonzini __direct_pte_prefetch(vcpu, sp, sptep); 3076c50d8ae3SPaolo Bonzini } 3077c50d8ae3SPaolo Bonzini 307865e3b446SSean Christopherson /* 307965e3b446SSean Christopherson * Lookup the mapping level for @gfn in the current mm. 308065e3b446SSean Christopherson * 308165e3b446SSean Christopherson * WARNING! Use of host_pfn_mapping_level() requires the caller and the end 308265e3b446SSean Christopherson * consumer to be tied into KVM's handlers for MMU notifier events! 308365e3b446SSean Christopherson * 308465e3b446SSean Christopherson * There are several ways to safely use this helper: 308565e3b446SSean Christopherson * 308620ec3ebdSChao Peng * - Check mmu_invalidate_retry_hva() after grabbing the mapping level, before 308765e3b446SSean Christopherson * consuming it. In this case, mmu_lock doesn't need to be held during the 308865e3b446SSean Christopherson * lookup, but it does need to be held while checking the MMU notifier. 308965e3b446SSean Christopherson * 309065e3b446SSean Christopherson * - Hold mmu_lock AND ensure there is no in-progress MMU notifier invalidation 309165e3b446SSean Christopherson * event for the hva. This can be done by explicit checking the MMU notifier 309265e3b446SSean Christopherson * or by ensuring that KVM already has a valid mapping that covers the hva. 309365e3b446SSean Christopherson * 309465e3b446SSean Christopherson * - Do not use the result to install new mappings, e.g. use the host mapping 309565e3b446SSean Christopherson * level only to decide whether or not to zap an entry. In this case, it's 309665e3b446SSean Christopherson * not required to hold mmu_lock (though it's highly likely the caller will 309765e3b446SSean Christopherson * want to hold mmu_lock anyways, e.g. to modify SPTEs). 309865e3b446SSean Christopherson * 309965e3b446SSean Christopherson * Note! The lookup can still race with modifications to host page tables, but 310065e3b446SSean Christopherson * the above "rules" ensure KVM will not _consume_ the result of the walk if a 310165e3b446SSean Christopherson * race with the primary MMU occurs. 310265e3b446SSean Christopherson */ 3103a8ac499bSSean Christopherson static int host_pfn_mapping_level(struct kvm *kvm, gfn_t gfn, 31048ca6f063SBen Gardon const struct kvm_memory_slot *slot) 3105db543216SSean Christopherson { 3106284dc493SSean Christopherson int level = PG_LEVEL_4K; 3107db543216SSean Christopherson unsigned long hva; 310844187235SMingwei Zhang unsigned long flags; 310944187235SMingwei Zhang pgd_t pgd; 311044187235SMingwei Zhang p4d_t p4d; 311144187235SMingwei Zhang pud_t pud; 311244187235SMingwei Zhang pmd_t pmd; 3113db543216SSean Christopherson 31145d49f08cSSean Christopherson /* 3115293e306eSSean Christopherson * Note, using the already-retrieved memslot and __gfn_to_hva_memslot() 3116293e306eSSean Christopherson * is not solely for performance, it's also necessary to avoid the 3117293e306eSSean Christopherson * "writable" check in __gfn_to_hva_many(), which will always fail on 3118293e306eSSean Christopherson * read-only memslots due to gfn_to_hva() assuming writes. Earlier 3119293e306eSSean Christopherson * page fault steps have already verified the guest isn't writing a 3120293e306eSSean Christopherson * read-only memslot. 3121293e306eSSean Christopherson */ 3122db543216SSean Christopherson hva = __gfn_to_hva_memslot(slot, gfn); 3123db543216SSean Christopherson 312444187235SMingwei Zhang /* 312565e3b446SSean Christopherson * Disable IRQs to prevent concurrent tear down of host page tables, 312665e3b446SSean Christopherson * e.g. if the primary MMU promotes a P*D to a huge page and then frees 312765e3b446SSean Christopherson * the original page table. 312844187235SMingwei Zhang */ 312944187235SMingwei Zhang local_irq_save(flags); 3130db543216SSean Christopherson 313165e3b446SSean Christopherson /* 313265e3b446SSean Christopherson * Read each entry once. As above, a non-leaf entry can be promoted to 313365e3b446SSean Christopherson * a huge page _during_ this walk. Re-reading the entry could send the 313465e3b446SSean Christopherson * walk into the weeks, e.g. p*d_large() returns false (sees the old 313565e3b446SSean Christopherson * value) and then p*d_offset() walks into the target huge page instead 313665e3b446SSean Christopherson * of the old page table (sees the new value). 313765e3b446SSean Christopherson */ 313844187235SMingwei Zhang pgd = READ_ONCE(*pgd_offset(kvm->mm, hva)); 313944187235SMingwei Zhang if (pgd_none(pgd)) 314044187235SMingwei Zhang goto out; 314144187235SMingwei Zhang 314244187235SMingwei Zhang p4d = READ_ONCE(*p4d_offset(&pgd, hva)); 314344187235SMingwei Zhang if (p4d_none(p4d) || !p4d_present(p4d)) 314444187235SMingwei Zhang goto out; 314544187235SMingwei Zhang 314644187235SMingwei Zhang pud = READ_ONCE(*pud_offset(&p4d, hva)); 314744187235SMingwei Zhang if (pud_none(pud) || !pud_present(pud)) 314844187235SMingwei Zhang goto out; 314944187235SMingwei Zhang 315044187235SMingwei Zhang if (pud_large(pud)) { 315144187235SMingwei Zhang level = PG_LEVEL_1G; 315244187235SMingwei Zhang goto out; 315344187235SMingwei Zhang } 315444187235SMingwei Zhang 315544187235SMingwei Zhang pmd = READ_ONCE(*pmd_offset(&pud, hva)); 315644187235SMingwei Zhang if (pmd_none(pmd) || !pmd_present(pmd)) 315744187235SMingwei Zhang goto out; 315844187235SMingwei Zhang 315944187235SMingwei Zhang if (pmd_large(pmd)) 316044187235SMingwei Zhang level = PG_LEVEL_2M; 316144187235SMingwei Zhang 316244187235SMingwei Zhang out: 316344187235SMingwei Zhang local_irq_restore(flags); 3164db543216SSean Christopherson return level; 3165db543216SSean Christopherson } 3166db543216SSean Christopherson 31678ca6f063SBen Gardon int kvm_mmu_max_mapping_level(struct kvm *kvm, 31688ca6f063SBen Gardon const struct kvm_memory_slot *slot, gfn_t gfn, 3169a8ac499bSSean Christopherson int max_level) 31701b6d9d9eSSean Christopherson { 31711b6d9d9eSSean Christopherson struct kvm_lpage_info *linfo; 3172ec607a56SPaolo Bonzini int host_level; 31731b6d9d9eSSean Christopherson 31741b6d9d9eSSean Christopherson max_level = min(max_level, max_huge_page_level); 31751b6d9d9eSSean Christopherson for ( ; max_level > PG_LEVEL_4K; max_level--) { 31761b6d9d9eSSean Christopherson linfo = lpage_info_slot(gfn, slot, max_level); 31771b6d9d9eSSean Christopherson if (!linfo->disallow_lpage) 31781b6d9d9eSSean Christopherson break; 31791b6d9d9eSSean Christopherson } 31801b6d9d9eSSean Christopherson 31811b6d9d9eSSean Christopherson if (max_level == PG_LEVEL_4K) 31821b6d9d9eSSean Christopherson return PG_LEVEL_4K; 31831b6d9d9eSSean Christopherson 3184a8ac499bSSean Christopherson host_level = host_pfn_mapping_level(kvm, gfn, slot); 3185ec607a56SPaolo Bonzini return min(host_level, max_level); 31861b6d9d9eSSean Christopherson } 31871b6d9d9eSSean Christopherson 318873a3c659SPaolo Bonzini void kvm_mmu_hugepage_adjust(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 31890885904dSSean Christopherson { 3190e710c5f6SDavid Matlack struct kvm_memory_slot *slot = fault->slot; 319117eff019SSean Christopherson kvm_pfn_t mask; 31920885904dSSean Christopherson 319373a3c659SPaolo Bonzini fault->huge_page_disallowed = fault->exec && fault->nx_huge_page_workaround_enabled; 31943cf06612SSean Christopherson 319573a3c659SPaolo Bonzini if (unlikely(fault->max_level == PG_LEVEL_4K)) 319673a3c659SPaolo Bonzini return; 319717eff019SSean Christopherson 31985d49f08cSSean Christopherson if (is_error_noslot_pfn(fault->pfn)) 319973a3c659SPaolo Bonzini return; 320017eff019SSean Christopherson 3201e710c5f6SDavid Matlack if (kvm_slot_dirty_track_enabled(slot)) 320273a3c659SPaolo Bonzini return; 3203293e306eSSean Christopherson 32043cf06612SSean Christopherson /* 32053cf06612SSean Christopherson * Enforce the iTLB multihit workaround after capturing the requested 32063cf06612SSean Christopherson * level, which will be used to do precise, accurate accounting. 32073cf06612SSean Christopherson */ 320873a3c659SPaolo Bonzini fault->req_level = kvm_mmu_max_mapping_level(vcpu->kvm, slot, 3209a8ac499bSSean Christopherson fault->gfn, fault->max_level); 321073a3c659SPaolo Bonzini if (fault->req_level == PG_LEVEL_4K || fault->huge_page_disallowed) 321173a3c659SPaolo Bonzini return; 32124cd071d1SSean Christopherson 32130885904dSSean Christopherson /* 321420ec3ebdSChao Peng * mmu_invalidate_retry() was successful and mmu_lock is held, so 32154cd071d1SSean Christopherson * the pmd can't be split from under us. 32160885904dSSean Christopherson */ 321773a3c659SPaolo Bonzini fault->goal_level = fault->req_level; 321873a3c659SPaolo Bonzini mask = KVM_PAGES_PER_HPAGE(fault->goal_level) - 1; 321973a3c659SPaolo Bonzini VM_BUG_ON((fault->gfn & mask) != (fault->pfn & mask)); 322073a3c659SPaolo Bonzini fault->pfn &= ~mask; 32210885904dSSean Christopherson } 32220885904dSSean Christopherson 3223536f0e6aSPaolo Bonzini void disallowed_hugepage_adjust(struct kvm_page_fault *fault, u64 spte, int cur_level) 3224c50d8ae3SPaolo Bonzini { 3225536f0e6aSPaolo Bonzini if (cur_level > PG_LEVEL_4K && 3226536f0e6aSPaolo Bonzini cur_level == fault->goal_level && 3227c50d8ae3SPaolo Bonzini is_shadow_present_pte(spte) && 322876901e56SMingwei Zhang !is_large_pte(spte) && 322976901e56SMingwei Zhang spte_to_child_sp(spte)->nx_huge_page_disallowed) { 3230c50d8ae3SPaolo Bonzini /* 32316c882ef4SDavid Matlack * A small SPTE exists for this pfn, but FNAME(fetch), 32326c882ef4SDavid Matlack * direct_map(), or kvm_tdp_mmu_map() would like to create a 32336c882ef4SDavid Matlack * large PTE instead: just force them to go down another level, 32346c882ef4SDavid Matlack * patching back for them into pfn the next 9 bits of the 32356c882ef4SDavid Matlack * address. 3236c50d8ae3SPaolo Bonzini */ 3237536f0e6aSPaolo Bonzini u64 page_mask = KVM_PAGES_PER_HPAGE(cur_level) - 3238536f0e6aSPaolo Bonzini KVM_PAGES_PER_HPAGE(cur_level - 1); 3239536f0e6aSPaolo Bonzini fault->pfn |= fault->gfn & page_mask; 3240536f0e6aSPaolo Bonzini fault->goal_level--; 3241c50d8ae3SPaolo Bonzini } 3242c50d8ae3SPaolo Bonzini } 3243c50d8ae3SPaolo Bonzini 32446c882ef4SDavid Matlack static int direct_map(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 3245c50d8ae3SPaolo Bonzini { 3246c50d8ae3SPaolo Bonzini struct kvm_shadow_walk_iterator it; 3247c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 324873a3c659SPaolo Bonzini int ret; 324943b74355SPaolo Bonzini gfn_t base_gfn = fault->gfn; 3250c50d8ae3SPaolo Bonzini 325173a3c659SPaolo Bonzini kvm_mmu_hugepage_adjust(vcpu, fault); 32524cd071d1SSean Christopherson 3253f0066d94SPaolo Bonzini trace_kvm_mmu_spte_requested(fault); 325443b74355SPaolo Bonzini for_each_shadow_entry(vcpu, fault->addr, it) { 3255c50d8ae3SPaolo Bonzini /* 3256c50d8ae3SPaolo Bonzini * We cannot overwrite existing page tables with an NX 3257c50d8ae3SPaolo Bonzini * large page, as the leaf could be executable. 3258c50d8ae3SPaolo Bonzini */ 325973a3c659SPaolo Bonzini if (fault->nx_huge_page_workaround_enabled) 3260536f0e6aSPaolo Bonzini disallowed_hugepage_adjust(fault, *it.sptep, it.level); 3261c50d8ae3SPaolo Bonzini 3262c667a3baSHou Wenlong base_gfn = gfn_round_for_level(fault->gfn, it.level); 326373a3c659SPaolo Bonzini if (it.level == fault->goal_level) 3264c50d8ae3SPaolo Bonzini break; 3265c50d8ae3SPaolo Bonzini 32662e65e842SDavid Matlack sp = kvm_mmu_get_child_sp(vcpu, it.sptep, base_gfn, true, ACC_ALL); 32670cd8dc73SPaolo Bonzini if (sp == ERR_PTR(-EEXIST)) 32680cd8dc73SPaolo Bonzini continue; 3269c50d8ae3SPaolo Bonzini 3270c50d8ae3SPaolo Bonzini link_shadow_page(vcpu, it.sptep, sp); 3271b5b0977fSSean Christopherson if (fault->huge_page_disallowed) 327255c510e2SSean Christopherson account_nx_huge_page(vcpu->kvm, sp, 3273428e9216SSean Christopherson fault->req_level >= it.level); 3274c50d8ae3SPaolo Bonzini } 3275c50d8ae3SPaolo Bonzini 3276b1a429fbSSean Christopherson if (WARN_ON_ONCE(it.level != fault->goal_level)) 3277b1a429fbSSean Christopherson return -EFAULT; 3278b1a429fbSSean Christopherson 32798a9f566aSDavid Matlack ret = mmu_set_spte(vcpu, fault->slot, it.sptep, ACC_ALL, 3280a12f4381SPaolo Bonzini base_gfn, fault->pfn, fault); 328112703759SSean Christopherson if (ret == RET_PF_SPURIOUS) 328212703759SSean Christopherson return ret; 328312703759SSean Christopherson 3284c50d8ae3SPaolo Bonzini direct_pte_prefetch(vcpu, it.sptep); 3285c50d8ae3SPaolo Bonzini return ret; 3286c50d8ae3SPaolo Bonzini } 3287c50d8ae3SPaolo Bonzini 3288cd08d178SDavid Matlack static void kvm_send_hwpoison_signal(struct kvm_memory_slot *slot, gfn_t gfn) 3289c50d8ae3SPaolo Bonzini { 3290cd08d178SDavid Matlack unsigned long hva = gfn_to_hva_memslot(slot, gfn); 3291cd08d178SDavid Matlack 3292cd08d178SDavid Matlack send_sig_mceerr(BUS_MCEERR_AR, (void __user *)hva, PAGE_SHIFT, current); 3293c50d8ae3SPaolo Bonzini } 3294c50d8ae3SPaolo Bonzini 3295cd08d178SDavid Matlack static int kvm_handle_error_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 3296c50d8ae3SPaolo Bonzini { 3297cd08d178SDavid Matlack if (is_sigpending_pfn(fault->pfn)) { 329876657687SPeter Xu kvm_handle_signal_exit(vcpu); 329976657687SPeter Xu return -EINTR; 330076657687SPeter Xu } 330176657687SPeter Xu 3302c50d8ae3SPaolo Bonzini /* 3303c50d8ae3SPaolo Bonzini * Do not cache the mmio info caused by writing the readonly gfn 3304c50d8ae3SPaolo Bonzini * into the spte otherwise read access on readonly gfn also can 3305c50d8ae3SPaolo Bonzini * caused mmio page fault and treat it as mmio access. 3306c50d8ae3SPaolo Bonzini */ 3307cd08d178SDavid Matlack if (fault->pfn == KVM_PFN_ERR_RO_FAULT) 3308c50d8ae3SPaolo Bonzini return RET_PF_EMULATE; 3309c50d8ae3SPaolo Bonzini 3310cd08d178SDavid Matlack if (fault->pfn == KVM_PFN_ERR_HWPOISON) { 3311cd08d178SDavid Matlack kvm_send_hwpoison_signal(fault->slot, fault->gfn); 3312c50d8ae3SPaolo Bonzini return RET_PF_RETRY; 3313c50d8ae3SPaolo Bonzini } 3314c50d8ae3SPaolo Bonzini 3315c50d8ae3SPaolo Bonzini return -EFAULT; 3316c50d8ae3SPaolo Bonzini } 3317c50d8ae3SPaolo Bonzini 3318354c908cSDavid Matlack static int kvm_handle_noslot_fault(struct kvm_vcpu *vcpu, 3319354c908cSDavid Matlack struct kvm_page_fault *fault, 33205276c616SSean Christopherson unsigned int access) 3321c50d8ae3SPaolo Bonzini { 33223a13f4feSPaolo Bonzini gva_t gva = fault->is_tdp ? 0 : fault->addr; 33233a13f4feSPaolo Bonzini 33243a13f4feSPaolo Bonzini vcpu_cache_mmio_info(vcpu, gva, fault->gfn, 3325c50d8ae3SPaolo Bonzini access & shadow_mmio_access_mask); 3326354c908cSDavid Matlack 332730ab5901SSean Christopherson /* 332830ab5901SSean Christopherson * If MMIO caching is disabled, emulate immediately without 332930ab5901SSean Christopherson * touching the shadow page tables as attempting to install an 3330354c908cSDavid Matlack * MMIO SPTE will just be an expensive nop. 333130ab5901SSean Christopherson */ 3332354c908cSDavid Matlack if (unlikely(!enable_mmio_caching)) 33335276c616SSean Christopherson return RET_PF_EMULATE; 3334354c908cSDavid Matlack 3335354c908cSDavid Matlack /* 3336354c908cSDavid Matlack * Do not create an MMIO SPTE for a gfn greater than host.MAXPHYADDR, 3337354c908cSDavid Matlack * any guest that generates such gfns is running nested and is being 3338354c908cSDavid Matlack * tricked by L0 userspace (you can observe gfn > L1.MAXPHYADDR if and 3339354c908cSDavid Matlack * only if L1's MAXPHYADDR is inaccurate with respect to the 3340354c908cSDavid Matlack * hardware's). 3341354c908cSDavid Matlack */ 3342354c908cSDavid Matlack if (unlikely(fault->gfn > kvm_mmu_max_gfn())) 3343354c908cSDavid Matlack return RET_PF_EMULATE; 3344c50d8ae3SPaolo Bonzini 33455276c616SSean Christopherson return RET_PF_CONTINUE; 3346c50d8ae3SPaolo Bonzini } 3347c50d8ae3SPaolo Bonzini 33483c8ad5a6SPaolo Bonzini static bool page_fault_can_be_fast(struct kvm_page_fault *fault) 3349c50d8ae3SPaolo Bonzini { 3350c50d8ae3SPaolo Bonzini /* 33515c64aba5SSean Christopherson * Page faults with reserved bits set, i.e. faults on MMIO SPTEs, only 33525c64aba5SSean Christopherson * reach the common page fault handler if the SPTE has an invalid MMIO 33535c64aba5SSean Christopherson * generation number. Refreshing the MMIO generation needs to go down 33545c64aba5SSean Christopherson * the slow path. Note, EPT Misconfigs do NOT set the PRESENT flag! 3355c50d8ae3SPaolo Bonzini */ 33563c8ad5a6SPaolo Bonzini if (fault->rsvd) 3357c50d8ae3SPaolo Bonzini return false; 3358c50d8ae3SPaolo Bonzini 3359c50d8ae3SPaolo Bonzini /* 3360c50d8ae3SPaolo Bonzini * #PF can be fast if: 3361c50d8ae3SPaolo Bonzini * 336254275f74SSean Christopherson * 1. The shadow page table entry is not present and A/D bits are 336354275f74SSean Christopherson * disabled _by KVM_, which could mean that the fault is potentially 336454275f74SSean Christopherson * caused by access tracking (if enabled). If A/D bits are enabled 336554275f74SSean Christopherson * by KVM, but disabled by L1 for L2, KVM is forced to disable A/D 336654275f74SSean Christopherson * bits for L2 and employ access tracking, but the fast page fault 336754275f74SSean Christopherson * mechanism only supports direct MMUs. 336854275f74SSean Christopherson * 2. The shadow page table entry is present, the access is a write, 336954275f74SSean Christopherson * and no reserved bits are set (MMIO SPTEs cannot be "fixed"), i.e. 337054275f74SSean Christopherson * the fault was caused by a write-protection violation. If the 337154275f74SSean Christopherson * SPTE is MMU-writable (determined later), the fault can be fixed 337254275f74SSean Christopherson * by setting the Writable bit, which can be done out of mmu_lock. 3373c50d8ae3SPaolo Bonzini */ 33745c64aba5SSean Christopherson if (!fault->present) 33755c64aba5SSean Christopherson return !kvm_ad_enabled(); 33765c64aba5SSean Christopherson 33775c64aba5SSean Christopherson /* 33785c64aba5SSean Christopherson * Note, instruction fetches and writes are mutually exclusive, ignore 33795c64aba5SSean Christopherson * the "exec" flag. 33805c64aba5SSean Christopherson */ 33815c64aba5SSean Christopherson return fault->write; 3382c50d8ae3SPaolo Bonzini } 3383c50d8ae3SPaolo Bonzini 3384c50d8ae3SPaolo Bonzini /* 3385c50d8ae3SPaolo Bonzini * Returns true if the SPTE was fixed successfully. Otherwise, 3386c50d8ae3SPaolo Bonzini * someone else modified the SPTE from its original value. 3387c50d8ae3SPaolo Bonzini */ 3388f3d90f90SSean Christopherson static bool fast_pf_fix_direct_spte(struct kvm_vcpu *vcpu, 3389f3d90f90SSean Christopherson struct kvm_page_fault *fault, 3390c50d8ae3SPaolo Bonzini u64 *sptep, u64 old_spte, u64 new_spte) 3391c50d8ae3SPaolo Bonzini { 3392c50d8ae3SPaolo Bonzini /* 3393c50d8ae3SPaolo Bonzini * Theoretically we could also set dirty bit (and flush TLB) here in 3394c50d8ae3SPaolo Bonzini * order to eliminate unnecessary PML logging. See comments in 3395c50d8ae3SPaolo Bonzini * set_spte. But fast_page_fault is very unlikely to happen with PML 3396c50d8ae3SPaolo Bonzini * enabled, so we do not do this. This might result in the same GPA 3397c50d8ae3SPaolo Bonzini * to be logged in PML buffer again when the write really happens, and 3398c50d8ae3SPaolo Bonzini * eventually to be called by mark_page_dirty twice. But it's also no 3399c50d8ae3SPaolo Bonzini * harm. This also avoids the TLB flush needed after setting dirty bit 3400c50d8ae3SPaolo Bonzini * so non-PML cases won't be impacted. 3401c50d8ae3SPaolo Bonzini * 3402c50d8ae3SPaolo Bonzini * Compare with set_spte where instead shadow_dirty_mask is set. 3403c50d8ae3SPaolo Bonzini */ 34042db2f46fSUros Bizjak if (!try_cmpxchg64(sptep, &old_spte, new_spte)) 3405c50d8ae3SPaolo Bonzini return false; 3406c50d8ae3SPaolo Bonzini 3407e710c5f6SDavid Matlack if (is_writable_pte(new_spte) && !is_writable_pte(old_spte)) 3408e710c5f6SDavid Matlack mark_page_dirty_in_slot(vcpu->kvm, fault->slot, fault->gfn); 3409c50d8ae3SPaolo Bonzini 3410c50d8ae3SPaolo Bonzini return true; 3411c50d8ae3SPaolo Bonzini } 3412c50d8ae3SPaolo Bonzini 34133c8ad5a6SPaolo Bonzini static bool is_access_allowed(struct kvm_page_fault *fault, u64 spte) 3414c50d8ae3SPaolo Bonzini { 34153c8ad5a6SPaolo Bonzini if (fault->exec) 3416c50d8ae3SPaolo Bonzini return is_executable_pte(spte); 3417c50d8ae3SPaolo Bonzini 34183c8ad5a6SPaolo Bonzini if (fault->write) 3419c50d8ae3SPaolo Bonzini return is_writable_pte(spte); 3420c50d8ae3SPaolo Bonzini 3421c50d8ae3SPaolo Bonzini /* Fault was on Read access */ 3422c50d8ae3SPaolo Bonzini return spte & PT_PRESENT_MASK; 3423c50d8ae3SPaolo Bonzini } 3424c50d8ae3SPaolo Bonzini 3425c50d8ae3SPaolo Bonzini /* 34266e8eb206SDavid Matlack * Returns the last level spte pointer of the shadow page walk for the given 34276e8eb206SDavid Matlack * gpa, and sets *spte to the spte value. This spte may be non-preset. If no 34286e8eb206SDavid Matlack * walk could be performed, returns NULL and *spte does not contain valid data. 34296e8eb206SDavid Matlack * 34306e8eb206SDavid Matlack * Contract: 34316e8eb206SDavid Matlack * - Must be called between walk_shadow_page_lockless_{begin,end}. 34326e8eb206SDavid Matlack * - The returned sptep must not be used after walk_shadow_page_lockless_end. 34336e8eb206SDavid Matlack */ 34346e8eb206SDavid Matlack static u64 *fast_pf_get_last_sptep(struct kvm_vcpu *vcpu, gpa_t gpa, u64 *spte) 34356e8eb206SDavid Matlack { 34366e8eb206SDavid Matlack struct kvm_shadow_walk_iterator iterator; 34376e8eb206SDavid Matlack u64 old_spte; 34386e8eb206SDavid Matlack u64 *sptep = NULL; 34396e8eb206SDavid Matlack 34406e8eb206SDavid Matlack for_each_shadow_entry_lockless(vcpu, gpa, iterator, old_spte) { 34416e8eb206SDavid Matlack sptep = iterator.sptep; 34426e8eb206SDavid Matlack *spte = old_spte; 34436e8eb206SDavid Matlack } 34446e8eb206SDavid Matlack 34456e8eb206SDavid Matlack return sptep; 34466e8eb206SDavid Matlack } 34476e8eb206SDavid Matlack 34486e8eb206SDavid Matlack /* 3449c4371c2aSSean Christopherson * Returns one of RET_PF_INVALID, RET_PF_FIXED or RET_PF_SPURIOUS. 3450c50d8ae3SPaolo Bonzini */ 34513c8ad5a6SPaolo Bonzini static int fast_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 3452c50d8ae3SPaolo Bonzini { 3453c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 3454c4371c2aSSean Christopherson int ret = RET_PF_INVALID; 3455c50d8ae3SPaolo Bonzini u64 spte = 0ull; 34566e8eb206SDavid Matlack u64 *sptep = NULL; 3457c50d8ae3SPaolo Bonzini uint retry_count = 0; 3458c50d8ae3SPaolo Bonzini 34593c8ad5a6SPaolo Bonzini if (!page_fault_can_be_fast(fault)) 3460c4371c2aSSean Christopherson return ret; 3461c50d8ae3SPaolo Bonzini 3462c50d8ae3SPaolo Bonzini walk_shadow_page_lockless_begin(vcpu); 3463c50d8ae3SPaolo Bonzini 3464c50d8ae3SPaolo Bonzini do { 3465c50d8ae3SPaolo Bonzini u64 new_spte; 3466c50d8ae3SPaolo Bonzini 3467dfe0ecc6SSean Christopherson if (tdp_mmu_enabled) 34683c8ad5a6SPaolo Bonzini sptep = kvm_tdp_mmu_fast_pf_get_last_sptep(vcpu, fault->addr, &spte); 34696e8eb206SDavid Matlack else 34703c8ad5a6SPaolo Bonzini sptep = fast_pf_get_last_sptep(vcpu, fault->addr, &spte); 3471c50d8ae3SPaolo Bonzini 3472ec89e643SSean Christopherson if (!is_shadow_present_pte(spte)) 3473ec89e643SSean Christopherson break; 3474ec89e643SSean Christopherson 34756e8eb206SDavid Matlack sp = sptep_to_sp(sptep); 3476c50d8ae3SPaolo Bonzini if (!is_last_spte(spte, sp->role.level)) 3477c50d8ae3SPaolo Bonzini break; 3478c50d8ae3SPaolo Bonzini 3479c50d8ae3SPaolo Bonzini /* 3480c50d8ae3SPaolo Bonzini * Check whether the memory access that caused the fault would 3481c50d8ae3SPaolo Bonzini * still cause it if it were to be performed right now. If not, 3482c50d8ae3SPaolo Bonzini * then this is a spurious fault caused by TLB lazily flushed, 3483c50d8ae3SPaolo Bonzini * or some other CPU has already fixed the PTE after the 3484c50d8ae3SPaolo Bonzini * current CPU took the fault. 3485c50d8ae3SPaolo Bonzini * 3486c50d8ae3SPaolo Bonzini * Need not check the access of upper level table entries since 3487c50d8ae3SPaolo Bonzini * they are always ACC_ALL. 3488c50d8ae3SPaolo Bonzini */ 34893c8ad5a6SPaolo Bonzini if (is_access_allowed(fault, spte)) { 3490c4371c2aSSean Christopherson ret = RET_PF_SPURIOUS; 3491c50d8ae3SPaolo Bonzini break; 3492c50d8ae3SPaolo Bonzini } 3493c50d8ae3SPaolo Bonzini 3494c50d8ae3SPaolo Bonzini new_spte = spte; 3495c50d8ae3SPaolo Bonzini 349654275f74SSean Christopherson /* 349754275f74SSean Christopherson * KVM only supports fixing page faults outside of MMU lock for 349854275f74SSean Christopherson * direct MMUs, nested MMUs are always indirect, and KVM always 349954275f74SSean Christopherson * uses A/D bits for non-nested MMUs. Thus, if A/D bits are 350054275f74SSean Christopherson * enabled, the SPTE can't be an access-tracked SPTE. 350154275f74SSean Christopherson */ 350254275f74SSean Christopherson if (unlikely(!kvm_ad_enabled()) && is_access_track_spte(spte)) 3503c50d8ae3SPaolo Bonzini new_spte = restore_acc_track_spte(new_spte); 3504c50d8ae3SPaolo Bonzini 3505c50d8ae3SPaolo Bonzini /* 350654275f74SSean Christopherson * To keep things simple, only SPTEs that are MMU-writable can 350754275f74SSean Christopherson * be made fully writable outside of mmu_lock, e.g. only SPTEs 350854275f74SSean Christopherson * that were write-protected for dirty-logging or access 350954275f74SSean Christopherson * tracking are handled here. Don't bother checking if the 351054275f74SSean Christopherson * SPTE is writable to prioritize running with A/D bits enabled. 351154275f74SSean Christopherson * The is_access_allowed() check above handles the common case 351254275f74SSean Christopherson * of the fault being spurious, and the SPTE is known to be 351354275f74SSean Christopherson * shadow-present, i.e. except for access tracking restoration 351454275f74SSean Christopherson * making the new SPTE writable, the check is wasteful. 3515c50d8ae3SPaolo Bonzini */ 3516706c9c55SSean Christopherson if (fault->write && is_mmu_writable_spte(spte)) { 3517c50d8ae3SPaolo Bonzini new_spte |= PT_WRITABLE_MASK; 3518c50d8ae3SPaolo Bonzini 3519c50d8ae3SPaolo Bonzini /* 352010c30de0SJunaid Shahid * Do not fix write-permission on the large spte when 352110c30de0SJunaid Shahid * dirty logging is enabled. Since we only dirty the 352210c30de0SJunaid Shahid * first page into the dirty-bitmap in 3523c50d8ae3SPaolo Bonzini * fast_pf_fix_direct_spte(), other pages are missed 3524c50d8ae3SPaolo Bonzini * if its slot has dirty logging enabled. 3525c50d8ae3SPaolo Bonzini * 3526c50d8ae3SPaolo Bonzini * Instead, we let the slow page fault path create a 3527c50d8ae3SPaolo Bonzini * normal spte to fix the access. 3528c50d8ae3SPaolo Bonzini */ 352910c30de0SJunaid Shahid if (sp->role.level > PG_LEVEL_4K && 353010c30de0SJunaid Shahid kvm_slot_dirty_track_enabled(fault->slot)) 3531c50d8ae3SPaolo Bonzini break; 3532c50d8ae3SPaolo Bonzini } 3533c50d8ae3SPaolo Bonzini 3534c50d8ae3SPaolo Bonzini /* Verify that the fault can be handled in the fast path */ 3535c50d8ae3SPaolo Bonzini if (new_spte == spte || 35363c8ad5a6SPaolo Bonzini !is_access_allowed(fault, new_spte)) 3537c50d8ae3SPaolo Bonzini break; 3538c50d8ae3SPaolo Bonzini 3539c50d8ae3SPaolo Bonzini /* 3540c50d8ae3SPaolo Bonzini * Currently, fast page fault only works for direct mapping 3541c50d8ae3SPaolo Bonzini * since the gfn is not stable for indirect shadow page. See 35423ecad8c2SMauro Carvalho Chehab * Documentation/virt/kvm/locking.rst to get more detail. 3543c50d8ae3SPaolo Bonzini */ 3544e710c5f6SDavid Matlack if (fast_pf_fix_direct_spte(vcpu, fault, sptep, spte, new_spte)) { 3545c4371c2aSSean Christopherson ret = RET_PF_FIXED; 3546c50d8ae3SPaolo Bonzini break; 3547c4371c2aSSean Christopherson } 3548c50d8ae3SPaolo Bonzini 3549c50d8ae3SPaolo Bonzini if (++retry_count > 4) { 35508d20bd63SSean Christopherson pr_warn_once("Fast #PF retrying more than 4 times.\n"); 3551c50d8ae3SPaolo Bonzini break; 3552c50d8ae3SPaolo Bonzini } 3553c50d8ae3SPaolo Bonzini 3554c50d8ae3SPaolo Bonzini } while (true); 3555c50d8ae3SPaolo Bonzini 3556f0066d94SPaolo Bonzini trace_fast_page_fault(vcpu, fault, sptep, spte, ret); 3557c50d8ae3SPaolo Bonzini walk_shadow_page_lockless_end(vcpu); 3558c50d8ae3SPaolo Bonzini 35591075d41eSSean Christopherson if (ret != RET_PF_INVALID) 35601075d41eSSean Christopherson vcpu->stat.pf_fast++; 35611075d41eSSean Christopherson 3562c4371c2aSSean Christopherson return ret; 3563c50d8ae3SPaolo Bonzini } 3564c50d8ae3SPaolo Bonzini 3565c50d8ae3SPaolo Bonzini static void mmu_free_root_page(struct kvm *kvm, hpa_t *root_hpa, 3566c50d8ae3SPaolo Bonzini struct list_head *invalid_list) 3567c50d8ae3SPaolo Bonzini { 3568c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 3569c50d8ae3SPaolo Bonzini 3570c50d8ae3SPaolo Bonzini if (!VALID_PAGE(*root_hpa)) 3571c50d8ae3SPaolo Bonzini return; 3572c50d8ae3SPaolo Bonzini 35735e3edd7eSSean Christopherson /* 35745e3edd7eSSean Christopherson * The "root" may be a special root, e.g. a PAE entry, treat it as a 35755e3edd7eSSean Christopherson * SPTE to ensure any non-PA bits are dropped. 35765e3edd7eSSean Christopherson */ 35775e3edd7eSSean Christopherson sp = spte_to_child_sp(*root_hpa); 35789191b8f0SPaolo Bonzini if (WARN_ON(!sp)) 35799191b8f0SPaolo Bonzini return; 358002c00b3aSBen Gardon 3581897218ffSPaolo Bonzini if (is_tdp_mmu_page(sp)) 35826103bc07SBen Gardon kvm_tdp_mmu_put_root(kvm, sp, false); 358376eb54e7SBen Gardon else if (!--sp->root_count && sp->role.invalid) 3584c50d8ae3SPaolo Bonzini kvm_mmu_prepare_zap_page(kvm, sp, invalid_list); 3585c50d8ae3SPaolo Bonzini 3586c50d8ae3SPaolo Bonzini *root_hpa = INVALID_PAGE; 3587c50d8ae3SPaolo Bonzini } 3588c50d8ae3SPaolo Bonzini 3589c50d8ae3SPaolo Bonzini /* roots_to_free must be some combination of the KVM_MMU_ROOT_* flags */ 35900c1c92f1SPaolo Bonzini void kvm_mmu_free_roots(struct kvm *kvm, struct kvm_mmu *mmu, 3591c50d8ae3SPaolo Bonzini ulong roots_to_free) 3592c50d8ae3SPaolo Bonzini { 3593c50d8ae3SPaolo Bonzini int i; 3594c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 3595594bef79SPaolo Bonzini bool free_active_root; 3596c50d8ae3SPaolo Bonzini 3597f94db0c8SSean Christopherson WARN_ON_ONCE(roots_to_free & ~KVM_MMU_ROOTS_ALL); 3598f94db0c8SSean Christopherson 3599c50d8ae3SPaolo Bonzini BUILD_BUG_ON(KVM_MMU_NUM_PREV_ROOTS >= BITS_PER_LONG); 3600c50d8ae3SPaolo Bonzini 3601c50d8ae3SPaolo Bonzini /* Before acquiring the MMU lock, see if we need to do any real work. */ 3602594bef79SPaolo Bonzini free_active_root = (roots_to_free & KVM_MMU_ROOT_CURRENT) 3603594bef79SPaolo Bonzini && VALID_PAGE(mmu->root.hpa); 3604594bef79SPaolo Bonzini 3605594bef79SPaolo Bonzini if (!free_active_root) { 3606c50d8ae3SPaolo Bonzini for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) 3607c50d8ae3SPaolo Bonzini if ((roots_to_free & KVM_MMU_ROOT_PREVIOUS(i)) && 3608c50d8ae3SPaolo Bonzini VALID_PAGE(mmu->prev_roots[i].hpa)) 3609c50d8ae3SPaolo Bonzini break; 3610c50d8ae3SPaolo Bonzini 3611c50d8ae3SPaolo Bonzini if (i == KVM_MMU_NUM_PREV_ROOTS) 3612c50d8ae3SPaolo Bonzini return; 3613c50d8ae3SPaolo Bonzini } 3614c50d8ae3SPaolo Bonzini 3615531810caSBen Gardon write_lock(&kvm->mmu_lock); 3616c50d8ae3SPaolo Bonzini 3617c50d8ae3SPaolo Bonzini for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) 3618c50d8ae3SPaolo Bonzini if (roots_to_free & KVM_MMU_ROOT_PREVIOUS(i)) 36194d710de9SSean Christopherson mmu_free_root_page(kvm, &mmu->prev_roots[i].hpa, 3620c50d8ae3SPaolo Bonzini &invalid_list); 3621c50d8ae3SPaolo Bonzini 3622c50d8ae3SPaolo Bonzini if (free_active_root) { 3623594bef79SPaolo Bonzini if (to_shadow_page(mmu->root.hpa)) { 3624b9e5603cSPaolo Bonzini mmu_free_root_page(kvm, &mmu->root.hpa, &invalid_list); 362504d45551SSean Christopherson } else if (mmu->pae_root) { 3626c834e5e4SSean Christopherson for (i = 0; i < 4; ++i) { 3627c834e5e4SSean Christopherson if (!IS_VALID_PAE_ROOT(mmu->pae_root[i])) 3628c834e5e4SSean Christopherson continue; 3629c834e5e4SSean Christopherson 3630c834e5e4SSean Christopherson mmu_free_root_page(kvm, &mmu->pae_root[i], 3631c50d8ae3SPaolo Bonzini &invalid_list); 3632c834e5e4SSean Christopherson mmu->pae_root[i] = INVALID_PAE_ROOT; 3633c50d8ae3SPaolo Bonzini } 3634c50d8ae3SPaolo Bonzini } 3635b9e5603cSPaolo Bonzini mmu->root.hpa = INVALID_PAGE; 3636b9e5603cSPaolo Bonzini mmu->root.pgd = 0; 3637c50d8ae3SPaolo Bonzini } 3638c50d8ae3SPaolo Bonzini 36394d710de9SSean Christopherson kvm_mmu_commit_zap_page(kvm, &invalid_list); 3640531810caSBen Gardon write_unlock(&kvm->mmu_lock); 3641c50d8ae3SPaolo Bonzini } 3642c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_mmu_free_roots); 3643c50d8ae3SPaolo Bonzini 36440c1c92f1SPaolo Bonzini void kvm_mmu_free_guest_mode_roots(struct kvm *kvm, struct kvm_mmu *mmu) 364525b62c62SSean Christopherson { 364625b62c62SSean Christopherson unsigned long roots_to_free = 0; 364725b62c62SSean Christopherson hpa_t root_hpa; 364825b62c62SSean Christopherson int i; 364925b62c62SSean Christopherson 365025b62c62SSean Christopherson /* 365125b62c62SSean Christopherson * This should not be called while L2 is active, L2 can't invalidate 365225b62c62SSean Christopherson * _only_ its own roots, e.g. INVVPID unconditionally exits. 365325b62c62SSean Christopherson */ 36547a458f0eSPaolo Bonzini WARN_ON_ONCE(mmu->root_role.guest_mode); 365525b62c62SSean Christopherson 365625b62c62SSean Christopherson for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) { 365725b62c62SSean Christopherson root_hpa = mmu->prev_roots[i].hpa; 365825b62c62SSean Christopherson if (!VALID_PAGE(root_hpa)) 365925b62c62SSean Christopherson continue; 366025b62c62SSean Christopherson 366125b62c62SSean Christopherson if (!to_shadow_page(root_hpa) || 366225b62c62SSean Christopherson to_shadow_page(root_hpa)->role.guest_mode) 366325b62c62SSean Christopherson roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i); 366425b62c62SSean Christopherson } 366525b62c62SSean Christopherson 36660c1c92f1SPaolo Bonzini kvm_mmu_free_roots(kvm, mmu, roots_to_free); 366725b62c62SSean Christopherson } 366825b62c62SSean Christopherson EXPORT_SYMBOL_GPL(kvm_mmu_free_guest_mode_roots); 366925b62c62SSean Christopherson 367025b62c62SSean Christopherson 3671c50d8ae3SPaolo Bonzini static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn) 3672c50d8ae3SPaolo Bonzini { 3673c50d8ae3SPaolo Bonzini int ret = 0; 3674c50d8ae3SPaolo Bonzini 3675995decb6SVitaly Kuznetsov if (!kvm_vcpu_is_visible_gfn(vcpu, root_gfn)) { 3676c50d8ae3SPaolo Bonzini kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu); 3677c50d8ae3SPaolo Bonzini ret = 1; 3678c50d8ae3SPaolo Bonzini } 3679c50d8ae3SPaolo Bonzini 3680c50d8ae3SPaolo Bonzini return ret; 3681c50d8ae3SPaolo Bonzini } 3682c50d8ae3SPaolo Bonzini 36832e65e842SDavid Matlack static hpa_t mmu_alloc_root(struct kvm_vcpu *vcpu, gfn_t gfn, int quadrant, 368486938ab6SDavid Matlack u8 level) 3685c50d8ae3SPaolo Bonzini { 36862e65e842SDavid Matlack union kvm_mmu_page_role role = vcpu->arch.mmu->root_role; 3687c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 36888123f265SSean Christopherson 36892e65e842SDavid Matlack role.level = level; 36902e65e842SDavid Matlack role.quadrant = quadrant; 36912e65e842SDavid Matlack 36927f497775SDavid Matlack WARN_ON_ONCE(quadrant && !role.has_4_byte_gpte); 36937f497775SDavid Matlack WARN_ON_ONCE(role.direct && role.has_4_byte_gpte); 36947f497775SDavid Matlack 369587654643SDavid Matlack sp = kvm_mmu_get_shadow_page(vcpu, gfn, role); 36968123f265SSean Christopherson ++sp->root_count; 36978123f265SSean Christopherson 36988123f265SSean Christopherson return __pa(sp->spt); 36998123f265SSean Christopherson } 37008123f265SSean Christopherson 37018123f265SSean Christopherson static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu) 37028123f265SSean Christopherson { 3703b37233c9SSean Christopherson struct kvm_mmu *mmu = vcpu->arch.mmu; 3704a972e29cSPaolo Bonzini u8 shadow_root_level = mmu->root_role.level; 37058123f265SSean Christopherson hpa_t root; 3706c50d8ae3SPaolo Bonzini unsigned i; 37074a38162eSPaolo Bonzini int r; 37084a38162eSPaolo Bonzini 37094a38162eSPaolo Bonzini write_lock(&vcpu->kvm->mmu_lock); 37104a38162eSPaolo Bonzini r = make_mmu_pages_available(vcpu); 37114a38162eSPaolo Bonzini if (r < 0) 37124a38162eSPaolo Bonzini goto out_unlock; 3713c50d8ae3SPaolo Bonzini 37141f98f2bdSDavid Matlack if (tdp_mmu_enabled) { 371502c00b3aSBen Gardon root = kvm_tdp_mmu_get_vcpu_root_hpa(vcpu); 3716b9e5603cSPaolo Bonzini mmu->root.hpa = root; 371702c00b3aSBen Gardon } else if (shadow_root_level >= PT64_ROOT_4LEVEL) { 371886938ab6SDavid Matlack root = mmu_alloc_root(vcpu, 0, 0, shadow_root_level); 3719b9e5603cSPaolo Bonzini mmu->root.hpa = root; 37208123f265SSean Christopherson } else if (shadow_root_level == PT32E_ROOT_LEVEL) { 37214a38162eSPaolo Bonzini if (WARN_ON_ONCE(!mmu->pae_root)) { 37224a38162eSPaolo Bonzini r = -EIO; 37234a38162eSPaolo Bonzini goto out_unlock; 37244a38162eSPaolo Bonzini } 372573ad1606SSean Christopherson 3726c50d8ae3SPaolo Bonzini for (i = 0; i < 4; ++i) { 3727c834e5e4SSean Christopherson WARN_ON_ONCE(IS_VALID_PAE_ROOT(mmu->pae_root[i])); 3728c50d8ae3SPaolo Bonzini 37297f497775SDavid Matlack root = mmu_alloc_root(vcpu, i << (30 - PAGE_SHIFT), 0, 37302e65e842SDavid Matlack PT32_ROOT_LEVEL); 373117e368d9SSean Christopherson mmu->pae_root[i] = root | PT_PRESENT_MASK | 3732d2263de1SYuan Yao shadow_me_value; 3733c50d8ae3SPaolo Bonzini } 3734b9e5603cSPaolo Bonzini mmu->root.hpa = __pa(mmu->pae_root); 373573ad1606SSean Christopherson } else { 373673ad1606SSean Christopherson WARN_ONCE(1, "Bad TDP root level = %d\n", shadow_root_level); 37374a38162eSPaolo Bonzini r = -EIO; 37384a38162eSPaolo Bonzini goto out_unlock; 373973ad1606SSean Christopherson } 37403651c7fcSSean Christopherson 3741b9e5603cSPaolo Bonzini /* root.pgd is ignored for direct MMUs. */ 3742b9e5603cSPaolo Bonzini mmu->root.pgd = 0; 37434a38162eSPaolo Bonzini out_unlock: 37444a38162eSPaolo Bonzini write_unlock(&vcpu->kvm->mmu_lock); 37454a38162eSPaolo Bonzini return r; 3746c50d8ae3SPaolo Bonzini } 3747c50d8ae3SPaolo Bonzini 37481e76a3ceSDavid Stevens static int mmu_first_shadow_root_alloc(struct kvm *kvm) 37491e76a3ceSDavid Stevens { 37501e76a3ceSDavid Stevens struct kvm_memslots *slots; 37511e76a3ceSDavid Stevens struct kvm_memory_slot *slot; 3752a54d8066SMaciej S. Szmigiero int r = 0, i, bkt; 37531e76a3ceSDavid Stevens 37541e76a3ceSDavid Stevens /* 37551e76a3ceSDavid Stevens * Check if this is the first shadow root being allocated before 37561e76a3ceSDavid Stevens * taking the lock. 37571e76a3ceSDavid Stevens */ 37581e76a3ceSDavid Stevens if (kvm_shadow_root_allocated(kvm)) 37591e76a3ceSDavid Stevens return 0; 37601e76a3ceSDavid Stevens 37611e76a3ceSDavid Stevens mutex_lock(&kvm->slots_arch_lock); 37621e76a3ceSDavid Stevens 37631e76a3ceSDavid Stevens /* Recheck, under the lock, whether this is the first shadow root. */ 37641e76a3ceSDavid Stevens if (kvm_shadow_root_allocated(kvm)) 37651e76a3ceSDavid Stevens goto out_unlock; 37661e76a3ceSDavid Stevens 37671e76a3ceSDavid Stevens /* 37681e76a3ceSDavid Stevens * Check if anything actually needs to be allocated, e.g. all metadata 37691e76a3ceSDavid Stevens * will be allocated upfront if TDP is disabled. 37701e76a3ceSDavid Stevens */ 37711e76a3ceSDavid Stevens if (kvm_memslots_have_rmaps(kvm) && 37721e76a3ceSDavid Stevens kvm_page_track_write_tracking_enabled(kvm)) 37731e76a3ceSDavid Stevens goto out_success; 37741e76a3ceSDavid Stevens 37751e76a3ceSDavid Stevens for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) { 37761e76a3ceSDavid Stevens slots = __kvm_memslots(kvm, i); 3777a54d8066SMaciej S. Szmigiero kvm_for_each_memslot(slot, bkt, slots) { 37781e76a3ceSDavid Stevens /* 37791e76a3ceSDavid Stevens * Both of these functions are no-ops if the target is 37801e76a3ceSDavid Stevens * already allocated, so unconditionally calling both 37811e76a3ceSDavid Stevens * is safe. Intentionally do NOT free allocations on 37821e76a3ceSDavid Stevens * failure to avoid having to track which allocations 37831e76a3ceSDavid Stevens * were made now versus when the memslot was created. 37841e76a3ceSDavid Stevens * The metadata is guaranteed to be freed when the slot 37851e76a3ceSDavid Stevens * is freed, and will be kept/used if userspace retries 37861e76a3ceSDavid Stevens * KVM_RUN instead of killing the VM. 37871e76a3ceSDavid Stevens */ 37881e76a3ceSDavid Stevens r = memslot_rmap_alloc(slot, slot->npages); 37891e76a3ceSDavid Stevens if (r) 37901e76a3ceSDavid Stevens goto out_unlock; 37911e76a3ceSDavid Stevens r = kvm_page_track_write_tracking_alloc(slot); 37921e76a3ceSDavid Stevens if (r) 37931e76a3ceSDavid Stevens goto out_unlock; 37941e76a3ceSDavid Stevens } 37951e76a3ceSDavid Stevens } 37961e76a3ceSDavid Stevens 37971e76a3ceSDavid Stevens /* 37981e76a3ceSDavid Stevens * Ensure that shadow_root_allocated becomes true strictly after 37991e76a3ceSDavid Stevens * all the related pointers are set. 38001e76a3ceSDavid Stevens */ 38011e76a3ceSDavid Stevens out_success: 38021e76a3ceSDavid Stevens smp_store_release(&kvm->arch.shadow_root_allocated, true); 38031e76a3ceSDavid Stevens 38041e76a3ceSDavid Stevens out_unlock: 38051e76a3ceSDavid Stevens mutex_unlock(&kvm->slots_arch_lock); 38061e76a3ceSDavid Stevens return r; 38071e76a3ceSDavid Stevens } 38081e76a3ceSDavid Stevens 3809c50d8ae3SPaolo Bonzini static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu) 3810c50d8ae3SPaolo Bonzini { 3811b37233c9SSean Christopherson struct kvm_mmu *mmu = vcpu->arch.mmu; 38126e0918aeSSean Christopherson u64 pdptrs[4], pm_mask; 3813be01e8e2SSean Christopherson gfn_t root_gfn, root_pgd; 38147f497775SDavid Matlack int quadrant, i, r; 38158123f265SSean Christopherson hpa_t root; 3816c50d8ae3SPaolo Bonzini 38172fdcc1b3SPaolo Bonzini root_pgd = kvm_mmu_get_guest_pgd(vcpu, mmu); 3818be01e8e2SSean Christopherson root_gfn = root_pgd >> PAGE_SHIFT; 3819c50d8ae3SPaolo Bonzini 3820c50d8ae3SPaolo Bonzini if (mmu_check_root(vcpu, root_gfn)) 3821c50d8ae3SPaolo Bonzini return 1; 3822c50d8ae3SPaolo Bonzini 3823c50d8ae3SPaolo Bonzini /* 38244a38162eSPaolo Bonzini * On SVM, reading PDPTRs might access guest memory, which might fault 38254a38162eSPaolo Bonzini * and thus might sleep. Grab the PDPTRs before acquiring mmu_lock. 38264a38162eSPaolo Bonzini */ 38274d25502aSPaolo Bonzini if (mmu->cpu_role.base.level == PT32E_ROOT_LEVEL) { 38286e0918aeSSean Christopherson for (i = 0; i < 4; ++i) { 38296e0918aeSSean Christopherson pdptrs[i] = mmu->get_pdptr(vcpu, i); 38306e0918aeSSean Christopherson if (!(pdptrs[i] & PT_PRESENT_MASK)) 38316e0918aeSSean Christopherson continue; 38326e0918aeSSean Christopherson 38336e0918aeSSean Christopherson if (mmu_check_root(vcpu, pdptrs[i] >> PAGE_SHIFT)) 38346e0918aeSSean Christopherson return 1; 38356e0918aeSSean Christopherson } 38366e0918aeSSean Christopherson } 38376e0918aeSSean Christopherson 38381e76a3ceSDavid Stevens r = mmu_first_shadow_root_alloc(vcpu->kvm); 3839d501f747SBen Gardon if (r) 3840d501f747SBen Gardon return r; 3841d501f747SBen Gardon 38424a38162eSPaolo Bonzini write_lock(&vcpu->kvm->mmu_lock); 38434a38162eSPaolo Bonzini r = make_mmu_pages_available(vcpu); 38444a38162eSPaolo Bonzini if (r < 0) 38454a38162eSPaolo Bonzini goto out_unlock; 38464a38162eSPaolo Bonzini 3847c50d8ae3SPaolo Bonzini /* 3848c50d8ae3SPaolo Bonzini * Do we shadow a long mode page table? If so we need to 3849c50d8ae3SPaolo Bonzini * write-protect the guests page table root. 3850c50d8ae3SPaolo Bonzini */ 38514d25502aSPaolo Bonzini if (mmu->cpu_role.base.level >= PT64_ROOT_4LEVEL) { 38528123f265SSean Christopherson root = mmu_alloc_root(vcpu, root_gfn, 0, 385386938ab6SDavid Matlack mmu->root_role.level); 3854b9e5603cSPaolo Bonzini mmu->root.hpa = root; 3855be01e8e2SSean Christopherson goto set_root_pgd; 3856c50d8ae3SPaolo Bonzini } 3857c50d8ae3SPaolo Bonzini 38584a38162eSPaolo Bonzini if (WARN_ON_ONCE(!mmu->pae_root)) { 38594a38162eSPaolo Bonzini r = -EIO; 38604a38162eSPaolo Bonzini goto out_unlock; 38614a38162eSPaolo Bonzini } 386273ad1606SSean Christopherson 3863c50d8ae3SPaolo Bonzini /* 3864c50d8ae3SPaolo Bonzini * We shadow a 32 bit page table. This may be a legacy 2-level 3865c50d8ae3SPaolo Bonzini * or a PAE 3-level page table. In either case we need to be aware that 3866c50d8ae3SPaolo Bonzini * the shadow page table may be a PAE or a long mode page table. 3867c50d8ae3SPaolo Bonzini */ 3868e54f1ff2SKai Huang pm_mask = PT_PRESENT_MASK | shadow_me_value; 3869a972e29cSPaolo Bonzini if (mmu->root_role.level >= PT64_ROOT_4LEVEL) { 3870c50d8ae3SPaolo Bonzini pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK; 3871c50d8ae3SPaolo Bonzini 387203ca4589SSean Christopherson if (WARN_ON_ONCE(!mmu->pml4_root)) { 38734a38162eSPaolo Bonzini r = -EIO; 38744a38162eSPaolo Bonzini goto out_unlock; 38754a38162eSPaolo Bonzini } 387603ca4589SSean Christopherson mmu->pml4_root[0] = __pa(mmu->pae_root) | pm_mask; 3877cb0f722aSWei Huang 3878a972e29cSPaolo Bonzini if (mmu->root_role.level == PT64_ROOT_5LEVEL) { 3879cb0f722aSWei Huang if (WARN_ON_ONCE(!mmu->pml5_root)) { 3880cb0f722aSWei Huang r = -EIO; 3881cb0f722aSWei Huang goto out_unlock; 3882cb0f722aSWei Huang } 3883cb0f722aSWei Huang mmu->pml5_root[0] = __pa(mmu->pml4_root) | pm_mask; 3884cb0f722aSWei Huang } 388504d45551SSean Christopherson } 388604d45551SSean Christopherson 3887c50d8ae3SPaolo Bonzini for (i = 0; i < 4; ++i) { 3888c834e5e4SSean Christopherson WARN_ON_ONCE(IS_VALID_PAE_ROOT(mmu->pae_root[i])); 38896e6ec584SSean Christopherson 38904d25502aSPaolo Bonzini if (mmu->cpu_role.base.level == PT32E_ROOT_LEVEL) { 38916e0918aeSSean Christopherson if (!(pdptrs[i] & PT_PRESENT_MASK)) { 3892c834e5e4SSean Christopherson mmu->pae_root[i] = INVALID_PAE_ROOT; 3893c50d8ae3SPaolo Bonzini continue; 3894c50d8ae3SPaolo Bonzini } 38956e0918aeSSean Christopherson root_gfn = pdptrs[i] >> PAGE_SHIFT; 3896c50d8ae3SPaolo Bonzini } 3897c50d8ae3SPaolo Bonzini 38987f497775SDavid Matlack /* 38997f497775SDavid Matlack * If shadowing 32-bit non-PAE page tables, each PAE page 39007f497775SDavid Matlack * directory maps one quarter of the guest's non-PAE page 39017f497775SDavid Matlack * directory. Othwerise each PAE page direct shadows one guest 39027f497775SDavid Matlack * PAE page directory so that quadrant should be 0. 39037f497775SDavid Matlack */ 39047f497775SDavid Matlack quadrant = (mmu->cpu_role.base.level == PT32_ROOT_LEVEL) ? i : 0; 39057f497775SDavid Matlack 39067f497775SDavid Matlack root = mmu_alloc_root(vcpu, root_gfn, quadrant, PT32_ROOT_LEVEL); 3907b37233c9SSean Christopherson mmu->pae_root[i] = root | pm_mask; 3908c50d8ae3SPaolo Bonzini } 3909c50d8ae3SPaolo Bonzini 3910a972e29cSPaolo Bonzini if (mmu->root_role.level == PT64_ROOT_5LEVEL) 3911b9e5603cSPaolo Bonzini mmu->root.hpa = __pa(mmu->pml5_root); 3912a972e29cSPaolo Bonzini else if (mmu->root_role.level == PT64_ROOT_4LEVEL) 3913b9e5603cSPaolo Bonzini mmu->root.hpa = __pa(mmu->pml4_root); 3914ba0a194fSSean Christopherson else 3915b9e5603cSPaolo Bonzini mmu->root.hpa = __pa(mmu->pae_root); 3916c50d8ae3SPaolo Bonzini 3917be01e8e2SSean Christopherson set_root_pgd: 3918b9e5603cSPaolo Bonzini mmu->root.pgd = root_pgd; 39194a38162eSPaolo Bonzini out_unlock: 39204a38162eSPaolo Bonzini write_unlock(&vcpu->kvm->mmu_lock); 3921c50d8ae3SPaolo Bonzini 3922c6c937d6SLike Xu return r; 3923c50d8ae3SPaolo Bonzini } 3924c50d8ae3SPaolo Bonzini 3925748e52b9SSean Christopherson static int mmu_alloc_special_roots(struct kvm_vcpu *vcpu) 3926c50d8ae3SPaolo Bonzini { 3927748e52b9SSean Christopherson struct kvm_mmu *mmu = vcpu->arch.mmu; 3928a972e29cSPaolo Bonzini bool need_pml5 = mmu->root_role.level > PT64_ROOT_4LEVEL; 3929cb0f722aSWei Huang u64 *pml5_root = NULL; 3930cb0f722aSWei Huang u64 *pml4_root = NULL; 3931cb0f722aSWei Huang u64 *pae_root; 3932748e52b9SSean Christopherson 3933748e52b9SSean Christopherson /* 3934748e52b9SSean Christopherson * When shadowing 32-bit or PAE NPT with 64-bit NPT, the PML4 and PDP 3935748e52b9SSean Christopherson * tables are allocated and initialized at root creation as there is no 3936748e52b9SSean Christopherson * equivalent level in the guest's NPT to shadow. Allocate the tables 3937748e52b9SSean Christopherson * on demand, as running a 32-bit L1 VMM on 64-bit KVM is very rare. 3938748e52b9SSean Christopherson */ 3939347a0d0dSPaolo Bonzini if (mmu->root_role.direct || 3940347a0d0dSPaolo Bonzini mmu->cpu_role.base.level >= PT64_ROOT_4LEVEL || 3941a972e29cSPaolo Bonzini mmu->root_role.level < PT64_ROOT_4LEVEL) 3942748e52b9SSean Christopherson return 0; 3943748e52b9SSean Christopherson 3944a717a780SSean Christopherson /* 3945a717a780SSean Christopherson * NPT, the only paging mode that uses this horror, uses a fixed number 3946a717a780SSean Christopherson * of levels for the shadow page tables, e.g. all MMUs are 4-level or 3947a717a780SSean Christopherson * all MMus are 5-level. Thus, this can safely require that pml5_root 3948a717a780SSean Christopherson * is allocated if the other roots are valid and pml5 is needed, as any 3949a717a780SSean Christopherson * prior MMU would also have required pml5. 3950a717a780SSean Christopherson */ 3951a717a780SSean Christopherson if (mmu->pae_root && mmu->pml4_root && (!need_pml5 || mmu->pml5_root)) 3952748e52b9SSean Christopherson return 0; 3953748e52b9SSean Christopherson 3954748e52b9SSean Christopherson /* 3955748e52b9SSean Christopherson * The special roots should always be allocated in concert. Yell and 3956748e52b9SSean Christopherson * bail if KVM ends up in a state where only one of the roots is valid. 3957748e52b9SSean Christopherson */ 3958cb0f722aSWei Huang if (WARN_ON_ONCE(!tdp_enabled || mmu->pae_root || mmu->pml4_root || 3959a717a780SSean Christopherson (need_pml5 && mmu->pml5_root))) 3960748e52b9SSean Christopherson return -EIO; 3961748e52b9SSean Christopherson 39624a98623dSSean Christopherson /* 39634a98623dSSean Christopherson * Unlike 32-bit NPT, the PDP table doesn't need to be in low mem, and 39644a98623dSSean Christopherson * doesn't need to be decrypted. 39654a98623dSSean Christopherson */ 3966748e52b9SSean Christopherson pae_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT); 3967748e52b9SSean Christopherson if (!pae_root) 3968748e52b9SSean Christopherson return -ENOMEM; 3969748e52b9SSean Christopherson 3970cb0f722aSWei Huang #ifdef CONFIG_X86_64 397103ca4589SSean Christopherson pml4_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT); 3972cb0f722aSWei Huang if (!pml4_root) 3973cb0f722aSWei Huang goto err_pml4; 3974cb0f722aSWei Huang 3975a717a780SSean Christopherson if (need_pml5) { 3976cb0f722aSWei Huang pml5_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT); 3977cb0f722aSWei Huang if (!pml5_root) 3978cb0f722aSWei Huang goto err_pml5; 3979748e52b9SSean Christopherson } 3980cb0f722aSWei Huang #endif 3981748e52b9SSean Christopherson 3982748e52b9SSean Christopherson mmu->pae_root = pae_root; 398303ca4589SSean Christopherson mmu->pml4_root = pml4_root; 3984cb0f722aSWei Huang mmu->pml5_root = pml5_root; 3985748e52b9SSean Christopherson 3986748e52b9SSean Christopherson return 0; 3987cb0f722aSWei Huang 3988cb0f722aSWei Huang #ifdef CONFIG_X86_64 3989cb0f722aSWei Huang err_pml5: 3990cb0f722aSWei Huang free_page((unsigned long)pml4_root); 3991cb0f722aSWei Huang err_pml4: 3992cb0f722aSWei Huang free_page((unsigned long)pae_root); 3993cb0f722aSWei Huang return -ENOMEM; 3994cb0f722aSWei Huang #endif 3995c50d8ae3SPaolo Bonzini } 3996c50d8ae3SPaolo Bonzini 3997264d3dc1SLai Jiangshan static bool is_unsync_root(hpa_t root) 3998264d3dc1SLai Jiangshan { 3999264d3dc1SLai Jiangshan struct kvm_mmu_page *sp; 4000264d3dc1SLai Jiangshan 400161b05a9fSLai Jiangshan if (!VALID_PAGE(root)) 400261b05a9fSLai Jiangshan return false; 400361b05a9fSLai Jiangshan 4004264d3dc1SLai Jiangshan /* 4005264d3dc1SLai Jiangshan * The read barrier orders the CPU's read of SPTE.W during the page table 4006264d3dc1SLai Jiangshan * walk before the reads of sp->unsync/sp->unsync_children here. 4007264d3dc1SLai Jiangshan * 4008264d3dc1SLai Jiangshan * Even if another CPU was marking the SP as unsync-ed simultaneously, 4009264d3dc1SLai Jiangshan * any guest page table changes are not guaranteed to be visible anyway 4010264d3dc1SLai Jiangshan * until this VCPU issues a TLB flush strictly after those changes are 4011264d3dc1SLai Jiangshan * made. We only need to ensure that the other CPU sets these flags 4012264d3dc1SLai Jiangshan * before any actual changes to the page tables are made. The comments 4013264d3dc1SLai Jiangshan * in mmu_try_to_unsync_pages() describe what could go wrong if this 4014264d3dc1SLai Jiangshan * requirement isn't satisfied. 4015264d3dc1SLai Jiangshan */ 4016264d3dc1SLai Jiangshan smp_rmb(); 4017264d3dc1SLai Jiangshan sp = to_shadow_page(root); 40185d6a3221SSean Christopherson 40195d6a3221SSean Christopherson /* 40205d6a3221SSean Christopherson * PAE roots (somewhat arbitrarily) aren't backed by shadow pages, the 40215d6a3221SSean Christopherson * PDPTEs for a given PAE root need to be synchronized individually. 40225d6a3221SSean Christopherson */ 40235d6a3221SSean Christopherson if (WARN_ON_ONCE(!sp)) 40245d6a3221SSean Christopherson return false; 40255d6a3221SSean Christopherson 4026264d3dc1SLai Jiangshan if (sp->unsync || sp->unsync_children) 4027264d3dc1SLai Jiangshan return true; 4028264d3dc1SLai Jiangshan 4029264d3dc1SLai Jiangshan return false; 4030264d3dc1SLai Jiangshan } 4031264d3dc1SLai Jiangshan 4032c50d8ae3SPaolo Bonzini void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu) 4033c50d8ae3SPaolo Bonzini { 4034c50d8ae3SPaolo Bonzini int i; 4035c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 4036c50d8ae3SPaolo Bonzini 4037347a0d0dSPaolo Bonzini if (vcpu->arch.mmu->root_role.direct) 4038c50d8ae3SPaolo Bonzini return; 4039c50d8ae3SPaolo Bonzini 4040b9e5603cSPaolo Bonzini if (!VALID_PAGE(vcpu->arch.mmu->root.hpa)) 4041c50d8ae3SPaolo Bonzini return; 4042c50d8ae3SPaolo Bonzini 4043c50d8ae3SPaolo Bonzini vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY); 4044c50d8ae3SPaolo Bonzini 40454d25502aSPaolo Bonzini if (vcpu->arch.mmu->cpu_role.base.level >= PT64_ROOT_4LEVEL) { 4046b9e5603cSPaolo Bonzini hpa_t root = vcpu->arch.mmu->root.hpa; 4047e47c4aeeSSean Christopherson sp = to_shadow_page(root); 4048c50d8ae3SPaolo Bonzini 4049264d3dc1SLai Jiangshan if (!is_unsync_root(root)) 4050c50d8ae3SPaolo Bonzini return; 4051c50d8ae3SPaolo Bonzini 4052531810caSBen Gardon write_lock(&vcpu->kvm->mmu_lock); 405365855ed8SLai Jiangshan mmu_sync_children(vcpu, sp, true); 4054531810caSBen Gardon write_unlock(&vcpu->kvm->mmu_lock); 4055c50d8ae3SPaolo Bonzini return; 4056c50d8ae3SPaolo Bonzini } 4057c50d8ae3SPaolo Bonzini 4058531810caSBen Gardon write_lock(&vcpu->kvm->mmu_lock); 4059c50d8ae3SPaolo Bonzini 4060c50d8ae3SPaolo Bonzini for (i = 0; i < 4; ++i) { 4061c50d8ae3SPaolo Bonzini hpa_t root = vcpu->arch.mmu->pae_root[i]; 4062c50d8ae3SPaolo Bonzini 4063c834e5e4SSean Christopherson if (IS_VALID_PAE_ROOT(root)) { 40645e3edd7eSSean Christopherson sp = spte_to_child_sp(root); 406565855ed8SLai Jiangshan mmu_sync_children(vcpu, sp, true); 4066c50d8ae3SPaolo Bonzini } 4067c50d8ae3SPaolo Bonzini } 4068c50d8ae3SPaolo Bonzini 4069531810caSBen Gardon write_unlock(&vcpu->kvm->mmu_lock); 4070c50d8ae3SPaolo Bonzini } 4071c50d8ae3SPaolo Bonzini 407261b05a9fSLai Jiangshan void kvm_mmu_sync_prev_roots(struct kvm_vcpu *vcpu) 407361b05a9fSLai Jiangshan { 407461b05a9fSLai Jiangshan unsigned long roots_to_free = 0; 407561b05a9fSLai Jiangshan int i; 407661b05a9fSLai Jiangshan 407761b05a9fSLai Jiangshan for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) 407861b05a9fSLai Jiangshan if (is_unsync_root(vcpu->arch.mmu->prev_roots[i].hpa)) 407961b05a9fSLai Jiangshan roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i); 408061b05a9fSLai Jiangshan 408161b05a9fSLai Jiangshan /* sync prev_roots by simply freeing them */ 40820c1c92f1SPaolo Bonzini kvm_mmu_free_roots(vcpu->kvm, vcpu->arch.mmu, roots_to_free); 408361b05a9fSLai Jiangshan } 408461b05a9fSLai Jiangshan 40851f5a21eeSLai Jiangshan static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu, 40865b22bbe7SLai Jiangshan gpa_t vaddr, u64 access, 4087c50d8ae3SPaolo Bonzini struct x86_exception *exception) 4088c50d8ae3SPaolo Bonzini { 4089c50d8ae3SPaolo Bonzini if (exception) 4090c50d8ae3SPaolo Bonzini exception->error_code = 0; 4091c59a0f57SLai Jiangshan return kvm_translate_gpa(vcpu, mmu, vaddr, access, exception); 4092c50d8ae3SPaolo Bonzini } 4093c50d8ae3SPaolo Bonzini 4094c50d8ae3SPaolo Bonzini static bool mmio_info_in_cache(struct kvm_vcpu *vcpu, u64 addr, bool direct) 4095c50d8ae3SPaolo Bonzini { 4096c50d8ae3SPaolo Bonzini /* 4097c50d8ae3SPaolo Bonzini * A nested guest cannot use the MMIO cache if it is using nested 4098c50d8ae3SPaolo Bonzini * page tables, because cr2 is a nGPA while the cache stores GPAs. 4099c50d8ae3SPaolo Bonzini */ 4100c50d8ae3SPaolo Bonzini if (mmu_is_nested(vcpu)) 4101c50d8ae3SPaolo Bonzini return false; 4102c50d8ae3SPaolo Bonzini 4103c50d8ae3SPaolo Bonzini if (direct) 4104c50d8ae3SPaolo Bonzini return vcpu_match_mmio_gpa(vcpu, addr); 4105c50d8ae3SPaolo Bonzini 4106c50d8ae3SPaolo Bonzini return vcpu_match_mmio_gva(vcpu, addr); 4107c50d8ae3SPaolo Bonzini } 4108c50d8ae3SPaolo Bonzini 410995fb5b02SBen Gardon /* 411095fb5b02SBen Gardon * Return the level of the lowest level SPTE added to sptes. 411195fb5b02SBen Gardon * That SPTE may be non-present. 4112c5c8c7c5SDavid Matlack * 4113c5c8c7c5SDavid Matlack * Must be called between walk_shadow_page_lockless_{begin,end}. 411495fb5b02SBen Gardon */ 411539b4d43eSSean Christopherson static int get_walk(struct kvm_vcpu *vcpu, u64 addr, u64 *sptes, int *root_level) 4116c50d8ae3SPaolo Bonzini { 4117c50d8ae3SPaolo Bonzini struct kvm_shadow_walk_iterator iterator; 41182aa07893SSean Christopherson int leaf = -1; 411995fb5b02SBen Gardon u64 spte; 4120c50d8ae3SPaolo Bonzini 412139b4d43eSSean Christopherson for (shadow_walk_init(&iterator, vcpu, addr), 412239b4d43eSSean Christopherson *root_level = iterator.level; 4123c50d8ae3SPaolo Bonzini shadow_walk_okay(&iterator); 4124c50d8ae3SPaolo Bonzini __shadow_walk_next(&iterator, spte)) { 412595fb5b02SBen Gardon leaf = iterator.level; 4126c50d8ae3SPaolo Bonzini spte = mmu_spte_get_lockless(iterator.sptep); 4127c50d8ae3SPaolo Bonzini 4128dde81f94SSean Christopherson sptes[leaf] = spte; 412995fb5b02SBen Gardon } 413095fb5b02SBen Gardon 413195fb5b02SBen Gardon return leaf; 413295fb5b02SBen Gardon } 413395fb5b02SBen Gardon 41349aa41879SSean Christopherson /* return true if reserved bit(s) are detected on a valid, non-MMIO SPTE. */ 413595fb5b02SBen Gardon static bool get_mmio_spte(struct kvm_vcpu *vcpu, u64 addr, u64 *sptep) 413695fb5b02SBen Gardon { 4137dde81f94SSean Christopherson u64 sptes[PT64_ROOT_MAX_LEVEL + 1]; 413895fb5b02SBen Gardon struct rsvd_bits_validate *rsvd_check; 413939b4d43eSSean Christopherson int root, leaf, level; 414095fb5b02SBen Gardon bool reserved = false; 414195fb5b02SBen Gardon 4142c5c8c7c5SDavid Matlack walk_shadow_page_lockless_begin(vcpu); 4143c5c8c7c5SDavid Matlack 414478fdd2f0SSean Christopherson if (is_tdp_mmu_active(vcpu)) 414539b4d43eSSean Christopherson leaf = kvm_tdp_mmu_get_walk(vcpu, addr, sptes, &root); 414695fb5b02SBen Gardon else 414739b4d43eSSean Christopherson leaf = get_walk(vcpu, addr, sptes, &root); 414895fb5b02SBen Gardon 4149c5c8c7c5SDavid Matlack walk_shadow_page_lockless_end(vcpu); 4150c5c8c7c5SDavid Matlack 41512aa07893SSean Christopherson if (unlikely(leaf < 0)) { 41522aa07893SSean Christopherson *sptep = 0ull; 41532aa07893SSean Christopherson return reserved; 41542aa07893SSean Christopherson } 41552aa07893SSean Christopherson 41569aa41879SSean Christopherson *sptep = sptes[leaf]; 41579aa41879SSean Christopherson 41589aa41879SSean Christopherson /* 41599aa41879SSean Christopherson * Skip reserved bits checks on the terminal leaf if it's not a valid 41609aa41879SSean Christopherson * SPTE. Note, this also (intentionally) skips MMIO SPTEs, which, by 41619aa41879SSean Christopherson * design, always have reserved bits set. The purpose of the checks is 41629aa41879SSean Christopherson * to detect reserved bits on non-MMIO SPTEs. i.e. buggy SPTEs. 41639aa41879SSean Christopherson */ 41649aa41879SSean Christopherson if (!is_shadow_present_pte(sptes[leaf])) 41659aa41879SSean Christopherson leaf++; 416695fb5b02SBen Gardon 416795fb5b02SBen Gardon rsvd_check = &vcpu->arch.mmu->shadow_zero_check; 416895fb5b02SBen Gardon 41699aa41879SSean Christopherson for (level = root; level >= leaf; level--) 4170961f8445SSean Christopherson reserved |= is_rsvd_spte(rsvd_check, sptes[level], level); 4171c50d8ae3SPaolo Bonzini 4172c50d8ae3SPaolo Bonzini if (reserved) { 4173bb4cdf3aSSean Christopherson pr_err("%s: reserved bits set on MMU-present spte, addr 0x%llx, hierarchy:\n", 4174c50d8ae3SPaolo Bonzini __func__, addr); 417595fb5b02SBen Gardon for (level = root; level >= leaf; level--) 4176bb4cdf3aSSean Christopherson pr_err("------ spte = 0x%llx level = %d, rsvd bits = 0x%llx", 4177bb4cdf3aSSean Christopherson sptes[level], level, 4178961f8445SSean Christopherson get_rsvd_bits(rsvd_check, sptes[level], level)); 4179c50d8ae3SPaolo Bonzini } 4180ddce6208SSean Christopherson 4181c50d8ae3SPaolo Bonzini return reserved; 4182c50d8ae3SPaolo Bonzini } 4183c50d8ae3SPaolo Bonzini 4184c50d8ae3SPaolo Bonzini static int handle_mmio_page_fault(struct kvm_vcpu *vcpu, u64 addr, bool direct) 4185c50d8ae3SPaolo Bonzini { 4186c50d8ae3SPaolo Bonzini u64 spte; 4187c50d8ae3SPaolo Bonzini bool reserved; 4188c50d8ae3SPaolo Bonzini 4189c50d8ae3SPaolo Bonzini if (mmio_info_in_cache(vcpu, addr, direct)) 4190c50d8ae3SPaolo Bonzini return RET_PF_EMULATE; 4191c50d8ae3SPaolo Bonzini 419295fb5b02SBen Gardon reserved = get_mmio_spte(vcpu, addr, &spte); 4193c50d8ae3SPaolo Bonzini if (WARN_ON(reserved)) 4194c50d8ae3SPaolo Bonzini return -EINVAL; 4195c50d8ae3SPaolo Bonzini 4196c50d8ae3SPaolo Bonzini if (is_mmio_spte(spte)) { 4197c50d8ae3SPaolo Bonzini gfn_t gfn = get_mmio_spte_gfn(spte); 41980a2b64c5SBen Gardon unsigned int access = get_mmio_spte_access(spte); 4199c50d8ae3SPaolo Bonzini 4200c50d8ae3SPaolo Bonzini if (!check_mmio_spte(vcpu, spte)) 4201c50d8ae3SPaolo Bonzini return RET_PF_INVALID; 4202c50d8ae3SPaolo Bonzini 4203c50d8ae3SPaolo Bonzini if (direct) 4204c50d8ae3SPaolo Bonzini addr = 0; 4205c50d8ae3SPaolo Bonzini 4206c50d8ae3SPaolo Bonzini trace_handle_mmio_page_fault(addr, gfn, access); 4207c50d8ae3SPaolo Bonzini vcpu_cache_mmio_info(vcpu, addr, gfn, access); 4208c50d8ae3SPaolo Bonzini return RET_PF_EMULATE; 4209c50d8ae3SPaolo Bonzini } 4210c50d8ae3SPaolo Bonzini 4211c50d8ae3SPaolo Bonzini /* 4212c50d8ae3SPaolo Bonzini * If the page table is zapped by other cpus, let CPU fault again on 4213c50d8ae3SPaolo Bonzini * the address. 4214c50d8ae3SPaolo Bonzini */ 4215c50d8ae3SPaolo Bonzini return RET_PF_RETRY; 4216c50d8ae3SPaolo Bonzini } 4217c50d8ae3SPaolo Bonzini 4218c50d8ae3SPaolo Bonzini static bool page_fault_handle_page_track(struct kvm_vcpu *vcpu, 4219b8a5d551SPaolo Bonzini struct kvm_page_fault *fault) 4220c50d8ae3SPaolo Bonzini { 4221b8a5d551SPaolo Bonzini if (unlikely(fault->rsvd)) 4222c50d8ae3SPaolo Bonzini return false; 4223c50d8ae3SPaolo Bonzini 4224b8a5d551SPaolo Bonzini if (!fault->present || !fault->write) 4225c50d8ae3SPaolo Bonzini return false; 4226c50d8ae3SPaolo Bonzini 4227c50d8ae3SPaolo Bonzini /* 4228c50d8ae3SPaolo Bonzini * guest is writing the page which is write tracked which can 4229c50d8ae3SPaolo Bonzini * not be fixed by page fault handler. 4230c50d8ae3SPaolo Bonzini */ 42319d395a0aSBen Gardon if (kvm_slot_page_track_is_active(vcpu->kvm, fault->slot, fault->gfn, KVM_PAGE_TRACK_WRITE)) 4232c50d8ae3SPaolo Bonzini return true; 4233c50d8ae3SPaolo Bonzini 4234c50d8ae3SPaolo Bonzini return false; 4235c50d8ae3SPaolo Bonzini } 4236c50d8ae3SPaolo Bonzini 4237c50d8ae3SPaolo Bonzini static void shadow_page_table_clear_flood(struct kvm_vcpu *vcpu, gva_t addr) 4238c50d8ae3SPaolo Bonzini { 4239c50d8ae3SPaolo Bonzini struct kvm_shadow_walk_iterator iterator; 4240c50d8ae3SPaolo Bonzini u64 spte; 4241c50d8ae3SPaolo Bonzini 4242c50d8ae3SPaolo Bonzini walk_shadow_page_lockless_begin(vcpu); 42433e44dce4SLai Jiangshan for_each_shadow_entry_lockless(vcpu, addr, iterator, spte) 4244c50d8ae3SPaolo Bonzini clear_sp_write_flooding_count(iterator.sptep); 4245c50d8ae3SPaolo Bonzini walk_shadow_page_lockless_end(vcpu); 4246c50d8ae3SPaolo Bonzini } 4247c50d8ae3SPaolo Bonzini 42486f3c1fc5SLiang Zhang static u32 alloc_apf_token(struct kvm_vcpu *vcpu) 42496f3c1fc5SLiang Zhang { 42506f3c1fc5SLiang Zhang /* make sure the token value is not 0 */ 42516f3c1fc5SLiang Zhang u32 id = vcpu->arch.apf.id; 42526f3c1fc5SLiang Zhang 42536f3c1fc5SLiang Zhang if (id << 12 == 0) 42546f3c1fc5SLiang Zhang vcpu->arch.apf.id = 1; 42556f3c1fc5SLiang Zhang 42566f3c1fc5SLiang Zhang return (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id; 42576f3c1fc5SLiang Zhang } 42586f3c1fc5SLiang Zhang 4259e8c22266SVitaly Kuznetsov static bool kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, 42609f1a8526SSean Christopherson gfn_t gfn) 4261c50d8ae3SPaolo Bonzini { 4262c50d8ae3SPaolo Bonzini struct kvm_arch_async_pf arch; 4263c50d8ae3SPaolo Bonzini 42646f3c1fc5SLiang Zhang arch.token = alloc_apf_token(vcpu); 4265c50d8ae3SPaolo Bonzini arch.gfn = gfn; 4266347a0d0dSPaolo Bonzini arch.direct_map = vcpu->arch.mmu->root_role.direct; 42672fdcc1b3SPaolo Bonzini arch.cr3 = kvm_mmu_get_guest_pgd(vcpu, vcpu->arch.mmu); 4268c50d8ae3SPaolo Bonzini 42699f1a8526SSean Christopherson return kvm_setup_async_pf(vcpu, cr2_or_gpa, 42709f1a8526SSean Christopherson kvm_vcpu_gfn_to_hva(vcpu, gfn), &arch); 4271c50d8ae3SPaolo Bonzini } 4272c50d8ae3SPaolo Bonzini 42738a009d5bSSean Christopherson void kvm_arch_async_page_ready(struct kvm_vcpu *vcpu, struct kvm_async_pf *work) 42748a009d5bSSean Christopherson { 42758a009d5bSSean Christopherson int r; 42768a009d5bSSean Christopherson 42778a009d5bSSean Christopherson if ((vcpu->arch.mmu->root_role.direct != work->arch.direct_map) || 42788a009d5bSSean Christopherson work->wakeup_all) 42798a009d5bSSean Christopherson return; 42808a009d5bSSean Christopherson 42818a009d5bSSean Christopherson r = kvm_mmu_reload(vcpu); 42828a009d5bSSean Christopherson if (unlikely(r)) 42838a009d5bSSean Christopherson return; 42848a009d5bSSean Christopherson 42858a009d5bSSean Christopherson if (!vcpu->arch.mmu->root_role.direct && 42862fdcc1b3SPaolo Bonzini work->arch.cr3 != kvm_mmu_get_guest_pgd(vcpu, vcpu->arch.mmu)) 42878a009d5bSSean Christopherson return; 42888a009d5bSSean Christopherson 4289258d985fSSean Christopherson kvm_mmu_do_page_fault(vcpu, work->cr2_or_gpa, 0, true, NULL); 42908a009d5bSSean Christopherson } 42918a009d5bSSean Christopherson 4292ba6e3fe2SDavid Matlack static int __kvm_faultin_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 4293c50d8ae3SPaolo Bonzini { 4294e710c5f6SDavid Matlack struct kvm_memory_slot *slot = fault->slot; 4295c50d8ae3SPaolo Bonzini bool async; 4296c50d8ae3SPaolo Bonzini 4297e0c37868SSean Christopherson /* 4298e0c37868SSean Christopherson * Retry the page fault if the gfn hit a memslot that is being deleted 4299e0c37868SSean Christopherson * or moved. This ensures any existing SPTEs for the old memslot will 4300e0c37868SSean Christopherson * be zapped before KVM inserts a new MMIO SPTE for the gfn. 4301e0c37868SSean Christopherson */ 4302e0c37868SSean Christopherson if (slot && (slot->flags & KVM_MEMSLOT_INVALID)) 43035276c616SSean Christopherson return RET_PF_RETRY; 4304e0c37868SSean Christopherson 43059cc13d60SMaxim Levitsky if (!kvm_is_visible_memslot(slot)) { 4306c36b7150SPaolo Bonzini /* Don't expose private memslots to L2. */ 43079cc13d60SMaxim Levitsky if (is_guest_mode(vcpu)) { 4308e710c5f6SDavid Matlack fault->slot = NULL; 43093647cd04SPaolo Bonzini fault->pfn = KVM_PFN_NOSLOT; 43103647cd04SPaolo Bonzini fault->map_writable = false; 43115276c616SSean Christopherson return RET_PF_CONTINUE; 4312c50d8ae3SPaolo Bonzini } 43139cc13d60SMaxim Levitsky /* 43149cc13d60SMaxim Levitsky * If the APIC access page exists but is disabled, go directly 43159cc13d60SMaxim Levitsky * to emulation without caching the MMIO access or creating a 43169cc13d60SMaxim Levitsky * MMIO SPTE. That way the cache doesn't need to be purged 43179cc13d60SMaxim Levitsky * when the AVIC is re-enabled. 43189cc13d60SMaxim Levitsky */ 43199cc13d60SMaxim Levitsky if (slot && slot->id == APIC_ACCESS_PAGE_PRIVATE_MEMSLOT && 43205276c616SSean Christopherson !kvm_apicv_activated(vcpu->kvm)) 43215276c616SSean Christopherson return RET_PF_EMULATE; 43229cc13d60SMaxim Levitsky } 4323c50d8ae3SPaolo Bonzini 4324c50d8ae3SPaolo Bonzini async = false; 4325c8b88b33SPeter Xu fault->pfn = __gfn_to_pfn_memslot(slot, fault->gfn, false, false, &async, 43263647cd04SPaolo Bonzini fault->write, &fault->map_writable, 43273647cd04SPaolo Bonzini &fault->hva); 4328c50d8ae3SPaolo Bonzini if (!async) 43295276c616SSean Christopherson return RET_PF_CONTINUE; /* *pfn has correct page already */ 4330c50d8ae3SPaolo Bonzini 43312839180cSPaolo Bonzini if (!fault->prefetch && kvm_can_do_async_pf(vcpu)) { 43323647cd04SPaolo Bonzini trace_kvm_try_async_get_page(fault->addr, fault->gfn); 43333647cd04SPaolo Bonzini if (kvm_find_async_pf_gfn(vcpu, fault->gfn)) { 43341685c0f3SMingwei Zhang trace_kvm_async_pf_repeated_fault(fault->addr, fault->gfn); 4335c50d8ae3SPaolo Bonzini kvm_make_request(KVM_REQ_APF_HALT, vcpu); 43365276c616SSean Christopherson return RET_PF_RETRY; 43375276c616SSean Christopherson } else if (kvm_arch_setup_async_pf(vcpu, fault->addr, fault->gfn)) { 43385276c616SSean Christopherson return RET_PF_RETRY; 43395276c616SSean Christopherson } 4340c50d8ae3SPaolo Bonzini } 4341c50d8ae3SPaolo Bonzini 434276657687SPeter Xu /* 434376657687SPeter Xu * Allow gup to bail on pending non-fatal signals when it's also allowed 434476657687SPeter Xu * to wait for IO. Note, gup always bails if it is unable to quickly 434576657687SPeter Xu * get a page and a fatal signal, i.e. SIGKILL, is pending. 434676657687SPeter Xu */ 434776657687SPeter Xu fault->pfn = __gfn_to_pfn_memslot(slot, fault->gfn, false, true, NULL, 43483647cd04SPaolo Bonzini fault->write, &fault->map_writable, 43493647cd04SPaolo Bonzini &fault->hva); 43505276c616SSean Christopherson return RET_PF_CONTINUE; 4351c50d8ae3SPaolo Bonzini } 4352c50d8ae3SPaolo Bonzini 4353354c908cSDavid Matlack static int kvm_faultin_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault, 4354354c908cSDavid Matlack unsigned int access) 4355ba6e3fe2SDavid Matlack { 435656c3a4e4SDavid Matlack int ret; 435756c3a4e4SDavid Matlack 4358ba6e3fe2SDavid Matlack fault->mmu_seq = vcpu->kvm->mmu_invalidate_seq; 4359ba6e3fe2SDavid Matlack smp_rmb(); 4360ba6e3fe2SDavid Matlack 436156c3a4e4SDavid Matlack ret = __kvm_faultin_pfn(vcpu, fault); 436256c3a4e4SDavid Matlack if (ret != RET_PF_CONTINUE) 436356c3a4e4SDavid Matlack return ret; 436456c3a4e4SDavid Matlack 436556c3a4e4SDavid Matlack if (unlikely(is_error_pfn(fault->pfn))) 4366cd08d178SDavid Matlack return kvm_handle_error_pfn(vcpu, fault); 436756c3a4e4SDavid Matlack 4368354c908cSDavid Matlack if (unlikely(!fault->slot)) 4369354c908cSDavid Matlack return kvm_handle_noslot_fault(vcpu, fault, access); 4370354c908cSDavid Matlack 437156c3a4e4SDavid Matlack return RET_PF_CONTINUE; 4372ba6e3fe2SDavid Matlack } 4373ba6e3fe2SDavid Matlack 4374a955cad8SSean Christopherson /* 4375a955cad8SSean Christopherson * Returns true if the page fault is stale and needs to be retried, i.e. if the 4376a955cad8SSean Christopherson * root was invalidated by a memslot update or a relevant mmu_notifier fired. 4377a955cad8SSean Christopherson */ 4378a955cad8SSean Christopherson static bool is_page_fault_stale(struct kvm_vcpu *vcpu, 4379ba6e3fe2SDavid Matlack struct kvm_page_fault *fault) 4380a955cad8SSean Christopherson { 4381b9e5603cSPaolo Bonzini struct kvm_mmu_page *sp = to_shadow_page(vcpu->arch.mmu->root.hpa); 438218c841e1SSean Christopherson 438318c841e1SSean Christopherson /* Special roots, e.g. pae_root, are not backed by shadow pages. */ 438418c841e1SSean Christopherson if (sp && is_obsolete_sp(vcpu->kvm, sp)) 438518c841e1SSean Christopherson return true; 438618c841e1SSean Christopherson 438718c841e1SSean Christopherson /* 438818c841e1SSean Christopherson * Roots without an associated shadow page are considered invalid if 438918c841e1SSean Christopherson * there is a pending request to free obsolete roots. The request is 439018c841e1SSean Christopherson * only a hint that the current root _may_ be obsolete and needs to be 439118c841e1SSean Christopherson * reloaded, e.g. if the guest frees a PGD that KVM is tracking as a 439218c841e1SSean Christopherson * previous root, then __kvm_mmu_prepare_zap_page() signals all vCPUs 439318c841e1SSean Christopherson * to reload even if no vCPU is actively using the root. 439418c841e1SSean Christopherson */ 4395527d5cd7SSean Christopherson if (!sp && kvm_test_request(KVM_REQ_MMU_FREE_OBSOLETE_ROOTS, vcpu)) 4396a955cad8SSean Christopherson return true; 4397a955cad8SSean Christopherson 4398a955cad8SSean Christopherson return fault->slot && 4399ba6e3fe2SDavid Matlack mmu_invalidate_retry_hva(vcpu->kvm, fault->mmu_seq, fault->hva); 4400a955cad8SSean Christopherson } 4401a955cad8SSean Christopherson 44024326e57eSPaolo Bonzini static int direct_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 4403c50d8ae3SPaolo Bonzini { 440483f06fa7SSean Christopherson int r; 4405c50d8ae3SPaolo Bonzini 4406b8a5d551SPaolo Bonzini if (page_fault_handle_page_track(vcpu, fault)) 4407c50d8ae3SPaolo Bonzini return RET_PF_EMULATE; 4408c50d8ae3SPaolo Bonzini 44093c8ad5a6SPaolo Bonzini r = fast_page_fault(vcpu, fault); 4410c4371c2aSSean Christopherson if (r != RET_PF_INVALID) 4411c4371c2aSSean Christopherson return r; 441283291445SSean Christopherson 4413378f5cd6SSean Christopherson r = mmu_topup_memory_caches(vcpu, false); 4414c50d8ae3SPaolo Bonzini if (r) 4415c50d8ae3SPaolo Bonzini return r; 4416c50d8ae3SPaolo Bonzini 4417354c908cSDavid Matlack r = kvm_faultin_pfn(vcpu, fault, ACC_ALL); 44185276c616SSean Christopherson if (r != RET_PF_CONTINUE) 4419367fd790SSean Christopherson return r; 4420367fd790SSean Christopherson 4421367fd790SSean Christopherson r = RET_PF_RETRY; 4422531810caSBen Gardon write_lock(&vcpu->kvm->mmu_lock); 4423a2855afcSBen Gardon 4424ba6e3fe2SDavid Matlack if (is_page_fault_stale(vcpu, fault)) 4425367fd790SSean Christopherson goto out_unlock; 4426a955cad8SSean Christopherson 44277bd7ded6SSean Christopherson r = make_mmu_pages_available(vcpu); 44287bd7ded6SSean Christopherson if (r) 4429367fd790SSean Christopherson goto out_unlock; 4430bb18842eSBen Gardon 44316c882ef4SDavid Matlack r = direct_map(vcpu, fault); 44320f90e1c1SSean Christopherson 4433367fd790SSean Christopherson out_unlock: 4434531810caSBen Gardon write_unlock(&vcpu->kvm->mmu_lock); 44353647cd04SPaolo Bonzini kvm_release_pfn_clean(fault->pfn); 4436367fd790SSean Christopherson return r; 4437c50d8ae3SPaolo Bonzini } 4438c50d8ae3SPaolo Bonzini 4439c501040aSPaolo Bonzini static int nonpaging_page_fault(struct kvm_vcpu *vcpu, 4440c501040aSPaolo Bonzini struct kvm_page_fault *fault) 44410f90e1c1SSean Christopherson { 44424326e57eSPaolo Bonzini pgprintk("%s: gva %lx error %x\n", __func__, fault->addr, fault->error_code); 44430f90e1c1SSean Christopherson 44440f90e1c1SSean Christopherson /* This path builds a PAE pagetable, we can map 2mb pages at maximum. */ 44454326e57eSPaolo Bonzini fault->max_level = PG_LEVEL_2M; 44464326e57eSPaolo Bonzini return direct_page_fault(vcpu, fault); 44470f90e1c1SSean Christopherson } 44480f90e1c1SSean Christopherson 4449c50d8ae3SPaolo Bonzini int kvm_handle_page_fault(struct kvm_vcpu *vcpu, u64 error_code, 4450c50d8ae3SPaolo Bonzini u64 fault_address, char *insn, int insn_len) 4451c50d8ae3SPaolo Bonzini { 4452c50d8ae3SPaolo Bonzini int r = 1; 44539ce372b3SVitaly Kuznetsov u32 flags = vcpu->arch.apf.host_apf_flags; 4454c50d8ae3SPaolo Bonzini 4455736c291cSSean Christopherson #ifndef CONFIG_X86_64 4456736c291cSSean Christopherson /* A 64-bit CR2 should be impossible on 32-bit KVM. */ 4457736c291cSSean Christopherson if (WARN_ON_ONCE(fault_address >> 32)) 4458736c291cSSean Christopherson return -EFAULT; 4459736c291cSSean Christopherson #endif 4460736c291cSSean Christopherson 4461c50d8ae3SPaolo Bonzini vcpu->arch.l1tf_flush_l1d = true; 44629ce372b3SVitaly Kuznetsov if (!flags) { 4463faa03b39SWonhyuk Yang trace_kvm_page_fault(vcpu, fault_address, error_code); 4464c50d8ae3SPaolo Bonzini 4465c50d8ae3SPaolo Bonzini if (kvm_event_needs_reinjection(vcpu)) 4466c50d8ae3SPaolo Bonzini kvm_mmu_unprotect_page_virt(vcpu, fault_address); 4467c50d8ae3SPaolo Bonzini r = kvm_mmu_page_fault(vcpu, fault_address, error_code, insn, 4468c50d8ae3SPaolo Bonzini insn_len); 44699ce372b3SVitaly Kuznetsov } else if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) { 447068fd66f1SVitaly Kuznetsov vcpu->arch.apf.host_apf_flags = 0; 4471c50d8ae3SPaolo Bonzini local_irq_disable(); 44726bca69adSThomas Gleixner kvm_async_pf_task_wait_schedule(fault_address); 4473c50d8ae3SPaolo Bonzini local_irq_enable(); 44749ce372b3SVitaly Kuznetsov } else { 44759ce372b3SVitaly Kuznetsov WARN_ONCE(1, "Unexpected host async PF flags: %x\n", flags); 4476c50d8ae3SPaolo Bonzini } 44779ce372b3SVitaly Kuznetsov 4478c50d8ae3SPaolo Bonzini return r; 4479c50d8ae3SPaolo Bonzini } 4480c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_handle_page_fault); 4481c50d8ae3SPaolo Bonzini 44829aa8ab43SDavid Matlack #ifdef CONFIG_X86_64 44839aa8ab43SDavid Matlack static int kvm_tdp_mmu_page_fault(struct kvm_vcpu *vcpu, 44849aa8ab43SDavid Matlack struct kvm_page_fault *fault) 44859aa8ab43SDavid Matlack { 44869aa8ab43SDavid Matlack int r; 44879aa8ab43SDavid Matlack 44889aa8ab43SDavid Matlack if (page_fault_handle_page_track(vcpu, fault)) 44899aa8ab43SDavid Matlack return RET_PF_EMULATE; 44909aa8ab43SDavid Matlack 44919aa8ab43SDavid Matlack r = fast_page_fault(vcpu, fault); 44929aa8ab43SDavid Matlack if (r != RET_PF_INVALID) 44939aa8ab43SDavid Matlack return r; 44949aa8ab43SDavid Matlack 44959aa8ab43SDavid Matlack r = mmu_topup_memory_caches(vcpu, false); 44969aa8ab43SDavid Matlack if (r) 44979aa8ab43SDavid Matlack return r; 44989aa8ab43SDavid Matlack 44999aa8ab43SDavid Matlack r = kvm_faultin_pfn(vcpu, fault, ACC_ALL); 45009aa8ab43SDavid Matlack if (r != RET_PF_CONTINUE) 45019aa8ab43SDavid Matlack return r; 45029aa8ab43SDavid Matlack 45039aa8ab43SDavid Matlack r = RET_PF_RETRY; 45049aa8ab43SDavid Matlack read_lock(&vcpu->kvm->mmu_lock); 45059aa8ab43SDavid Matlack 45069aa8ab43SDavid Matlack if (is_page_fault_stale(vcpu, fault)) 45079aa8ab43SDavid Matlack goto out_unlock; 45089aa8ab43SDavid Matlack 45099aa8ab43SDavid Matlack r = kvm_tdp_mmu_map(vcpu, fault); 45109aa8ab43SDavid Matlack 45119aa8ab43SDavid Matlack out_unlock: 45129aa8ab43SDavid Matlack read_unlock(&vcpu->kvm->mmu_lock); 45139aa8ab43SDavid Matlack kvm_release_pfn_clean(fault->pfn); 45149aa8ab43SDavid Matlack return r; 45159aa8ab43SDavid Matlack } 45169aa8ab43SDavid Matlack #endif 45179aa8ab43SDavid Matlack 4518c501040aSPaolo Bonzini int kvm_tdp_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) 4519c50d8ae3SPaolo Bonzini { 4520d5e90a69SSean Christopherson /* 4521d5e90a69SSean Christopherson * If the guest's MTRRs may be used to compute the "real" memtype, 4522d5e90a69SSean Christopherson * restrict the mapping level to ensure KVM uses a consistent memtype 4523d5e90a69SSean Christopherson * across the entire mapping. If the host MTRRs are ignored by TDP 4524d5e90a69SSean Christopherson * (shadow_memtype_mask is non-zero), and the VM has non-coherent DMA 4525d5e90a69SSean Christopherson * (DMA doesn't snoop CPU caches), KVM's ABI is to honor the memtype 4526d5e90a69SSean Christopherson * from the guest's MTRRs so that guest accesses to memory that is 4527d5e90a69SSean Christopherson * DMA'd aren't cached against the guest's wishes. 4528d5e90a69SSean Christopherson * 4529d5e90a69SSean Christopherson * Note, KVM may still ultimately ignore guest MTRRs for certain PFNs, 4530d5e90a69SSean Christopherson * e.g. KVM will force UC memtype for host MMIO. 4531d5e90a69SSean Christopherson */ 4532d5e90a69SSean Christopherson if (shadow_memtype_mask && kvm_arch_has_noncoherent_dma(vcpu->kvm)) { 4533d5e90a69SSean Christopherson for ( ; fault->max_level > PG_LEVEL_4K; --fault->max_level) { 45344326e57eSPaolo Bonzini int page_num = KVM_PAGES_PER_HPAGE(fault->max_level); 4535c667a3baSHou Wenlong gfn_t base = gfn_round_for_level(fault->gfn, 4536c667a3baSHou Wenlong fault->max_level); 4537c50d8ae3SPaolo Bonzini 4538cb9b88c6SSean Christopherson if (kvm_mtrr_check_gfn_range_consistency(vcpu, base, page_num)) 4539cb9b88c6SSean Christopherson break; 4540d5e90a69SSean Christopherson } 4541c50d8ae3SPaolo Bonzini } 4542c50d8ae3SPaolo Bonzini 45439aa8ab43SDavid Matlack #ifdef CONFIG_X86_64 45449aa8ab43SDavid Matlack if (tdp_mmu_enabled) 45459aa8ab43SDavid Matlack return kvm_tdp_mmu_page_fault(vcpu, fault); 45469aa8ab43SDavid Matlack #endif 45479aa8ab43SDavid Matlack 45484326e57eSPaolo Bonzini return direct_page_fault(vcpu, fault); 4549c50d8ae3SPaolo Bonzini } 4550c50d8ae3SPaolo Bonzini 455184a16226SSean Christopherson static void nonpaging_init_context(struct kvm_mmu *context) 4552c50d8ae3SPaolo Bonzini { 4553c50d8ae3SPaolo Bonzini context->page_fault = nonpaging_page_fault; 4554c50d8ae3SPaolo Bonzini context->gva_to_gpa = nonpaging_gva_to_gpa; 4555c3c6c9fcSLai Jiangshan context->sync_spte = NULL; 4556c50d8ae3SPaolo Bonzini } 4557c50d8ae3SPaolo Bonzini 4558be01e8e2SSean Christopherson static inline bool is_root_usable(struct kvm_mmu_root_info *root, gpa_t pgd, 45590be44352SSean Christopherson union kvm_mmu_page_role role) 45600be44352SSean Christopherson { 4561be01e8e2SSean Christopherson return (role.direct || pgd == root->pgd) && 45625499ea73SPaolo Bonzini VALID_PAGE(root->hpa) && 4563e47c4aeeSSean Christopherson role.word == to_shadow_page(root->hpa)->role.word; 45640be44352SSean Christopherson } 45650be44352SSean Christopherson 4566c50d8ae3SPaolo Bonzini /* 45675499ea73SPaolo Bonzini * Find out if a previously cached root matching the new pgd/role is available, 45685499ea73SPaolo Bonzini * and insert the current root as the MRU in the cache. 45695499ea73SPaolo Bonzini * If a matching root is found, it is assigned to kvm_mmu->root and 45705499ea73SPaolo Bonzini * true is returned. 45715499ea73SPaolo Bonzini * If no match is found, kvm_mmu->root is left invalid, the LRU root is 45725499ea73SPaolo Bonzini * evicted to make room for the current root, and false is returned. 4573c50d8ae3SPaolo Bonzini */ 45745499ea73SPaolo Bonzini static bool cached_root_find_and_keep_current(struct kvm *kvm, struct kvm_mmu *mmu, 45755499ea73SPaolo Bonzini gpa_t new_pgd, 4576c50d8ae3SPaolo Bonzini union kvm_mmu_page_role new_role) 4577c50d8ae3SPaolo Bonzini { 4578c50d8ae3SPaolo Bonzini uint i; 4579c50d8ae3SPaolo Bonzini 4580b9e5603cSPaolo Bonzini if (is_root_usable(&mmu->root, new_pgd, new_role)) 45810be44352SSean Christopherson return true; 45820be44352SSean Christopherson 4583c50d8ae3SPaolo Bonzini for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) { 45845499ea73SPaolo Bonzini /* 45855499ea73SPaolo Bonzini * The swaps end up rotating the cache like this: 45865499ea73SPaolo Bonzini * C 0 1 2 3 (on entry to the function) 45875499ea73SPaolo Bonzini * 0 C 1 2 3 45885499ea73SPaolo Bonzini * 1 C 0 2 3 45895499ea73SPaolo Bonzini * 2 C 0 1 3 45905499ea73SPaolo Bonzini * 3 C 0 1 2 (on exit from the loop) 45915499ea73SPaolo Bonzini */ 4592b9e5603cSPaolo Bonzini swap(mmu->root, mmu->prev_roots[i]); 4593b9e5603cSPaolo Bonzini if (is_root_usable(&mmu->root, new_pgd, new_role)) 45945499ea73SPaolo Bonzini return true; 4595c50d8ae3SPaolo Bonzini } 4596c50d8ae3SPaolo Bonzini 45975499ea73SPaolo Bonzini kvm_mmu_free_roots(kvm, mmu, KVM_MMU_ROOT_CURRENT); 45985499ea73SPaolo Bonzini return false; 4599c50d8ae3SPaolo Bonzini } 4600c50d8ae3SPaolo Bonzini 4601c50d8ae3SPaolo Bonzini /* 46025499ea73SPaolo Bonzini * Find out if a previously cached root matching the new pgd/role is available. 46035499ea73SPaolo Bonzini * On entry, mmu->root is invalid. 46045499ea73SPaolo Bonzini * If a matching root is found, it is assigned to kvm_mmu->root, the LRU entry 46055499ea73SPaolo Bonzini * of the cache becomes invalid, and true is returned. 46065499ea73SPaolo Bonzini * If no match is found, kvm_mmu->root is left invalid and false is returned. 46075499ea73SPaolo Bonzini */ 46085499ea73SPaolo Bonzini static bool cached_root_find_without_current(struct kvm *kvm, struct kvm_mmu *mmu, 46095499ea73SPaolo Bonzini gpa_t new_pgd, 46105499ea73SPaolo Bonzini union kvm_mmu_page_role new_role) 46115499ea73SPaolo Bonzini { 46125499ea73SPaolo Bonzini uint i; 46135499ea73SPaolo Bonzini 46145499ea73SPaolo Bonzini for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) 46155499ea73SPaolo Bonzini if (is_root_usable(&mmu->prev_roots[i], new_pgd, new_role)) 46165499ea73SPaolo Bonzini goto hit; 46175499ea73SPaolo Bonzini 46185499ea73SPaolo Bonzini return false; 46195499ea73SPaolo Bonzini 46205499ea73SPaolo Bonzini hit: 46215499ea73SPaolo Bonzini swap(mmu->root, mmu->prev_roots[i]); 46225499ea73SPaolo Bonzini /* Bubble up the remaining roots. */ 46235499ea73SPaolo Bonzini for (; i < KVM_MMU_NUM_PREV_ROOTS - 1; i++) 46245499ea73SPaolo Bonzini mmu->prev_roots[i] = mmu->prev_roots[i + 1]; 46255499ea73SPaolo Bonzini mmu->prev_roots[i].hpa = INVALID_PAGE; 46265499ea73SPaolo Bonzini return true; 46275499ea73SPaolo Bonzini } 46285499ea73SPaolo Bonzini 46295499ea73SPaolo Bonzini static bool fast_pgd_switch(struct kvm *kvm, struct kvm_mmu *mmu, 46305499ea73SPaolo Bonzini gpa_t new_pgd, union kvm_mmu_page_role new_role) 46315499ea73SPaolo Bonzini { 46325499ea73SPaolo Bonzini /* 46335499ea73SPaolo Bonzini * For now, limit the caching to 64-bit hosts+VMs in order to avoid 4634c50d8ae3SPaolo Bonzini * having to deal with PDPTEs. We may add support for 32-bit hosts/VMs 4635c50d8ae3SPaolo Bonzini * later if necessary. 4636c50d8ae3SPaolo Bonzini */ 46375499ea73SPaolo Bonzini if (VALID_PAGE(mmu->root.hpa) && !to_shadow_page(mmu->root.hpa)) 46385499ea73SPaolo Bonzini kvm_mmu_free_roots(kvm, mmu, KVM_MMU_ROOT_CURRENT); 4639c50d8ae3SPaolo Bonzini 46405499ea73SPaolo Bonzini if (VALID_PAGE(mmu->root.hpa)) 46415499ea73SPaolo Bonzini return cached_root_find_and_keep_current(kvm, mmu, new_pgd, new_role); 46425499ea73SPaolo Bonzini else 46435499ea73SPaolo Bonzini return cached_root_find_without_current(kvm, mmu, new_pgd, new_role); 4644c50d8ae3SPaolo Bonzini } 4645c50d8ae3SPaolo Bonzini 4646d2e5f333SPaolo Bonzini void kvm_mmu_new_pgd(struct kvm_vcpu *vcpu, gpa_t new_pgd) 4647c50d8ae3SPaolo Bonzini { 46480c1c92f1SPaolo Bonzini struct kvm_mmu *mmu = vcpu->arch.mmu; 46497a458f0eSPaolo Bonzini union kvm_mmu_page_role new_role = mmu->root_role; 46500c1c92f1SPaolo Bonzini 4651a7e48ef7SWei Liu /* 4652a7e48ef7SWei Liu * Return immediately if no usable root was found, kvm_mmu_reload() 4653a7e48ef7SWei Liu * will establish a valid root prior to the next VM-Enter. 4654a7e48ef7SWei Liu */ 4655a7e48ef7SWei Liu if (!fast_pgd_switch(vcpu->kvm, mmu, new_pgd, new_role)) 4656b869855bSSean Christopherson return; 4657c50d8ae3SPaolo Bonzini 4658c50d8ae3SPaolo Bonzini /* 4659b869855bSSean Christopherson * It's possible that the cached previous root page is obsolete because 4660b869855bSSean Christopherson * of a change in the MMU generation number. However, changing the 4661527d5cd7SSean Christopherson * generation number is accompanied by KVM_REQ_MMU_FREE_OBSOLETE_ROOTS, 4662527d5cd7SSean Christopherson * which will free the root set here and allocate a new one. 4663b869855bSSean Christopherson */ 4664b869855bSSean Christopherson kvm_make_request(KVM_REQ_LOAD_MMU_PGD, vcpu); 4665b869855bSSean Christopherson 4666b5129100SSean Christopherson if (force_flush_and_sync_on_reuse) { 4667b869855bSSean Christopherson kvm_make_request(KVM_REQ_MMU_SYNC, vcpu); 4668b869855bSSean Christopherson kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu); 4669b5129100SSean Christopherson } 4670b869855bSSean Christopherson 4671b869855bSSean Christopherson /* 4672b869855bSSean Christopherson * The last MMIO access's GVA and GPA are cached in the VCPU. When 4673b869855bSSean Christopherson * switching to a new CR3, that GVA->GPA mapping may no longer be 4674b869855bSSean Christopherson * valid. So clear any cached MMIO info even when we don't need to sync 4675b869855bSSean Christopherson * the shadow page tables. 4676c50d8ae3SPaolo Bonzini */ 4677c50d8ae3SPaolo Bonzini vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY); 4678c50d8ae3SPaolo Bonzini 4679daa5b6c1SBen Gardon /* 4680daa5b6c1SBen Gardon * If this is a direct root page, it doesn't have a write flooding 4681daa5b6c1SBen Gardon * count. Otherwise, clear the write flooding count. 4682daa5b6c1SBen Gardon */ 4683daa5b6c1SBen Gardon if (!new_role.direct) 4684daa5b6c1SBen Gardon __clear_sp_write_flooding_count( 4685b9e5603cSPaolo Bonzini to_shadow_page(vcpu->arch.mmu->root.hpa)); 4686c50d8ae3SPaolo Bonzini } 4687be01e8e2SSean Christopherson EXPORT_SYMBOL_GPL(kvm_mmu_new_pgd); 4688c50d8ae3SPaolo Bonzini 4689c50d8ae3SPaolo Bonzini static bool sync_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn, 4690c3e5e415SLai Jiangshan unsigned int access) 4691c50d8ae3SPaolo Bonzini { 4692c50d8ae3SPaolo Bonzini if (unlikely(is_mmio_spte(*sptep))) { 4693c50d8ae3SPaolo Bonzini if (gfn != get_mmio_spte_gfn(*sptep)) { 4694c50d8ae3SPaolo Bonzini mmu_spte_clear_no_track(sptep); 4695c50d8ae3SPaolo Bonzini return true; 4696c50d8ae3SPaolo Bonzini } 4697c50d8ae3SPaolo Bonzini 4698c50d8ae3SPaolo Bonzini mark_mmio_spte(vcpu, sptep, gfn, access); 4699c50d8ae3SPaolo Bonzini return true; 4700c50d8ae3SPaolo Bonzini } 4701c50d8ae3SPaolo Bonzini 4702c50d8ae3SPaolo Bonzini return false; 4703c50d8ae3SPaolo Bonzini } 4704c50d8ae3SPaolo Bonzini 4705c50d8ae3SPaolo Bonzini #define PTTYPE_EPT 18 /* arbitrary */ 4706c50d8ae3SPaolo Bonzini #define PTTYPE PTTYPE_EPT 4707c50d8ae3SPaolo Bonzini #include "paging_tmpl.h" 4708c50d8ae3SPaolo Bonzini #undef PTTYPE 4709c50d8ae3SPaolo Bonzini 4710c50d8ae3SPaolo Bonzini #define PTTYPE 64 4711c50d8ae3SPaolo Bonzini #include "paging_tmpl.h" 4712c50d8ae3SPaolo Bonzini #undef PTTYPE 4713c50d8ae3SPaolo Bonzini 4714c50d8ae3SPaolo Bonzini #define PTTYPE 32 4715c50d8ae3SPaolo Bonzini #include "paging_tmpl.h" 4716c50d8ae3SPaolo Bonzini #undef PTTYPE 4717c50d8ae3SPaolo Bonzini 4718f3d90f90SSean Christopherson static void __reset_rsvds_bits_mask(struct rsvd_bits_validate *rsvd_check, 4719f3d90f90SSean Christopherson u64 pa_bits_rsvd, int level, bool nx, 4720f3d90f90SSean Christopherson bool gbpages, bool pse, bool amd) 4721c50d8ae3SPaolo Bonzini { 4722c50d8ae3SPaolo Bonzini u64 gbpages_bit_rsvd = 0; 4723c50d8ae3SPaolo Bonzini u64 nonleaf_bit8_rsvd = 0; 47245b7f575cSSean Christopherson u64 high_bits_rsvd; 4725c50d8ae3SPaolo Bonzini 4726c50d8ae3SPaolo Bonzini rsvd_check->bad_mt_xwr = 0; 4727c50d8ae3SPaolo Bonzini 4728c50d8ae3SPaolo Bonzini if (!gbpages) 4729c50d8ae3SPaolo Bonzini gbpages_bit_rsvd = rsvd_bits(7, 7); 4730c50d8ae3SPaolo Bonzini 47315b7f575cSSean Christopherson if (level == PT32E_ROOT_LEVEL) 47325b7f575cSSean Christopherson high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 62); 47335b7f575cSSean Christopherson else 47345b7f575cSSean Christopherson high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 51); 47355b7f575cSSean Christopherson 47365b7f575cSSean Christopherson /* Note, NX doesn't exist in PDPTEs, this is handled below. */ 47375b7f575cSSean Christopherson if (!nx) 47385b7f575cSSean Christopherson high_bits_rsvd |= rsvd_bits(63, 63); 47395b7f575cSSean Christopherson 4740c50d8ae3SPaolo Bonzini /* 4741c50d8ae3SPaolo Bonzini * Non-leaf PML4Es and PDPEs reserve bit 8 (which would be the G bit for 4742c50d8ae3SPaolo Bonzini * leaf entries) on AMD CPUs only. 4743c50d8ae3SPaolo Bonzini */ 4744c50d8ae3SPaolo Bonzini if (amd) 4745c50d8ae3SPaolo Bonzini nonleaf_bit8_rsvd = rsvd_bits(8, 8); 4746c50d8ae3SPaolo Bonzini 4747c50d8ae3SPaolo Bonzini switch (level) { 4748c50d8ae3SPaolo Bonzini case PT32_ROOT_LEVEL: 4749c50d8ae3SPaolo Bonzini /* no rsvd bits for 2 level 4K page table entries */ 4750c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][1] = 0; 4751c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][0] = 0; 4752c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][0] = 4753c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][0]; 4754c50d8ae3SPaolo Bonzini 4755c50d8ae3SPaolo Bonzini if (!pse) { 4756c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][1] = 0; 4757c50d8ae3SPaolo Bonzini break; 4758c50d8ae3SPaolo Bonzini } 4759c50d8ae3SPaolo Bonzini 4760c50d8ae3SPaolo Bonzini if (is_cpuid_PSE36()) 4761c50d8ae3SPaolo Bonzini /* 36bits PSE 4MB page */ 4762c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(17, 21); 4763c50d8ae3SPaolo Bonzini else 4764c50d8ae3SPaolo Bonzini /* 32 bits PSE 4MB page */ 4765c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(13, 21); 4766c50d8ae3SPaolo Bonzini break; 4767c50d8ae3SPaolo Bonzini case PT32E_ROOT_LEVEL: 47685b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][2] = rsvd_bits(63, 63) | 47695b7f575cSSean Christopherson high_bits_rsvd | 47705b7f575cSSean Christopherson rsvd_bits(5, 8) | 47715b7f575cSSean Christopherson rsvd_bits(1, 2); /* PDPTE */ 47725b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd; /* PDE */ 47735b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd; /* PTE */ 47745b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd | 4775c50d8ae3SPaolo Bonzini rsvd_bits(13, 20); /* large page */ 4776c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][0] = 4777c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][0]; 4778c50d8ae3SPaolo Bonzini break; 4779c50d8ae3SPaolo Bonzini case PT64_ROOT_5LEVEL: 47805b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][4] = high_bits_rsvd | 47815b7f575cSSean Christopherson nonleaf_bit8_rsvd | 47825b7f575cSSean Christopherson rsvd_bits(7, 7); 4783c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][4] = 4784c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][4]; 4785df561f66SGustavo A. R. Silva fallthrough; 4786c50d8ae3SPaolo Bonzini case PT64_ROOT_4LEVEL: 47875b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][3] = high_bits_rsvd | 47885b7f575cSSean Christopherson nonleaf_bit8_rsvd | 47895b7f575cSSean Christopherson rsvd_bits(7, 7); 47905b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][2] = high_bits_rsvd | 47915b7f575cSSean Christopherson gbpages_bit_rsvd; 47925b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd; 47935b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd; 4794c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][3] = 4795c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][3]; 47965b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[1][2] = high_bits_rsvd | 47975b7f575cSSean Christopherson gbpages_bit_rsvd | 4798c50d8ae3SPaolo Bonzini rsvd_bits(13, 29); 47995b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd | 4800c50d8ae3SPaolo Bonzini rsvd_bits(13, 20); /* large page */ 4801c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][0] = 4802c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[0][0]; 4803c50d8ae3SPaolo Bonzini break; 4804c50d8ae3SPaolo Bonzini } 4805c50d8ae3SPaolo Bonzini } 4806c50d8ae3SPaolo Bonzini 480727de9250SSean Christopherson static bool guest_can_use_gbpages(struct kvm_vcpu *vcpu) 480827de9250SSean Christopherson { 480927de9250SSean Christopherson /* 481027de9250SSean Christopherson * If TDP is enabled, let the guest use GBPAGES if they're supported in 481127de9250SSean Christopherson * hardware. The hardware page walker doesn't let KVM disable GBPAGES, 481227de9250SSean Christopherson * i.e. won't treat them as reserved, and KVM doesn't redo the GVA->GPA 481327de9250SSean Christopherson * walk for performance and complexity reasons. Not to mention KVM 481427de9250SSean Christopherson * _can't_ solve the problem because GVA->GPA walks aren't visible to 481527de9250SSean Christopherson * KVM once a TDP translation is installed. Mimic hardware behavior so 481627de9250SSean Christopherson * that KVM's is at least consistent, i.e. doesn't randomly inject #PF. 481727de9250SSean Christopherson */ 481827de9250SSean Christopherson return tdp_enabled ? boot_cpu_has(X86_FEATURE_GBPAGES) : 481927de9250SSean Christopherson guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES); 482027de9250SSean Christopherson } 482127de9250SSean Christopherson 4822c919e881SKai Huang static void reset_guest_rsvds_bits_mask(struct kvm_vcpu *vcpu, 4823c50d8ae3SPaolo Bonzini struct kvm_mmu *context) 4824c50d8ae3SPaolo Bonzini { 4825b705a277SSean Christopherson __reset_rsvds_bits_mask(&context->guest_rsvd_check, 48265b7f575cSSean Christopherson vcpu->arch.reserved_gpa_bits, 48274d25502aSPaolo Bonzini context->cpu_role.base.level, is_efer_nx(context), 482827de9250SSean Christopherson guest_can_use_gbpages(vcpu), 48294e9c0d80SSean Christopherson is_cr4_pse(context), 483023493d0aSSean Christopherson guest_cpuid_is_amd_or_hygon(vcpu)); 4831c50d8ae3SPaolo Bonzini } 4832c50d8ae3SPaolo Bonzini 4833f3d90f90SSean Christopherson static void __reset_rsvds_bits_mask_ept(struct rsvd_bits_validate *rsvd_check, 4834f3d90f90SSean Christopherson u64 pa_bits_rsvd, bool execonly, 4835f3d90f90SSean Christopherson int huge_page_level) 4836c50d8ae3SPaolo Bonzini { 48375b7f575cSSean Christopherson u64 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 51); 483884ea5c09SLai Jiangshan u64 large_1g_rsvd = 0, large_2m_rsvd = 0; 4839c50d8ae3SPaolo Bonzini u64 bad_mt_xwr; 4840c50d8ae3SPaolo Bonzini 484184ea5c09SLai Jiangshan if (huge_page_level < PG_LEVEL_1G) 484284ea5c09SLai Jiangshan large_1g_rsvd = rsvd_bits(7, 7); 484384ea5c09SLai Jiangshan if (huge_page_level < PG_LEVEL_2M) 484484ea5c09SLai Jiangshan large_2m_rsvd = rsvd_bits(7, 7); 484584ea5c09SLai Jiangshan 48465b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][4] = high_bits_rsvd | rsvd_bits(3, 7); 48475b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][3] = high_bits_rsvd | rsvd_bits(3, 7); 484884ea5c09SLai Jiangshan rsvd_check->rsvd_bits_mask[0][2] = high_bits_rsvd | rsvd_bits(3, 6) | large_1g_rsvd; 484984ea5c09SLai Jiangshan rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd | rsvd_bits(3, 6) | large_2m_rsvd; 48505b7f575cSSean Christopherson rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd; 4851c50d8ae3SPaolo Bonzini 4852c50d8ae3SPaolo Bonzini /* large page */ 4853c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][4] = rsvd_check->rsvd_bits_mask[0][4]; 4854c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][3] = rsvd_check->rsvd_bits_mask[0][3]; 485584ea5c09SLai Jiangshan rsvd_check->rsvd_bits_mask[1][2] = high_bits_rsvd | rsvd_bits(12, 29) | large_1g_rsvd; 485684ea5c09SLai Jiangshan rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd | rsvd_bits(12, 20) | large_2m_rsvd; 4857c50d8ae3SPaolo Bonzini rsvd_check->rsvd_bits_mask[1][0] = rsvd_check->rsvd_bits_mask[0][0]; 4858c50d8ae3SPaolo Bonzini 4859c50d8ae3SPaolo Bonzini bad_mt_xwr = 0xFFull << (2 * 8); /* bits 3..5 must not be 2 */ 4860c50d8ae3SPaolo Bonzini bad_mt_xwr |= 0xFFull << (3 * 8); /* bits 3..5 must not be 3 */ 4861c50d8ae3SPaolo Bonzini bad_mt_xwr |= 0xFFull << (7 * 8); /* bits 3..5 must not be 7 */ 4862c50d8ae3SPaolo Bonzini bad_mt_xwr |= REPEAT_BYTE(1ull << 2); /* bits 0..2 must not be 010 */ 4863c50d8ae3SPaolo Bonzini bad_mt_xwr |= REPEAT_BYTE(1ull << 6); /* bits 0..2 must not be 110 */ 4864c50d8ae3SPaolo Bonzini if (!execonly) { 4865c50d8ae3SPaolo Bonzini /* bits 0..2 must not be 100 unless VMX capabilities allow it */ 4866c50d8ae3SPaolo Bonzini bad_mt_xwr |= REPEAT_BYTE(1ull << 4); 4867c50d8ae3SPaolo Bonzini } 4868c50d8ae3SPaolo Bonzini rsvd_check->bad_mt_xwr = bad_mt_xwr; 4869c50d8ae3SPaolo Bonzini } 4870c50d8ae3SPaolo Bonzini 4871c50d8ae3SPaolo Bonzini static void reset_rsvds_bits_mask_ept(struct kvm_vcpu *vcpu, 487284ea5c09SLai Jiangshan struct kvm_mmu *context, bool execonly, int huge_page_level) 4873c50d8ae3SPaolo Bonzini { 4874c50d8ae3SPaolo Bonzini __reset_rsvds_bits_mask_ept(&context->guest_rsvd_check, 487584ea5c09SLai Jiangshan vcpu->arch.reserved_gpa_bits, execonly, 487684ea5c09SLai Jiangshan huge_page_level); 4877c50d8ae3SPaolo Bonzini } 4878c50d8ae3SPaolo Bonzini 48796f8e65a6SSean Christopherson static inline u64 reserved_hpa_bits(void) 48806f8e65a6SSean Christopherson { 48816f8e65a6SSean Christopherson return rsvd_bits(shadow_phys_bits, 63); 48826f8e65a6SSean Christopherson } 48836f8e65a6SSean Christopherson 4884c50d8ae3SPaolo Bonzini /* 4885c50d8ae3SPaolo Bonzini * the page table on host is the shadow page table for the page 4886c50d8ae3SPaolo Bonzini * table in guest or amd nested guest, its mmu features completely 4887c50d8ae3SPaolo Bonzini * follow the features in guest. 4888c50d8ae3SPaolo Bonzini */ 488916be1d12SSean Christopherson static void reset_shadow_zero_bits_mask(struct kvm_vcpu *vcpu, 489016be1d12SSean Christopherson struct kvm_mmu *context) 4891c50d8ae3SPaolo Bonzini { 48928c985b2dSSean Christopherson /* @amd adds a check on bit of SPTEs, which KVM shouldn't use anyways. */ 48938c985b2dSSean Christopherson bool is_amd = true; 48948c985b2dSSean Christopherson /* KVM doesn't use 2-level page tables for the shadow MMU. */ 48958c985b2dSSean Christopherson bool is_pse = false; 4896c50d8ae3SPaolo Bonzini struct rsvd_bits_validate *shadow_zero_check; 4897c50d8ae3SPaolo Bonzini int i; 4898c50d8ae3SPaolo Bonzini 4899a972e29cSPaolo Bonzini WARN_ON_ONCE(context->root_role.level < PT32E_ROOT_LEVEL); 49008c985b2dSSean Christopherson 4901c50d8ae3SPaolo Bonzini shadow_zero_check = &context->shadow_zero_check; 4902b705a277SSean Christopherson __reset_rsvds_bits_mask(shadow_zero_check, reserved_hpa_bits(), 4903a972e29cSPaolo Bonzini context->root_role.level, 49047a458f0eSPaolo Bonzini context->root_role.efer_nx, 490527de9250SSean Christopherson guest_can_use_gbpages(vcpu), is_pse, is_amd); 4906c50d8ae3SPaolo Bonzini 4907c50d8ae3SPaolo Bonzini if (!shadow_me_mask) 4908c50d8ae3SPaolo Bonzini return; 4909c50d8ae3SPaolo Bonzini 4910a972e29cSPaolo Bonzini for (i = context->root_role.level; --i >= 0;) { 4911e54f1ff2SKai Huang /* 4912e54f1ff2SKai Huang * So far shadow_me_value is a constant during KVM's life 4913e54f1ff2SKai Huang * time. Bits in shadow_me_value are allowed to be set. 4914e54f1ff2SKai Huang * Bits in shadow_me_mask but not in shadow_me_value are 4915e54f1ff2SKai Huang * not allowed to be set. 4916e54f1ff2SKai Huang */ 4917e54f1ff2SKai Huang shadow_zero_check->rsvd_bits_mask[0][i] |= shadow_me_mask; 4918e54f1ff2SKai Huang shadow_zero_check->rsvd_bits_mask[1][i] |= shadow_me_mask; 4919e54f1ff2SKai Huang shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_value; 4920e54f1ff2SKai Huang shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_value; 4921c50d8ae3SPaolo Bonzini } 4922c50d8ae3SPaolo Bonzini 4923c50d8ae3SPaolo Bonzini } 4924c50d8ae3SPaolo Bonzini 4925c50d8ae3SPaolo Bonzini static inline bool boot_cpu_is_amd(void) 4926c50d8ae3SPaolo Bonzini { 4927c50d8ae3SPaolo Bonzini WARN_ON_ONCE(!tdp_enabled); 4928c50d8ae3SPaolo Bonzini return shadow_x_mask == 0; 4929c50d8ae3SPaolo Bonzini } 4930c50d8ae3SPaolo Bonzini 4931c50d8ae3SPaolo Bonzini /* 4932c50d8ae3SPaolo Bonzini * the direct page table on host, use as much mmu features as 4933c50d8ae3SPaolo Bonzini * possible, however, kvm currently does not do execution-protection. 4934c50d8ae3SPaolo Bonzini */ 4935f3d90f90SSean Christopherson static void reset_tdp_shadow_zero_bits_mask(struct kvm_mmu *context) 4936c50d8ae3SPaolo Bonzini { 4937c50d8ae3SPaolo Bonzini struct rsvd_bits_validate *shadow_zero_check; 4938c50d8ae3SPaolo Bonzini int i; 4939c50d8ae3SPaolo Bonzini 4940c50d8ae3SPaolo Bonzini shadow_zero_check = &context->shadow_zero_check; 4941c50d8ae3SPaolo Bonzini 4942c50d8ae3SPaolo Bonzini if (boot_cpu_is_amd()) 4943b705a277SSean Christopherson __reset_rsvds_bits_mask(shadow_zero_check, reserved_hpa_bits(), 49446c6ab524SSean Christopherson context->root_role.level, true, 4945c50d8ae3SPaolo Bonzini boot_cpu_has(X86_FEATURE_GBPAGES), 49468c985b2dSSean Christopherson false, true); 4947c50d8ae3SPaolo Bonzini else 4948c50d8ae3SPaolo Bonzini __reset_rsvds_bits_mask_ept(shadow_zero_check, 494984ea5c09SLai Jiangshan reserved_hpa_bits(), false, 495084ea5c09SLai Jiangshan max_huge_page_level); 4951c50d8ae3SPaolo Bonzini 4952c50d8ae3SPaolo Bonzini if (!shadow_me_mask) 4953c50d8ae3SPaolo Bonzini return; 4954c50d8ae3SPaolo Bonzini 4955a972e29cSPaolo Bonzini for (i = context->root_role.level; --i >= 0;) { 4956c50d8ae3SPaolo Bonzini shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask; 4957c50d8ae3SPaolo Bonzini shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask; 4958c50d8ae3SPaolo Bonzini } 4959c50d8ae3SPaolo Bonzini } 4960c50d8ae3SPaolo Bonzini 4961c50d8ae3SPaolo Bonzini /* 4962c50d8ae3SPaolo Bonzini * as the comments in reset_shadow_zero_bits_mask() except it 4963c50d8ae3SPaolo Bonzini * is the shadow page table for intel nested guest. 4964c50d8ae3SPaolo Bonzini */ 4965c50d8ae3SPaolo Bonzini static void 4966e8f6e738SJinrong Liang reset_ept_shadow_zero_bits_mask(struct kvm_mmu *context, bool execonly) 4967c50d8ae3SPaolo Bonzini { 4968c50d8ae3SPaolo Bonzini __reset_rsvds_bits_mask_ept(&context->shadow_zero_check, 496984ea5c09SLai Jiangshan reserved_hpa_bits(), execonly, 497084ea5c09SLai Jiangshan max_huge_page_level); 4971c50d8ae3SPaolo Bonzini } 4972c50d8ae3SPaolo Bonzini 4973c50d8ae3SPaolo Bonzini #define BYTE_MASK(access) \ 4974c50d8ae3SPaolo Bonzini ((1 & (access) ? 2 : 0) | \ 4975c50d8ae3SPaolo Bonzini (2 & (access) ? 4 : 0) | \ 4976c50d8ae3SPaolo Bonzini (3 & (access) ? 8 : 0) | \ 4977c50d8ae3SPaolo Bonzini (4 & (access) ? 16 : 0) | \ 4978c50d8ae3SPaolo Bonzini (5 & (access) ? 32 : 0) | \ 4979c50d8ae3SPaolo Bonzini (6 & (access) ? 64 : 0) | \ 4980c50d8ae3SPaolo Bonzini (7 & (access) ? 128 : 0)) 4981c50d8ae3SPaolo Bonzini 4982c50d8ae3SPaolo Bonzini 4983c596f147SSean Christopherson static void update_permission_bitmask(struct kvm_mmu *mmu, bool ept) 4984c50d8ae3SPaolo Bonzini { 4985c50d8ae3SPaolo Bonzini unsigned byte; 4986c50d8ae3SPaolo Bonzini 4987c50d8ae3SPaolo Bonzini const u8 x = BYTE_MASK(ACC_EXEC_MASK); 4988c50d8ae3SPaolo Bonzini const u8 w = BYTE_MASK(ACC_WRITE_MASK); 4989c50d8ae3SPaolo Bonzini const u8 u = BYTE_MASK(ACC_USER_MASK); 4990c50d8ae3SPaolo Bonzini 4991c596f147SSean Christopherson bool cr4_smep = is_cr4_smep(mmu); 4992c596f147SSean Christopherson bool cr4_smap = is_cr4_smap(mmu); 4993c596f147SSean Christopherson bool cr0_wp = is_cr0_wp(mmu); 499490599c28SSean Christopherson bool efer_nx = is_efer_nx(mmu); 4995c50d8ae3SPaolo Bonzini 4996c50d8ae3SPaolo Bonzini for (byte = 0; byte < ARRAY_SIZE(mmu->permissions); ++byte) { 4997c50d8ae3SPaolo Bonzini unsigned pfec = byte << 1; 4998c50d8ae3SPaolo Bonzini 4999c50d8ae3SPaolo Bonzini /* 5000c50d8ae3SPaolo Bonzini * Each "*f" variable has a 1 bit for each UWX value 5001c50d8ae3SPaolo Bonzini * that causes a fault with the given PFEC. 5002c50d8ae3SPaolo Bonzini */ 5003c50d8ae3SPaolo Bonzini 5004c50d8ae3SPaolo Bonzini /* Faults from writes to non-writable pages */ 5005c50d8ae3SPaolo Bonzini u8 wf = (pfec & PFERR_WRITE_MASK) ? (u8)~w : 0; 5006c50d8ae3SPaolo Bonzini /* Faults from user mode accesses to supervisor pages */ 5007c50d8ae3SPaolo Bonzini u8 uf = (pfec & PFERR_USER_MASK) ? (u8)~u : 0; 5008c50d8ae3SPaolo Bonzini /* Faults from fetches of non-executable pages*/ 5009c50d8ae3SPaolo Bonzini u8 ff = (pfec & PFERR_FETCH_MASK) ? (u8)~x : 0; 5010c50d8ae3SPaolo Bonzini /* Faults from kernel mode fetches of user pages */ 5011c50d8ae3SPaolo Bonzini u8 smepf = 0; 5012c50d8ae3SPaolo Bonzini /* Faults from kernel mode accesses of user pages */ 5013c50d8ae3SPaolo Bonzini u8 smapf = 0; 5014c50d8ae3SPaolo Bonzini 5015c50d8ae3SPaolo Bonzini if (!ept) { 5016c50d8ae3SPaolo Bonzini /* Faults from kernel mode accesses to user pages */ 5017c50d8ae3SPaolo Bonzini u8 kf = (pfec & PFERR_USER_MASK) ? 0 : u; 5018c50d8ae3SPaolo Bonzini 5019c50d8ae3SPaolo Bonzini /* Not really needed: !nx will cause pte.nx to fault */ 502090599c28SSean Christopherson if (!efer_nx) 5021c50d8ae3SPaolo Bonzini ff = 0; 5022c50d8ae3SPaolo Bonzini 5023c50d8ae3SPaolo Bonzini /* Allow supervisor writes if !cr0.wp */ 5024c50d8ae3SPaolo Bonzini if (!cr0_wp) 5025c50d8ae3SPaolo Bonzini wf = (pfec & PFERR_USER_MASK) ? wf : 0; 5026c50d8ae3SPaolo Bonzini 5027c50d8ae3SPaolo Bonzini /* Disallow supervisor fetches of user code if cr4.smep */ 5028c50d8ae3SPaolo Bonzini if (cr4_smep) 5029c50d8ae3SPaolo Bonzini smepf = (pfec & PFERR_FETCH_MASK) ? kf : 0; 5030c50d8ae3SPaolo Bonzini 5031c50d8ae3SPaolo Bonzini /* 5032c50d8ae3SPaolo Bonzini * SMAP:kernel-mode data accesses from user-mode 5033c50d8ae3SPaolo Bonzini * mappings should fault. A fault is considered 5034c50d8ae3SPaolo Bonzini * as a SMAP violation if all of the following 5035c50d8ae3SPaolo Bonzini * conditions are true: 5036c50d8ae3SPaolo Bonzini * - X86_CR4_SMAP is set in CR4 5037c50d8ae3SPaolo Bonzini * - A user page is accessed 5038c50d8ae3SPaolo Bonzini * - The access is not a fetch 50394f4aa80eSLai Jiangshan * - The access is supervisor mode 50404f4aa80eSLai Jiangshan * - If implicit supervisor access or X86_EFLAGS_AC is clear 5041c50d8ae3SPaolo Bonzini * 504294b4a2f1SLai Jiangshan * Here, we cover the first four conditions. 504394b4a2f1SLai Jiangshan * The fifth is computed dynamically in permission_fault(); 5044c50d8ae3SPaolo Bonzini * PFERR_RSVD_MASK bit will be set in PFEC if the access is 5045c50d8ae3SPaolo Bonzini * *not* subject to SMAP restrictions. 5046c50d8ae3SPaolo Bonzini */ 5047c50d8ae3SPaolo Bonzini if (cr4_smap) 5048c50d8ae3SPaolo Bonzini smapf = (pfec & (PFERR_RSVD_MASK|PFERR_FETCH_MASK)) ? 0 : kf; 5049c50d8ae3SPaolo Bonzini } 5050c50d8ae3SPaolo Bonzini 5051c50d8ae3SPaolo Bonzini mmu->permissions[byte] = ff | uf | wf | smepf | smapf; 5052c50d8ae3SPaolo Bonzini } 5053c50d8ae3SPaolo Bonzini } 5054c50d8ae3SPaolo Bonzini 5055c50d8ae3SPaolo Bonzini /* 5056c50d8ae3SPaolo Bonzini * PKU is an additional mechanism by which the paging controls access to 5057c50d8ae3SPaolo Bonzini * user-mode addresses based on the value in the PKRU register. Protection 5058c50d8ae3SPaolo Bonzini * key violations are reported through a bit in the page fault error code. 5059c50d8ae3SPaolo Bonzini * Unlike other bits of the error code, the PK bit is not known at the 5060c50d8ae3SPaolo Bonzini * call site of e.g. gva_to_gpa; it must be computed directly in 5061c50d8ae3SPaolo Bonzini * permission_fault based on two bits of PKRU, on some machine state (CR4, 5062c50d8ae3SPaolo Bonzini * CR0, EFER, CPL), and on other bits of the error code and the page tables. 5063c50d8ae3SPaolo Bonzini * 5064c50d8ae3SPaolo Bonzini * In particular the following conditions come from the error code, the 5065c50d8ae3SPaolo Bonzini * page tables and the machine state: 5066c50d8ae3SPaolo Bonzini * - PK is always zero unless CR4.PKE=1 and EFER.LMA=1 5067c50d8ae3SPaolo Bonzini * - PK is always zero if RSVD=1 (reserved bit set) or F=1 (instruction fetch) 5068c50d8ae3SPaolo Bonzini * - PK is always zero if U=0 in the page tables 5069c50d8ae3SPaolo Bonzini * - PKRU.WD is ignored if CR0.WP=0 and the access is a supervisor access. 5070c50d8ae3SPaolo Bonzini * 5071c50d8ae3SPaolo Bonzini * The PKRU bitmask caches the result of these four conditions. The error 5072c50d8ae3SPaolo Bonzini * code (minus the P bit) and the page table's U bit form an index into the 5073c50d8ae3SPaolo Bonzini * PKRU bitmask. Two bits of the PKRU bitmask are then extracted and ANDed 5074c50d8ae3SPaolo Bonzini * with the two bits of the PKRU register corresponding to the protection key. 5075c50d8ae3SPaolo Bonzini * For the first three conditions above the bits will be 00, thus masking 5076c50d8ae3SPaolo Bonzini * away both AD and WD. For all reads or if the last condition holds, WD 5077c50d8ae3SPaolo Bonzini * only will be masked away. 5078c50d8ae3SPaolo Bonzini */ 50792e4c0661SSean Christopherson static void update_pkru_bitmask(struct kvm_mmu *mmu) 5080c50d8ae3SPaolo Bonzini { 5081c50d8ae3SPaolo Bonzini unsigned bit; 5082c50d8ae3SPaolo Bonzini bool wp; 5083c50d8ae3SPaolo Bonzini 5084c50d8ae3SPaolo Bonzini mmu->pkru_mask = 0; 5085a3ca5281SChenyi Qiang 5086a3ca5281SChenyi Qiang if (!is_cr4_pke(mmu)) 5087c50d8ae3SPaolo Bonzini return; 5088c50d8ae3SPaolo Bonzini 50892e4c0661SSean Christopherson wp = is_cr0_wp(mmu); 5090c50d8ae3SPaolo Bonzini 5091c50d8ae3SPaolo Bonzini for (bit = 0; bit < ARRAY_SIZE(mmu->permissions); ++bit) { 5092c50d8ae3SPaolo Bonzini unsigned pfec, pkey_bits; 5093c50d8ae3SPaolo Bonzini bool check_pkey, check_write, ff, uf, wf, pte_user; 5094c50d8ae3SPaolo Bonzini 5095c50d8ae3SPaolo Bonzini pfec = bit << 1; 5096c50d8ae3SPaolo Bonzini ff = pfec & PFERR_FETCH_MASK; 5097c50d8ae3SPaolo Bonzini uf = pfec & PFERR_USER_MASK; 5098c50d8ae3SPaolo Bonzini wf = pfec & PFERR_WRITE_MASK; 5099c50d8ae3SPaolo Bonzini 5100c50d8ae3SPaolo Bonzini /* PFEC.RSVD is replaced by ACC_USER_MASK. */ 5101c50d8ae3SPaolo Bonzini pte_user = pfec & PFERR_RSVD_MASK; 5102c50d8ae3SPaolo Bonzini 5103c50d8ae3SPaolo Bonzini /* 5104c50d8ae3SPaolo Bonzini * Only need to check the access which is not an 5105c50d8ae3SPaolo Bonzini * instruction fetch and is to a user page. 5106c50d8ae3SPaolo Bonzini */ 5107c50d8ae3SPaolo Bonzini check_pkey = (!ff && pte_user); 5108c50d8ae3SPaolo Bonzini /* 5109c50d8ae3SPaolo Bonzini * write access is controlled by PKRU if it is a 5110c50d8ae3SPaolo Bonzini * user access or CR0.WP = 1. 5111c50d8ae3SPaolo Bonzini */ 5112c50d8ae3SPaolo Bonzini check_write = check_pkey && wf && (uf || wp); 5113c50d8ae3SPaolo Bonzini 5114c50d8ae3SPaolo Bonzini /* PKRU.AD stops both read and write access. */ 5115c50d8ae3SPaolo Bonzini pkey_bits = !!check_pkey; 5116c50d8ae3SPaolo Bonzini /* PKRU.WD stops write access. */ 5117c50d8ae3SPaolo Bonzini pkey_bits |= (!!check_write) << 1; 5118c50d8ae3SPaolo Bonzini 5119c50d8ae3SPaolo Bonzini mmu->pkru_mask |= (pkey_bits & 3) << pfec; 5120c50d8ae3SPaolo Bonzini } 5121c50d8ae3SPaolo Bonzini } 5122c50d8ae3SPaolo Bonzini 5123533f9a4bSSean Christopherson static void reset_guest_paging_metadata(struct kvm_vcpu *vcpu, 5124533f9a4bSSean Christopherson struct kvm_mmu *mmu) 5125c50d8ae3SPaolo Bonzini { 5126533f9a4bSSean Christopherson if (!is_cr0_pg(mmu)) 5127533f9a4bSSean Christopherson return; 5128c50d8ae3SPaolo Bonzini 5129c919e881SKai Huang reset_guest_rsvds_bits_mask(vcpu, mmu); 5130533f9a4bSSean Christopherson update_permission_bitmask(mmu, false); 5131533f9a4bSSean Christopherson update_pkru_bitmask(mmu); 5132c50d8ae3SPaolo Bonzini } 5133c50d8ae3SPaolo Bonzini 5134fe660f72SSean Christopherson static void paging64_init_context(struct kvm_mmu *context) 5135c50d8ae3SPaolo Bonzini { 5136c50d8ae3SPaolo Bonzini context->page_fault = paging64_page_fault; 5137c50d8ae3SPaolo Bonzini context->gva_to_gpa = paging64_gva_to_gpa; 5138c3c6c9fcSLai Jiangshan context->sync_spte = paging64_sync_spte; 5139c50d8ae3SPaolo Bonzini } 5140c50d8ae3SPaolo Bonzini 514184a16226SSean Christopherson static void paging32_init_context(struct kvm_mmu *context) 5142c50d8ae3SPaolo Bonzini { 5143c50d8ae3SPaolo Bonzini context->page_fault = paging32_page_fault; 5144c50d8ae3SPaolo Bonzini context->gva_to_gpa = paging32_gva_to_gpa; 5145c3c6c9fcSLai Jiangshan context->sync_spte = paging32_sync_spte; 5146c50d8ae3SPaolo Bonzini } 5147c50d8ae3SPaolo Bonzini 5148f3d90f90SSean Christopherson static union kvm_cpu_role kvm_calc_cpu_role(struct kvm_vcpu *vcpu, 5149f3d90f90SSean Christopherson const struct kvm_mmu_role_regs *regs) 5150e5ed0fb0SPaolo Bonzini { 51517a7ae829SPaolo Bonzini union kvm_cpu_role role = {0}; 5152e5ed0fb0SPaolo Bonzini 5153e5ed0fb0SPaolo Bonzini role.base.access = ACC_ALL; 5154e5ed0fb0SPaolo Bonzini role.base.smm = is_smm(vcpu); 5155e5ed0fb0SPaolo Bonzini role.base.guest_mode = is_guest_mode(vcpu); 5156e5ed0fb0SPaolo Bonzini role.ext.valid = 1; 5157e5ed0fb0SPaolo Bonzini 5158e5ed0fb0SPaolo Bonzini if (!____is_cr0_pg(regs)) { 5159e5ed0fb0SPaolo Bonzini role.base.direct = 1; 5160e5ed0fb0SPaolo Bonzini return role; 5161e5ed0fb0SPaolo Bonzini } 5162e5ed0fb0SPaolo Bonzini 5163e5ed0fb0SPaolo Bonzini role.base.efer_nx = ____is_efer_nx(regs); 5164e5ed0fb0SPaolo Bonzini role.base.cr0_wp = ____is_cr0_wp(regs); 5165e5ed0fb0SPaolo Bonzini role.base.smep_andnot_wp = ____is_cr4_smep(regs) && !____is_cr0_wp(regs); 5166e5ed0fb0SPaolo Bonzini role.base.smap_andnot_wp = ____is_cr4_smap(regs) && !____is_cr0_wp(regs); 5167e5ed0fb0SPaolo Bonzini role.base.has_4_byte_gpte = !____is_cr4_pae(regs); 516860f3cb60SPaolo Bonzini 516960f3cb60SPaolo Bonzini if (____is_efer_lma(regs)) 517060f3cb60SPaolo Bonzini role.base.level = ____is_cr4_la57(regs) ? PT64_ROOT_5LEVEL 517160f3cb60SPaolo Bonzini : PT64_ROOT_4LEVEL; 517260f3cb60SPaolo Bonzini else if (____is_cr4_pae(regs)) 517360f3cb60SPaolo Bonzini role.base.level = PT32E_ROOT_LEVEL; 517460f3cb60SPaolo Bonzini else 517560f3cb60SPaolo Bonzini role.base.level = PT32_ROOT_LEVEL; 5176e5ed0fb0SPaolo Bonzini 5177e5ed0fb0SPaolo Bonzini role.ext.cr4_smep = ____is_cr4_smep(regs); 5178e5ed0fb0SPaolo Bonzini role.ext.cr4_smap = ____is_cr4_smap(regs); 5179e5ed0fb0SPaolo Bonzini role.ext.cr4_pse = ____is_cr4_pse(regs); 5180e5ed0fb0SPaolo Bonzini 5181e5ed0fb0SPaolo Bonzini /* PKEY and LA57 are active iff long mode is active. */ 5182e5ed0fb0SPaolo Bonzini role.ext.cr4_pke = ____is_efer_lma(regs) && ____is_cr4_pke(regs); 5183e5ed0fb0SPaolo Bonzini role.ext.cr4_la57 = ____is_efer_lma(regs) && ____is_cr4_la57(regs); 5184e5ed0fb0SPaolo Bonzini role.ext.efer_lma = ____is_efer_lma(regs); 5185e5ed0fb0SPaolo Bonzini return role; 5186e5ed0fb0SPaolo Bonzini } 5187e5ed0fb0SPaolo Bonzini 5188cf9f4c0eSSean Christopherson void __kvm_mmu_refresh_passthrough_bits(struct kvm_vcpu *vcpu, 5189cf9f4c0eSSean Christopherson struct kvm_mmu *mmu) 5190cf9f4c0eSSean Christopherson { 5191cf9f4c0eSSean Christopherson const bool cr0_wp = kvm_is_cr0_bit_set(vcpu, X86_CR0_WP); 5192cf9f4c0eSSean Christopherson 5193cf9f4c0eSSean Christopherson BUILD_BUG_ON((KVM_MMU_CR0_ROLE_BITS & KVM_POSSIBLE_CR0_GUEST_BITS) != X86_CR0_WP); 5194cf9f4c0eSSean Christopherson BUILD_BUG_ON((KVM_MMU_CR4_ROLE_BITS & KVM_POSSIBLE_CR4_GUEST_BITS)); 5195cf9f4c0eSSean Christopherson 5196cf9f4c0eSSean Christopherson if (is_cr0_wp(mmu) == cr0_wp) 5197cf9f4c0eSSean Christopherson return; 5198cf9f4c0eSSean Christopherson 5199cf9f4c0eSSean Christopherson mmu->cpu_role.base.cr0_wp = cr0_wp; 5200cf9f4c0eSSean Christopherson reset_guest_paging_metadata(vcpu, mmu); 5201cf9f4c0eSSean Christopherson } 5202cf9f4c0eSSean Christopherson 5203d468d94bSSean Christopherson static inline int kvm_mmu_get_tdp_level(struct kvm_vcpu *vcpu) 5204d468d94bSSean Christopherson { 5205746700d2SWei Huang /* tdp_root_level is architecture forced level, use it if nonzero */ 5206746700d2SWei Huang if (tdp_root_level) 5207746700d2SWei Huang return tdp_root_level; 5208746700d2SWei Huang 5209d468d94bSSean Christopherson /* Use 5-level TDP if and only if it's useful/necessary. */ 521083013059SSean Christopherson if (max_tdp_level == 5 && cpuid_maxphyaddr(vcpu) <= 48) 5211d468d94bSSean Christopherson return 4; 5212d468d94bSSean Christopherson 521383013059SSean Christopherson return max_tdp_level; 5214d468d94bSSean Christopherson } 5215d468d94bSSean Christopherson 52167a458f0eSPaolo Bonzini static union kvm_mmu_page_role 52178626c120SSean Christopherson kvm_calc_tdp_mmu_root_page_role(struct kvm_vcpu *vcpu, 52187a7ae829SPaolo Bonzini union kvm_cpu_role cpu_role) 5219c50d8ae3SPaolo Bonzini { 52207a458f0eSPaolo Bonzini union kvm_mmu_page_role role = {0}; 5221c50d8ae3SPaolo Bonzini 52227a458f0eSPaolo Bonzini role.access = ACC_ALL; 52237a458f0eSPaolo Bonzini role.cr0_wp = true; 52247a458f0eSPaolo Bonzini role.efer_nx = true; 52257a458f0eSPaolo Bonzini role.smm = cpu_role.base.smm; 52267a458f0eSPaolo Bonzini role.guest_mode = cpu_role.base.guest_mode; 522754275f74SSean Christopherson role.ad_disabled = !kvm_ad_enabled(); 52287a458f0eSPaolo Bonzini role.level = kvm_mmu_get_tdp_level(vcpu); 52297a458f0eSPaolo Bonzini role.direct = true; 52307a458f0eSPaolo Bonzini role.has_4_byte_gpte = false; 5231c50d8ae3SPaolo Bonzini 5232c50d8ae3SPaolo Bonzini return role; 5233c50d8ae3SPaolo Bonzini } 5234c50d8ae3SPaolo Bonzini 523539e7e2bfSPaolo Bonzini static void init_kvm_tdp_mmu(struct kvm_vcpu *vcpu, 5236a7f1de9bSPaolo Bonzini union kvm_cpu_role cpu_role) 5237c50d8ae3SPaolo Bonzini { 52388c008659SPaolo Bonzini struct kvm_mmu *context = &vcpu->arch.root_mmu; 52397a458f0eSPaolo Bonzini union kvm_mmu_page_role root_role = kvm_calc_tdp_mmu_root_page_role(vcpu, cpu_role); 5240c50d8ae3SPaolo Bonzini 5241e5ed0fb0SPaolo Bonzini if (cpu_role.as_u64 == context->cpu_role.as_u64 && 52427a458f0eSPaolo Bonzini root_role.word == context->root_role.word) 5243c50d8ae3SPaolo Bonzini return; 5244c50d8ae3SPaolo Bonzini 5245e5ed0fb0SPaolo Bonzini context->cpu_role.as_u64 = cpu_role.as_u64; 52467a458f0eSPaolo Bonzini context->root_role.word = root_role.word; 52477a02674dSSean Christopherson context->page_fault = kvm_tdp_page_fault; 5248c3c6c9fcSLai Jiangshan context->sync_spte = NULL; 52492fdcc1b3SPaolo Bonzini context->get_guest_pgd = get_guest_cr3; 5250c50d8ae3SPaolo Bonzini context->get_pdptr = kvm_pdptr_read; 5251c50d8ae3SPaolo Bonzini context->inject_page_fault = kvm_inject_page_fault; 5252c50d8ae3SPaolo Bonzini 525336f26787SSean Christopherson if (!is_cr0_pg(context)) 5254c50d8ae3SPaolo Bonzini context->gva_to_gpa = nonpaging_gva_to_gpa; 525536f26787SSean Christopherson else if (is_cr4_pae(context)) 5256c50d8ae3SPaolo Bonzini context->gva_to_gpa = paging64_gva_to_gpa; 5257f4bd6f73SSean Christopherson else 5258c50d8ae3SPaolo Bonzini context->gva_to_gpa = paging32_gva_to_gpa; 5259c50d8ae3SPaolo Bonzini 5260533f9a4bSSean Christopherson reset_guest_paging_metadata(vcpu, context); 5261e8f6e738SJinrong Liang reset_tdp_shadow_zero_bits_mask(context); 5262c50d8ae3SPaolo Bonzini } 5263c50d8ae3SPaolo Bonzini 52648c008659SPaolo Bonzini static void shadow_mmu_init_context(struct kvm_vcpu *vcpu, struct kvm_mmu *context, 52657a7ae829SPaolo Bonzini union kvm_cpu_role cpu_role, 52667a458f0eSPaolo Bonzini union kvm_mmu_page_role root_role) 5267c50d8ae3SPaolo Bonzini { 5268e5ed0fb0SPaolo Bonzini if (cpu_role.as_u64 == context->cpu_role.as_u64 && 52697a458f0eSPaolo Bonzini root_role.word == context->root_role.word) 527018db1b17SSean Christopherson return; 5271c50d8ae3SPaolo Bonzini 5272e5ed0fb0SPaolo Bonzini context->cpu_role.as_u64 = cpu_role.as_u64; 52737a458f0eSPaolo Bonzini context->root_role.word = root_role.word; 527418db1b17SSean Christopherson 527536f26787SSean Christopherson if (!is_cr0_pg(context)) 527684a16226SSean Christopherson nonpaging_init_context(context); 527736f26787SSean Christopherson else if (is_cr4_pae(context)) 5278fe660f72SSean Christopherson paging64_init_context(context); 5279c50d8ae3SPaolo Bonzini else 528084a16226SSean Christopherson paging32_init_context(context); 5281c50d8ae3SPaolo Bonzini 5282533f9a4bSSean Christopherson reset_guest_paging_metadata(vcpu, context); 5283c50d8ae3SPaolo Bonzini reset_shadow_zero_bits_mask(vcpu, context); 5284c50d8ae3SPaolo Bonzini } 52850f04a2acSVitaly Kuznetsov 5286594e91a1SSean Christopherson static void kvm_init_shadow_mmu(struct kvm_vcpu *vcpu, 5287a7f1de9bSPaolo Bonzini union kvm_cpu_role cpu_role) 52880f04a2acSVitaly Kuznetsov { 52898c008659SPaolo Bonzini struct kvm_mmu *context = &vcpu->arch.root_mmu; 529056b321f9SPaolo Bonzini union kvm_mmu_page_role root_role; 529156b321f9SPaolo Bonzini 529256b321f9SPaolo Bonzini root_role = cpu_role.base; 529356b321f9SPaolo Bonzini 529456b321f9SPaolo Bonzini /* KVM uses PAE paging whenever the guest isn't using 64-bit paging. */ 529556b321f9SPaolo Bonzini root_role.level = max_t(u32, root_role.level, PT32E_ROOT_LEVEL); 529656b321f9SPaolo Bonzini 529756b321f9SPaolo Bonzini /* 529856b321f9SPaolo Bonzini * KVM forces EFER.NX=1 when TDP is disabled, reflect it in the MMU role. 529956b321f9SPaolo Bonzini * KVM uses NX when TDP is disabled to handle a variety of scenarios, 530056b321f9SPaolo Bonzini * notably for huge SPTEs if iTLB multi-hit mitigation is enabled and 530156b321f9SPaolo Bonzini * to generate correct permissions for CR0.WP=0/CR4.SMEP=1/EFER.NX=0. 530256b321f9SPaolo Bonzini * The iTLB multi-hit workaround can be toggled at any time, so assume 530356b321f9SPaolo Bonzini * NX can be used by any non-nested shadow MMU to avoid having to reset 530456b321f9SPaolo Bonzini * MMU contexts. 530556b321f9SPaolo Bonzini */ 530656b321f9SPaolo Bonzini root_role.efer_nx = true; 53070f04a2acSVitaly Kuznetsov 53087a458f0eSPaolo Bonzini shadow_mmu_init_context(vcpu, context, cpu_role, root_role); 53090f04a2acSVitaly Kuznetsov } 53100f04a2acSVitaly Kuznetsov 5311dbc4739bSSean Christopherson void kvm_init_shadow_npt_mmu(struct kvm_vcpu *vcpu, unsigned long cr0, 5312dbc4739bSSean Christopherson unsigned long cr4, u64 efer, gpa_t nested_cr3) 53130f04a2acSVitaly Kuznetsov { 53148c008659SPaolo Bonzini struct kvm_mmu *context = &vcpu->arch.guest_mmu; 5315594e91a1SSean Christopherson struct kvm_mmu_role_regs regs = { 5316594e91a1SSean Christopherson .cr0 = cr0, 531728f091bcSPaolo Bonzini .cr4 = cr4 & ~X86_CR4_PKE, 5318594e91a1SSean Christopherson .efer = efer, 5319594e91a1SSean Christopherson }; 53207a7ae829SPaolo Bonzini union kvm_cpu_role cpu_role = kvm_calc_cpu_role(vcpu, ®s); 532156b321f9SPaolo Bonzini union kvm_mmu_page_role root_role; 532256b321f9SPaolo Bonzini 532356b321f9SPaolo Bonzini /* NPT requires CR0.PG=1. */ 532456b321f9SPaolo Bonzini WARN_ON_ONCE(cpu_role.base.direct); 532556b321f9SPaolo Bonzini 532656b321f9SPaolo Bonzini root_role = cpu_role.base; 532756b321f9SPaolo Bonzini root_role.level = kvm_mmu_get_tdp_level(vcpu); 532884e5ffd0SLai Jiangshan if (root_role.level == PT64_ROOT_5LEVEL && 532984e5ffd0SLai Jiangshan cpu_role.base.level == PT64_ROOT_4LEVEL) 533084e5ffd0SLai Jiangshan root_role.passthrough = 1; 53310f04a2acSVitaly Kuznetsov 53327a458f0eSPaolo Bonzini shadow_mmu_init_context(vcpu, context, cpu_role, root_role); 5333d2e5f333SPaolo Bonzini kvm_mmu_new_pgd(vcpu, nested_cr3); 53340f04a2acSVitaly Kuznetsov } 53350f04a2acSVitaly Kuznetsov EXPORT_SYMBOL_GPL(kvm_init_shadow_npt_mmu); 5336c50d8ae3SPaolo Bonzini 53377a7ae829SPaolo Bonzini static union kvm_cpu_role 5338c50d8ae3SPaolo Bonzini kvm_calc_shadow_ept_root_page_role(struct kvm_vcpu *vcpu, bool accessed_dirty, 5339bb1fcc70SSean Christopherson bool execonly, u8 level) 5340c50d8ae3SPaolo Bonzini { 53417a7ae829SPaolo Bonzini union kvm_cpu_role role = {0}; 5342c50d8ae3SPaolo Bonzini 5343daed87b8SPaolo Bonzini /* 5344daed87b8SPaolo Bonzini * KVM does not support SMM transfer monitors, and consequently does not 5345daed87b8SPaolo Bonzini * support the "entry to SMM" control either. role.base.smm is always 0. 5346daed87b8SPaolo Bonzini */ 5347daed87b8SPaolo Bonzini WARN_ON_ONCE(is_smm(vcpu)); 5348bb1fcc70SSean Christopherson role.base.level = level; 5349bb3b394dSLai Jiangshan role.base.has_4_byte_gpte = false; 5350c50d8ae3SPaolo Bonzini role.base.direct = false; 5351c50d8ae3SPaolo Bonzini role.base.ad_disabled = !accessed_dirty; 5352c50d8ae3SPaolo Bonzini role.base.guest_mode = true; 5353c50d8ae3SPaolo Bonzini role.base.access = ACC_ALL; 5354c50d8ae3SPaolo Bonzini 5355cd6767c3SSean Christopherson role.ext.word = 0; 5356c50d8ae3SPaolo Bonzini role.ext.execonly = execonly; 5357cd6767c3SSean Christopherson role.ext.valid = 1; 5358c50d8ae3SPaolo Bonzini 5359c50d8ae3SPaolo Bonzini return role; 5360c50d8ae3SPaolo Bonzini } 5361c50d8ae3SPaolo Bonzini 5362c50d8ae3SPaolo Bonzini void kvm_init_shadow_ept_mmu(struct kvm_vcpu *vcpu, bool execonly, 5363cc022ae1SLai Jiangshan int huge_page_level, bool accessed_dirty, 5364cc022ae1SLai Jiangshan gpa_t new_eptp) 5365c50d8ae3SPaolo Bonzini { 53668c008659SPaolo Bonzini struct kvm_mmu *context = &vcpu->arch.guest_mmu; 5367bb1fcc70SSean Christopherson u8 level = vmx_eptp_page_walk_level(new_eptp); 53687a7ae829SPaolo Bonzini union kvm_cpu_role new_mode = 5369c50d8ae3SPaolo Bonzini kvm_calc_shadow_ept_root_page_role(vcpu, accessed_dirty, 5370bb1fcc70SSean Christopherson execonly, level); 5371c50d8ae3SPaolo Bonzini 5372e5ed0fb0SPaolo Bonzini if (new_mode.as_u64 != context->cpu_role.as_u64) { 5373e5ed0fb0SPaolo Bonzini /* EPT, and thus nested EPT, does not consume CR0, CR4, nor EFER. */ 5374e5ed0fb0SPaolo Bonzini context->cpu_role.as_u64 = new_mode.as_u64; 53757a458f0eSPaolo Bonzini context->root_role.word = new_mode.base.word; 537618db1b17SSean Christopherson 5377c50d8ae3SPaolo Bonzini context->page_fault = ept_page_fault; 5378c50d8ae3SPaolo Bonzini context->gva_to_gpa = ept_gva_to_gpa; 5379c3c6c9fcSLai Jiangshan context->sync_spte = ept_sync_spte; 5380347a0d0dSPaolo Bonzini 5381c596f147SSean Christopherson update_permission_bitmask(context, true); 538228f091bcSPaolo Bonzini context->pkru_mask = 0; 5383cc022ae1SLai Jiangshan reset_rsvds_bits_mask_ept(vcpu, context, execonly, huge_page_level); 5384e8f6e738SJinrong Liang reset_ept_shadow_zero_bits_mask(context, execonly); 5385c50d8ae3SPaolo Bonzini } 53863cffc89dSPaolo Bonzini 5387d2e5f333SPaolo Bonzini kvm_mmu_new_pgd(vcpu, new_eptp); 53883cffc89dSPaolo Bonzini } 5389c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_init_shadow_ept_mmu); 5390c50d8ae3SPaolo Bonzini 539139e7e2bfSPaolo Bonzini static void init_kvm_softmmu(struct kvm_vcpu *vcpu, 5392a7f1de9bSPaolo Bonzini union kvm_cpu_role cpu_role) 5393c50d8ae3SPaolo Bonzini { 53948c008659SPaolo Bonzini struct kvm_mmu *context = &vcpu->arch.root_mmu; 5395c50d8ae3SPaolo Bonzini 5396a7f1de9bSPaolo Bonzini kvm_init_shadow_mmu(vcpu, cpu_role); 5397929d1cfaSPaolo Bonzini 53982fdcc1b3SPaolo Bonzini context->get_guest_pgd = get_guest_cr3; 5399c50d8ae3SPaolo Bonzini context->get_pdptr = kvm_pdptr_read; 5400c50d8ae3SPaolo Bonzini context->inject_page_fault = kvm_inject_page_fault; 5401c50d8ae3SPaolo Bonzini } 5402c50d8ae3SPaolo Bonzini 540339e7e2bfSPaolo Bonzini static void init_kvm_nested_mmu(struct kvm_vcpu *vcpu, 5404a7f1de9bSPaolo Bonzini union kvm_cpu_role new_mode) 5405c50d8ae3SPaolo Bonzini { 5406c50d8ae3SPaolo Bonzini struct kvm_mmu *g_context = &vcpu->arch.nested_mmu; 5407c50d8ae3SPaolo Bonzini 5408e5ed0fb0SPaolo Bonzini if (new_mode.as_u64 == g_context->cpu_role.as_u64) 5409c50d8ae3SPaolo Bonzini return; 5410c50d8ae3SPaolo Bonzini 5411e5ed0fb0SPaolo Bonzini g_context->cpu_role.as_u64 = new_mode.as_u64; 54122fdcc1b3SPaolo Bonzini g_context->get_guest_pgd = get_guest_cr3; 5413c50d8ae3SPaolo Bonzini g_context->get_pdptr = kvm_pdptr_read; 5414c50d8ae3SPaolo Bonzini g_context->inject_page_fault = kvm_inject_page_fault; 5415c50d8ae3SPaolo Bonzini 5416c50d8ae3SPaolo Bonzini /* 54175efac074SPaolo Bonzini * L2 page tables are never shadowed, so there is no need to sync 54185efac074SPaolo Bonzini * SPTEs. 54195efac074SPaolo Bonzini */ 54209fd4a4e3SLai Jiangshan g_context->sync_spte = NULL; 54215efac074SPaolo Bonzini 54225efac074SPaolo Bonzini /* 5423c50d8ae3SPaolo Bonzini * Note that arch.mmu->gva_to_gpa translates l2_gpa to l1_gpa using 5424c50d8ae3SPaolo Bonzini * L1's nested page tables (e.g. EPT12). The nested translation 5425c50d8ae3SPaolo Bonzini * of l2_gva to l1_gpa is done by arch.nested_mmu.gva_to_gpa using 5426c50d8ae3SPaolo Bonzini * L2's page tables as the first level of translation and L1's 5427c50d8ae3SPaolo Bonzini * nested page tables as the second level of translation. Basically 5428c50d8ae3SPaolo Bonzini * the gva_to_gpa functions between mmu and nested_mmu are swapped. 5429c50d8ae3SPaolo Bonzini */ 5430fa4b5588SSean Christopherson if (!is_paging(vcpu)) 54311f5a21eeSLai Jiangshan g_context->gva_to_gpa = nonpaging_gva_to_gpa; 5432fa4b5588SSean Christopherson else if (is_long_mode(vcpu)) 54331f5a21eeSLai Jiangshan g_context->gva_to_gpa = paging64_gva_to_gpa; 5434fa4b5588SSean Christopherson else if (is_pae(vcpu)) 54351f5a21eeSLai Jiangshan g_context->gva_to_gpa = paging64_gva_to_gpa; 5436fa4b5588SSean Christopherson else 54371f5a21eeSLai Jiangshan g_context->gva_to_gpa = paging32_gva_to_gpa; 5438fa4b5588SSean Christopherson 5439533f9a4bSSean Christopherson reset_guest_paging_metadata(vcpu, g_context); 5440c50d8ae3SPaolo Bonzini } 5441c50d8ae3SPaolo Bonzini 5442c9060662SSean Christopherson void kvm_init_mmu(struct kvm_vcpu *vcpu) 5443c50d8ae3SPaolo Bonzini { 544439e7e2bfSPaolo Bonzini struct kvm_mmu_role_regs regs = vcpu_to_role_regs(vcpu); 5445a7f1de9bSPaolo Bonzini union kvm_cpu_role cpu_role = kvm_calc_cpu_role(vcpu, ®s); 544639e7e2bfSPaolo Bonzini 5447c50d8ae3SPaolo Bonzini if (mmu_is_nested(vcpu)) 5448a7f1de9bSPaolo Bonzini init_kvm_nested_mmu(vcpu, cpu_role); 5449c50d8ae3SPaolo Bonzini else if (tdp_enabled) 5450a7f1de9bSPaolo Bonzini init_kvm_tdp_mmu(vcpu, cpu_role); 5451c50d8ae3SPaolo Bonzini else 5452a7f1de9bSPaolo Bonzini init_kvm_softmmu(vcpu, cpu_role); 5453c50d8ae3SPaolo Bonzini } 5454c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_init_mmu); 5455c50d8ae3SPaolo Bonzini 545649c6f875SSean Christopherson void kvm_mmu_after_set_cpuid(struct kvm_vcpu *vcpu) 545749c6f875SSean Christopherson { 545849c6f875SSean Christopherson /* 545949c6f875SSean Christopherson * Invalidate all MMU roles to force them to reinitialize as CPUID 546049c6f875SSean Christopherson * information is factored into reserved bit calculations. 5461feb627e8SVitaly Kuznetsov * 5462feb627e8SVitaly Kuznetsov * Correctly handling multiple vCPU models with respect to paging and 5463feb627e8SVitaly Kuznetsov * physical address properties) in a single VM would require tracking 5464feb627e8SVitaly Kuznetsov * all relevant CPUID information in kvm_mmu_page_role. That is very 5465feb627e8SVitaly Kuznetsov * undesirable as it would increase the memory requirements for 5466feb627e8SVitaly Kuznetsov * gfn_track (see struct kvm_mmu_page_role comments). For now that 5467feb627e8SVitaly Kuznetsov * problem is swept under the rug; KVM's CPUID API is horrific and 5468feb627e8SVitaly Kuznetsov * it's all but impossible to solve it without introducing a new API. 546949c6f875SSean Christopherson */ 54707a458f0eSPaolo Bonzini vcpu->arch.root_mmu.root_role.word = 0; 54717a458f0eSPaolo Bonzini vcpu->arch.guest_mmu.root_role.word = 0; 54727a458f0eSPaolo Bonzini vcpu->arch.nested_mmu.root_role.word = 0; 5473e5ed0fb0SPaolo Bonzini vcpu->arch.root_mmu.cpu_role.ext.valid = 0; 5474e5ed0fb0SPaolo Bonzini vcpu->arch.guest_mmu.cpu_role.ext.valid = 0; 5475e5ed0fb0SPaolo Bonzini vcpu->arch.nested_mmu.cpu_role.ext.valid = 0; 547649c6f875SSean Christopherson kvm_mmu_reset_context(vcpu); 547763f5a190SSean Christopherson 547863f5a190SSean Christopherson /* 5479feb627e8SVitaly Kuznetsov * Changing guest CPUID after KVM_RUN is forbidden, see the comment in 5480feb627e8SVitaly Kuznetsov * kvm_arch_vcpu_ioctl(). 548163f5a190SSean Christopherson */ 5482fb3146b4SSean Christopherson KVM_BUG_ON(kvm_vcpu_has_run(vcpu), vcpu->kvm); 548349c6f875SSean Christopherson } 548449c6f875SSean Christopherson 5485c50d8ae3SPaolo Bonzini void kvm_mmu_reset_context(struct kvm_vcpu *vcpu) 5486c50d8ae3SPaolo Bonzini { 5487c50d8ae3SPaolo Bonzini kvm_mmu_unload(vcpu); 5488c9060662SSean Christopherson kvm_init_mmu(vcpu); 5489c50d8ae3SPaolo Bonzini } 5490c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_mmu_reset_context); 5491c50d8ae3SPaolo Bonzini 5492c50d8ae3SPaolo Bonzini int kvm_mmu_load(struct kvm_vcpu *vcpu) 5493c50d8ae3SPaolo Bonzini { 5494c50d8ae3SPaolo Bonzini int r; 5495c50d8ae3SPaolo Bonzini 5496347a0d0dSPaolo Bonzini r = mmu_topup_memory_caches(vcpu, !vcpu->arch.mmu->root_role.direct); 5497c50d8ae3SPaolo Bonzini if (r) 5498c50d8ae3SPaolo Bonzini goto out; 5499748e52b9SSean Christopherson r = mmu_alloc_special_roots(vcpu); 5500c50d8ae3SPaolo Bonzini if (r) 5501c50d8ae3SPaolo Bonzini goto out; 5502347a0d0dSPaolo Bonzini if (vcpu->arch.mmu->root_role.direct) 55036e6ec584SSean Christopherson r = mmu_alloc_direct_roots(vcpu); 55046e6ec584SSean Christopherson else 55056e6ec584SSean Christopherson r = mmu_alloc_shadow_roots(vcpu); 5506c50d8ae3SPaolo Bonzini if (r) 5507c50d8ae3SPaolo Bonzini goto out; 5508a91f387bSSean Christopherson 5509a91f387bSSean Christopherson kvm_mmu_sync_roots(vcpu); 5510a91f387bSSean Christopherson 5511727a7e27SPaolo Bonzini kvm_mmu_load_pgd(vcpu); 5512db01416bSSean Christopherson 5513db01416bSSean Christopherson /* 5514db01416bSSean Christopherson * Flush any TLB entries for the new root, the provenance of the root 5515db01416bSSean Christopherson * is unknown. Even if KVM ensures there are no stale TLB entries 5516db01416bSSean Christopherson * for a freed root, in theory another hypervisor could have left 5517db01416bSSean Christopherson * stale entries. Flushing on alloc also allows KVM to skip the TLB 5518db01416bSSean Christopherson * flush when freeing a root (see kvm_tdp_mmu_put_root()). 5519db01416bSSean Christopherson */ 5520e27bc044SSean Christopherson static_call(kvm_x86_flush_tlb_current)(vcpu); 5521c50d8ae3SPaolo Bonzini out: 5522c50d8ae3SPaolo Bonzini return r; 5523c50d8ae3SPaolo Bonzini } 5524c50d8ae3SPaolo Bonzini 5525c50d8ae3SPaolo Bonzini void kvm_mmu_unload(struct kvm_vcpu *vcpu) 5526c50d8ae3SPaolo Bonzini { 55270c1c92f1SPaolo Bonzini struct kvm *kvm = vcpu->kvm; 55280c1c92f1SPaolo Bonzini 55290c1c92f1SPaolo Bonzini kvm_mmu_free_roots(kvm, &vcpu->arch.root_mmu, KVM_MMU_ROOTS_ALL); 5530b9e5603cSPaolo Bonzini WARN_ON(VALID_PAGE(vcpu->arch.root_mmu.root.hpa)); 55310c1c92f1SPaolo Bonzini kvm_mmu_free_roots(kvm, &vcpu->arch.guest_mmu, KVM_MMU_ROOTS_ALL); 5532b9e5603cSPaolo Bonzini WARN_ON(VALID_PAGE(vcpu->arch.guest_mmu.root.hpa)); 55336d58f275SPaolo Bonzini vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY); 5534c50d8ae3SPaolo Bonzini } 5535c50d8ae3SPaolo Bonzini 5536527d5cd7SSean Christopherson static bool is_obsolete_root(struct kvm *kvm, hpa_t root_hpa) 5537527d5cd7SSean Christopherson { 5538527d5cd7SSean Christopherson struct kvm_mmu_page *sp; 5539527d5cd7SSean Christopherson 5540527d5cd7SSean Christopherson if (!VALID_PAGE(root_hpa)) 5541527d5cd7SSean Christopherson return false; 5542527d5cd7SSean Christopherson 5543527d5cd7SSean Christopherson /* 5544527d5cd7SSean Christopherson * When freeing obsolete roots, treat roots as obsolete if they don't 5545527d5cd7SSean Christopherson * have an associated shadow page. This does mean KVM will get false 5546527d5cd7SSean Christopherson * positives and free roots that don't strictly need to be freed, but 5547527d5cd7SSean Christopherson * such false positives are relatively rare: 5548527d5cd7SSean Christopherson * 5549527d5cd7SSean Christopherson * (a) only PAE paging and nested NPT has roots without shadow pages 5550527d5cd7SSean Christopherson * (b) remote reloads due to a memslot update obsoletes _all_ roots 5551527d5cd7SSean Christopherson * (c) KVM doesn't track previous roots for PAE paging, and the guest 5552527d5cd7SSean Christopherson * is unlikely to zap an in-use PGD. 5553527d5cd7SSean Christopherson */ 5554527d5cd7SSean Christopherson sp = to_shadow_page(root_hpa); 5555527d5cd7SSean Christopherson return !sp || is_obsolete_sp(kvm, sp); 5556527d5cd7SSean Christopherson } 5557527d5cd7SSean Christopherson 5558527d5cd7SSean Christopherson static void __kvm_mmu_free_obsolete_roots(struct kvm *kvm, struct kvm_mmu *mmu) 5559527d5cd7SSean Christopherson { 5560527d5cd7SSean Christopherson unsigned long roots_to_free = 0; 5561527d5cd7SSean Christopherson int i; 5562527d5cd7SSean Christopherson 5563527d5cd7SSean Christopherson if (is_obsolete_root(kvm, mmu->root.hpa)) 5564527d5cd7SSean Christopherson roots_to_free |= KVM_MMU_ROOT_CURRENT; 5565527d5cd7SSean Christopherson 5566527d5cd7SSean Christopherson for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) { 5567cf4a8693SShaoqin Huang if (is_obsolete_root(kvm, mmu->prev_roots[i].hpa)) 5568527d5cd7SSean Christopherson roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i); 5569527d5cd7SSean Christopherson } 5570527d5cd7SSean Christopherson 5571527d5cd7SSean Christopherson if (roots_to_free) 5572527d5cd7SSean Christopherson kvm_mmu_free_roots(kvm, mmu, roots_to_free); 5573527d5cd7SSean Christopherson } 5574527d5cd7SSean Christopherson 5575527d5cd7SSean Christopherson void kvm_mmu_free_obsolete_roots(struct kvm_vcpu *vcpu) 5576527d5cd7SSean Christopherson { 5577527d5cd7SSean Christopherson __kvm_mmu_free_obsolete_roots(vcpu->kvm, &vcpu->arch.root_mmu); 5578527d5cd7SSean Christopherson __kvm_mmu_free_obsolete_roots(vcpu->kvm, &vcpu->arch.guest_mmu); 5579527d5cd7SSean Christopherson } 5580527d5cd7SSean Christopherson 5581c50d8ae3SPaolo Bonzini static u64 mmu_pte_write_fetch_gpte(struct kvm_vcpu *vcpu, gpa_t *gpa, 5582c50d8ae3SPaolo Bonzini int *bytes) 5583c50d8ae3SPaolo Bonzini { 5584c50d8ae3SPaolo Bonzini u64 gentry = 0; 5585c50d8ae3SPaolo Bonzini int r; 5586c50d8ae3SPaolo Bonzini 5587c50d8ae3SPaolo Bonzini /* 5588c50d8ae3SPaolo Bonzini * Assume that the pte write on a page table of the same type 5589c50d8ae3SPaolo Bonzini * as the current vcpu paging mode since we update the sptes only 5590c50d8ae3SPaolo Bonzini * when they have the same mode. 5591c50d8ae3SPaolo Bonzini */ 5592c50d8ae3SPaolo Bonzini if (is_pae(vcpu) && *bytes == 4) { 5593c50d8ae3SPaolo Bonzini /* Handle a 32-bit guest writing two halves of a 64-bit gpte */ 5594c50d8ae3SPaolo Bonzini *gpa &= ~(gpa_t)7; 5595c50d8ae3SPaolo Bonzini *bytes = 8; 5596c50d8ae3SPaolo Bonzini } 5597c50d8ae3SPaolo Bonzini 5598c50d8ae3SPaolo Bonzini if (*bytes == 4 || *bytes == 8) { 5599c50d8ae3SPaolo Bonzini r = kvm_vcpu_read_guest_atomic(vcpu, *gpa, &gentry, *bytes); 5600c50d8ae3SPaolo Bonzini if (r) 5601c50d8ae3SPaolo Bonzini gentry = 0; 5602c50d8ae3SPaolo Bonzini } 5603c50d8ae3SPaolo Bonzini 5604c50d8ae3SPaolo Bonzini return gentry; 5605c50d8ae3SPaolo Bonzini } 5606c50d8ae3SPaolo Bonzini 5607c50d8ae3SPaolo Bonzini /* 5608c50d8ae3SPaolo Bonzini * If we're seeing too many writes to a page, it may no longer be a page table, 5609c50d8ae3SPaolo Bonzini * or we may be forking, in which case it is better to unmap the page. 5610c50d8ae3SPaolo Bonzini */ 5611c50d8ae3SPaolo Bonzini static bool detect_write_flooding(struct kvm_mmu_page *sp) 5612c50d8ae3SPaolo Bonzini { 5613c50d8ae3SPaolo Bonzini /* 5614c50d8ae3SPaolo Bonzini * Skip write-flooding detected for the sp whose level is 1, because 5615c50d8ae3SPaolo Bonzini * it can become unsync, then the guest page is not write-protected. 5616c50d8ae3SPaolo Bonzini */ 56173bae0459SSean Christopherson if (sp->role.level == PG_LEVEL_4K) 5618c50d8ae3SPaolo Bonzini return false; 5619c50d8ae3SPaolo Bonzini 5620c50d8ae3SPaolo Bonzini atomic_inc(&sp->write_flooding_count); 5621c50d8ae3SPaolo Bonzini return atomic_read(&sp->write_flooding_count) >= 3; 5622c50d8ae3SPaolo Bonzini } 5623c50d8ae3SPaolo Bonzini 5624c50d8ae3SPaolo Bonzini /* 5625c50d8ae3SPaolo Bonzini * Misaligned accesses are too much trouble to fix up; also, they usually 5626c50d8ae3SPaolo Bonzini * indicate a page is not used as a page table. 5627c50d8ae3SPaolo Bonzini */ 5628c50d8ae3SPaolo Bonzini static bool detect_write_misaligned(struct kvm_mmu_page *sp, gpa_t gpa, 5629c50d8ae3SPaolo Bonzini int bytes) 5630c50d8ae3SPaolo Bonzini { 5631c50d8ae3SPaolo Bonzini unsigned offset, pte_size, misaligned; 5632c50d8ae3SPaolo Bonzini 5633c50d8ae3SPaolo Bonzini pgprintk("misaligned: gpa %llx bytes %d role %x\n", 5634c50d8ae3SPaolo Bonzini gpa, bytes, sp->role.word); 5635c50d8ae3SPaolo Bonzini 5636c50d8ae3SPaolo Bonzini offset = offset_in_page(gpa); 5637bb3b394dSLai Jiangshan pte_size = sp->role.has_4_byte_gpte ? 4 : 8; 5638c50d8ae3SPaolo Bonzini 5639c50d8ae3SPaolo Bonzini /* 5640c50d8ae3SPaolo Bonzini * Sometimes, the OS only writes the last one bytes to update status 5641c50d8ae3SPaolo Bonzini * bits, for example, in linux, andb instruction is used in clear_bit(). 5642c50d8ae3SPaolo Bonzini */ 5643c50d8ae3SPaolo Bonzini if (!(offset & (pte_size - 1)) && bytes == 1) 5644c50d8ae3SPaolo Bonzini return false; 5645c50d8ae3SPaolo Bonzini 5646c50d8ae3SPaolo Bonzini misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1); 5647c50d8ae3SPaolo Bonzini misaligned |= bytes < 4; 5648c50d8ae3SPaolo Bonzini 5649c50d8ae3SPaolo Bonzini return misaligned; 5650c50d8ae3SPaolo Bonzini } 5651c50d8ae3SPaolo Bonzini 5652c50d8ae3SPaolo Bonzini static u64 *get_written_sptes(struct kvm_mmu_page *sp, gpa_t gpa, int *nspte) 5653c50d8ae3SPaolo Bonzini { 5654c50d8ae3SPaolo Bonzini unsigned page_offset, quadrant; 5655c50d8ae3SPaolo Bonzini u64 *spte; 5656c50d8ae3SPaolo Bonzini int level; 5657c50d8ae3SPaolo Bonzini 5658c50d8ae3SPaolo Bonzini page_offset = offset_in_page(gpa); 5659c50d8ae3SPaolo Bonzini level = sp->role.level; 5660c50d8ae3SPaolo Bonzini *nspte = 1; 5661bb3b394dSLai Jiangshan if (sp->role.has_4_byte_gpte) { 5662c50d8ae3SPaolo Bonzini page_offset <<= 1; /* 32->64 */ 5663c50d8ae3SPaolo Bonzini /* 5664c50d8ae3SPaolo Bonzini * A 32-bit pde maps 4MB while the shadow pdes map 5665c50d8ae3SPaolo Bonzini * only 2MB. So we need to double the offset again 5666c50d8ae3SPaolo Bonzini * and zap two pdes instead of one. 5667c50d8ae3SPaolo Bonzini */ 5668c50d8ae3SPaolo Bonzini if (level == PT32_ROOT_LEVEL) { 5669c50d8ae3SPaolo Bonzini page_offset &= ~7; /* kill rounding error */ 5670c50d8ae3SPaolo Bonzini page_offset <<= 1; 5671c50d8ae3SPaolo Bonzini *nspte = 2; 5672c50d8ae3SPaolo Bonzini } 5673c50d8ae3SPaolo Bonzini quadrant = page_offset >> PAGE_SHIFT; 5674c50d8ae3SPaolo Bonzini page_offset &= ~PAGE_MASK; 5675c50d8ae3SPaolo Bonzini if (quadrant != sp->role.quadrant) 5676c50d8ae3SPaolo Bonzini return NULL; 5677c50d8ae3SPaolo Bonzini } 5678c50d8ae3SPaolo Bonzini 5679c50d8ae3SPaolo Bonzini spte = &sp->spt[page_offset / sizeof(*spte)]; 5680c50d8ae3SPaolo Bonzini return spte; 5681c50d8ae3SPaolo Bonzini } 5682c50d8ae3SPaolo Bonzini 5683c50d8ae3SPaolo Bonzini static void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa, 5684c50d8ae3SPaolo Bonzini const u8 *new, int bytes, 5685c50d8ae3SPaolo Bonzini struct kvm_page_track_notifier_node *node) 5686c50d8ae3SPaolo Bonzini { 5687c50d8ae3SPaolo Bonzini gfn_t gfn = gpa >> PAGE_SHIFT; 5688c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 5689c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 5690c50d8ae3SPaolo Bonzini u64 entry, gentry, *spte; 5691c50d8ae3SPaolo Bonzini int npte; 569206152b2dSLai Jiangshan bool flush = false; 5693c50d8ae3SPaolo Bonzini 5694c50d8ae3SPaolo Bonzini /* 5695c50d8ae3SPaolo Bonzini * If we don't have indirect shadow pages, it means no page is 5696c50d8ae3SPaolo Bonzini * write-protected, so we can exit simply. 5697c50d8ae3SPaolo Bonzini */ 5698c50d8ae3SPaolo Bonzini if (!READ_ONCE(vcpu->kvm->arch.indirect_shadow_pages)) 5699c50d8ae3SPaolo Bonzini return; 5700c50d8ae3SPaolo Bonzini 5701c50d8ae3SPaolo Bonzini pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes); 5702c50d8ae3SPaolo Bonzini 5703531810caSBen Gardon write_lock(&vcpu->kvm->mmu_lock); 5704c50d8ae3SPaolo Bonzini 5705c50d8ae3SPaolo Bonzini gentry = mmu_pte_write_fetch_gpte(vcpu, &gpa, &bytes); 5706c50d8ae3SPaolo Bonzini 5707c50d8ae3SPaolo Bonzini ++vcpu->kvm->stat.mmu_pte_write; 5708c50d8ae3SPaolo Bonzini 5709767d8d8dSLai Jiangshan for_each_gfn_valid_sp_with_gptes(vcpu->kvm, sp, gfn) { 5710c50d8ae3SPaolo Bonzini if (detect_write_misaligned(sp, gpa, bytes) || 5711c50d8ae3SPaolo Bonzini detect_write_flooding(sp)) { 5712c50d8ae3SPaolo Bonzini kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list); 5713c50d8ae3SPaolo Bonzini ++vcpu->kvm->stat.mmu_flooded; 5714c50d8ae3SPaolo Bonzini continue; 5715c50d8ae3SPaolo Bonzini } 5716c50d8ae3SPaolo Bonzini 5717c50d8ae3SPaolo Bonzini spte = get_written_sptes(sp, gpa, &npte); 5718c50d8ae3SPaolo Bonzini if (!spte) 5719c50d8ae3SPaolo Bonzini continue; 5720c50d8ae3SPaolo Bonzini 5721c50d8ae3SPaolo Bonzini while (npte--) { 5722c50d8ae3SPaolo Bonzini entry = *spte; 57232de4085cSBen Gardon mmu_page_zap_pte(vcpu->kvm, sp, spte, NULL); 5724c5e2184dSSean Christopherson if (gentry && sp->role.level != PG_LEVEL_4K) 5725c5e2184dSSean Christopherson ++vcpu->kvm->stat.mmu_pde_zapped; 57261441ca14SJunaid Shahid if (is_shadow_present_pte(entry)) 572706152b2dSLai Jiangshan flush = true; 5728c50d8ae3SPaolo Bonzini ++spte; 5729c50d8ae3SPaolo Bonzini } 5730c50d8ae3SPaolo Bonzini } 573106152b2dSLai Jiangshan kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, flush); 5732531810caSBen Gardon write_unlock(&vcpu->kvm->mmu_lock); 5733c50d8ae3SPaolo Bonzini } 5734c50d8ae3SPaolo Bonzini 57351075d41eSSean Christopherson int noinline kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, u64 error_code, 5736c50d8ae3SPaolo Bonzini void *insn, int insn_len) 5737c50d8ae3SPaolo Bonzini { 573892daa48bSSean Christopherson int r, emulation_type = EMULTYPE_PF; 5739347a0d0dSPaolo Bonzini bool direct = vcpu->arch.mmu->root_role.direct; 5740c50d8ae3SPaolo Bonzini 5741b9e5603cSPaolo Bonzini if (WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root.hpa))) 5742ddce6208SSean Christopherson return RET_PF_RETRY; 5743ddce6208SSean Christopherson 5744c50d8ae3SPaolo Bonzini r = RET_PF_INVALID; 5745c50d8ae3SPaolo Bonzini if (unlikely(error_code & PFERR_RSVD_MASK)) { 5746736c291cSSean Christopherson r = handle_mmio_page_fault(vcpu, cr2_or_gpa, direct); 5747c50d8ae3SPaolo Bonzini if (r == RET_PF_EMULATE) 5748c50d8ae3SPaolo Bonzini goto emulate; 5749c50d8ae3SPaolo Bonzini } 5750c50d8ae3SPaolo Bonzini 5751c50d8ae3SPaolo Bonzini if (r == RET_PF_INVALID) { 57527a02674dSSean Christopherson r = kvm_mmu_do_page_fault(vcpu, cr2_or_gpa, 5753258d985fSSean Christopherson lower_32_bits(error_code), false, 5754258d985fSSean Christopherson &emulation_type); 575519025e7bSSean Christopherson if (KVM_BUG_ON(r == RET_PF_INVALID, vcpu->kvm)) 57567b367bc9SSean Christopherson return -EIO; 5757c50d8ae3SPaolo Bonzini } 5758c50d8ae3SPaolo Bonzini 5759c50d8ae3SPaolo Bonzini if (r < 0) 5760c50d8ae3SPaolo Bonzini return r; 576183a2ba4cSSean Christopherson if (r != RET_PF_EMULATE) 576283a2ba4cSSean Christopherson return 1; 5763c50d8ae3SPaolo Bonzini 5764c50d8ae3SPaolo Bonzini /* 5765c50d8ae3SPaolo Bonzini * Before emulating the instruction, check if the error code 5766c50d8ae3SPaolo Bonzini * was due to a RO violation while translating the guest page. 5767c50d8ae3SPaolo Bonzini * This can occur when using nested virtualization with nested 5768c50d8ae3SPaolo Bonzini * paging in both guests. If true, we simply unprotect the page 5769c50d8ae3SPaolo Bonzini * and resume the guest. 5770c50d8ae3SPaolo Bonzini */ 5771347a0d0dSPaolo Bonzini if (vcpu->arch.mmu->root_role.direct && 5772c50d8ae3SPaolo Bonzini (error_code & PFERR_NESTED_GUEST_PAGE) == PFERR_NESTED_GUEST_PAGE) { 5773736c291cSSean Christopherson kvm_mmu_unprotect_page(vcpu->kvm, gpa_to_gfn(cr2_or_gpa)); 5774c50d8ae3SPaolo Bonzini return 1; 5775c50d8ae3SPaolo Bonzini } 5776c50d8ae3SPaolo Bonzini 5777c50d8ae3SPaolo Bonzini /* 5778c50d8ae3SPaolo Bonzini * vcpu->arch.mmu.page_fault returned RET_PF_EMULATE, but we can still 5779c50d8ae3SPaolo Bonzini * optimistically try to just unprotect the page and let the processor 5780c50d8ae3SPaolo Bonzini * re-execute the instruction that caused the page fault. Do not allow 5781c50d8ae3SPaolo Bonzini * retrying MMIO emulation, as it's not only pointless but could also 5782c50d8ae3SPaolo Bonzini * cause us to enter an infinite loop because the processor will keep 5783c50d8ae3SPaolo Bonzini * faulting on the non-existent MMIO address. Retrying an instruction 5784c50d8ae3SPaolo Bonzini * from a nested guest is also pointless and dangerous as we are only 5785c50d8ae3SPaolo Bonzini * explicitly shadowing L1's page tables, i.e. unprotecting something 5786c50d8ae3SPaolo Bonzini * for L1 isn't going to magically fix whatever issue cause L2 to fail. 5787c50d8ae3SPaolo Bonzini */ 5788736c291cSSean Christopherson if (!mmio_info_in_cache(vcpu, cr2_or_gpa, direct) && !is_guest_mode(vcpu)) 578992daa48bSSean Christopherson emulation_type |= EMULTYPE_ALLOW_RETRY_PF; 5790c50d8ae3SPaolo Bonzini emulate: 5791736c291cSSean Christopherson return x86_emulate_instruction(vcpu, cr2_or_gpa, emulation_type, insn, 5792c50d8ae3SPaolo Bonzini insn_len); 5793c50d8ae3SPaolo Bonzini } 5794c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_mmu_page_fault); 5795c50d8ae3SPaolo Bonzini 57969fd4a4e3SLai Jiangshan static void __kvm_mmu_invalidate_addr(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu, 57979fd4a4e3SLai Jiangshan u64 addr, hpa_t root_hpa) 57989fd4a4e3SLai Jiangshan { 57999fd4a4e3SLai Jiangshan struct kvm_shadow_walk_iterator iterator; 58009fd4a4e3SLai Jiangshan 58019fd4a4e3SLai Jiangshan vcpu_clear_mmio_info(vcpu, addr); 58029fd4a4e3SLai Jiangshan 58039fd4a4e3SLai Jiangshan if (!VALID_PAGE(root_hpa)) 58049fd4a4e3SLai Jiangshan return; 58059fd4a4e3SLai Jiangshan 58069fd4a4e3SLai Jiangshan write_lock(&vcpu->kvm->mmu_lock); 58079fd4a4e3SLai Jiangshan for_each_shadow_entry_using_root(vcpu, root_hpa, addr, iterator) { 58089fd4a4e3SLai Jiangshan struct kvm_mmu_page *sp = sptep_to_sp(iterator.sptep); 58099fd4a4e3SLai Jiangshan 58109fd4a4e3SLai Jiangshan if (sp->unsync) { 581119ace7d6SLai Jiangshan int ret = kvm_sync_spte(vcpu, sp, iterator.index); 58129fd4a4e3SLai Jiangshan 58139fd4a4e3SLai Jiangshan if (ret < 0) 58149fd4a4e3SLai Jiangshan mmu_page_zap_pte(vcpu->kvm, sp, iterator.sptep, NULL); 58159fd4a4e3SLai Jiangshan if (ret) 58169fd4a4e3SLai Jiangshan kvm_flush_remote_tlbs_sptep(vcpu->kvm, iterator.sptep); 58179fd4a4e3SLai Jiangshan } 58189fd4a4e3SLai Jiangshan 58199fd4a4e3SLai Jiangshan if (!sp->unsync_children) 58209fd4a4e3SLai Jiangshan break; 58219fd4a4e3SLai Jiangshan } 58229fd4a4e3SLai Jiangshan write_unlock(&vcpu->kvm->mmu_lock); 58239fd4a4e3SLai Jiangshan } 58249fd4a4e3SLai Jiangshan 5825753b43c9SLai Jiangshan void kvm_mmu_invalidate_addr(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu, 5826cd42853eSLai Jiangshan u64 addr, unsigned long roots) 5827c50d8ae3SPaolo Bonzini { 5828c50d8ae3SPaolo Bonzini int i; 5829c50d8ae3SPaolo Bonzini 5830cd42853eSLai Jiangshan WARN_ON_ONCE(roots & ~KVM_MMU_ROOTS_ALL); 5831cd42853eSLai Jiangshan 58325efac074SPaolo Bonzini /* It's actually a GPA for vcpu->arch.guest_mmu. */ 58335efac074SPaolo Bonzini if (mmu != &vcpu->arch.guest_mmu) { 58345efac074SPaolo Bonzini /* INVLPG on a non-canonical address is a NOP according to the SDM. */ 5835753b43c9SLai Jiangshan if (is_noncanonical_address(addr, vcpu)) 5836c50d8ae3SPaolo Bonzini return; 5837c50d8ae3SPaolo Bonzini 5838753b43c9SLai Jiangshan static_call(kvm_x86_flush_tlb_gva)(vcpu, addr); 58395efac074SPaolo Bonzini } 58405efac074SPaolo Bonzini 58419fd4a4e3SLai Jiangshan if (!mmu->sync_spte) 58425efac074SPaolo Bonzini return; 58435efac074SPaolo Bonzini 5844cd42853eSLai Jiangshan if (roots & KVM_MMU_ROOT_CURRENT) 58459fd4a4e3SLai Jiangshan __kvm_mmu_invalidate_addr(vcpu, mmu, addr, mmu->root.hpa); 5846c50d8ae3SPaolo Bonzini 5847cd42853eSLai Jiangshan for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) { 5848ed335278SLai Jiangshan if (roots & KVM_MMU_ROOT_PREVIOUS(i)) 58499fd4a4e3SLai Jiangshan __kvm_mmu_invalidate_addr(vcpu, mmu, addr, mmu->prev_roots[i].hpa); 58505efac074SPaolo Bonzini } 58515efac074SPaolo Bonzini } 58522c86c444SLai Jiangshan EXPORT_SYMBOL_GPL(kvm_mmu_invalidate_addr); 5853c50d8ae3SPaolo Bonzini 58545efac074SPaolo Bonzini void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva) 58555efac074SPaolo Bonzini { 5856cd42853eSLai Jiangshan /* 5857cd42853eSLai Jiangshan * INVLPG is required to invalidate any global mappings for the VA, 5858cd42853eSLai Jiangshan * irrespective of PCID. Blindly sync all roots as it would take 5859cd42853eSLai Jiangshan * roughly the same amount of work/time to determine whether any of the 5860cd42853eSLai Jiangshan * previous roots have a global mapping. 5861cd42853eSLai Jiangshan * 5862cd42853eSLai Jiangshan * Mappings not reachable via the current or previous cached roots will 5863cd42853eSLai Jiangshan * be synced when switching to that new cr3, so nothing needs to be 5864cd42853eSLai Jiangshan * done here for them. 5865cd42853eSLai Jiangshan */ 5866cd42853eSLai Jiangshan kvm_mmu_invalidate_addr(vcpu, vcpu->arch.walk_mmu, gva, KVM_MMU_ROOTS_ALL); 5867c50d8ae3SPaolo Bonzini ++vcpu->stat.invlpg; 5868c50d8ae3SPaolo Bonzini } 5869c50d8ae3SPaolo Bonzini EXPORT_SYMBOL_GPL(kvm_mmu_invlpg); 5870c50d8ae3SPaolo Bonzini 58715efac074SPaolo Bonzini 5872c50d8ae3SPaolo Bonzini void kvm_mmu_invpcid_gva(struct kvm_vcpu *vcpu, gva_t gva, unsigned long pcid) 5873c50d8ae3SPaolo Bonzini { 5874c50d8ae3SPaolo Bonzini struct kvm_mmu *mmu = vcpu->arch.mmu; 58759ebc3f51SLai Jiangshan unsigned long roots = 0; 5876c50d8ae3SPaolo Bonzini uint i; 5877c50d8ae3SPaolo Bonzini 58789ebc3f51SLai Jiangshan if (pcid == kvm_get_active_pcid(vcpu)) 58799ebc3f51SLai Jiangshan roots |= KVM_MMU_ROOT_CURRENT; 5880c50d8ae3SPaolo Bonzini 5881c50d8ae3SPaolo Bonzini for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) { 5882c50d8ae3SPaolo Bonzini if (VALID_PAGE(mmu->prev_roots[i].hpa) && 58839ebc3f51SLai Jiangshan pcid == kvm_get_pcid(vcpu, mmu->prev_roots[i].pgd)) 58849ebc3f51SLai Jiangshan roots |= KVM_MMU_ROOT_PREVIOUS(i); 5885c50d8ae3SPaolo Bonzini } 5886c50d8ae3SPaolo Bonzini 58879ebc3f51SLai Jiangshan if (roots) 58889ebc3f51SLai Jiangshan kvm_mmu_invalidate_addr(vcpu, mmu, gva, roots); 5889c50d8ae3SPaolo Bonzini ++vcpu->stat.invlpg; 5890c50d8ae3SPaolo Bonzini 5891c50d8ae3SPaolo Bonzini /* 5892c50d8ae3SPaolo Bonzini * Mappings not reachable via the current cr3 or the prev_roots will be 5893c50d8ae3SPaolo Bonzini * synced when switching to that cr3, so nothing needs to be done here 5894c50d8ae3SPaolo Bonzini * for them. 5895c50d8ae3SPaolo Bonzini */ 5896c50d8ae3SPaolo Bonzini } 5897c50d8ae3SPaolo Bonzini 5898746700d2SWei Huang void kvm_configure_mmu(bool enable_tdp, int tdp_forced_root_level, 5899746700d2SWei Huang int tdp_max_root_level, int tdp_huge_page_level) 5900c50d8ae3SPaolo Bonzini { 5901bde77235SSean Christopherson tdp_enabled = enable_tdp; 5902746700d2SWei Huang tdp_root_level = tdp_forced_root_level; 590383013059SSean Christopherson max_tdp_level = tdp_max_root_level; 5904703c335dSSean Christopherson 59051f98f2bdSDavid Matlack #ifdef CONFIG_X86_64 59061f98f2bdSDavid Matlack tdp_mmu_enabled = tdp_mmu_allowed && tdp_enabled; 59071f98f2bdSDavid Matlack #endif 5908703c335dSSean Christopherson /* 59091d92d2e8SSean Christopherson * max_huge_page_level reflects KVM's MMU capabilities irrespective 5910703c335dSSean Christopherson * of kernel support, e.g. KVM may be capable of using 1GB pages when 5911703c335dSSean Christopherson * the kernel is not. But, KVM never creates a page size greater than 5912703c335dSSean Christopherson * what is used by the kernel for any given HVA, i.e. the kernel's 5913703c335dSSean Christopherson * capabilities are ultimately consulted by kvm_mmu_hugepage_adjust(). 5914703c335dSSean Christopherson */ 5915703c335dSSean Christopherson if (tdp_enabled) 59161d92d2e8SSean Christopherson max_huge_page_level = tdp_huge_page_level; 5917703c335dSSean Christopherson else if (boot_cpu_has(X86_FEATURE_GBPAGES)) 59181d92d2e8SSean Christopherson max_huge_page_level = PG_LEVEL_1G; 5919703c335dSSean Christopherson else 59201d92d2e8SSean Christopherson max_huge_page_level = PG_LEVEL_2M; 5921c50d8ae3SPaolo Bonzini } 5922bde77235SSean Christopherson EXPORT_SYMBOL_GPL(kvm_configure_mmu); 5923c50d8ae3SPaolo Bonzini 5924c50d8ae3SPaolo Bonzini /* The return value indicates if tlb flush on all vcpus is needed. */ 5925727ae377SSean Christopherson typedef bool (*slot_rmaps_handler) (struct kvm *kvm, 5926269e9552SHamza Mahfooz struct kvm_rmap_head *rmap_head, 5927269e9552SHamza Mahfooz const struct kvm_memory_slot *slot); 5928c50d8ae3SPaolo Bonzini 5929727ae377SSean Christopherson static __always_inline bool __walk_slot_rmaps(struct kvm *kvm, 5930727ae377SSean Christopherson const struct kvm_memory_slot *slot, 5931727ae377SSean Christopherson slot_rmaps_handler fn, 5932727ae377SSean Christopherson int start_level, int end_level, 5933727ae377SSean Christopherson gfn_t start_gfn, gfn_t end_gfn, 5934727ae377SSean Christopherson bool flush_on_yield, bool flush) 5935c50d8ae3SPaolo Bonzini { 5936c50d8ae3SPaolo Bonzini struct slot_rmap_walk_iterator iterator; 5937c50d8ae3SPaolo Bonzini 5938eddd9e83SSean Christopherson lockdep_assert_held_write(&kvm->mmu_lock); 5939eddd9e83SSean Christopherson 5940727ae377SSean Christopherson for_each_slot_rmap_range(slot, start_level, end_level, start_gfn, 5941c50d8ae3SPaolo Bonzini end_gfn, &iterator) { 5942c50d8ae3SPaolo Bonzini if (iterator.rmap) 5943727ae377SSean Christopherson flush |= fn(kvm, iterator.rmap, slot); 5944c50d8ae3SPaolo Bonzini 5945531810caSBen Gardon if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) { 5946302695a5SSean Christopherson if (flush && flush_on_yield) { 59478c63e8c2SDavid Matlack kvm_flush_remote_tlbs_range(kvm, start_gfn, 5948c50d8ae3SPaolo Bonzini iterator.gfn - start_gfn + 1); 5949c50d8ae3SPaolo Bonzini flush = false; 5950c50d8ae3SPaolo Bonzini } 5951531810caSBen Gardon cond_resched_rwlock_write(&kvm->mmu_lock); 5952c50d8ae3SPaolo Bonzini } 5953c50d8ae3SPaolo Bonzini } 5954c50d8ae3SPaolo Bonzini 5955c50d8ae3SPaolo Bonzini return flush; 5956c50d8ae3SPaolo Bonzini } 5957c50d8ae3SPaolo Bonzini 5958727ae377SSean Christopherson static __always_inline bool walk_slot_rmaps(struct kvm *kvm, 5959727ae377SSean Christopherson const struct kvm_memory_slot *slot, 5960727ae377SSean Christopherson slot_rmaps_handler fn, 5961727ae377SSean Christopherson int start_level, int end_level, 5962302695a5SSean Christopherson bool flush_on_yield) 5963c50d8ae3SPaolo Bonzini { 5964727ae377SSean Christopherson return __walk_slot_rmaps(kvm, slot, fn, start_level, end_level, 5965727ae377SSean Christopherson slot->base_gfn, slot->base_gfn + slot->npages - 1, 59661a61b7dbSSean Christopherson flush_on_yield, false); 5967c50d8ae3SPaolo Bonzini } 5968c50d8ae3SPaolo Bonzini 5969727ae377SSean Christopherson static __always_inline bool walk_slot_rmaps_4k(struct kvm *kvm, 5970727ae377SSean Christopherson const struct kvm_memory_slot *slot, 5971727ae377SSean Christopherson slot_rmaps_handler fn, 5972727ae377SSean Christopherson bool flush_on_yield) 5973c50d8ae3SPaolo Bonzini { 5974727ae377SSean Christopherson return walk_slot_rmaps(kvm, slot, fn, PG_LEVEL_4K, PG_LEVEL_4K, flush_on_yield); 5975c50d8ae3SPaolo Bonzini } 5976c50d8ae3SPaolo Bonzini 5977c50d8ae3SPaolo Bonzini static void free_mmu_pages(struct kvm_mmu *mmu) 5978c50d8ae3SPaolo Bonzini { 59794a98623dSSean Christopherson if (!tdp_enabled && mmu->pae_root) 59804a98623dSSean Christopherson set_memory_encrypted((unsigned long)mmu->pae_root, 1); 5981c50d8ae3SPaolo Bonzini free_page((unsigned long)mmu->pae_root); 598203ca4589SSean Christopherson free_page((unsigned long)mmu->pml4_root); 5983cb0f722aSWei Huang free_page((unsigned long)mmu->pml5_root); 5984c50d8ae3SPaolo Bonzini } 5985c50d8ae3SPaolo Bonzini 598604d28e37SSean Christopherson static int __kvm_mmu_create(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu) 5987c50d8ae3SPaolo Bonzini { 5988c50d8ae3SPaolo Bonzini struct page *page; 5989c50d8ae3SPaolo Bonzini int i; 5990c50d8ae3SPaolo Bonzini 5991b9e5603cSPaolo Bonzini mmu->root.hpa = INVALID_PAGE; 5992b9e5603cSPaolo Bonzini mmu->root.pgd = 0; 599304d28e37SSean Christopherson for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) 599404d28e37SSean Christopherson mmu->prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID; 599504d28e37SSean Christopherson 599627f4fca2SLai Jiangshan /* vcpu->arch.guest_mmu isn't used when !tdp_enabled. */ 599727f4fca2SLai Jiangshan if (!tdp_enabled && mmu == &vcpu->arch.guest_mmu) 599827f4fca2SLai Jiangshan return 0; 599927f4fca2SLai Jiangshan 6000c50d8ae3SPaolo Bonzini /* 6001c50d8ae3SPaolo Bonzini * When using PAE paging, the four PDPTEs are treated as 'root' pages, 6002c50d8ae3SPaolo Bonzini * while the PDP table is a per-vCPU construct that's allocated at MMU 6003c50d8ae3SPaolo Bonzini * creation. When emulating 32-bit mode, cr3 is only 32 bits even on 6004c50d8ae3SPaolo Bonzini * x86_64. Therefore we need to allocate the PDP table in the first 600504d45551SSean Christopherson * 4GB of memory, which happens to fit the DMA32 zone. TDP paging 600604d45551SSean Christopherson * generally doesn't use PAE paging and can skip allocating the PDP 600704d45551SSean Christopherson * table. The main exception, handled here, is SVM's 32-bit NPT. The 600804d45551SSean Christopherson * other exception is for shadowing L1's 32-bit or PAE NPT on 64-bit 600984432316SLai Jiangshan * KVM; that horror is handled on-demand by mmu_alloc_special_roots(). 6010c50d8ae3SPaolo Bonzini */ 6011d468d94bSSean Christopherson if (tdp_enabled && kvm_mmu_get_tdp_level(vcpu) > PT32E_ROOT_LEVEL) 6012c50d8ae3SPaolo Bonzini return 0; 6013c50d8ae3SPaolo Bonzini 6014c50d8ae3SPaolo Bonzini page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_DMA32); 6015c50d8ae3SPaolo Bonzini if (!page) 6016c50d8ae3SPaolo Bonzini return -ENOMEM; 6017c50d8ae3SPaolo Bonzini 6018c50d8ae3SPaolo Bonzini mmu->pae_root = page_address(page); 60194a98623dSSean Christopherson 60204a98623dSSean Christopherson /* 60214a98623dSSean Christopherson * CR3 is only 32 bits when PAE paging is used, thus it's impossible to 60224a98623dSSean Christopherson * get the CPU to treat the PDPTEs as encrypted. Decrypt the page so 60234a98623dSSean Christopherson * that KVM's writes and the CPU's reads get along. Note, this is 60244a98623dSSean Christopherson * only necessary when using shadow paging, as 64-bit NPT can get at 60254a98623dSSean Christopherson * the C-bit even when shadowing 32-bit NPT, and SME isn't supported 60264a98623dSSean Christopherson * by 32-bit kernels (when KVM itself uses 32-bit NPT). 60274a98623dSSean Christopherson */ 60284a98623dSSean Christopherson if (!tdp_enabled) 60294a98623dSSean Christopherson set_memory_decrypted((unsigned long)mmu->pae_root, 1); 60304a98623dSSean Christopherson else 6031e54f1ff2SKai Huang WARN_ON_ONCE(shadow_me_value); 60324a98623dSSean Christopherson 6033c50d8ae3SPaolo Bonzini for (i = 0; i < 4; ++i) 6034c834e5e4SSean Christopherson mmu->pae_root[i] = INVALID_PAE_ROOT; 6035c50d8ae3SPaolo Bonzini 6036c50d8ae3SPaolo Bonzini return 0; 6037c50d8ae3SPaolo Bonzini } 6038c50d8ae3SPaolo Bonzini 6039c50d8ae3SPaolo Bonzini int kvm_mmu_create(struct kvm_vcpu *vcpu) 6040c50d8ae3SPaolo Bonzini { 6041c50d8ae3SPaolo Bonzini int ret; 6042c50d8ae3SPaolo Bonzini 60435962bfb7SSean Christopherson vcpu->arch.mmu_pte_list_desc_cache.kmem_cache = pte_list_desc_cache; 60445f6078f9SSean Christopherson vcpu->arch.mmu_pte_list_desc_cache.gfp_zero = __GFP_ZERO; 60455f6078f9SSean Christopherson 60465962bfb7SSean Christopherson vcpu->arch.mmu_page_header_cache.kmem_cache = mmu_page_header_cache; 60475f6078f9SSean Christopherson vcpu->arch.mmu_page_header_cache.gfp_zero = __GFP_ZERO; 60485962bfb7SSean Christopherson 604996880883SSean Christopherson vcpu->arch.mmu_shadow_page_cache.gfp_zero = __GFP_ZERO; 605096880883SSean Christopherson 6051c50d8ae3SPaolo Bonzini vcpu->arch.mmu = &vcpu->arch.root_mmu; 6052c50d8ae3SPaolo Bonzini vcpu->arch.walk_mmu = &vcpu->arch.root_mmu; 6053c50d8ae3SPaolo Bonzini 605404d28e37SSean Christopherson ret = __kvm_mmu_create(vcpu, &vcpu->arch.guest_mmu); 6055c50d8ae3SPaolo Bonzini if (ret) 6056c50d8ae3SPaolo Bonzini return ret; 6057c50d8ae3SPaolo Bonzini 605804d28e37SSean Christopherson ret = __kvm_mmu_create(vcpu, &vcpu->arch.root_mmu); 6059c50d8ae3SPaolo Bonzini if (ret) 6060c50d8ae3SPaolo Bonzini goto fail_allocate_root; 6061c50d8ae3SPaolo Bonzini 6062c50d8ae3SPaolo Bonzini return ret; 6063c50d8ae3SPaolo Bonzini fail_allocate_root: 6064c50d8ae3SPaolo Bonzini free_mmu_pages(&vcpu->arch.guest_mmu); 6065c50d8ae3SPaolo Bonzini return ret; 6066c50d8ae3SPaolo Bonzini } 6067c50d8ae3SPaolo Bonzini 6068c50d8ae3SPaolo Bonzini #define BATCH_ZAP_PAGES 10 6069c50d8ae3SPaolo Bonzini static void kvm_zap_obsolete_pages(struct kvm *kvm) 6070c50d8ae3SPaolo Bonzini { 6071c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, *node; 6072c50d8ae3SPaolo Bonzini int nr_zapped, batch = 0; 6073b28cb0cdSSean Christopherson bool unstable; 6074c50d8ae3SPaolo Bonzini 6075c50d8ae3SPaolo Bonzini restart: 6076c50d8ae3SPaolo Bonzini list_for_each_entry_safe_reverse(sp, node, 6077c50d8ae3SPaolo Bonzini &kvm->arch.active_mmu_pages, link) { 6078c50d8ae3SPaolo Bonzini /* 6079c50d8ae3SPaolo Bonzini * No obsolete valid page exists before a newly created page 6080c50d8ae3SPaolo Bonzini * since active_mmu_pages is a FIFO list. 6081c50d8ae3SPaolo Bonzini */ 6082c50d8ae3SPaolo Bonzini if (!is_obsolete_sp(kvm, sp)) 6083c50d8ae3SPaolo Bonzini break; 6084c50d8ae3SPaolo Bonzini 6085c50d8ae3SPaolo Bonzini /* 6086f95eec9bSSean Christopherson * Invalid pages should never land back on the list of active 6087f95eec9bSSean Christopherson * pages. Skip the bogus page, otherwise we'll get stuck in an 6088f95eec9bSSean Christopherson * infinite loop if the page gets put back on the list (again). 6089c50d8ae3SPaolo Bonzini */ 6090f95eec9bSSean Christopherson if (WARN_ON(sp->role.invalid)) 6091c50d8ae3SPaolo Bonzini continue; 6092c50d8ae3SPaolo Bonzini 6093c50d8ae3SPaolo Bonzini /* 6094c50d8ae3SPaolo Bonzini * No need to flush the TLB since we're only zapping shadow 6095c50d8ae3SPaolo Bonzini * pages with an obsolete generation number and all vCPUS have 6096c50d8ae3SPaolo Bonzini * loaded a new root, i.e. the shadow pages being zapped cannot 6097c50d8ae3SPaolo Bonzini * be in active use by the guest. 6098c50d8ae3SPaolo Bonzini */ 6099c50d8ae3SPaolo Bonzini if (batch >= BATCH_ZAP_PAGES && 6100531810caSBen Gardon cond_resched_rwlock_write(&kvm->mmu_lock)) { 6101c50d8ae3SPaolo Bonzini batch = 0; 6102c50d8ae3SPaolo Bonzini goto restart; 6103c50d8ae3SPaolo Bonzini } 6104c50d8ae3SPaolo Bonzini 6105b28cb0cdSSean Christopherson unstable = __kvm_mmu_prepare_zap_page(kvm, sp, 6106b28cb0cdSSean Christopherson &kvm->arch.zapped_obsolete_pages, &nr_zapped); 6107c50d8ae3SPaolo Bonzini batch += nr_zapped; 6108b28cb0cdSSean Christopherson 6109b28cb0cdSSean Christopherson if (unstable) 6110c50d8ae3SPaolo Bonzini goto restart; 6111c50d8ae3SPaolo Bonzini } 6112c50d8ae3SPaolo Bonzini 6113c50d8ae3SPaolo Bonzini /* 61147ae5840eSSean Christopherson * Kick all vCPUs (via remote TLB flush) before freeing the page tables 61157ae5840eSSean Christopherson * to ensure KVM is not in the middle of a lockless shadow page table 61167ae5840eSSean Christopherson * walk, which may reference the pages. The remote TLB flush itself is 61177ae5840eSSean Christopherson * not required and is simply a convenient way to kick vCPUs as needed. 61187ae5840eSSean Christopherson * KVM performs a local TLB flush when allocating a new root (see 61197ae5840eSSean Christopherson * kvm_mmu_load()), and the reload in the caller ensure no vCPUs are 61207ae5840eSSean Christopherson * running with an obsolete MMU. 6121c50d8ae3SPaolo Bonzini */ 6122c50d8ae3SPaolo Bonzini kvm_mmu_commit_zap_page(kvm, &kvm->arch.zapped_obsolete_pages); 6123c50d8ae3SPaolo Bonzini } 6124c50d8ae3SPaolo Bonzini 6125c50d8ae3SPaolo Bonzini /* 6126c50d8ae3SPaolo Bonzini * Fast invalidate all shadow pages and use lock-break technique 6127c50d8ae3SPaolo Bonzini * to zap obsolete pages. 6128c50d8ae3SPaolo Bonzini * 6129c50d8ae3SPaolo Bonzini * It's required when memslot is being deleted or VM is being 6130c50d8ae3SPaolo Bonzini * destroyed, in these cases, we should ensure that KVM MMU does 6131c50d8ae3SPaolo Bonzini * not use any resource of the being-deleted slot or all slots 6132c50d8ae3SPaolo Bonzini * after calling the function. 6133c50d8ae3SPaolo Bonzini */ 6134c50d8ae3SPaolo Bonzini static void kvm_mmu_zap_all_fast(struct kvm *kvm) 6135c50d8ae3SPaolo Bonzini { 6136c50d8ae3SPaolo Bonzini lockdep_assert_held(&kvm->slots_lock); 6137c50d8ae3SPaolo Bonzini 6138531810caSBen Gardon write_lock(&kvm->mmu_lock); 6139c50d8ae3SPaolo Bonzini trace_kvm_mmu_zap_all_fast(kvm); 6140c50d8ae3SPaolo Bonzini 6141c50d8ae3SPaolo Bonzini /* 6142c50d8ae3SPaolo Bonzini * Toggle mmu_valid_gen between '0' and '1'. Because slots_lock is 6143c50d8ae3SPaolo Bonzini * held for the entire duration of zapping obsolete pages, it's 6144c50d8ae3SPaolo Bonzini * impossible for there to be multiple invalid generations associated 6145c50d8ae3SPaolo Bonzini * with *valid* shadow pages at any given time, i.e. there is exactly 6146c50d8ae3SPaolo Bonzini * one valid generation and (at most) one invalid generation. 6147c50d8ae3SPaolo Bonzini */ 6148c50d8ae3SPaolo Bonzini kvm->arch.mmu_valid_gen = kvm->arch.mmu_valid_gen ? 0 : 1; 6149c50d8ae3SPaolo Bonzini 61502f6f66ccSSean Christopherson /* 61512f6f66ccSSean Christopherson * In order to ensure all vCPUs drop their soon-to-be invalid roots, 61522f6f66ccSSean Christopherson * invalidating TDP MMU roots must be done while holding mmu_lock for 61532f6f66ccSSean Christopherson * write and in the same critical section as making the reload request, 61542f6f66ccSSean Christopherson * e.g. before kvm_zap_obsolete_pages() could drop mmu_lock and yield. 6155b7cccd39SBen Gardon */ 61561f98f2bdSDavid Matlack if (tdp_mmu_enabled) 6157b7cccd39SBen Gardon kvm_tdp_mmu_invalidate_all_roots(kvm); 6158b7cccd39SBen Gardon 6159c50d8ae3SPaolo Bonzini /* 6160c50d8ae3SPaolo Bonzini * Notify all vcpus to reload its shadow page table and flush TLB. 6161c50d8ae3SPaolo Bonzini * Then all vcpus will switch to new shadow page table with the new 6162c50d8ae3SPaolo Bonzini * mmu_valid_gen. 6163c50d8ae3SPaolo Bonzini * 6164c50d8ae3SPaolo Bonzini * Note: we need to do this under the protection of mmu_lock, 6165c50d8ae3SPaolo Bonzini * otherwise, vcpu would purge shadow page but miss tlb flush. 6166c50d8ae3SPaolo Bonzini */ 6167527d5cd7SSean Christopherson kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_FREE_OBSOLETE_ROOTS); 6168c50d8ae3SPaolo Bonzini 6169c50d8ae3SPaolo Bonzini kvm_zap_obsolete_pages(kvm); 6170faaf05b0SBen Gardon 6171531810caSBen Gardon write_unlock(&kvm->mmu_lock); 61724c6654bdSBen Gardon 6173f28e9c7fSSean Christopherson /* 6174f28e9c7fSSean Christopherson * Zap the invalidated TDP MMU roots, all SPTEs must be dropped before 6175f28e9c7fSSean Christopherson * returning to the caller, e.g. if the zap is in response to a memslot 6176f28e9c7fSSean Christopherson * deletion, mmu_notifier callbacks will be unable to reach the SPTEs 6177f28e9c7fSSean Christopherson * associated with the deleted memslot once the update completes, and 6178f28e9c7fSSean Christopherson * Deferring the zap until the final reference to the root is put would 6179f28e9c7fSSean Christopherson * lead to use-after-free. 6180f28e9c7fSSean Christopherson */ 61811f98f2bdSDavid Matlack if (tdp_mmu_enabled) 61824c6654bdSBen Gardon kvm_tdp_mmu_zap_invalidated_roots(kvm); 6183c50d8ae3SPaolo Bonzini } 6184c50d8ae3SPaolo Bonzini 6185c50d8ae3SPaolo Bonzini static bool kvm_has_zapped_obsolete_pages(struct kvm *kvm) 6186c50d8ae3SPaolo Bonzini { 6187c50d8ae3SPaolo Bonzini return unlikely(!list_empty_careful(&kvm->arch.zapped_obsolete_pages)); 6188c50d8ae3SPaolo Bonzini } 6189c50d8ae3SPaolo Bonzini 6190c50d8ae3SPaolo Bonzini static void kvm_mmu_invalidate_zap_pages_in_memslot(struct kvm *kvm, 6191c50d8ae3SPaolo Bonzini struct kvm_memory_slot *slot, 6192c50d8ae3SPaolo Bonzini struct kvm_page_track_notifier_node *node) 6193c50d8ae3SPaolo Bonzini { 6194c50d8ae3SPaolo Bonzini kvm_mmu_zap_all_fast(kvm); 6195c50d8ae3SPaolo Bonzini } 6196c50d8ae3SPaolo Bonzini 6197a1a39128SPaolo Bonzini int kvm_mmu_init_vm(struct kvm *kvm) 6198c50d8ae3SPaolo Bonzini { 6199c50d8ae3SPaolo Bonzini struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker; 6200a1a39128SPaolo Bonzini int r; 6201c50d8ae3SPaolo Bonzini 6202a1a39128SPaolo Bonzini INIT_LIST_HEAD(&kvm->arch.active_mmu_pages); 6203a1a39128SPaolo Bonzini INIT_LIST_HEAD(&kvm->arch.zapped_obsolete_pages); 620455c510e2SSean Christopherson INIT_LIST_HEAD(&kvm->arch.possible_nx_huge_pages); 6205ce25681dSSean Christopherson spin_lock_init(&kvm->arch.mmu_unsync_pages_lock); 6206ce25681dSSean Christopherson 620709732d2bSDavid Matlack if (tdp_mmu_enabled) { 6208a1a39128SPaolo Bonzini r = kvm_mmu_init_tdp_mmu(kvm); 6209a1a39128SPaolo Bonzini if (r < 0) 6210a1a39128SPaolo Bonzini return r; 621109732d2bSDavid Matlack } 6212fe5db27dSBen Gardon 6213c50d8ae3SPaolo Bonzini node->track_write = kvm_mmu_pte_write; 6214c50d8ae3SPaolo Bonzini node->track_flush_slot = kvm_mmu_invalidate_zap_pages_in_memslot; 6215c50d8ae3SPaolo Bonzini kvm_page_track_register_notifier(kvm, node); 6216ada51a9dSDavid Matlack 6217ada51a9dSDavid Matlack kvm->arch.split_page_header_cache.kmem_cache = mmu_page_header_cache; 6218ada51a9dSDavid Matlack kvm->arch.split_page_header_cache.gfp_zero = __GFP_ZERO; 6219ada51a9dSDavid Matlack 6220ada51a9dSDavid Matlack kvm->arch.split_shadow_page_cache.gfp_zero = __GFP_ZERO; 6221ada51a9dSDavid Matlack 6222ada51a9dSDavid Matlack kvm->arch.split_desc_cache.kmem_cache = pte_list_desc_cache; 6223ada51a9dSDavid Matlack kvm->arch.split_desc_cache.gfp_zero = __GFP_ZERO; 6224ada51a9dSDavid Matlack 6225a1a39128SPaolo Bonzini return 0; 6226c50d8ae3SPaolo Bonzini } 6227c50d8ae3SPaolo Bonzini 6228ada51a9dSDavid Matlack static void mmu_free_vm_memory_caches(struct kvm *kvm) 6229ada51a9dSDavid Matlack { 6230ada51a9dSDavid Matlack kvm_mmu_free_memory_cache(&kvm->arch.split_desc_cache); 6231ada51a9dSDavid Matlack kvm_mmu_free_memory_cache(&kvm->arch.split_page_header_cache); 6232ada51a9dSDavid Matlack kvm_mmu_free_memory_cache(&kvm->arch.split_shadow_page_cache); 6233ada51a9dSDavid Matlack } 6234ada51a9dSDavid Matlack 6235c50d8ae3SPaolo Bonzini void kvm_mmu_uninit_vm(struct kvm *kvm) 6236c50d8ae3SPaolo Bonzini { 6237c50d8ae3SPaolo Bonzini struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker; 6238c50d8ae3SPaolo Bonzini 6239c50d8ae3SPaolo Bonzini kvm_page_track_unregister_notifier(kvm, node); 6240fe5db27dSBen Gardon 624109732d2bSDavid Matlack if (tdp_mmu_enabled) 6242fe5db27dSBen Gardon kvm_mmu_uninit_tdp_mmu(kvm); 6243ada51a9dSDavid Matlack 6244ada51a9dSDavid Matlack mmu_free_vm_memory_caches(kvm); 6245c50d8ae3SPaolo Bonzini } 6246c50d8ae3SPaolo Bonzini 62472833eda0SSean Christopherson static bool kvm_rmap_zap_gfn_range(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end) 624821fa3246SSean Christopherson { 624921fa3246SSean Christopherson const struct kvm_memory_slot *memslot; 625021fa3246SSean Christopherson struct kvm_memslots *slots; 6251f4209439SMaciej S. Szmigiero struct kvm_memslot_iter iter; 625221fa3246SSean Christopherson bool flush = false; 625321fa3246SSean Christopherson gfn_t start, end; 6254f4209439SMaciej S. Szmigiero int i; 625521fa3246SSean Christopherson 625621fa3246SSean Christopherson if (!kvm_memslots_have_rmaps(kvm)) 625721fa3246SSean Christopherson return flush; 625821fa3246SSean Christopherson 625921fa3246SSean Christopherson for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) { 626021fa3246SSean Christopherson slots = __kvm_memslots(kvm, i); 6261f4209439SMaciej S. Szmigiero 6262f4209439SMaciej S. Szmigiero kvm_for_each_memslot_in_gfn_range(&iter, slots, gfn_start, gfn_end) { 6263f4209439SMaciej S. Szmigiero memslot = iter.slot; 626421fa3246SSean Christopherson start = max(gfn_start, memslot->base_gfn); 626521fa3246SSean Christopherson end = min(gfn_end, memslot->base_gfn + memslot->npages); 6266f4209439SMaciej S. Szmigiero if (WARN_ON_ONCE(start >= end)) 626721fa3246SSean Christopherson continue; 626821fa3246SSean Christopherson 6269727ae377SSean Christopherson flush = __walk_slot_rmaps(kvm, memslot, __kvm_zap_rmap, 627021fa3246SSean Christopherson PG_LEVEL_4K, KVM_MAX_HUGEPAGE_LEVEL, 627121fa3246SSean Christopherson start, end - 1, true, flush); 627221fa3246SSean Christopherson } 627321fa3246SSean Christopherson } 627421fa3246SSean Christopherson 627521fa3246SSean Christopherson return flush; 627621fa3246SSean Christopherson } 627721fa3246SSean Christopherson 627888f58535SMaxim Levitsky /* 627988f58535SMaxim Levitsky * Invalidate (zap) SPTEs that cover GFNs from gfn_start and up to gfn_end 628088f58535SMaxim Levitsky * (not including it) 628188f58535SMaxim Levitsky */ 6282c50d8ae3SPaolo Bonzini void kvm_zap_gfn_range(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end) 6283c50d8ae3SPaolo Bonzini { 628421fa3246SSean Christopherson bool flush; 6285c50d8ae3SPaolo Bonzini int i; 6286c50d8ae3SPaolo Bonzini 6287f4209439SMaciej S. Szmigiero if (WARN_ON_ONCE(gfn_end <= gfn_start)) 6288f4209439SMaciej S. Szmigiero return; 6289f4209439SMaciej S. Szmigiero 6290531810caSBen Gardon write_lock(&kvm->mmu_lock); 62915a324c24SSean Christopherson 62926d3085e4SSean Christopherson kvm_mmu_invalidate_begin(kvm, 0, -1ul); 6293edb298c6SMaxim Levitsky 62942833eda0SSean Christopherson flush = kvm_rmap_zap_gfn_range(kvm, gfn_start, gfn_end); 62956103bc07SBen Gardon 62961f98f2bdSDavid Matlack if (tdp_mmu_enabled) { 62976103bc07SBen Gardon for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) 6298f47e5bbbSSean Christopherson flush = kvm_tdp_mmu_zap_leafs(kvm, i, gfn_start, 6299f47e5bbbSSean Christopherson gfn_end, true, flush); 63006103bc07SBen Gardon } 63015a324c24SSean Christopherson 63025a324c24SSean Christopherson if (flush) 63038c63e8c2SDavid Matlack kvm_flush_remote_tlbs_range(kvm, gfn_start, gfn_end - gfn_start); 63045a324c24SSean Christopherson 63056d3085e4SSean Christopherson kvm_mmu_invalidate_end(kvm, 0, -1ul); 6306edb298c6SMaxim Levitsky 63075a324c24SSean Christopherson write_unlock(&kvm->mmu_lock); 6308c50d8ae3SPaolo Bonzini } 6309c50d8ae3SPaolo Bonzini 6310c50d8ae3SPaolo Bonzini static bool slot_rmap_write_protect(struct kvm *kvm, 63110a234f5dSSean Christopherson struct kvm_rmap_head *rmap_head, 6312269e9552SHamza Mahfooz const struct kvm_memory_slot *slot) 6313c50d8ae3SPaolo Bonzini { 63141346bbb6SDavid Matlack return rmap_write_protect(rmap_head, false); 6315c50d8ae3SPaolo Bonzini } 6316c50d8ae3SPaolo Bonzini 6317c50d8ae3SPaolo Bonzini void kvm_mmu_slot_remove_write_access(struct kvm *kvm, 6318269e9552SHamza Mahfooz const struct kvm_memory_slot *memslot, 63193c9bd400SJay Zhou int start_level) 6320c50d8ae3SPaolo Bonzini { 6321e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) { 6322531810caSBen Gardon write_lock(&kvm->mmu_lock); 6323727ae377SSean Christopherson walk_slot_rmaps(kvm, memslot, slot_rmap_write_protect, 6324b64d740eSJunaid Shahid start_level, KVM_MAX_HUGEPAGE_LEVEL, false); 6325531810caSBen Gardon write_unlock(&kvm->mmu_lock); 6326e2209710SBen Gardon } 6327c50d8ae3SPaolo Bonzini 63281f98f2bdSDavid Matlack if (tdp_mmu_enabled) { 632924ae4cfaSBen Gardon read_lock(&kvm->mmu_lock); 6330b64d740eSJunaid Shahid kvm_tdp_mmu_wrprot_slot(kvm, memslot, start_level); 633124ae4cfaSBen Gardon read_unlock(&kvm->mmu_lock); 633224ae4cfaSBen Gardon } 6333c50d8ae3SPaolo Bonzini } 6334c50d8ae3SPaolo Bonzini 6335ada51a9dSDavid Matlack static inline bool need_topup(struct kvm_mmu_memory_cache *cache, int min) 6336ada51a9dSDavid Matlack { 6337ada51a9dSDavid Matlack return kvm_mmu_memory_cache_nr_free_objects(cache) < min; 6338ada51a9dSDavid Matlack } 6339ada51a9dSDavid Matlack 6340ada51a9dSDavid Matlack static bool need_topup_split_caches_or_resched(struct kvm *kvm) 6341ada51a9dSDavid Matlack { 6342ada51a9dSDavid Matlack if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) 6343ada51a9dSDavid Matlack return true; 6344ada51a9dSDavid Matlack 6345ada51a9dSDavid Matlack /* 6346ada51a9dSDavid Matlack * In the worst case, SPLIT_DESC_CACHE_MIN_NR_OBJECTS descriptors are needed 6347ada51a9dSDavid Matlack * to split a single huge page. Calculating how many are actually needed 6348ada51a9dSDavid Matlack * is possible but not worth the complexity. 6349ada51a9dSDavid Matlack */ 6350ada51a9dSDavid Matlack return need_topup(&kvm->arch.split_desc_cache, SPLIT_DESC_CACHE_MIN_NR_OBJECTS) || 6351ada51a9dSDavid Matlack need_topup(&kvm->arch.split_page_header_cache, 1) || 6352ada51a9dSDavid Matlack need_topup(&kvm->arch.split_shadow_page_cache, 1); 6353ada51a9dSDavid Matlack } 6354ada51a9dSDavid Matlack 6355ada51a9dSDavid Matlack static int topup_split_caches(struct kvm *kvm) 6356ada51a9dSDavid Matlack { 6357b9b71f43SSean Christopherson /* 6358b9b71f43SSean Christopherson * Allocating rmap list entries when splitting huge pages for nested 6359dfd4eb44SSean Christopherson * MMUs is uncommon as KVM needs to use a list if and only if there is 6360b9b71f43SSean Christopherson * more than one rmap entry for a gfn, i.e. requires an L1 gfn to be 6361dfd4eb44SSean Christopherson * aliased by multiple L2 gfns and/or from multiple nested roots with 6362dfd4eb44SSean Christopherson * different roles. Aliasing gfns when using TDP is atypical for VMMs; 6363dfd4eb44SSean Christopherson * a few gfns are often aliased during boot, e.g. when remapping BIOS, 6364dfd4eb44SSean Christopherson * but aliasing rarely occurs post-boot or for many gfns. If there is 6365dfd4eb44SSean Christopherson * only one rmap entry, rmap->val points directly at that one entry and 6366dfd4eb44SSean Christopherson * doesn't need to allocate a list. Buffer the cache by the default 6367dfd4eb44SSean Christopherson * capacity so that KVM doesn't have to drop mmu_lock to topup if KVM 6368b9b71f43SSean Christopherson * encounters an aliased gfn or two. 6369b9b71f43SSean Christopherson */ 6370b9b71f43SSean Christopherson const int capacity = SPLIT_DESC_CACHE_MIN_NR_OBJECTS + 6371b9b71f43SSean Christopherson KVM_ARCH_NR_OBJS_PER_MEMORY_CACHE; 6372ada51a9dSDavid Matlack int r; 6373ada51a9dSDavid Matlack 6374ada51a9dSDavid Matlack lockdep_assert_held(&kvm->slots_lock); 6375ada51a9dSDavid Matlack 6376b9b71f43SSean Christopherson r = __kvm_mmu_topup_memory_cache(&kvm->arch.split_desc_cache, capacity, 6377ada51a9dSDavid Matlack SPLIT_DESC_CACHE_MIN_NR_OBJECTS); 6378ada51a9dSDavid Matlack if (r) 6379ada51a9dSDavid Matlack return r; 6380ada51a9dSDavid Matlack 6381ada51a9dSDavid Matlack r = kvm_mmu_topup_memory_cache(&kvm->arch.split_page_header_cache, 1); 6382ada51a9dSDavid Matlack if (r) 6383ada51a9dSDavid Matlack return r; 6384ada51a9dSDavid Matlack 6385ada51a9dSDavid Matlack return kvm_mmu_topup_memory_cache(&kvm->arch.split_shadow_page_cache, 1); 6386ada51a9dSDavid Matlack } 6387ada51a9dSDavid Matlack 6388ada51a9dSDavid Matlack static struct kvm_mmu_page *shadow_mmu_get_sp_for_split(struct kvm *kvm, u64 *huge_sptep) 6389ada51a9dSDavid Matlack { 6390ada51a9dSDavid Matlack struct kvm_mmu_page *huge_sp = sptep_to_sp(huge_sptep); 6391ada51a9dSDavid Matlack struct shadow_page_caches caches = {}; 6392ada51a9dSDavid Matlack union kvm_mmu_page_role role; 6393ada51a9dSDavid Matlack unsigned int access; 6394ada51a9dSDavid Matlack gfn_t gfn; 6395ada51a9dSDavid Matlack 639679e48cecSSean Christopherson gfn = kvm_mmu_page_get_gfn(huge_sp, spte_index(huge_sptep)); 639779e48cecSSean Christopherson access = kvm_mmu_page_get_access(huge_sp, spte_index(huge_sptep)); 6398ada51a9dSDavid Matlack 6399ada51a9dSDavid Matlack /* 6400ada51a9dSDavid Matlack * Note, huge page splitting always uses direct shadow pages, regardless 6401ada51a9dSDavid Matlack * of whether the huge page itself is mapped by a direct or indirect 6402ada51a9dSDavid Matlack * shadow page, since the huge page region itself is being directly 6403ada51a9dSDavid Matlack * mapped with smaller pages. 6404ada51a9dSDavid Matlack */ 6405ada51a9dSDavid Matlack role = kvm_mmu_child_role(huge_sptep, /*direct=*/true, access); 6406ada51a9dSDavid Matlack 6407ada51a9dSDavid Matlack /* Direct SPs do not require a shadowed_info_cache. */ 6408ada51a9dSDavid Matlack caches.page_header_cache = &kvm->arch.split_page_header_cache; 6409ada51a9dSDavid Matlack caches.shadow_page_cache = &kvm->arch.split_shadow_page_cache; 6410ada51a9dSDavid Matlack 6411ada51a9dSDavid Matlack /* Safe to pass NULL for vCPU since requesting a direct SP. */ 6412ada51a9dSDavid Matlack return __kvm_mmu_get_shadow_page(kvm, NULL, &caches, gfn, role); 6413ada51a9dSDavid Matlack } 6414ada51a9dSDavid Matlack 6415ada51a9dSDavid Matlack static void shadow_mmu_split_huge_page(struct kvm *kvm, 6416ada51a9dSDavid Matlack const struct kvm_memory_slot *slot, 6417ada51a9dSDavid Matlack u64 *huge_sptep) 6418ada51a9dSDavid Matlack 6419ada51a9dSDavid Matlack { 6420ada51a9dSDavid Matlack struct kvm_mmu_memory_cache *cache = &kvm->arch.split_desc_cache; 6421ada51a9dSDavid Matlack u64 huge_spte = READ_ONCE(*huge_sptep); 6422ada51a9dSDavid Matlack struct kvm_mmu_page *sp; 642303787394SPaolo Bonzini bool flush = false; 6424ada51a9dSDavid Matlack u64 *sptep, spte; 6425ada51a9dSDavid Matlack gfn_t gfn; 6426ada51a9dSDavid Matlack int index; 6427ada51a9dSDavid Matlack 6428ada51a9dSDavid Matlack sp = shadow_mmu_get_sp_for_split(kvm, huge_sptep); 6429ada51a9dSDavid Matlack 6430ada51a9dSDavid Matlack for (index = 0; index < SPTE_ENT_PER_PAGE; index++) { 6431ada51a9dSDavid Matlack sptep = &sp->spt[index]; 6432ada51a9dSDavid Matlack gfn = kvm_mmu_page_get_gfn(sp, index); 6433ada51a9dSDavid Matlack 6434ada51a9dSDavid Matlack /* 6435ada51a9dSDavid Matlack * The SP may already have populated SPTEs, e.g. if this huge 6436ada51a9dSDavid Matlack * page is aliased by multiple sptes with the same access 6437ada51a9dSDavid Matlack * permissions. These entries are guaranteed to map the same 6438ada51a9dSDavid Matlack * gfn-to-pfn translation since the SP is direct, so no need to 6439ada51a9dSDavid Matlack * modify them. 6440ada51a9dSDavid Matlack * 644103787394SPaolo Bonzini * However, if a given SPTE points to a lower level page table, 644203787394SPaolo Bonzini * that lower level page table may only be partially populated. 644303787394SPaolo Bonzini * Installing such SPTEs would effectively unmap a potion of the 644403787394SPaolo Bonzini * huge page. Unmapping guest memory always requires a TLB flush 644503787394SPaolo Bonzini * since a subsequent operation on the unmapped regions would 644603787394SPaolo Bonzini * fail to detect the need to flush. 6447ada51a9dSDavid Matlack */ 644803787394SPaolo Bonzini if (is_shadow_present_pte(*sptep)) { 644903787394SPaolo Bonzini flush |= !is_last_spte(*sptep, sp->role.level); 6450ada51a9dSDavid Matlack continue; 645103787394SPaolo Bonzini } 6452ada51a9dSDavid Matlack 6453ada51a9dSDavid Matlack spte = make_huge_page_split_spte(kvm, huge_spte, sp->role, index); 6454ada51a9dSDavid Matlack mmu_spte_set(sptep, spte); 6455ada51a9dSDavid Matlack __rmap_add(kvm, cache, slot, sptep, gfn, sp->role.access); 6456ada51a9dSDavid Matlack } 6457ada51a9dSDavid Matlack 645803787394SPaolo Bonzini __link_shadow_page(kvm, cache, huge_sptep, sp, flush); 6459ada51a9dSDavid Matlack } 6460ada51a9dSDavid Matlack 6461ada51a9dSDavid Matlack static int shadow_mmu_try_split_huge_page(struct kvm *kvm, 6462ada51a9dSDavid Matlack const struct kvm_memory_slot *slot, 6463ada51a9dSDavid Matlack u64 *huge_sptep) 6464ada51a9dSDavid Matlack { 6465ada51a9dSDavid Matlack struct kvm_mmu_page *huge_sp = sptep_to_sp(huge_sptep); 6466ada51a9dSDavid Matlack int level, r = 0; 6467ada51a9dSDavid Matlack gfn_t gfn; 6468ada51a9dSDavid Matlack u64 spte; 6469ada51a9dSDavid Matlack 6470ada51a9dSDavid Matlack /* Grab information for the tracepoint before dropping the MMU lock. */ 647179e48cecSSean Christopherson gfn = kvm_mmu_page_get_gfn(huge_sp, spte_index(huge_sptep)); 6472ada51a9dSDavid Matlack level = huge_sp->role.level; 6473ada51a9dSDavid Matlack spte = *huge_sptep; 6474ada51a9dSDavid Matlack 6475ada51a9dSDavid Matlack if (kvm_mmu_available_pages(kvm) <= KVM_MIN_FREE_MMU_PAGES) { 6476ada51a9dSDavid Matlack r = -ENOSPC; 6477ada51a9dSDavid Matlack goto out; 6478ada51a9dSDavid Matlack } 6479ada51a9dSDavid Matlack 6480ada51a9dSDavid Matlack if (need_topup_split_caches_or_resched(kvm)) { 6481ada51a9dSDavid Matlack write_unlock(&kvm->mmu_lock); 6482ada51a9dSDavid Matlack cond_resched(); 6483ada51a9dSDavid Matlack /* 6484ada51a9dSDavid Matlack * If the topup succeeds, return -EAGAIN to indicate that the 6485ada51a9dSDavid Matlack * rmap iterator should be restarted because the MMU lock was 6486ada51a9dSDavid Matlack * dropped. 6487ada51a9dSDavid Matlack */ 6488ada51a9dSDavid Matlack r = topup_split_caches(kvm) ?: -EAGAIN; 6489ada51a9dSDavid Matlack write_lock(&kvm->mmu_lock); 6490ada51a9dSDavid Matlack goto out; 6491ada51a9dSDavid Matlack } 6492ada51a9dSDavid Matlack 6493ada51a9dSDavid Matlack shadow_mmu_split_huge_page(kvm, slot, huge_sptep); 6494ada51a9dSDavid Matlack 6495ada51a9dSDavid Matlack out: 6496ada51a9dSDavid Matlack trace_kvm_mmu_split_huge_page(gfn, spte, level, r); 6497ada51a9dSDavid Matlack return r; 6498ada51a9dSDavid Matlack } 6499ada51a9dSDavid Matlack 6500ada51a9dSDavid Matlack static bool shadow_mmu_try_split_huge_pages(struct kvm *kvm, 6501ada51a9dSDavid Matlack struct kvm_rmap_head *rmap_head, 6502ada51a9dSDavid Matlack const struct kvm_memory_slot *slot) 6503ada51a9dSDavid Matlack { 6504ada51a9dSDavid Matlack struct rmap_iterator iter; 6505ada51a9dSDavid Matlack struct kvm_mmu_page *sp; 6506ada51a9dSDavid Matlack u64 *huge_sptep; 6507ada51a9dSDavid Matlack int r; 6508ada51a9dSDavid Matlack 6509ada51a9dSDavid Matlack restart: 6510ada51a9dSDavid Matlack for_each_rmap_spte(rmap_head, &iter, huge_sptep) { 6511ada51a9dSDavid Matlack sp = sptep_to_sp(huge_sptep); 6512ada51a9dSDavid Matlack 6513ada51a9dSDavid Matlack /* TDP MMU is enabled, so rmap only contains nested MMU SPs. */ 6514ada51a9dSDavid Matlack if (WARN_ON_ONCE(!sp->role.guest_mode)) 6515ada51a9dSDavid Matlack continue; 6516ada51a9dSDavid Matlack 6517ada51a9dSDavid Matlack /* The rmaps should never contain non-leaf SPTEs. */ 6518ada51a9dSDavid Matlack if (WARN_ON_ONCE(!is_large_pte(*huge_sptep))) 6519ada51a9dSDavid Matlack continue; 6520ada51a9dSDavid Matlack 6521ada51a9dSDavid Matlack /* SPs with level >PG_LEVEL_4K should never by unsync. */ 6522ada51a9dSDavid Matlack if (WARN_ON_ONCE(sp->unsync)) 6523ada51a9dSDavid Matlack continue; 6524ada51a9dSDavid Matlack 6525ada51a9dSDavid Matlack /* Don't bother splitting huge pages on invalid SPs. */ 6526ada51a9dSDavid Matlack if (sp->role.invalid) 6527ada51a9dSDavid Matlack continue; 6528ada51a9dSDavid Matlack 6529ada51a9dSDavid Matlack r = shadow_mmu_try_split_huge_page(kvm, slot, huge_sptep); 6530ada51a9dSDavid Matlack 6531ada51a9dSDavid Matlack /* 6532ada51a9dSDavid Matlack * The split succeeded or needs to be retried because the MMU 6533ada51a9dSDavid Matlack * lock was dropped. Either way, restart the iterator to get it 6534ada51a9dSDavid Matlack * back into a consistent state. 6535ada51a9dSDavid Matlack */ 6536ada51a9dSDavid Matlack if (!r || r == -EAGAIN) 6537ada51a9dSDavid Matlack goto restart; 6538ada51a9dSDavid Matlack 6539ada51a9dSDavid Matlack /* The split failed and shouldn't be retried (e.g. -ENOMEM). */ 6540ada51a9dSDavid Matlack break; 6541ada51a9dSDavid Matlack } 6542ada51a9dSDavid Matlack 6543ada51a9dSDavid Matlack return false; 6544ada51a9dSDavid Matlack } 6545ada51a9dSDavid Matlack 6546ada51a9dSDavid Matlack static void kvm_shadow_mmu_try_split_huge_pages(struct kvm *kvm, 6547ada51a9dSDavid Matlack const struct kvm_memory_slot *slot, 6548ada51a9dSDavid Matlack gfn_t start, gfn_t end, 6549ada51a9dSDavid Matlack int target_level) 6550ada51a9dSDavid Matlack { 6551ada51a9dSDavid Matlack int level; 6552ada51a9dSDavid Matlack 6553ada51a9dSDavid Matlack /* 6554ada51a9dSDavid Matlack * Split huge pages starting with KVM_MAX_HUGEPAGE_LEVEL and working 6555ada51a9dSDavid Matlack * down to the target level. This ensures pages are recursively split 6556ada51a9dSDavid Matlack * all the way to the target level. There's no need to split pages 6557ada51a9dSDavid Matlack * already at the target level. 6558ada51a9dSDavid Matlack */ 6559727ae377SSean Christopherson for (level = KVM_MAX_HUGEPAGE_LEVEL; level > target_level; level--) 6560727ae377SSean Christopherson __walk_slot_rmaps(kvm, slot, shadow_mmu_try_split_huge_pages, 6561ada51a9dSDavid Matlack level, level, start, end - 1, true, false); 6562ada51a9dSDavid Matlack } 6563ada51a9dSDavid Matlack 6564cb00a70bSDavid Matlack /* Must be called with the mmu_lock held in write-mode. */ 6565cb00a70bSDavid Matlack void kvm_mmu_try_split_huge_pages(struct kvm *kvm, 6566cb00a70bSDavid Matlack const struct kvm_memory_slot *memslot, 6567cb00a70bSDavid Matlack u64 start, u64 end, 6568cb00a70bSDavid Matlack int target_level) 6569cb00a70bSDavid Matlack { 65701f98f2bdSDavid Matlack if (!tdp_mmu_enabled) 6571ada51a9dSDavid Matlack return; 6572ada51a9dSDavid Matlack 6573ada51a9dSDavid Matlack if (kvm_memslots_have_rmaps(kvm)) 6574ada51a9dSDavid Matlack kvm_shadow_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level); 6575ada51a9dSDavid Matlack 6576ada51a9dSDavid Matlack kvm_tdp_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level, false); 6577cb00a70bSDavid Matlack 6578cb00a70bSDavid Matlack /* 6579cb00a70bSDavid Matlack * A TLB flush is unnecessary at this point for the same resons as in 6580cb00a70bSDavid Matlack * kvm_mmu_slot_try_split_huge_pages(). 6581cb00a70bSDavid Matlack */ 6582cb00a70bSDavid Matlack } 6583cb00a70bSDavid Matlack 6584a3fe5dbdSDavid Matlack void kvm_mmu_slot_try_split_huge_pages(struct kvm *kvm, 6585a3fe5dbdSDavid Matlack const struct kvm_memory_slot *memslot, 6586a3fe5dbdSDavid Matlack int target_level) 6587a3fe5dbdSDavid Matlack { 6588a3fe5dbdSDavid Matlack u64 start = memslot->base_gfn; 6589a3fe5dbdSDavid Matlack u64 end = start + memslot->npages; 6590a3fe5dbdSDavid Matlack 65911f98f2bdSDavid Matlack if (!tdp_mmu_enabled) 6592ada51a9dSDavid Matlack return; 6593ada51a9dSDavid Matlack 6594ada51a9dSDavid Matlack if (kvm_memslots_have_rmaps(kvm)) { 6595ada51a9dSDavid Matlack write_lock(&kvm->mmu_lock); 6596ada51a9dSDavid Matlack kvm_shadow_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level); 6597ada51a9dSDavid Matlack write_unlock(&kvm->mmu_lock); 6598ada51a9dSDavid Matlack } 6599ada51a9dSDavid Matlack 6600a3fe5dbdSDavid Matlack read_lock(&kvm->mmu_lock); 6601cb00a70bSDavid Matlack kvm_tdp_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level, true); 6602a3fe5dbdSDavid Matlack read_unlock(&kvm->mmu_lock); 6603a3fe5dbdSDavid Matlack 6604a3fe5dbdSDavid Matlack /* 6605a3fe5dbdSDavid Matlack * No TLB flush is necessary here. KVM will flush TLBs after 6606a3fe5dbdSDavid Matlack * write-protecting and/or clearing dirty on the newly split SPTEs to 6607a3fe5dbdSDavid Matlack * ensure that guest writes are reflected in the dirty log before the 6608a3fe5dbdSDavid Matlack * ioctl to enable dirty logging on this memslot completes. Since the 6609a3fe5dbdSDavid Matlack * split SPTEs retain the write and dirty bits of the huge SPTE, it is 6610a3fe5dbdSDavid Matlack * safe for KVM to decide if a TLB flush is necessary based on the split 6611a3fe5dbdSDavid Matlack * SPTEs. 6612a3fe5dbdSDavid Matlack */ 6613a3fe5dbdSDavid Matlack } 6614a3fe5dbdSDavid Matlack 6615c50d8ae3SPaolo Bonzini static bool kvm_mmu_zap_collapsible_spte(struct kvm *kvm, 66160a234f5dSSean Christopherson struct kvm_rmap_head *rmap_head, 6617269e9552SHamza Mahfooz const struct kvm_memory_slot *slot) 6618c50d8ae3SPaolo Bonzini { 6619c50d8ae3SPaolo Bonzini u64 *sptep; 6620c50d8ae3SPaolo Bonzini struct rmap_iterator iter; 6621c50d8ae3SPaolo Bonzini int need_tlb_flush = 0; 6622c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 6623c50d8ae3SPaolo Bonzini 6624c50d8ae3SPaolo Bonzini restart: 6625c50d8ae3SPaolo Bonzini for_each_rmap_spte(rmap_head, &iter, sptep) { 662657354682SSean Christopherson sp = sptep_to_sp(sptep); 6627c50d8ae3SPaolo Bonzini 6628c50d8ae3SPaolo Bonzini /* 6629c50d8ae3SPaolo Bonzini * We cannot do huge page mapping for indirect shadow pages, 6630c50d8ae3SPaolo Bonzini * which are found on the last rmap (level = 1) when not using 6631c50d8ae3SPaolo Bonzini * tdp; such shadow pages are synced with the page table in 6632c50d8ae3SPaolo Bonzini * the guest, and the guest page table is using 4K page size 6633c50d8ae3SPaolo Bonzini * mapping if the indirect sp has level = 1. 6634c50d8ae3SPaolo Bonzini */ 66355d49f08cSSean Christopherson if (sp->role.direct && 66369eba50f8SSean Christopherson sp->role.level < kvm_mmu_max_mapping_level(kvm, slot, sp->gfn, 6637a8ac499bSSean Christopherson PG_LEVEL_NUM)) { 66389202aee8SSean Christopherson kvm_zap_one_rmap_spte(kvm, rmap_head, sptep); 6639c50d8ae3SPaolo Bonzini 66408a1300ffSSean Christopherson if (kvm_available_flush_remote_tlbs_range()) 66411b2dc736SHou Wenlong kvm_flush_remote_tlbs_sptep(kvm, sptep); 6642c50d8ae3SPaolo Bonzini else 6643c50d8ae3SPaolo Bonzini need_tlb_flush = 1; 6644c50d8ae3SPaolo Bonzini 6645c50d8ae3SPaolo Bonzini goto restart; 6646c50d8ae3SPaolo Bonzini } 6647c50d8ae3SPaolo Bonzini } 6648c50d8ae3SPaolo Bonzini 6649c50d8ae3SPaolo Bonzini return need_tlb_flush; 6650c50d8ae3SPaolo Bonzini } 6651c50d8ae3SPaolo Bonzini 665220d49186SDavid Matlack static void kvm_rmap_zap_collapsible_sptes(struct kvm *kvm, 665320d49186SDavid Matlack const struct kvm_memory_slot *slot) 665420d49186SDavid Matlack { 665520d49186SDavid Matlack /* 665620d49186SDavid Matlack * Note, use KVM_MAX_HUGEPAGE_LEVEL - 1 since there's no need to zap 665720d49186SDavid Matlack * pages that are already mapped at the maximum hugepage level. 665820d49186SDavid Matlack */ 6659727ae377SSean Christopherson if (walk_slot_rmaps(kvm, slot, kvm_mmu_zap_collapsible_spte, 666020d49186SDavid Matlack PG_LEVEL_4K, KVM_MAX_HUGEPAGE_LEVEL - 1, true)) 666120d49186SDavid Matlack kvm_arch_flush_remote_tlbs_memslot(kvm, slot); 666220d49186SDavid Matlack } 666320d49186SDavid Matlack 6664c50d8ae3SPaolo Bonzini void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm, 6665269e9552SHamza Mahfooz const struct kvm_memory_slot *slot) 6666c50d8ae3SPaolo Bonzini { 6667e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) { 6668531810caSBen Gardon write_lock(&kvm->mmu_lock); 666920d49186SDavid Matlack kvm_rmap_zap_collapsible_sptes(kvm, slot); 6670531810caSBen Gardon write_unlock(&kvm->mmu_lock); 6671e2209710SBen Gardon } 66722db6f772SBen Gardon 66731f98f2bdSDavid Matlack if (tdp_mmu_enabled) { 66742db6f772SBen Gardon read_lock(&kvm->mmu_lock); 66754b85c921SSean Christopherson kvm_tdp_mmu_zap_collapsible_sptes(kvm, slot); 66762db6f772SBen Gardon read_unlock(&kvm->mmu_lock); 66772db6f772SBen Gardon } 6678c50d8ae3SPaolo Bonzini } 6679c50d8ae3SPaolo Bonzini 6680b3594ffbSSean Christopherson void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm, 66816c9dd6d2SPaolo Bonzini const struct kvm_memory_slot *memslot) 6682b3594ffbSSean Christopherson { 6683b3594ffbSSean Christopherson /* 66847f42aa76SSean Christopherson * All current use cases for flushing the TLBs for a specific memslot 6685302695a5SSean Christopherson * related to dirty logging, and many do the TLB flush out of mmu_lock. 66867f42aa76SSean Christopherson * The interaction between the various operations on memslot must be 66877f42aa76SSean Christopherson * serialized by slots_locks to ensure the TLB flush from one operation 66887f42aa76SSean Christopherson * is observed by any other operation on the same memslot. 6689b3594ffbSSean Christopherson */ 6690b3594ffbSSean Christopherson lockdep_assert_held(&kvm->slots_lock); 66918c63e8c2SDavid Matlack kvm_flush_remote_tlbs_range(kvm, memslot->base_gfn, memslot->npages); 6692b3594ffbSSean Christopherson } 6693b3594ffbSSean Christopherson 6694c50d8ae3SPaolo Bonzini void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm, 6695269e9552SHamza Mahfooz const struct kvm_memory_slot *memslot) 6696c50d8ae3SPaolo Bonzini { 6697e2209710SBen Gardon if (kvm_memslots_have_rmaps(kvm)) { 6698531810caSBen Gardon write_lock(&kvm->mmu_lock); 6699610265eaSDavid Matlack /* 6700610265eaSDavid Matlack * Clear dirty bits only on 4k SPTEs since the legacy MMU only 6701610265eaSDavid Matlack * support dirty logging at a 4k granularity. 6702610265eaSDavid Matlack */ 6703727ae377SSean Christopherson walk_slot_rmaps_4k(kvm, memslot, __rmap_clear_dirty, false); 6704531810caSBen Gardon write_unlock(&kvm->mmu_lock); 6705e2209710SBen Gardon } 6706c50d8ae3SPaolo Bonzini 67071f98f2bdSDavid Matlack if (tdp_mmu_enabled) { 670824ae4cfaSBen Gardon read_lock(&kvm->mmu_lock); 6709b64d740eSJunaid Shahid kvm_tdp_mmu_clear_dirty_slot(kvm, memslot); 671024ae4cfaSBen Gardon read_unlock(&kvm->mmu_lock); 671124ae4cfaSBen Gardon } 671224ae4cfaSBen Gardon 6713c50d8ae3SPaolo Bonzini /* 6714b64d740eSJunaid Shahid * The caller will flush the TLBs after this function returns. 6715b64d740eSJunaid Shahid * 6716c50d8ae3SPaolo Bonzini * It's also safe to flush TLBs out of mmu lock here as currently this 6717c50d8ae3SPaolo Bonzini * function is only used for dirty logging, in which case flushing TLB 6718c50d8ae3SPaolo Bonzini * out of mmu lock also guarantees no dirty pages will be lost in 6719c50d8ae3SPaolo Bonzini * dirty_bitmap. 6720c50d8ae3SPaolo Bonzini */ 6721c50d8ae3SPaolo Bonzini } 6722c50d8ae3SPaolo Bonzini 6723c50d8ae3SPaolo Bonzini void kvm_mmu_zap_all(struct kvm *kvm) 6724c50d8ae3SPaolo Bonzini { 6725c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp, *node; 6726c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 6727c50d8ae3SPaolo Bonzini int ign; 6728c50d8ae3SPaolo Bonzini 6729531810caSBen Gardon write_lock(&kvm->mmu_lock); 6730c50d8ae3SPaolo Bonzini restart: 6731c50d8ae3SPaolo Bonzini list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link) { 6732f95eec9bSSean Christopherson if (WARN_ON(sp->role.invalid)) 6733c50d8ae3SPaolo Bonzini continue; 6734c50d8ae3SPaolo Bonzini if (__kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list, &ign)) 6735c50d8ae3SPaolo Bonzini goto restart; 6736531810caSBen Gardon if (cond_resched_rwlock_write(&kvm->mmu_lock)) 6737c50d8ae3SPaolo Bonzini goto restart; 6738c50d8ae3SPaolo Bonzini } 6739c50d8ae3SPaolo Bonzini 6740c50d8ae3SPaolo Bonzini kvm_mmu_commit_zap_page(kvm, &invalid_list); 6741faaf05b0SBen Gardon 67421f98f2bdSDavid Matlack if (tdp_mmu_enabled) 6743faaf05b0SBen Gardon kvm_tdp_mmu_zap_all(kvm); 6744faaf05b0SBen Gardon 6745531810caSBen Gardon write_unlock(&kvm->mmu_lock); 6746c50d8ae3SPaolo Bonzini } 6747c50d8ae3SPaolo Bonzini 6748c50d8ae3SPaolo Bonzini void kvm_mmu_invalidate_mmio_sptes(struct kvm *kvm, u64 gen) 6749c50d8ae3SPaolo Bonzini { 6750c50d8ae3SPaolo Bonzini WARN_ON(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS); 6751c50d8ae3SPaolo Bonzini 6752c50d8ae3SPaolo Bonzini gen &= MMIO_SPTE_GEN_MASK; 6753c50d8ae3SPaolo Bonzini 6754c50d8ae3SPaolo Bonzini /* 6755c50d8ae3SPaolo Bonzini * Generation numbers are incremented in multiples of the number of 6756c50d8ae3SPaolo Bonzini * address spaces in order to provide unique generations across all 6757c50d8ae3SPaolo Bonzini * address spaces. Strip what is effectively the address space 6758c50d8ae3SPaolo Bonzini * modifier prior to checking for a wrap of the MMIO generation so 6759c50d8ae3SPaolo Bonzini * that a wrap in any address space is detected. 6760c50d8ae3SPaolo Bonzini */ 6761c50d8ae3SPaolo Bonzini gen &= ~((u64)KVM_ADDRESS_SPACE_NUM - 1); 6762c50d8ae3SPaolo Bonzini 6763c50d8ae3SPaolo Bonzini /* 6764c50d8ae3SPaolo Bonzini * The very rare case: if the MMIO generation number has wrapped, 6765c50d8ae3SPaolo Bonzini * zap all shadow pages. 6766c50d8ae3SPaolo Bonzini */ 6767c50d8ae3SPaolo Bonzini if (unlikely(gen == 0)) { 67688d20bd63SSean Christopherson kvm_debug_ratelimited("zapping shadow pages for mmio generation wraparound\n"); 6769c50d8ae3SPaolo Bonzini kvm_mmu_zap_all_fast(kvm); 6770c50d8ae3SPaolo Bonzini } 6771c50d8ae3SPaolo Bonzini } 6772c50d8ae3SPaolo Bonzini 6773f3d90f90SSean Christopherson static unsigned long mmu_shrink_scan(struct shrinker *shrink, 6774f3d90f90SSean Christopherson struct shrink_control *sc) 6775c50d8ae3SPaolo Bonzini { 6776c50d8ae3SPaolo Bonzini struct kvm *kvm; 6777c50d8ae3SPaolo Bonzini int nr_to_scan = sc->nr_to_scan; 6778c50d8ae3SPaolo Bonzini unsigned long freed = 0; 6779c50d8ae3SPaolo Bonzini 6780c50d8ae3SPaolo Bonzini mutex_lock(&kvm_lock); 6781c50d8ae3SPaolo Bonzini 6782c50d8ae3SPaolo Bonzini list_for_each_entry(kvm, &vm_list, vm_list) { 6783c50d8ae3SPaolo Bonzini int idx; 6784c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 6785c50d8ae3SPaolo Bonzini 6786c50d8ae3SPaolo Bonzini /* 6787c50d8ae3SPaolo Bonzini * Never scan more than sc->nr_to_scan VM instances. 6788c50d8ae3SPaolo Bonzini * Will not hit this condition practically since we do not try 6789c50d8ae3SPaolo Bonzini * to shrink more than one VM and it is very unlikely to see 6790c50d8ae3SPaolo Bonzini * !n_used_mmu_pages so many times. 6791c50d8ae3SPaolo Bonzini */ 6792c50d8ae3SPaolo Bonzini if (!nr_to_scan--) 6793c50d8ae3SPaolo Bonzini break; 6794c50d8ae3SPaolo Bonzini /* 6795c50d8ae3SPaolo Bonzini * n_used_mmu_pages is accessed without holding kvm->mmu_lock 6796c50d8ae3SPaolo Bonzini * here. We may skip a VM instance errorneosly, but we do not 6797c50d8ae3SPaolo Bonzini * want to shrink a VM that only started to populate its MMU 6798c50d8ae3SPaolo Bonzini * anyway. 6799c50d8ae3SPaolo Bonzini */ 6800c50d8ae3SPaolo Bonzini if (!kvm->arch.n_used_mmu_pages && 6801c50d8ae3SPaolo Bonzini !kvm_has_zapped_obsolete_pages(kvm)) 6802c50d8ae3SPaolo Bonzini continue; 6803c50d8ae3SPaolo Bonzini 6804c50d8ae3SPaolo Bonzini idx = srcu_read_lock(&kvm->srcu); 6805531810caSBen Gardon write_lock(&kvm->mmu_lock); 6806c50d8ae3SPaolo Bonzini 6807c50d8ae3SPaolo Bonzini if (kvm_has_zapped_obsolete_pages(kvm)) { 6808c50d8ae3SPaolo Bonzini kvm_mmu_commit_zap_page(kvm, 6809c50d8ae3SPaolo Bonzini &kvm->arch.zapped_obsolete_pages); 6810c50d8ae3SPaolo Bonzini goto unlock; 6811c50d8ae3SPaolo Bonzini } 6812c50d8ae3SPaolo Bonzini 6813ebdb292dSSean Christopherson freed = kvm_mmu_zap_oldest_mmu_pages(kvm, sc->nr_to_scan); 6814c50d8ae3SPaolo Bonzini 6815c50d8ae3SPaolo Bonzini unlock: 6816531810caSBen Gardon write_unlock(&kvm->mmu_lock); 6817c50d8ae3SPaolo Bonzini srcu_read_unlock(&kvm->srcu, idx); 6818c50d8ae3SPaolo Bonzini 6819c50d8ae3SPaolo Bonzini /* 6820c50d8ae3SPaolo Bonzini * unfair on small ones 6821c50d8ae3SPaolo Bonzini * per-vm shrinkers cry out 6822c50d8ae3SPaolo Bonzini * sadness comes quickly 6823c50d8ae3SPaolo Bonzini */ 6824c50d8ae3SPaolo Bonzini list_move_tail(&kvm->vm_list, &vm_list); 6825c50d8ae3SPaolo Bonzini break; 6826c50d8ae3SPaolo Bonzini } 6827c50d8ae3SPaolo Bonzini 6828c50d8ae3SPaolo Bonzini mutex_unlock(&kvm_lock); 6829c50d8ae3SPaolo Bonzini return freed; 6830c50d8ae3SPaolo Bonzini } 6831c50d8ae3SPaolo Bonzini 6832f3d90f90SSean Christopherson static unsigned long mmu_shrink_count(struct shrinker *shrink, 6833f3d90f90SSean Christopherson struct shrink_control *sc) 6834c50d8ae3SPaolo Bonzini { 6835c50d8ae3SPaolo Bonzini return percpu_counter_read_positive(&kvm_total_used_mmu_pages); 6836c50d8ae3SPaolo Bonzini } 6837c50d8ae3SPaolo Bonzini 6838c50d8ae3SPaolo Bonzini static struct shrinker mmu_shrinker = { 6839c50d8ae3SPaolo Bonzini .count_objects = mmu_shrink_count, 6840c50d8ae3SPaolo Bonzini .scan_objects = mmu_shrink_scan, 6841c50d8ae3SPaolo Bonzini .seeks = DEFAULT_SEEKS * 10, 6842c50d8ae3SPaolo Bonzini }; 6843c50d8ae3SPaolo Bonzini 6844c50d8ae3SPaolo Bonzini static void mmu_destroy_caches(void) 6845c50d8ae3SPaolo Bonzini { 6846c50d8ae3SPaolo Bonzini kmem_cache_destroy(pte_list_desc_cache); 6847c50d8ae3SPaolo Bonzini kmem_cache_destroy(mmu_page_header_cache); 6848c50d8ae3SPaolo Bonzini } 6849c50d8ae3SPaolo Bonzini 6850c50d8ae3SPaolo Bonzini static bool get_nx_auto_mode(void) 6851c50d8ae3SPaolo Bonzini { 6852c50d8ae3SPaolo Bonzini /* Return true when CPU has the bug, and mitigations are ON */ 6853c50d8ae3SPaolo Bonzini return boot_cpu_has_bug(X86_BUG_ITLB_MULTIHIT) && !cpu_mitigations_off(); 6854c50d8ae3SPaolo Bonzini } 6855c50d8ae3SPaolo Bonzini 6856c50d8ae3SPaolo Bonzini static void __set_nx_huge_pages(bool val) 6857c50d8ae3SPaolo Bonzini { 6858c50d8ae3SPaolo Bonzini nx_huge_pages = itlb_multihit_kvm_mitigation = val; 6859c50d8ae3SPaolo Bonzini } 6860c50d8ae3SPaolo Bonzini 6861c50d8ae3SPaolo Bonzini static int set_nx_huge_pages(const char *val, const struct kernel_param *kp) 6862c50d8ae3SPaolo Bonzini { 6863c50d8ae3SPaolo Bonzini bool old_val = nx_huge_pages; 6864c50d8ae3SPaolo Bonzini bool new_val; 6865c50d8ae3SPaolo Bonzini 6866c50d8ae3SPaolo Bonzini /* In "auto" mode deploy workaround only if CPU has the bug. */ 6867c50d8ae3SPaolo Bonzini if (sysfs_streq(val, "off")) 6868c50d8ae3SPaolo Bonzini new_val = 0; 6869c50d8ae3SPaolo Bonzini else if (sysfs_streq(val, "force")) 6870c50d8ae3SPaolo Bonzini new_val = 1; 6871c50d8ae3SPaolo Bonzini else if (sysfs_streq(val, "auto")) 6872c50d8ae3SPaolo Bonzini new_val = get_nx_auto_mode(); 687311b36fe7SChristophe JAILLET else if (kstrtobool(val, &new_val) < 0) 6874c50d8ae3SPaolo Bonzini return -EINVAL; 6875c50d8ae3SPaolo Bonzini 6876c50d8ae3SPaolo Bonzini __set_nx_huge_pages(new_val); 6877c50d8ae3SPaolo Bonzini 6878c50d8ae3SPaolo Bonzini if (new_val != old_val) { 6879c50d8ae3SPaolo Bonzini struct kvm *kvm; 6880c50d8ae3SPaolo Bonzini 6881c50d8ae3SPaolo Bonzini mutex_lock(&kvm_lock); 6882c50d8ae3SPaolo Bonzini 6883c50d8ae3SPaolo Bonzini list_for_each_entry(kvm, &vm_list, vm_list) { 6884c50d8ae3SPaolo Bonzini mutex_lock(&kvm->slots_lock); 6885c50d8ae3SPaolo Bonzini kvm_mmu_zap_all_fast(kvm); 6886c50d8ae3SPaolo Bonzini mutex_unlock(&kvm->slots_lock); 6887c50d8ae3SPaolo Bonzini 688855c510e2SSean Christopherson wake_up_process(kvm->arch.nx_huge_page_recovery_thread); 6889c50d8ae3SPaolo Bonzini } 6890c50d8ae3SPaolo Bonzini mutex_unlock(&kvm_lock); 6891c50d8ae3SPaolo Bonzini } 6892c50d8ae3SPaolo Bonzini 6893c50d8ae3SPaolo Bonzini return 0; 6894c50d8ae3SPaolo Bonzini } 6895c50d8ae3SPaolo Bonzini 68961d0e8480SSean Christopherson /* 68971d0e8480SSean Christopherson * nx_huge_pages needs to be resolved to true/false when kvm.ko is loaded, as 68981d0e8480SSean Christopherson * its default value of -1 is technically undefined behavior for a boolean. 6899c3e0c8c2SSean Christopherson * Forward the module init call to SPTE code so that it too can handle module 6900c3e0c8c2SSean Christopherson * params that need to be resolved/snapshot. 69011d0e8480SSean Christopherson */ 6902982bae43SSean Christopherson void __init kvm_mmu_x86_module_init(void) 6903c50d8ae3SPaolo Bonzini { 6904c50d8ae3SPaolo Bonzini if (nx_huge_pages == -1) 6905c50d8ae3SPaolo Bonzini __set_nx_huge_pages(get_nx_auto_mode()); 6906c3e0c8c2SSean Christopherson 69071f98f2bdSDavid Matlack /* 69081f98f2bdSDavid Matlack * Snapshot userspace's desire to enable the TDP MMU. Whether or not the 69091f98f2bdSDavid Matlack * TDP MMU is actually enabled is determined in kvm_configure_mmu() 69101f98f2bdSDavid Matlack * when the vendor module is loaded. 69111f98f2bdSDavid Matlack */ 69121f98f2bdSDavid Matlack tdp_mmu_allowed = tdp_mmu_enabled; 69131f98f2bdSDavid Matlack 6914c3e0c8c2SSean Christopherson kvm_mmu_spte_module_init(); 69151d0e8480SSean Christopherson } 69161d0e8480SSean Christopherson 69171d0e8480SSean Christopherson /* 69181d0e8480SSean Christopherson * The bulk of the MMU initialization is deferred until the vendor module is 69191d0e8480SSean Christopherson * loaded as many of the masks/values may be modified by VMX or SVM, i.e. need 69201d0e8480SSean Christopherson * to be reset when a potentially different vendor module is loaded. 69211d0e8480SSean Christopherson */ 69221d0e8480SSean Christopherson int kvm_mmu_vendor_module_init(void) 69231d0e8480SSean Christopherson { 69241d0e8480SSean Christopherson int ret = -ENOMEM; 6925c50d8ae3SPaolo Bonzini 6926c50d8ae3SPaolo Bonzini /* 6927c50d8ae3SPaolo Bonzini * MMU roles use union aliasing which is, generally speaking, an 6928c50d8ae3SPaolo Bonzini * undefined behavior. However, we supposedly know how compilers behave 6929c50d8ae3SPaolo Bonzini * and the current status quo is unlikely to change. Guardians below are 6930c50d8ae3SPaolo Bonzini * supposed to let us know if the assumption becomes false. 6931c50d8ae3SPaolo Bonzini */ 6932c50d8ae3SPaolo Bonzini BUILD_BUG_ON(sizeof(union kvm_mmu_page_role) != sizeof(u32)); 6933c50d8ae3SPaolo Bonzini BUILD_BUG_ON(sizeof(union kvm_mmu_extended_role) != sizeof(u32)); 69347a7ae829SPaolo Bonzini BUILD_BUG_ON(sizeof(union kvm_cpu_role) != sizeof(u64)); 6935c50d8ae3SPaolo Bonzini 6936c50d8ae3SPaolo Bonzini kvm_mmu_reset_all_pte_masks(); 6937c50d8ae3SPaolo Bonzini 6938c50d8ae3SPaolo Bonzini pte_list_desc_cache = kmem_cache_create("pte_list_desc", 6939c50d8ae3SPaolo Bonzini sizeof(struct pte_list_desc), 6940c50d8ae3SPaolo Bonzini 0, SLAB_ACCOUNT, NULL); 6941c50d8ae3SPaolo Bonzini if (!pte_list_desc_cache) 6942c50d8ae3SPaolo Bonzini goto out; 6943c50d8ae3SPaolo Bonzini 6944c50d8ae3SPaolo Bonzini mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header", 6945c50d8ae3SPaolo Bonzini sizeof(struct kvm_mmu_page), 6946c50d8ae3SPaolo Bonzini 0, SLAB_ACCOUNT, NULL); 6947c50d8ae3SPaolo Bonzini if (!mmu_page_header_cache) 6948c50d8ae3SPaolo Bonzini goto out; 6949c50d8ae3SPaolo Bonzini 6950c50d8ae3SPaolo Bonzini if (percpu_counter_init(&kvm_total_used_mmu_pages, 0, GFP_KERNEL)) 6951c50d8ae3SPaolo Bonzini goto out; 6952c50d8ae3SPaolo Bonzini 6953e33c267aSRoman Gushchin ret = register_shrinker(&mmu_shrinker, "x86-mmu"); 6954c50d8ae3SPaolo Bonzini if (ret) 6955d7c9bfb9SMiaohe Lin goto out_shrinker; 6956c50d8ae3SPaolo Bonzini 6957c50d8ae3SPaolo Bonzini return 0; 6958c50d8ae3SPaolo Bonzini 6959d7c9bfb9SMiaohe Lin out_shrinker: 6960d7c9bfb9SMiaohe Lin percpu_counter_destroy(&kvm_total_used_mmu_pages); 6961c50d8ae3SPaolo Bonzini out: 6962c50d8ae3SPaolo Bonzini mmu_destroy_caches(); 6963c50d8ae3SPaolo Bonzini return ret; 6964c50d8ae3SPaolo Bonzini } 6965c50d8ae3SPaolo Bonzini 6966c50d8ae3SPaolo Bonzini void kvm_mmu_destroy(struct kvm_vcpu *vcpu) 6967c50d8ae3SPaolo Bonzini { 6968c50d8ae3SPaolo Bonzini kvm_mmu_unload(vcpu); 6969c50d8ae3SPaolo Bonzini free_mmu_pages(&vcpu->arch.root_mmu); 6970c50d8ae3SPaolo Bonzini free_mmu_pages(&vcpu->arch.guest_mmu); 6971c50d8ae3SPaolo Bonzini mmu_free_memory_caches(vcpu); 6972c50d8ae3SPaolo Bonzini } 6973c50d8ae3SPaolo Bonzini 69741d0e8480SSean Christopherson void kvm_mmu_vendor_module_exit(void) 6975c50d8ae3SPaolo Bonzini { 6976c50d8ae3SPaolo Bonzini mmu_destroy_caches(); 6977c50d8ae3SPaolo Bonzini percpu_counter_destroy(&kvm_total_used_mmu_pages); 6978c50d8ae3SPaolo Bonzini unregister_shrinker(&mmu_shrinker); 6979c50d8ae3SPaolo Bonzini } 6980c50d8ae3SPaolo Bonzini 6981f47491d7SSean Christopherson /* 6982f47491d7SSean Christopherson * Calculate the effective recovery period, accounting for '0' meaning "let KVM 6983f47491d7SSean Christopherson * select a halving time of 1 hour". Returns true if recovery is enabled. 6984f47491d7SSean Christopherson */ 6985f47491d7SSean Christopherson static bool calc_nx_huge_pages_recovery_period(uint *period) 6986f47491d7SSean Christopherson { 6987f47491d7SSean Christopherson /* 6988f47491d7SSean Christopherson * Use READ_ONCE to get the params, this may be called outside of the 6989f47491d7SSean Christopherson * param setters, e.g. by the kthread to compute its next timeout. 6990f47491d7SSean Christopherson */ 6991f47491d7SSean Christopherson bool enabled = READ_ONCE(nx_huge_pages); 6992f47491d7SSean Christopherson uint ratio = READ_ONCE(nx_huge_pages_recovery_ratio); 6993f47491d7SSean Christopherson 6994f47491d7SSean Christopherson if (!enabled || !ratio) 6995f47491d7SSean Christopherson return false; 6996f47491d7SSean Christopherson 6997f47491d7SSean Christopherson *period = READ_ONCE(nx_huge_pages_recovery_period_ms); 6998f47491d7SSean Christopherson if (!*period) { 6999f47491d7SSean Christopherson /* Make sure the period is not less than one second. */ 7000f47491d7SSean Christopherson ratio = min(ratio, 3600u); 7001f47491d7SSean Christopherson *period = 60 * 60 * 1000 / ratio; 7002f47491d7SSean Christopherson } 7003f47491d7SSean Christopherson return true; 7004f47491d7SSean Christopherson } 7005f47491d7SSean Christopherson 70064dfe4f40SJunaid Shahid static int set_nx_huge_pages_recovery_param(const char *val, const struct kernel_param *kp) 7007c50d8ae3SPaolo Bonzini { 70084dfe4f40SJunaid Shahid bool was_recovery_enabled, is_recovery_enabled; 70094dfe4f40SJunaid Shahid uint old_period, new_period; 7010c50d8ae3SPaolo Bonzini int err; 7011c50d8ae3SPaolo Bonzini 7012f47491d7SSean Christopherson was_recovery_enabled = calc_nx_huge_pages_recovery_period(&old_period); 70134dfe4f40SJunaid Shahid 7014c50d8ae3SPaolo Bonzini err = param_set_uint(val, kp); 7015c50d8ae3SPaolo Bonzini if (err) 7016c50d8ae3SPaolo Bonzini return err; 7017c50d8ae3SPaolo Bonzini 7018f47491d7SSean Christopherson is_recovery_enabled = calc_nx_huge_pages_recovery_period(&new_period); 70194dfe4f40SJunaid Shahid 7020f47491d7SSean Christopherson if (is_recovery_enabled && 70214dfe4f40SJunaid Shahid (!was_recovery_enabled || old_period > new_period)) { 7022c50d8ae3SPaolo Bonzini struct kvm *kvm; 7023c50d8ae3SPaolo Bonzini 7024c50d8ae3SPaolo Bonzini mutex_lock(&kvm_lock); 7025c50d8ae3SPaolo Bonzini 7026c50d8ae3SPaolo Bonzini list_for_each_entry(kvm, &vm_list, vm_list) 702755c510e2SSean Christopherson wake_up_process(kvm->arch.nx_huge_page_recovery_thread); 7028c50d8ae3SPaolo Bonzini 7029c50d8ae3SPaolo Bonzini mutex_unlock(&kvm_lock); 7030c50d8ae3SPaolo Bonzini } 7031c50d8ae3SPaolo Bonzini 7032c50d8ae3SPaolo Bonzini return err; 7033c50d8ae3SPaolo Bonzini } 7034c50d8ae3SPaolo Bonzini 703555c510e2SSean Christopherson static void kvm_recover_nx_huge_pages(struct kvm *kvm) 7036c50d8ae3SPaolo Bonzini { 7037ade74e14SSean Christopherson unsigned long nx_lpage_splits = kvm->stat.nx_lpage_splits; 7038eb298605SDavid Matlack struct kvm_memory_slot *slot; 7039c50d8ae3SPaolo Bonzini int rcu_idx; 7040c50d8ae3SPaolo Bonzini struct kvm_mmu_page *sp; 7041c50d8ae3SPaolo Bonzini unsigned int ratio; 7042c50d8ae3SPaolo Bonzini LIST_HEAD(invalid_list); 7043048f4980SSean Christopherson bool flush = false; 7044c50d8ae3SPaolo Bonzini ulong to_zap; 7045c50d8ae3SPaolo Bonzini 7046c50d8ae3SPaolo Bonzini rcu_idx = srcu_read_lock(&kvm->srcu); 7047531810caSBen Gardon write_lock(&kvm->mmu_lock); 7048c50d8ae3SPaolo Bonzini 7049bb95dfb9SSean Christopherson /* 7050bb95dfb9SSean Christopherson * Zapping TDP MMU shadow pages, including the remote TLB flush, must 7051bb95dfb9SSean Christopherson * be done under RCU protection, because the pages are freed via RCU 7052bb95dfb9SSean Christopherson * callback. 7053bb95dfb9SSean Christopherson */ 7054bb95dfb9SSean Christopherson rcu_read_lock(); 7055bb95dfb9SSean Christopherson 7056c50d8ae3SPaolo Bonzini ratio = READ_ONCE(nx_huge_pages_recovery_ratio); 7057ade74e14SSean Christopherson to_zap = ratio ? DIV_ROUND_UP(nx_lpage_splits, ratio) : 0; 70587d919c7aSSean Christopherson for ( ; to_zap; --to_zap) { 705955c510e2SSean Christopherson if (list_empty(&kvm->arch.possible_nx_huge_pages)) 70607d919c7aSSean Christopherson break; 70617d919c7aSSean Christopherson 7062c50d8ae3SPaolo Bonzini /* 7063c50d8ae3SPaolo Bonzini * We use a separate list instead of just using active_mmu_pages 706455c510e2SSean Christopherson * because the number of shadow pages that be replaced with an 706555c510e2SSean Christopherson * NX huge page is expected to be relatively small compared to 706655c510e2SSean Christopherson * the total number of shadow pages. And because the TDP MMU 706755c510e2SSean Christopherson * doesn't use active_mmu_pages. 7068c50d8ae3SPaolo Bonzini */ 706955c510e2SSean Christopherson sp = list_first_entry(&kvm->arch.possible_nx_huge_pages, 7070c50d8ae3SPaolo Bonzini struct kvm_mmu_page, 707155c510e2SSean Christopherson possible_nx_huge_page_link); 707255c510e2SSean Christopherson WARN_ON_ONCE(!sp->nx_huge_page_disallowed); 7073eb298605SDavid Matlack WARN_ON_ONCE(!sp->role.direct); 7074eb298605SDavid Matlack 7075eb298605SDavid Matlack /* 7076eb298605SDavid Matlack * Unaccount and do not attempt to recover any NX Huge Pages 7077eb298605SDavid Matlack * that are being dirty tracked, as they would just be faulted 7078eb298605SDavid Matlack * back in as 4KiB pages. The NX Huge Pages in this slot will be 7079eb298605SDavid Matlack * recovered, along with all the other huge pages in the slot, 7080eb298605SDavid Matlack * when dirty logging is disabled. 70816c7b2202SPaolo Bonzini * 70826c7b2202SPaolo Bonzini * Since gfn_to_memslot() is relatively expensive, it helps to 70836c7b2202SPaolo Bonzini * skip it if it the test cannot possibly return true. On the 70846c7b2202SPaolo Bonzini * other hand, if any memslot has logging enabled, chances are 70856c7b2202SPaolo Bonzini * good that all of them do, in which case unaccount_nx_huge_page() 70866c7b2202SPaolo Bonzini * is much cheaper than zapping the page. 70876c7b2202SPaolo Bonzini * 70886c7b2202SPaolo Bonzini * If a memslot update is in progress, reading an incorrect value 70896c7b2202SPaolo Bonzini * of kvm->nr_memslots_dirty_logging is not a problem: if it is 70906c7b2202SPaolo Bonzini * becoming zero, gfn_to_memslot() will be done unnecessarily; if 70916c7b2202SPaolo Bonzini * it is becoming nonzero, the page will be zapped unnecessarily. 70926c7b2202SPaolo Bonzini * Either way, this only affects efficiency in racy situations, 70936c7b2202SPaolo Bonzini * and not correctness. 7094eb298605SDavid Matlack */ 70956c7b2202SPaolo Bonzini slot = NULL; 70966c7b2202SPaolo Bonzini if (atomic_read(&kvm->nr_memslots_dirty_logging)) { 70976c7b2202SPaolo Bonzini slot = gfn_to_memslot(kvm, sp->gfn); 70986c7b2202SPaolo Bonzini WARN_ON_ONCE(!slot); 709929cf0f50SBen Gardon } 7100c50d8ae3SPaolo Bonzini 7101eb298605SDavid Matlack if (slot && kvm_slot_dirty_track_enabled(slot)) 7102eb298605SDavid Matlack unaccount_nx_huge_page(kvm, sp); 7103eb298605SDavid Matlack else if (is_tdp_mmu_page(sp)) 7104c50d8ae3SPaolo Bonzini flush |= kvm_tdp_mmu_zap_sp(kvm, sp); 71053a056757SSean Christopherson else 7106c50d8ae3SPaolo Bonzini kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list); 710755c510e2SSean Christopherson WARN_ON_ONCE(sp->nx_huge_page_disallowed); 7108c50d8ae3SPaolo Bonzini 7109531810caSBen Gardon if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) { 7110048f4980SSean Christopherson kvm_mmu_remote_flush_or_zap(kvm, &invalid_list, flush); 7111bb95dfb9SSean Christopherson rcu_read_unlock(); 7112bb95dfb9SSean Christopherson 7113531810caSBen Gardon cond_resched_rwlock_write(&kvm->mmu_lock); 7114048f4980SSean Christopherson flush = false; 7115bb95dfb9SSean Christopherson 7116bb95dfb9SSean Christopherson rcu_read_lock(); 7117c50d8ae3SPaolo Bonzini } 7118c50d8ae3SPaolo Bonzini } 7119048f4980SSean Christopherson kvm_mmu_remote_flush_or_zap(kvm, &invalid_list, flush); 7120c50d8ae3SPaolo Bonzini 7121bb95dfb9SSean Christopherson rcu_read_unlock(); 7122bb95dfb9SSean Christopherson 7123531810caSBen Gardon write_unlock(&kvm->mmu_lock); 7124c50d8ae3SPaolo Bonzini srcu_read_unlock(&kvm->srcu, rcu_idx); 7125c50d8ae3SPaolo Bonzini } 7126c50d8ae3SPaolo Bonzini 712755c510e2SSean Christopherson static long get_nx_huge_page_recovery_timeout(u64 start_time) 7128c50d8ae3SPaolo Bonzini { 7129f47491d7SSean Christopherson bool enabled; 7130f47491d7SSean Christopherson uint period; 71314dfe4f40SJunaid Shahid 7132f47491d7SSean Christopherson enabled = calc_nx_huge_pages_recovery_period(&period); 71334dfe4f40SJunaid Shahid 7134f47491d7SSean Christopherson return enabled ? start_time + msecs_to_jiffies(period) - get_jiffies_64() 7135c50d8ae3SPaolo Bonzini : MAX_SCHEDULE_TIMEOUT; 7136c50d8ae3SPaolo Bonzini } 7137c50d8ae3SPaolo Bonzini 713855c510e2SSean Christopherson static int kvm_nx_huge_page_recovery_worker(struct kvm *kvm, uintptr_t data) 7139c50d8ae3SPaolo Bonzini { 7140c50d8ae3SPaolo Bonzini u64 start_time; 7141c50d8ae3SPaolo Bonzini long remaining_time; 7142c50d8ae3SPaolo Bonzini 7143c50d8ae3SPaolo Bonzini while (true) { 7144c50d8ae3SPaolo Bonzini start_time = get_jiffies_64(); 714555c510e2SSean Christopherson remaining_time = get_nx_huge_page_recovery_timeout(start_time); 7146c50d8ae3SPaolo Bonzini 7147c50d8ae3SPaolo Bonzini set_current_state(TASK_INTERRUPTIBLE); 7148c50d8ae3SPaolo Bonzini while (!kthread_should_stop() && remaining_time > 0) { 7149c50d8ae3SPaolo Bonzini schedule_timeout(remaining_time); 715055c510e2SSean Christopherson remaining_time = get_nx_huge_page_recovery_timeout(start_time); 7151c50d8ae3SPaolo Bonzini set_current_state(TASK_INTERRUPTIBLE); 7152c50d8ae3SPaolo Bonzini } 7153c50d8ae3SPaolo Bonzini 7154c50d8ae3SPaolo Bonzini set_current_state(TASK_RUNNING); 7155c50d8ae3SPaolo Bonzini 7156c50d8ae3SPaolo Bonzini if (kthread_should_stop()) 7157c50d8ae3SPaolo Bonzini return 0; 7158c50d8ae3SPaolo Bonzini 715955c510e2SSean Christopherson kvm_recover_nx_huge_pages(kvm); 7160c50d8ae3SPaolo Bonzini } 7161c50d8ae3SPaolo Bonzini } 7162c50d8ae3SPaolo Bonzini 7163c50d8ae3SPaolo Bonzini int kvm_mmu_post_init_vm(struct kvm *kvm) 7164c50d8ae3SPaolo Bonzini { 7165c50d8ae3SPaolo Bonzini int err; 7166c50d8ae3SPaolo Bonzini 716755c510e2SSean Christopherson err = kvm_vm_create_worker_thread(kvm, kvm_nx_huge_page_recovery_worker, 0, 7168c50d8ae3SPaolo Bonzini "kvm-nx-lpage-recovery", 716955c510e2SSean Christopherson &kvm->arch.nx_huge_page_recovery_thread); 7170c50d8ae3SPaolo Bonzini if (!err) 717155c510e2SSean Christopherson kthread_unpark(kvm->arch.nx_huge_page_recovery_thread); 7172c50d8ae3SPaolo Bonzini 7173c50d8ae3SPaolo Bonzini return err; 7174c50d8ae3SPaolo Bonzini } 7175c50d8ae3SPaolo Bonzini 7176c50d8ae3SPaolo Bonzini void kvm_mmu_pre_destroy_vm(struct kvm *kvm) 7177c50d8ae3SPaolo Bonzini { 717855c510e2SSean Christopherson if (kvm->arch.nx_huge_page_recovery_thread) 717955c510e2SSean Christopherson kthread_stop(kvm->arch.nx_huge_page_recovery_thread); 7180c50d8ae3SPaolo Bonzini } 7181