1 // SPDX-License-Identifier: GPL-2.0-only
3 * Kernel-based Virtual Machine driver for Linux
5 * This module enables machines with Intel VT-x extensions to run virtual
6 * machines without emulation or binary translation.
10 * Copyright (C) 2006 Qumranet, Inc.
11 * Copyright 2010 Red Hat, Inc. and/or its affiliates.
14 * Yaniv Kamay <yaniv@qumranet.com>
15 * Avi Kivity <avi@qumranet.com>
21 #include "mmu_internal.h"
24 #include "kvm_cache_regs.h"
25 #include "kvm_emulate.h"
29 #include <linux/kvm_host.h>
30 #include <linux/types.h>
31 #include <linux/string.h>
33 #include <linux/highmem.h>
34 #include <linux/moduleparam.h>
35 #include <linux/export.h>
36 #include <linux/swap.h>
37 #include <linux/hugetlb.h>
38 #include <linux/compiler.h>
39 #include <linux/srcu.h>
40 #include <linux/slab.h>
41 #include <linux/sched/signal.h>
42 #include <linux/uaccess.h>
43 #include <linux/hash.h>
44 #include <linux/kern_levels.h>
45 #include <linux/kthread.h>
48 #include <asm/memtype.h>
49 #include <asm/cmpxchg.h>
51 #include <asm/set_memory.h>
53 #include <asm/kvm_page_track.h>
56 extern bool itlb_multihit_kvm_mitigation;
58 int __read_mostly nx_huge_pages = -1;
59 static uint __read_mostly nx_huge_pages_recovery_period_ms;
60 #ifdef CONFIG_PREEMPT_RT
61 /* Recovery can cause latency spikes, disable it for PREEMPT_RT. */
62 static uint __read_mostly nx_huge_pages_recovery_ratio = 0;
64 static uint __read_mostly nx_huge_pages_recovery_ratio = 60;
67 static int set_nx_huge_pages(const char *val, const struct kernel_param *kp);
68 static int set_nx_huge_pages_recovery_param(const char *val, const struct kernel_param *kp);
70 static const struct kernel_param_ops nx_huge_pages_ops = {
71 .set = set_nx_huge_pages,
72 .get = param_get_bool,
75 static const struct kernel_param_ops nx_huge_pages_recovery_param_ops = {
76 .set = set_nx_huge_pages_recovery_param,
77 .get = param_get_uint,
80 module_param_cb(nx_huge_pages, &nx_huge_pages_ops, &nx_huge_pages, 0644);
81 __MODULE_PARM_TYPE(nx_huge_pages, "bool");
82 module_param_cb(nx_huge_pages_recovery_ratio, &nx_huge_pages_recovery_param_ops,
83 &nx_huge_pages_recovery_ratio, 0644);
84 __MODULE_PARM_TYPE(nx_huge_pages_recovery_ratio, "uint");
85 module_param_cb(nx_huge_pages_recovery_period_ms, &nx_huge_pages_recovery_param_ops,
86 &nx_huge_pages_recovery_period_ms, 0644);
87 __MODULE_PARM_TYPE(nx_huge_pages_recovery_period_ms, "uint");
89 static bool __read_mostly force_flush_and_sync_on_reuse;
90 module_param_named(flush_on_reuse, force_flush_and_sync_on_reuse, bool, 0644);
93 * When setting this variable to true it enables Two-Dimensional-Paging
94 * where the hardware walks 2 page tables:
95 * 1. the guest-virtual to guest-physical
96 * 2. while doing 1. it walks guest-physical to host-physical
97 * If the hardware supports that we don't need to do shadow paging.
99 bool tdp_enabled = false;
101 static int max_huge_page_level __read_mostly;
102 static int tdp_root_level __read_mostly;
103 static int max_tdp_level __read_mostly;
107 module_param(dbg, bool, 0644);
110 #define PTE_PREFETCH_NUM 8
112 #include <trace/events/kvm.h>
114 /* make pte_list_desc fit well in cache lines */
115 #define PTE_LIST_EXT 14
118 * Slight optimization of cacheline layout, by putting `more' and `spte_count'
119 * at the start; then accessing it will only use one single cacheline for
120 * either full (entries==PTE_LIST_EXT) case or entries<=6.
122 struct pte_list_desc {
123 struct pte_list_desc *more;
125 * Stores number of entries stored in the pte_list_desc. No need to be
126 * u64 but just for easier alignment. When PTE_LIST_EXT, means full.
129 u64 *sptes[PTE_LIST_EXT];
132 struct kvm_shadow_walk_iterator {
140 #define for_each_shadow_entry_using_root(_vcpu, _root, _addr, _walker) \
141 for (shadow_walk_init_using_root(&(_walker), (_vcpu), \
143 shadow_walk_okay(&(_walker)); \
144 shadow_walk_next(&(_walker)))
146 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
147 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
148 shadow_walk_okay(&(_walker)); \
149 shadow_walk_next(&(_walker)))
151 #define for_each_shadow_entry_lockless(_vcpu, _addr, _walker, spte) \
152 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
153 shadow_walk_okay(&(_walker)) && \
154 ({ spte = mmu_spte_get_lockless(_walker.sptep); 1; }); \
155 __shadow_walk_next(&(_walker), spte))
157 static struct kmem_cache *pte_list_desc_cache;
158 struct kmem_cache *mmu_page_header_cache;
159 static struct percpu_counter kvm_total_used_mmu_pages;
161 static void mmu_spte_set(u64 *sptep, u64 spte);
163 struct kvm_mmu_role_regs {
164 const unsigned long cr0;
165 const unsigned long cr4;
169 #define CREATE_TRACE_POINTS
170 #include "mmutrace.h"
173 * Yes, lot's of underscores. They're a hint that you probably shouldn't be
174 * reading from the role_regs. Once the root_role is constructed, it becomes
175 * the single source of truth for the MMU's state.
177 #define BUILD_MMU_ROLE_REGS_ACCESSOR(reg, name, flag) \
178 static inline bool __maybe_unused \
179 ____is_##reg##_##name(const struct kvm_mmu_role_regs *regs) \
181 return !!(regs->reg & flag); \
183 BUILD_MMU_ROLE_REGS_ACCESSOR(cr0, pg, X86_CR0_PG);
184 BUILD_MMU_ROLE_REGS_ACCESSOR(cr0, wp, X86_CR0_WP);
185 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pse, X86_CR4_PSE);
186 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pae, X86_CR4_PAE);
187 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, smep, X86_CR4_SMEP);
188 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, smap, X86_CR4_SMAP);
189 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pke, X86_CR4_PKE);
190 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, la57, X86_CR4_LA57);
191 BUILD_MMU_ROLE_REGS_ACCESSOR(efer, nx, EFER_NX);
192 BUILD_MMU_ROLE_REGS_ACCESSOR(efer, lma, EFER_LMA);
195 * The MMU itself (with a valid role) is the single source of truth for the
196 * MMU. Do not use the regs used to build the MMU/role, nor the vCPU. The
197 * regs don't account for dependencies, e.g. clearing CR4 bits if CR0.PG=1,
198 * and the vCPU may be incorrect/irrelevant.
200 #define BUILD_MMU_ROLE_ACCESSOR(base_or_ext, reg, name) \
201 static inline bool __maybe_unused is_##reg##_##name(struct kvm_mmu *mmu) \
203 return !!(mmu->cpu_role. base_or_ext . reg##_##name); \
205 BUILD_MMU_ROLE_ACCESSOR(base, cr0, wp);
206 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, pse);
207 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, smep);
208 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, smap);
209 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, pke);
210 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, la57);
211 BUILD_MMU_ROLE_ACCESSOR(base, efer, nx);
212 BUILD_MMU_ROLE_ACCESSOR(ext, efer, lma);
214 static inline bool is_cr0_pg(struct kvm_mmu *mmu)
216 return mmu->cpu_role.base.level > 0;
219 static inline bool is_cr4_pae(struct kvm_mmu *mmu)
221 return !mmu->cpu_role.base.has_4_byte_gpte;
224 static struct kvm_mmu_role_regs vcpu_to_role_regs(struct kvm_vcpu *vcpu)
226 struct kvm_mmu_role_regs regs = {
227 .cr0 = kvm_read_cr0_bits(vcpu, KVM_MMU_CR0_ROLE_BITS),
228 .cr4 = kvm_read_cr4_bits(vcpu, KVM_MMU_CR4_ROLE_BITS),
229 .efer = vcpu->arch.efer,
235 static inline bool kvm_available_flush_tlb_with_range(void)
237 return kvm_x86_ops.tlb_remote_flush_with_range;
240 static void kvm_flush_remote_tlbs_with_range(struct kvm *kvm,
241 struct kvm_tlb_range *range)
245 if (range && kvm_x86_ops.tlb_remote_flush_with_range)
246 ret = static_call(kvm_x86_tlb_remote_flush_with_range)(kvm, range);
249 kvm_flush_remote_tlbs(kvm);
252 void kvm_flush_remote_tlbs_with_address(struct kvm *kvm,
253 u64 start_gfn, u64 pages)
255 struct kvm_tlb_range range;
257 range.start_gfn = start_gfn;
260 kvm_flush_remote_tlbs_with_range(kvm, &range);
263 static void mark_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, u64 gfn,
266 u64 spte = make_mmio_spte(vcpu, gfn, access);
268 trace_mark_mmio_spte(sptep, gfn, spte);
269 mmu_spte_set(sptep, spte);
272 static gfn_t get_mmio_spte_gfn(u64 spte)
274 u64 gpa = spte & shadow_nonpresent_or_rsvd_lower_gfn_mask;
276 gpa |= (spte >> SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)
277 & shadow_nonpresent_or_rsvd_mask;
279 return gpa >> PAGE_SHIFT;
282 static unsigned get_mmio_spte_access(u64 spte)
284 return spte & shadow_mmio_access_mask;
287 static bool check_mmio_spte(struct kvm_vcpu *vcpu, u64 spte)
289 u64 kvm_gen, spte_gen, gen;
291 gen = kvm_vcpu_memslots(vcpu)->generation;
292 if (unlikely(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS))
295 kvm_gen = gen & MMIO_SPTE_GEN_MASK;
296 spte_gen = get_mmio_spte_generation(spte);
298 trace_check_mmio_spte(spte, kvm_gen, spte_gen);
299 return likely(kvm_gen == spte_gen);
302 static int is_cpuid_PSE36(void)
308 static void __set_spte(u64 *sptep, u64 spte)
310 WRITE_ONCE(*sptep, spte);
313 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
315 WRITE_ONCE(*sptep, spte);
318 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
320 return xchg(sptep, spte);
323 static u64 __get_spte_lockless(u64 *sptep)
325 return READ_ONCE(*sptep);
336 static void count_spte_clear(u64 *sptep, u64 spte)
338 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
340 if (is_shadow_present_pte(spte))
343 /* Ensure the spte is completely set before we increase the count */
345 sp->clear_spte_count++;
348 static void __set_spte(u64 *sptep, u64 spte)
350 union split_spte *ssptep, sspte;
352 ssptep = (union split_spte *)sptep;
353 sspte = (union split_spte)spte;
355 ssptep->spte_high = sspte.spte_high;
358 * If we map the spte from nonpresent to present, We should store
359 * the high bits firstly, then set present bit, so cpu can not
360 * fetch this spte while we are setting the spte.
364 WRITE_ONCE(ssptep->spte_low, sspte.spte_low);
367 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
369 union split_spte *ssptep, sspte;
371 ssptep = (union split_spte *)sptep;
372 sspte = (union split_spte)spte;
374 WRITE_ONCE(ssptep->spte_low, sspte.spte_low);
377 * If we map the spte from present to nonpresent, we should clear
378 * present bit firstly to avoid vcpu fetch the old high bits.
382 ssptep->spte_high = sspte.spte_high;
383 count_spte_clear(sptep, spte);
386 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
388 union split_spte *ssptep, sspte, orig;
390 ssptep = (union split_spte *)sptep;
391 sspte = (union split_spte)spte;
393 /* xchg acts as a barrier before the setting of the high bits */
394 orig.spte_low = xchg(&ssptep->spte_low, sspte.spte_low);
395 orig.spte_high = ssptep->spte_high;
396 ssptep->spte_high = sspte.spte_high;
397 count_spte_clear(sptep, spte);
403 * The idea using the light way get the spte on x86_32 guest is from
404 * gup_get_pte (mm/gup.c).
406 * An spte tlb flush may be pending, because kvm_set_pte_rmap
407 * coalesces them and we are running out of the MMU lock. Therefore
408 * we need to protect against in-progress updates of the spte.
410 * Reading the spte while an update is in progress may get the old value
411 * for the high part of the spte. The race is fine for a present->non-present
412 * change (because the high part of the spte is ignored for non-present spte),
413 * but for a present->present change we must reread the spte.
415 * All such changes are done in two steps (present->non-present and
416 * non-present->present), hence it is enough to count the number of
417 * present->non-present updates: if it changed while reading the spte,
418 * we might have hit the race. This is done using clear_spte_count.
420 static u64 __get_spte_lockless(u64 *sptep)
422 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
423 union split_spte spte, *orig = (union split_spte *)sptep;
427 count = sp->clear_spte_count;
430 spte.spte_low = orig->spte_low;
433 spte.spte_high = orig->spte_high;
436 if (unlikely(spte.spte_low != orig->spte_low ||
437 count != sp->clear_spte_count))
444 /* Rules for using mmu_spte_set:
445 * Set the sptep from nonpresent to present.
446 * Note: the sptep being assigned *must* be either not present
447 * or in a state where the hardware will not attempt to update
450 static void mmu_spte_set(u64 *sptep, u64 new_spte)
452 WARN_ON(is_shadow_present_pte(*sptep));
453 __set_spte(sptep, new_spte);
457 * Update the SPTE (excluding the PFN), but do not track changes in its
458 * accessed/dirty status.
460 static u64 mmu_spte_update_no_track(u64 *sptep, u64 new_spte)
462 u64 old_spte = *sptep;
464 WARN_ON(!is_shadow_present_pte(new_spte));
465 check_spte_writable_invariants(new_spte);
467 if (!is_shadow_present_pte(old_spte)) {
468 mmu_spte_set(sptep, new_spte);
472 if (!spte_has_volatile_bits(old_spte))
473 __update_clear_spte_fast(sptep, new_spte);
475 old_spte = __update_clear_spte_slow(sptep, new_spte);
477 WARN_ON(spte_to_pfn(old_spte) != spte_to_pfn(new_spte));
482 /* Rules for using mmu_spte_update:
483 * Update the state bits, it means the mapped pfn is not changed.
485 * Whenever an MMU-writable SPTE is overwritten with a read-only SPTE, remote
486 * TLBs must be flushed. Otherwise rmap_write_protect will find a read-only
487 * spte, even though the writable spte might be cached on a CPU's TLB.
489 * Returns true if the TLB needs to be flushed
491 static bool mmu_spte_update(u64 *sptep, u64 new_spte)
494 u64 old_spte = mmu_spte_update_no_track(sptep, new_spte);
496 if (!is_shadow_present_pte(old_spte))
500 * For the spte updated out of mmu-lock is safe, since
501 * we always atomically update it, see the comments in
502 * spte_has_volatile_bits().
504 if (is_mmu_writable_spte(old_spte) &&
505 !is_writable_pte(new_spte))
509 * Flush TLB when accessed/dirty states are changed in the page tables,
510 * to guarantee consistency between TLB and page tables.
513 if (is_accessed_spte(old_spte) && !is_accessed_spte(new_spte)) {
515 kvm_set_pfn_accessed(spte_to_pfn(old_spte));
518 if (is_dirty_spte(old_spte) && !is_dirty_spte(new_spte)) {
520 kvm_set_pfn_dirty(spte_to_pfn(old_spte));
527 * Rules for using mmu_spte_clear_track_bits:
528 * It sets the sptep from present to nonpresent, and track the
529 * state bits, it is used to clear the last level sptep.
530 * Returns the old PTE.
532 static u64 mmu_spte_clear_track_bits(struct kvm *kvm, u64 *sptep)
535 u64 old_spte = *sptep;
536 int level = sptep_to_sp(sptep)->role.level;
539 if (!is_shadow_present_pte(old_spte) ||
540 !spte_has_volatile_bits(old_spte))
541 __update_clear_spte_fast(sptep, 0ull);
543 old_spte = __update_clear_spte_slow(sptep, 0ull);
545 if (!is_shadow_present_pte(old_spte))
548 kvm_update_page_stats(kvm, level, -1);
550 pfn = spte_to_pfn(old_spte);
553 * KVM doesn't hold a reference to any pages mapped into the guest, and
554 * instead uses the mmu_notifier to ensure that KVM unmaps any pages
555 * before they are reclaimed. Sanity check that, if the pfn is backed
556 * by a refcounted page, the refcount is elevated.
558 page = kvm_pfn_to_refcounted_page(pfn);
559 WARN_ON(page && !page_count(page));
561 if (is_accessed_spte(old_spte))
562 kvm_set_pfn_accessed(pfn);
564 if (is_dirty_spte(old_spte))
565 kvm_set_pfn_dirty(pfn);
571 * Rules for using mmu_spte_clear_no_track:
572 * Directly clear spte without caring the state bits of sptep,
573 * it is used to set the upper level spte.
575 static void mmu_spte_clear_no_track(u64 *sptep)
577 __update_clear_spte_fast(sptep, 0ull);
580 static u64 mmu_spte_get_lockless(u64 *sptep)
582 return __get_spte_lockless(sptep);
585 /* Returns the Accessed status of the PTE and resets it at the same time. */
586 static bool mmu_spte_age(u64 *sptep)
588 u64 spte = mmu_spte_get_lockless(sptep);
590 if (!is_accessed_spte(spte))
593 if (spte_ad_enabled(spte)) {
594 clear_bit((ffs(shadow_accessed_mask) - 1),
595 (unsigned long *)sptep);
598 * Capture the dirty status of the page, so that it doesn't get
599 * lost when the SPTE is marked for access tracking.
601 if (is_writable_pte(spte))
602 kvm_set_pfn_dirty(spte_to_pfn(spte));
604 spte = mark_spte_for_access_track(spte);
605 mmu_spte_update_no_track(sptep, spte);
611 static void walk_shadow_page_lockless_begin(struct kvm_vcpu *vcpu)
613 if (is_tdp_mmu(vcpu->arch.mmu)) {
614 kvm_tdp_mmu_walk_lockless_begin();
617 * Prevent page table teardown by making any free-er wait during
618 * kvm_flush_remote_tlbs() IPI to all active vcpus.
623 * Make sure a following spte read is not reordered ahead of the write
626 smp_store_mb(vcpu->mode, READING_SHADOW_PAGE_TABLES);
630 static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
632 if (is_tdp_mmu(vcpu->arch.mmu)) {
633 kvm_tdp_mmu_walk_lockless_end();
636 * Make sure the write to vcpu->mode is not reordered in front of
637 * reads to sptes. If it does, kvm_mmu_commit_zap_page() can see us
638 * OUTSIDE_GUEST_MODE and proceed to free the shadow page table.
640 smp_store_release(&vcpu->mode, OUTSIDE_GUEST_MODE);
645 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu, bool maybe_indirect)
649 /* 1 rmap, 1 parent PTE per level, and the prefetched rmaps. */
650 r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
651 1 + PT64_ROOT_MAX_LEVEL + PTE_PREFETCH_NUM);
654 r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_shadow_page_cache,
655 PT64_ROOT_MAX_LEVEL);
658 if (maybe_indirect) {
659 r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_shadowed_info_cache,
660 PT64_ROOT_MAX_LEVEL);
664 return kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
665 PT64_ROOT_MAX_LEVEL);
668 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
670 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache);
671 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_shadow_page_cache);
672 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_shadowed_info_cache);
673 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
676 static void mmu_free_pte_list_desc(struct pte_list_desc *pte_list_desc)
678 kmem_cache_free(pte_list_desc_cache, pte_list_desc);
681 static bool sp_has_gptes(struct kvm_mmu_page *sp);
683 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
685 if (sp->role.passthrough)
688 if (!sp->role.direct)
689 return sp->shadowed_translation[index] >> PAGE_SHIFT;
691 return sp->gfn + (index << ((sp->role.level - 1) * SPTE_LEVEL_BITS));
695 * For leaf SPTEs, fetch the *guest* access permissions being shadowed. Note
696 * that the SPTE itself may have a more constrained access permissions that
697 * what the guest enforces. For example, a guest may create an executable
698 * huge PTE but KVM may disallow execution to mitigate iTLB multihit.
700 static u32 kvm_mmu_page_get_access(struct kvm_mmu_page *sp, int index)
702 if (sp_has_gptes(sp))
703 return sp->shadowed_translation[index] & ACC_ALL;
706 * For direct MMUs (e.g. TDP or non-paging guests) or passthrough SPs,
707 * KVM is not shadowing any guest page tables, so the "guest access
708 * permissions" are just ACC_ALL.
710 * For direct SPs in indirect MMUs (shadow paging), i.e. when KVM
711 * is shadowing a guest huge page with small pages, the guest access
712 * permissions being shadowed are the access permissions of the huge
715 * In both cases, sp->role.access contains the correct access bits.
717 return sp->role.access;
720 static void kvm_mmu_page_set_translation(struct kvm_mmu_page *sp, int index,
721 gfn_t gfn, unsigned int access)
723 if (sp_has_gptes(sp)) {
724 sp->shadowed_translation[index] = (gfn << PAGE_SHIFT) | access;
728 WARN_ONCE(access != kvm_mmu_page_get_access(sp, index),
729 "access mismatch under %s page %llx (expected %u, got %u)\n",
730 sp->role.passthrough ? "passthrough" : "direct",
731 sp->gfn, kvm_mmu_page_get_access(sp, index), access);
733 WARN_ONCE(gfn != kvm_mmu_page_get_gfn(sp, index),
734 "gfn mismatch under %s page %llx (expected %llx, got %llx)\n",
735 sp->role.passthrough ? "passthrough" : "direct",
736 sp->gfn, kvm_mmu_page_get_gfn(sp, index), gfn);
739 static void kvm_mmu_page_set_access(struct kvm_mmu_page *sp, int index,
742 gfn_t gfn = kvm_mmu_page_get_gfn(sp, index);
744 kvm_mmu_page_set_translation(sp, index, gfn, access);
748 * Return the pointer to the large page information for a given gfn,
749 * handling slots that are not large page aligned.
751 static struct kvm_lpage_info *lpage_info_slot(gfn_t gfn,
752 const struct kvm_memory_slot *slot, int level)
756 idx = gfn_to_index(gfn, slot->base_gfn, level);
757 return &slot->arch.lpage_info[level - 2][idx];
760 static void update_gfn_disallow_lpage_count(const struct kvm_memory_slot *slot,
761 gfn_t gfn, int count)
763 struct kvm_lpage_info *linfo;
766 for (i = PG_LEVEL_2M; i <= KVM_MAX_HUGEPAGE_LEVEL; ++i) {
767 linfo = lpage_info_slot(gfn, slot, i);
768 linfo->disallow_lpage += count;
769 WARN_ON(linfo->disallow_lpage < 0);
773 void kvm_mmu_gfn_disallow_lpage(const struct kvm_memory_slot *slot, gfn_t gfn)
775 update_gfn_disallow_lpage_count(slot, gfn, 1);
778 void kvm_mmu_gfn_allow_lpage(const struct kvm_memory_slot *slot, gfn_t gfn)
780 update_gfn_disallow_lpage_count(slot, gfn, -1);
783 static void account_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
785 struct kvm_memslots *slots;
786 struct kvm_memory_slot *slot;
789 kvm->arch.indirect_shadow_pages++;
791 slots = kvm_memslots_for_spte_role(kvm, sp->role);
792 slot = __gfn_to_memslot(slots, gfn);
794 /* the non-leaf shadow pages are keeping readonly. */
795 if (sp->role.level > PG_LEVEL_4K)
796 return kvm_slot_page_track_add_page(kvm, slot, gfn,
797 KVM_PAGE_TRACK_WRITE);
799 kvm_mmu_gfn_disallow_lpage(slot, gfn);
801 if (kvm_mmu_slot_gfn_write_protect(kvm, slot, gfn, PG_LEVEL_4K))
802 kvm_flush_remote_tlbs_with_address(kvm, gfn, 1);
805 void account_huge_nx_page(struct kvm *kvm, struct kvm_mmu_page *sp)
807 if (sp->lpage_disallowed)
810 ++kvm->stat.nx_lpage_splits;
811 list_add_tail(&sp->lpage_disallowed_link,
812 &kvm->arch.lpage_disallowed_mmu_pages);
813 sp->lpage_disallowed = true;
816 static void unaccount_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
818 struct kvm_memslots *slots;
819 struct kvm_memory_slot *slot;
822 kvm->arch.indirect_shadow_pages--;
824 slots = kvm_memslots_for_spte_role(kvm, sp->role);
825 slot = __gfn_to_memslot(slots, gfn);
826 if (sp->role.level > PG_LEVEL_4K)
827 return kvm_slot_page_track_remove_page(kvm, slot, gfn,
828 KVM_PAGE_TRACK_WRITE);
830 kvm_mmu_gfn_allow_lpage(slot, gfn);
833 void unaccount_huge_nx_page(struct kvm *kvm, struct kvm_mmu_page *sp)
835 --kvm->stat.nx_lpage_splits;
836 sp->lpage_disallowed = false;
837 list_del(&sp->lpage_disallowed_link);
840 static struct kvm_memory_slot *
841 gfn_to_memslot_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t gfn,
844 struct kvm_memory_slot *slot;
846 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
847 if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
849 if (no_dirty_log && kvm_slot_dirty_track_enabled(slot))
856 * About rmap_head encoding:
858 * If the bit zero of rmap_head->val is clear, then it points to the only spte
859 * in this rmap chain. Otherwise, (rmap_head->val & ~1) points to a struct
860 * pte_list_desc containing more mappings.
864 * Returns the number of pointers in the rmap chain, not counting the new one.
866 static int pte_list_add(struct kvm_mmu_memory_cache *cache, u64 *spte,
867 struct kvm_rmap_head *rmap_head)
869 struct pte_list_desc *desc;
872 if (!rmap_head->val) {
873 rmap_printk("%p %llx 0->1\n", spte, *spte);
874 rmap_head->val = (unsigned long)spte;
875 } else if (!(rmap_head->val & 1)) {
876 rmap_printk("%p %llx 1->many\n", spte, *spte);
877 desc = kvm_mmu_memory_cache_alloc(cache);
878 desc->sptes[0] = (u64 *)rmap_head->val;
879 desc->sptes[1] = spte;
880 desc->spte_count = 2;
881 rmap_head->val = (unsigned long)desc | 1;
884 rmap_printk("%p %llx many->many\n", spte, *spte);
885 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
886 while (desc->spte_count == PTE_LIST_EXT) {
887 count += PTE_LIST_EXT;
889 desc->more = kvm_mmu_memory_cache_alloc(cache);
891 desc->spte_count = 0;
896 count += desc->spte_count;
897 desc->sptes[desc->spte_count++] = spte;
903 pte_list_desc_remove_entry(struct kvm_rmap_head *rmap_head,
904 struct pte_list_desc *desc, int i,
905 struct pte_list_desc *prev_desc)
907 int j = desc->spte_count - 1;
909 desc->sptes[i] = desc->sptes[j];
910 desc->sptes[j] = NULL;
912 if (desc->spte_count)
914 if (!prev_desc && !desc->more)
918 prev_desc->more = desc->more;
920 rmap_head->val = (unsigned long)desc->more | 1;
921 mmu_free_pte_list_desc(desc);
924 static void pte_list_remove(u64 *spte, struct kvm_rmap_head *rmap_head)
926 struct pte_list_desc *desc;
927 struct pte_list_desc *prev_desc;
930 if (!rmap_head->val) {
931 pr_err("%s: %p 0->BUG\n", __func__, spte);
933 } else if (!(rmap_head->val & 1)) {
934 rmap_printk("%p 1->0\n", spte);
935 if ((u64 *)rmap_head->val != spte) {
936 pr_err("%s: %p 1->BUG\n", __func__, spte);
941 rmap_printk("%p many->many\n", spte);
942 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
945 for (i = 0; i < desc->spte_count; ++i) {
946 if (desc->sptes[i] == spte) {
947 pte_list_desc_remove_entry(rmap_head,
955 pr_err("%s: %p many->many\n", __func__, spte);
960 static void kvm_zap_one_rmap_spte(struct kvm *kvm,
961 struct kvm_rmap_head *rmap_head, u64 *sptep)
963 mmu_spte_clear_track_bits(kvm, sptep);
964 pte_list_remove(sptep, rmap_head);
967 /* Return true if at least one SPTE was zapped, false otherwise */
968 static bool kvm_zap_all_rmap_sptes(struct kvm *kvm,
969 struct kvm_rmap_head *rmap_head)
971 struct pte_list_desc *desc, *next;
977 if (!(rmap_head->val & 1)) {
978 mmu_spte_clear_track_bits(kvm, (u64 *)rmap_head->val);
982 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
984 for (; desc; desc = next) {
985 for (i = 0; i < desc->spte_count; i++)
986 mmu_spte_clear_track_bits(kvm, desc->sptes[i]);
988 mmu_free_pte_list_desc(desc);
991 /* rmap_head is meaningless now, remember to reset it */
996 unsigned int pte_list_count(struct kvm_rmap_head *rmap_head)
998 struct pte_list_desc *desc;
999 unsigned int count = 0;
1001 if (!rmap_head->val)
1003 else if (!(rmap_head->val & 1))
1006 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
1009 count += desc->spte_count;
1016 static struct kvm_rmap_head *gfn_to_rmap(gfn_t gfn, int level,
1017 const struct kvm_memory_slot *slot)
1021 idx = gfn_to_index(gfn, slot->base_gfn, level);
1022 return &slot->arch.rmap[level - PG_LEVEL_4K][idx];
1025 static bool rmap_can_add(struct kvm_vcpu *vcpu)
1027 struct kvm_mmu_memory_cache *mc;
1029 mc = &vcpu->arch.mmu_pte_list_desc_cache;
1030 return kvm_mmu_memory_cache_nr_free_objects(mc);
1033 static void rmap_remove(struct kvm *kvm, u64 *spte)
1035 struct kvm_memslots *slots;
1036 struct kvm_memory_slot *slot;
1037 struct kvm_mmu_page *sp;
1039 struct kvm_rmap_head *rmap_head;
1041 sp = sptep_to_sp(spte);
1042 gfn = kvm_mmu_page_get_gfn(sp, spte_index(spte));
1045 * Unlike rmap_add, rmap_remove does not run in the context of a vCPU
1046 * so we have to determine which memslots to use based on context
1047 * information in sp->role.
1049 slots = kvm_memslots_for_spte_role(kvm, sp->role);
1051 slot = __gfn_to_memslot(slots, gfn);
1052 rmap_head = gfn_to_rmap(gfn, sp->role.level, slot);
1054 pte_list_remove(spte, rmap_head);
1058 * Used by the following functions to iterate through the sptes linked by a
1059 * rmap. All fields are private and not assumed to be used outside.
1061 struct rmap_iterator {
1062 /* private fields */
1063 struct pte_list_desc *desc; /* holds the sptep if not NULL */
1064 int pos; /* index of the sptep */
1068 * Iteration must be started by this function. This should also be used after
1069 * removing/dropping sptes from the rmap link because in such cases the
1070 * information in the iterator may not be valid.
1072 * Returns sptep if found, NULL otherwise.
1074 static u64 *rmap_get_first(struct kvm_rmap_head *rmap_head,
1075 struct rmap_iterator *iter)
1079 if (!rmap_head->val)
1082 if (!(rmap_head->val & 1)) {
1084 sptep = (u64 *)rmap_head->val;
1088 iter->desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
1090 sptep = iter->desc->sptes[iter->pos];
1092 BUG_ON(!is_shadow_present_pte(*sptep));
1097 * Must be used with a valid iterator: e.g. after rmap_get_first().
1099 * Returns sptep if found, NULL otherwise.
1101 static u64 *rmap_get_next(struct rmap_iterator *iter)
1106 if (iter->pos < PTE_LIST_EXT - 1) {
1108 sptep = iter->desc->sptes[iter->pos];
1113 iter->desc = iter->desc->more;
1117 /* desc->sptes[0] cannot be NULL */
1118 sptep = iter->desc->sptes[iter->pos];
1125 BUG_ON(!is_shadow_present_pte(*sptep));
1129 #define for_each_rmap_spte(_rmap_head_, _iter_, _spte_) \
1130 for (_spte_ = rmap_get_first(_rmap_head_, _iter_); \
1131 _spte_; _spte_ = rmap_get_next(_iter_))
1133 static void drop_spte(struct kvm *kvm, u64 *sptep)
1135 u64 old_spte = mmu_spte_clear_track_bits(kvm, sptep);
1137 if (is_shadow_present_pte(old_spte))
1138 rmap_remove(kvm, sptep);
1141 static void drop_large_spte(struct kvm *kvm, u64 *sptep, bool flush)
1143 struct kvm_mmu_page *sp;
1145 sp = sptep_to_sp(sptep);
1146 WARN_ON(sp->role.level == PG_LEVEL_4K);
1148 drop_spte(kvm, sptep);
1151 kvm_flush_remote_tlbs_with_address(kvm, sp->gfn,
1152 KVM_PAGES_PER_HPAGE(sp->role.level));
1156 * Write-protect on the specified @sptep, @pt_protect indicates whether
1157 * spte write-protection is caused by protecting shadow page table.
1159 * Note: write protection is difference between dirty logging and spte
1161 * - for dirty logging, the spte can be set to writable at anytime if
1162 * its dirty bitmap is properly set.
1163 * - for spte protection, the spte can be writable only after unsync-ing
1166 * Return true if tlb need be flushed.
1168 static bool spte_write_protect(u64 *sptep, bool pt_protect)
1172 if (!is_writable_pte(spte) &&
1173 !(pt_protect && is_mmu_writable_spte(spte)))
1176 rmap_printk("spte %p %llx\n", sptep, *sptep);
1179 spte &= ~shadow_mmu_writable_mask;
1180 spte = spte & ~PT_WRITABLE_MASK;
1182 return mmu_spte_update(sptep, spte);
1185 static bool rmap_write_protect(struct kvm_rmap_head *rmap_head,
1189 struct rmap_iterator iter;
1192 for_each_rmap_spte(rmap_head, &iter, sptep)
1193 flush |= spte_write_protect(sptep, pt_protect);
1198 static bool spte_clear_dirty(u64 *sptep)
1202 rmap_printk("spte %p %llx\n", sptep, *sptep);
1204 MMU_WARN_ON(!spte_ad_enabled(spte));
1205 spte &= ~shadow_dirty_mask;
1206 return mmu_spte_update(sptep, spte);
1209 static bool spte_wrprot_for_clear_dirty(u64 *sptep)
1211 bool was_writable = test_and_clear_bit(PT_WRITABLE_SHIFT,
1212 (unsigned long *)sptep);
1213 if (was_writable && !spte_ad_enabled(*sptep))
1214 kvm_set_pfn_dirty(spte_to_pfn(*sptep));
1216 return was_writable;
1220 * Gets the GFN ready for another round of dirty logging by clearing the
1221 * - D bit on ad-enabled SPTEs, and
1222 * - W bit on ad-disabled SPTEs.
1223 * Returns true iff any D or W bits were cleared.
1225 static bool __rmap_clear_dirty(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1226 const struct kvm_memory_slot *slot)
1229 struct rmap_iterator iter;
1232 for_each_rmap_spte(rmap_head, &iter, sptep)
1233 if (spte_ad_need_write_protect(*sptep))
1234 flush |= spte_wrprot_for_clear_dirty(sptep);
1236 flush |= spte_clear_dirty(sptep);
1242 * kvm_mmu_write_protect_pt_masked - write protect selected PT level pages
1243 * @kvm: kvm instance
1244 * @slot: slot to protect
1245 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1246 * @mask: indicates which pages we should protect
1248 * Used when we do not need to care about huge page mappings.
1250 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1251 struct kvm_memory_slot *slot,
1252 gfn_t gfn_offset, unsigned long mask)
1254 struct kvm_rmap_head *rmap_head;
1256 if (is_tdp_mmu_enabled(kvm))
1257 kvm_tdp_mmu_clear_dirty_pt_masked(kvm, slot,
1258 slot->base_gfn + gfn_offset, mask, true);
1260 if (!kvm_memslots_have_rmaps(kvm))
1264 rmap_head = gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1266 rmap_write_protect(rmap_head, false);
1268 /* clear the first set bit */
1274 * kvm_mmu_clear_dirty_pt_masked - clear MMU D-bit for PT level pages, or write
1275 * protect the page if the D-bit isn't supported.
1276 * @kvm: kvm instance
1277 * @slot: slot to clear D-bit
1278 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1279 * @mask: indicates which pages we should clear D-bit
1281 * Used for PML to re-log the dirty GPAs after userspace querying dirty_bitmap.
1283 static void kvm_mmu_clear_dirty_pt_masked(struct kvm *kvm,
1284 struct kvm_memory_slot *slot,
1285 gfn_t gfn_offset, unsigned long mask)
1287 struct kvm_rmap_head *rmap_head;
1289 if (is_tdp_mmu_enabled(kvm))
1290 kvm_tdp_mmu_clear_dirty_pt_masked(kvm, slot,
1291 slot->base_gfn + gfn_offset, mask, false);
1293 if (!kvm_memslots_have_rmaps(kvm))
1297 rmap_head = gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1299 __rmap_clear_dirty(kvm, rmap_head, slot);
1301 /* clear the first set bit */
1307 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1310 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1311 * enable dirty logging for them.
1313 * We need to care about huge page mappings: e.g. during dirty logging we may
1314 * have such mappings.
1316 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1317 struct kvm_memory_slot *slot,
1318 gfn_t gfn_offset, unsigned long mask)
1321 * Huge pages are NOT write protected when we start dirty logging in
1322 * initially-all-set mode; must write protect them here so that they
1323 * are split to 4K on the first write.
1325 * The gfn_offset is guaranteed to be aligned to 64, but the base_gfn
1326 * of memslot has no such restriction, so the range can cross two large
1329 if (kvm_dirty_log_manual_protect_and_init_set(kvm)) {
1330 gfn_t start = slot->base_gfn + gfn_offset + __ffs(mask);
1331 gfn_t end = slot->base_gfn + gfn_offset + __fls(mask);
1333 if (READ_ONCE(eager_page_split))
1334 kvm_mmu_try_split_huge_pages(kvm, slot, start, end, PG_LEVEL_4K);
1336 kvm_mmu_slot_gfn_write_protect(kvm, slot, start, PG_LEVEL_2M);
1338 /* Cross two large pages? */
1339 if (ALIGN(start << PAGE_SHIFT, PMD_SIZE) !=
1340 ALIGN(end << PAGE_SHIFT, PMD_SIZE))
1341 kvm_mmu_slot_gfn_write_protect(kvm, slot, end,
1345 /* Now handle 4K PTEs. */
1346 if (kvm_x86_ops.cpu_dirty_log_size)
1347 kvm_mmu_clear_dirty_pt_masked(kvm, slot, gfn_offset, mask);
1349 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1352 int kvm_cpu_dirty_log_size(void)
1354 return kvm_x86_ops.cpu_dirty_log_size;
1357 bool kvm_mmu_slot_gfn_write_protect(struct kvm *kvm,
1358 struct kvm_memory_slot *slot, u64 gfn,
1361 struct kvm_rmap_head *rmap_head;
1363 bool write_protected = false;
1365 if (kvm_memslots_have_rmaps(kvm)) {
1366 for (i = min_level; i <= KVM_MAX_HUGEPAGE_LEVEL; ++i) {
1367 rmap_head = gfn_to_rmap(gfn, i, slot);
1368 write_protected |= rmap_write_protect(rmap_head, true);
1372 if (is_tdp_mmu_enabled(kvm))
1374 kvm_tdp_mmu_write_protect_gfn(kvm, slot, gfn, min_level);
1376 return write_protected;
1379 static bool kvm_vcpu_write_protect_gfn(struct kvm_vcpu *vcpu, u64 gfn)
1381 struct kvm_memory_slot *slot;
1383 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1384 return kvm_mmu_slot_gfn_write_protect(vcpu->kvm, slot, gfn, PG_LEVEL_4K);
1387 static bool __kvm_zap_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1388 const struct kvm_memory_slot *slot)
1390 return kvm_zap_all_rmap_sptes(kvm, rmap_head);
1393 static bool kvm_zap_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1394 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1397 return __kvm_zap_rmap(kvm, rmap_head, slot);
1400 static bool kvm_set_pte_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1401 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1405 struct rmap_iterator iter;
1406 bool need_flush = false;
1410 WARN_ON(pte_huge(pte));
1411 new_pfn = pte_pfn(pte);
1414 for_each_rmap_spte(rmap_head, &iter, sptep) {
1415 rmap_printk("spte %p %llx gfn %llx (%d)\n",
1416 sptep, *sptep, gfn, level);
1420 if (pte_write(pte)) {
1421 kvm_zap_one_rmap_spte(kvm, rmap_head, sptep);
1424 new_spte = kvm_mmu_changed_pte_notifier_make_spte(
1427 mmu_spte_clear_track_bits(kvm, sptep);
1428 mmu_spte_set(sptep, new_spte);
1432 if (need_flush && kvm_available_flush_tlb_with_range()) {
1433 kvm_flush_remote_tlbs_with_address(kvm, gfn, 1);
1440 struct slot_rmap_walk_iterator {
1442 const struct kvm_memory_slot *slot;
1448 /* output fields. */
1450 struct kvm_rmap_head *rmap;
1453 /* private field. */
1454 struct kvm_rmap_head *end_rmap;
1458 rmap_walk_init_level(struct slot_rmap_walk_iterator *iterator, int level)
1460 iterator->level = level;
1461 iterator->gfn = iterator->start_gfn;
1462 iterator->rmap = gfn_to_rmap(iterator->gfn, level, iterator->slot);
1463 iterator->end_rmap = gfn_to_rmap(iterator->end_gfn, level, iterator->slot);
1467 slot_rmap_walk_init(struct slot_rmap_walk_iterator *iterator,
1468 const struct kvm_memory_slot *slot, int start_level,
1469 int end_level, gfn_t start_gfn, gfn_t end_gfn)
1471 iterator->slot = slot;
1472 iterator->start_level = start_level;
1473 iterator->end_level = end_level;
1474 iterator->start_gfn = start_gfn;
1475 iterator->end_gfn = end_gfn;
1477 rmap_walk_init_level(iterator, iterator->start_level);
1480 static bool slot_rmap_walk_okay(struct slot_rmap_walk_iterator *iterator)
1482 return !!iterator->rmap;
1485 static void slot_rmap_walk_next(struct slot_rmap_walk_iterator *iterator)
1487 while (++iterator->rmap <= iterator->end_rmap) {
1488 iterator->gfn += (1UL << KVM_HPAGE_GFN_SHIFT(iterator->level));
1490 if (iterator->rmap->val)
1494 if (++iterator->level > iterator->end_level) {
1495 iterator->rmap = NULL;
1499 rmap_walk_init_level(iterator, iterator->level);
1502 #define for_each_slot_rmap_range(_slot_, _start_level_, _end_level_, \
1503 _start_gfn, _end_gfn, _iter_) \
1504 for (slot_rmap_walk_init(_iter_, _slot_, _start_level_, \
1505 _end_level_, _start_gfn, _end_gfn); \
1506 slot_rmap_walk_okay(_iter_); \
1507 slot_rmap_walk_next(_iter_))
1509 typedef bool (*rmap_handler_t)(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1510 struct kvm_memory_slot *slot, gfn_t gfn,
1511 int level, pte_t pte);
1513 static __always_inline bool kvm_handle_gfn_range(struct kvm *kvm,
1514 struct kvm_gfn_range *range,
1515 rmap_handler_t handler)
1517 struct slot_rmap_walk_iterator iterator;
1520 for_each_slot_rmap_range(range->slot, PG_LEVEL_4K, KVM_MAX_HUGEPAGE_LEVEL,
1521 range->start, range->end - 1, &iterator)
1522 ret |= handler(kvm, iterator.rmap, range->slot, iterator.gfn,
1523 iterator.level, range->pte);
1528 bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
1532 if (kvm_memslots_have_rmaps(kvm))
1533 flush = kvm_handle_gfn_range(kvm, range, kvm_zap_rmap);
1535 if (is_tdp_mmu_enabled(kvm))
1536 flush = kvm_tdp_mmu_unmap_gfn_range(kvm, range, flush);
1541 bool kvm_set_spte_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1545 if (kvm_memslots_have_rmaps(kvm))
1546 flush = kvm_handle_gfn_range(kvm, range, kvm_set_pte_rmap);
1548 if (is_tdp_mmu_enabled(kvm))
1549 flush |= kvm_tdp_mmu_set_spte_gfn(kvm, range);
1554 static bool kvm_age_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1555 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1559 struct rmap_iterator iter;
1562 for_each_rmap_spte(rmap_head, &iter, sptep)
1563 young |= mmu_spte_age(sptep);
1568 static bool kvm_test_age_rmap(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1569 struct kvm_memory_slot *slot, gfn_t gfn,
1570 int level, pte_t unused)
1573 struct rmap_iterator iter;
1575 for_each_rmap_spte(rmap_head, &iter, sptep)
1576 if (is_accessed_spte(*sptep))
1581 #define RMAP_RECYCLE_THRESHOLD 1000
1583 static void __rmap_add(struct kvm *kvm,
1584 struct kvm_mmu_memory_cache *cache,
1585 const struct kvm_memory_slot *slot,
1586 u64 *spte, gfn_t gfn, unsigned int access)
1588 struct kvm_mmu_page *sp;
1589 struct kvm_rmap_head *rmap_head;
1592 sp = sptep_to_sp(spte);
1593 kvm_mmu_page_set_translation(sp, spte_index(spte), gfn, access);
1594 kvm_update_page_stats(kvm, sp->role.level, 1);
1596 rmap_head = gfn_to_rmap(gfn, sp->role.level, slot);
1597 rmap_count = pte_list_add(cache, spte, rmap_head);
1599 if (rmap_count > RMAP_RECYCLE_THRESHOLD) {
1600 kvm_zap_all_rmap_sptes(kvm, rmap_head);
1601 kvm_flush_remote_tlbs_with_address(
1602 kvm, sp->gfn, KVM_PAGES_PER_HPAGE(sp->role.level));
1606 static void rmap_add(struct kvm_vcpu *vcpu, const struct kvm_memory_slot *slot,
1607 u64 *spte, gfn_t gfn, unsigned int access)
1609 struct kvm_mmu_memory_cache *cache = &vcpu->arch.mmu_pte_list_desc_cache;
1611 __rmap_add(vcpu->kvm, cache, slot, spte, gfn, access);
1614 bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1618 if (kvm_memslots_have_rmaps(kvm))
1619 young = kvm_handle_gfn_range(kvm, range, kvm_age_rmap);
1621 if (is_tdp_mmu_enabled(kvm))
1622 young |= kvm_tdp_mmu_age_gfn_range(kvm, range);
1627 bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1631 if (kvm_memslots_have_rmaps(kvm))
1632 young = kvm_handle_gfn_range(kvm, range, kvm_test_age_rmap);
1634 if (is_tdp_mmu_enabled(kvm))
1635 young |= kvm_tdp_mmu_test_age_gfn(kvm, range);
1641 static int is_empty_shadow_page(u64 *spt)
1646 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
1647 if (is_shadow_present_pte(*pos)) {
1648 printk(KERN_ERR "%s: %p %llx\n", __func__,
1657 * This value is the sum of all of the kvm instances's
1658 * kvm->arch.n_used_mmu_pages values. We need a global,
1659 * aggregate version in order to make the slab shrinker
1662 static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, long nr)
1664 kvm->arch.n_used_mmu_pages += nr;
1665 percpu_counter_add(&kvm_total_used_mmu_pages, nr);
1668 static void kvm_mmu_free_shadow_page(struct kvm_mmu_page *sp)
1670 MMU_WARN_ON(!is_empty_shadow_page(sp->spt));
1671 hlist_del(&sp->hash_link);
1672 list_del(&sp->link);
1673 free_page((unsigned long)sp->spt);
1674 if (!sp->role.direct)
1675 free_page((unsigned long)sp->shadowed_translation);
1676 kmem_cache_free(mmu_page_header_cache, sp);
1679 static unsigned kvm_page_table_hashfn(gfn_t gfn)
1681 return hash_64(gfn, KVM_MMU_HASH_SHIFT);
1684 static void mmu_page_add_parent_pte(struct kvm_mmu_memory_cache *cache,
1685 struct kvm_mmu_page *sp, u64 *parent_pte)
1690 pte_list_add(cache, parent_pte, &sp->parent_ptes);
1693 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
1696 pte_list_remove(parent_pte, &sp->parent_ptes);
1699 static void drop_parent_pte(struct kvm_mmu_page *sp,
1702 mmu_page_remove_parent_pte(sp, parent_pte);
1703 mmu_spte_clear_no_track(parent_pte);
1706 static void mark_unsync(u64 *spte);
1707 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1710 struct rmap_iterator iter;
1712 for_each_rmap_spte(&sp->parent_ptes, &iter, sptep) {
1717 static void mark_unsync(u64 *spte)
1719 struct kvm_mmu_page *sp;
1721 sp = sptep_to_sp(spte);
1722 if (__test_and_set_bit(spte_index(spte), sp->unsync_child_bitmap))
1724 if (sp->unsync_children++)
1726 kvm_mmu_mark_parents_unsync(sp);
1729 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1730 struct kvm_mmu_page *sp)
1735 #define KVM_PAGE_ARRAY_NR 16
1737 struct kvm_mmu_pages {
1738 struct mmu_page_and_offset {
1739 struct kvm_mmu_page *sp;
1741 } page[KVM_PAGE_ARRAY_NR];
1745 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1751 for (i=0; i < pvec->nr; i++)
1752 if (pvec->page[i].sp == sp)
1755 pvec->page[pvec->nr].sp = sp;
1756 pvec->page[pvec->nr].idx = idx;
1758 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1761 static inline void clear_unsync_child_bit(struct kvm_mmu_page *sp, int idx)
1763 --sp->unsync_children;
1764 WARN_ON((int)sp->unsync_children < 0);
1765 __clear_bit(idx, sp->unsync_child_bitmap);
1768 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1769 struct kvm_mmu_pages *pvec)
1771 int i, ret, nr_unsync_leaf = 0;
1773 for_each_set_bit(i, sp->unsync_child_bitmap, 512) {
1774 struct kvm_mmu_page *child;
1775 u64 ent = sp->spt[i];
1777 if (!is_shadow_present_pte(ent) || is_large_pte(ent)) {
1778 clear_unsync_child_bit(sp, i);
1782 child = to_shadow_page(ent & SPTE_BASE_ADDR_MASK);
1784 if (child->unsync_children) {
1785 if (mmu_pages_add(pvec, child, i))
1788 ret = __mmu_unsync_walk(child, pvec);
1790 clear_unsync_child_bit(sp, i);
1792 } else if (ret > 0) {
1793 nr_unsync_leaf += ret;
1796 } else if (child->unsync) {
1798 if (mmu_pages_add(pvec, child, i))
1801 clear_unsync_child_bit(sp, i);
1804 return nr_unsync_leaf;
1807 #define INVALID_INDEX (-1)
1809 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1810 struct kvm_mmu_pages *pvec)
1813 if (!sp->unsync_children)
1816 mmu_pages_add(pvec, sp, INVALID_INDEX);
1817 return __mmu_unsync_walk(sp, pvec);
1820 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1822 WARN_ON(!sp->unsync);
1823 trace_kvm_mmu_sync_page(sp);
1825 --kvm->stat.mmu_unsync;
1828 static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1829 struct list_head *invalid_list);
1830 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1831 struct list_head *invalid_list);
1833 static bool sp_has_gptes(struct kvm_mmu_page *sp)
1835 if (sp->role.direct)
1838 if (sp->role.passthrough)
1844 #define for_each_valid_sp(_kvm, _sp, _list) \
1845 hlist_for_each_entry(_sp, _list, hash_link) \
1846 if (is_obsolete_sp((_kvm), (_sp))) { \
1849 #define for_each_gfn_valid_sp_with_gptes(_kvm, _sp, _gfn) \
1850 for_each_valid_sp(_kvm, _sp, \
1851 &(_kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(_gfn)]) \
1852 if ((_sp)->gfn != (_gfn) || !sp_has_gptes(_sp)) {} else
1854 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1855 struct list_head *invalid_list)
1857 int ret = vcpu->arch.mmu->sync_page(vcpu, sp);
1860 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1864 static bool kvm_mmu_remote_flush_or_zap(struct kvm *kvm,
1865 struct list_head *invalid_list,
1868 if (!remote_flush && list_empty(invalid_list))
1871 if (!list_empty(invalid_list))
1872 kvm_mmu_commit_zap_page(kvm, invalid_list);
1874 kvm_flush_remote_tlbs(kvm);
1878 static bool is_obsolete_sp(struct kvm *kvm, struct kvm_mmu_page *sp)
1880 if (sp->role.invalid)
1883 /* TDP MMU pages due not use the MMU generation. */
1884 return !sp->tdp_mmu_page &&
1885 unlikely(sp->mmu_valid_gen != kvm->arch.mmu_valid_gen);
1888 struct mmu_page_path {
1889 struct kvm_mmu_page *parent[PT64_ROOT_MAX_LEVEL];
1890 unsigned int idx[PT64_ROOT_MAX_LEVEL];
1893 #define for_each_sp(pvec, sp, parents, i) \
1894 for (i = mmu_pages_first(&pvec, &parents); \
1895 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1896 i = mmu_pages_next(&pvec, &parents, i))
1898 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1899 struct mmu_page_path *parents,
1904 for (n = i+1; n < pvec->nr; n++) {
1905 struct kvm_mmu_page *sp = pvec->page[n].sp;
1906 unsigned idx = pvec->page[n].idx;
1907 int level = sp->role.level;
1909 parents->idx[level-1] = idx;
1910 if (level == PG_LEVEL_4K)
1913 parents->parent[level-2] = sp;
1919 static int mmu_pages_first(struct kvm_mmu_pages *pvec,
1920 struct mmu_page_path *parents)
1922 struct kvm_mmu_page *sp;
1928 WARN_ON(pvec->page[0].idx != INVALID_INDEX);
1930 sp = pvec->page[0].sp;
1931 level = sp->role.level;
1932 WARN_ON(level == PG_LEVEL_4K);
1934 parents->parent[level-2] = sp;
1936 /* Also set up a sentinel. Further entries in pvec are all
1937 * children of sp, so this element is never overwritten.
1939 parents->parent[level-1] = NULL;
1940 return mmu_pages_next(pvec, parents, 0);
1943 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1945 struct kvm_mmu_page *sp;
1946 unsigned int level = 0;
1949 unsigned int idx = parents->idx[level];
1950 sp = parents->parent[level];
1954 WARN_ON(idx == INVALID_INDEX);
1955 clear_unsync_child_bit(sp, idx);
1957 } while (!sp->unsync_children);
1960 static int mmu_sync_children(struct kvm_vcpu *vcpu,
1961 struct kvm_mmu_page *parent, bool can_yield)
1964 struct kvm_mmu_page *sp;
1965 struct mmu_page_path parents;
1966 struct kvm_mmu_pages pages;
1967 LIST_HEAD(invalid_list);
1970 while (mmu_unsync_walk(parent, &pages)) {
1971 bool protected = false;
1973 for_each_sp(pages, sp, parents, i)
1974 protected |= kvm_vcpu_write_protect_gfn(vcpu, sp->gfn);
1977 kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, true);
1981 for_each_sp(pages, sp, parents, i) {
1982 kvm_unlink_unsync_page(vcpu->kvm, sp);
1983 flush |= kvm_sync_page(vcpu, sp, &invalid_list) > 0;
1984 mmu_pages_clear_parents(&parents);
1986 if (need_resched() || rwlock_needbreak(&vcpu->kvm->mmu_lock)) {
1987 kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, flush);
1989 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
1993 cond_resched_rwlock_write(&vcpu->kvm->mmu_lock);
1998 kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, flush);
2002 static void __clear_sp_write_flooding_count(struct kvm_mmu_page *sp)
2004 atomic_set(&sp->write_flooding_count, 0);
2007 static void clear_sp_write_flooding_count(u64 *spte)
2009 __clear_sp_write_flooding_count(sptep_to_sp(spte));
2013 * The vCPU is required when finding indirect shadow pages; the shadow
2014 * page may already exist and syncing it needs the vCPU pointer in
2015 * order to read guest page tables. Direct shadow pages are never
2016 * unsync, thus @vcpu can be NULL if @role.direct is true.
2018 static struct kvm_mmu_page *kvm_mmu_find_shadow_page(struct kvm *kvm,
2019 struct kvm_vcpu *vcpu,
2021 struct hlist_head *sp_list,
2022 union kvm_mmu_page_role role)
2024 struct kvm_mmu_page *sp;
2027 LIST_HEAD(invalid_list);
2029 for_each_valid_sp(kvm, sp, sp_list) {
2030 if (sp->gfn != gfn) {
2035 if (sp->role.word != role.word) {
2037 * If the guest is creating an upper-level page, zap
2038 * unsync pages for the same gfn. While it's possible
2039 * the guest is using recursive page tables, in all
2040 * likelihood the guest has stopped using the unsync
2041 * page and is installing a completely unrelated page.
2042 * Unsync pages must not be left as is, because the new
2043 * upper-level page will be write-protected.
2045 if (role.level > PG_LEVEL_4K && sp->unsync)
2046 kvm_mmu_prepare_zap_page(kvm, sp,
2051 /* unsync and write-flooding only apply to indirect SPs. */
2052 if (sp->role.direct)
2056 if (KVM_BUG_ON(!vcpu, kvm))
2060 * The page is good, but is stale. kvm_sync_page does
2061 * get the latest guest state, but (unlike mmu_unsync_children)
2062 * it doesn't write-protect the page or mark it synchronized!
2063 * This way the validity of the mapping is ensured, but the
2064 * overhead of write protection is not incurred until the
2065 * guest invalidates the TLB mapping. This allows multiple
2066 * SPs for a single gfn to be unsync.
2068 * If the sync fails, the page is zapped. If so, break
2069 * in order to rebuild it.
2071 ret = kvm_sync_page(vcpu, sp, &invalid_list);
2075 WARN_ON(!list_empty(&invalid_list));
2077 kvm_flush_remote_tlbs(kvm);
2080 __clear_sp_write_flooding_count(sp);
2086 ++kvm->stat.mmu_cache_miss;
2089 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2091 if (collisions > kvm->stat.max_mmu_page_hash_collisions)
2092 kvm->stat.max_mmu_page_hash_collisions = collisions;
2096 /* Caches used when allocating a new shadow page. */
2097 struct shadow_page_caches {
2098 struct kvm_mmu_memory_cache *page_header_cache;
2099 struct kvm_mmu_memory_cache *shadow_page_cache;
2100 struct kvm_mmu_memory_cache *shadowed_info_cache;
2103 static struct kvm_mmu_page *kvm_mmu_alloc_shadow_page(struct kvm *kvm,
2104 struct shadow_page_caches *caches,
2106 struct hlist_head *sp_list,
2107 union kvm_mmu_page_role role)
2109 struct kvm_mmu_page *sp;
2111 sp = kvm_mmu_memory_cache_alloc(caches->page_header_cache);
2112 sp->spt = kvm_mmu_memory_cache_alloc(caches->shadow_page_cache);
2114 sp->shadowed_translation = kvm_mmu_memory_cache_alloc(caches->shadowed_info_cache);
2116 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
2119 * active_mmu_pages must be a FIFO list, as kvm_zap_obsolete_pages()
2120 * depends on valid pages being added to the head of the list. See
2121 * comments in kvm_zap_obsolete_pages().
2123 sp->mmu_valid_gen = kvm->arch.mmu_valid_gen;
2124 list_add(&sp->link, &kvm->arch.active_mmu_pages);
2125 kvm_mod_used_mmu_pages(kvm, +1);
2129 hlist_add_head(&sp->hash_link, sp_list);
2130 if (sp_has_gptes(sp))
2131 account_shadowed(kvm, sp);
2136 /* Note, @vcpu may be NULL if @role.direct is true; see kvm_mmu_find_shadow_page. */
2137 static struct kvm_mmu_page *__kvm_mmu_get_shadow_page(struct kvm *kvm,
2138 struct kvm_vcpu *vcpu,
2139 struct shadow_page_caches *caches,
2141 union kvm_mmu_page_role role)
2143 struct hlist_head *sp_list;
2144 struct kvm_mmu_page *sp;
2145 bool created = false;
2147 sp_list = &kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)];
2149 sp = kvm_mmu_find_shadow_page(kvm, vcpu, gfn, sp_list, role);
2152 sp = kvm_mmu_alloc_shadow_page(kvm, caches, gfn, sp_list, role);
2155 trace_kvm_mmu_get_page(sp, created);
2159 static struct kvm_mmu_page *kvm_mmu_get_shadow_page(struct kvm_vcpu *vcpu,
2161 union kvm_mmu_page_role role)
2163 struct shadow_page_caches caches = {
2164 .page_header_cache = &vcpu->arch.mmu_page_header_cache,
2165 .shadow_page_cache = &vcpu->arch.mmu_shadow_page_cache,
2166 .shadowed_info_cache = &vcpu->arch.mmu_shadowed_info_cache,
2169 return __kvm_mmu_get_shadow_page(vcpu->kvm, vcpu, &caches, gfn, role);
2172 static union kvm_mmu_page_role kvm_mmu_child_role(u64 *sptep, bool direct,
2173 unsigned int access)
2175 struct kvm_mmu_page *parent_sp = sptep_to_sp(sptep);
2176 union kvm_mmu_page_role role;
2178 role = parent_sp->role;
2180 role.access = access;
2181 role.direct = direct;
2182 role.passthrough = 0;
2185 * If the guest has 4-byte PTEs then that means it's using 32-bit,
2186 * 2-level, non-PAE paging. KVM shadows such guests with PAE paging
2187 * (i.e. 8-byte PTEs). The difference in PTE size means that KVM must
2188 * shadow each guest page table with multiple shadow page tables, which
2189 * requires extra bookkeeping in the role.
2191 * Specifically, to shadow the guest's page directory (which covers a
2192 * 4GiB address space), KVM uses 4 PAE page directories, each mapping
2193 * 1GiB of the address space. @role.quadrant encodes which quarter of
2194 * the address space each maps.
2196 * To shadow the guest's page tables (which each map a 4MiB region), KVM
2197 * uses 2 PAE page tables, each mapping a 2MiB region. For these,
2198 * @role.quadrant encodes which half of the region they map.
2200 * Concretely, a 4-byte PDE consumes bits 31:22, while an 8-byte PDE
2201 * consumes bits 29:21. To consume bits 31:30, KVM's uses 4 shadow
2202 * PDPTEs; those 4 PAE page directories are pre-allocated and their
2203 * quadrant is assigned in mmu_alloc_root(). A 4-byte PTE consumes
2204 * bits 21:12, while an 8-byte PTE consumes bits 20:12. To consume
2205 * bit 21 in the PTE (the child here), KVM propagates that bit to the
2206 * quadrant, i.e. sets quadrant to '0' or '1'. The parent 8-byte PDE
2207 * covers bit 21 (see above), thus the quadrant is calculated from the
2208 * _least_ significant bit of the PDE index.
2210 if (role.has_4_byte_gpte) {
2211 WARN_ON_ONCE(role.level != PG_LEVEL_4K);
2212 role.quadrant = spte_index(sptep) & 1;
2218 static struct kvm_mmu_page *kvm_mmu_get_child_sp(struct kvm_vcpu *vcpu,
2219 u64 *sptep, gfn_t gfn,
2220 bool direct, unsigned int access)
2222 union kvm_mmu_page_role role;
2224 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep))
2225 return ERR_PTR(-EEXIST);
2227 role = kvm_mmu_child_role(sptep, direct, access);
2228 return kvm_mmu_get_shadow_page(vcpu, gfn, role);
2231 static void shadow_walk_init_using_root(struct kvm_shadow_walk_iterator *iterator,
2232 struct kvm_vcpu *vcpu, hpa_t root,
2235 iterator->addr = addr;
2236 iterator->shadow_addr = root;
2237 iterator->level = vcpu->arch.mmu->root_role.level;
2239 if (iterator->level >= PT64_ROOT_4LEVEL &&
2240 vcpu->arch.mmu->cpu_role.base.level < PT64_ROOT_4LEVEL &&
2241 !vcpu->arch.mmu->root_role.direct)
2242 iterator->level = PT32E_ROOT_LEVEL;
2244 if (iterator->level == PT32E_ROOT_LEVEL) {
2246 * prev_root is currently only used for 64-bit hosts. So only
2247 * the active root_hpa is valid here.
2249 BUG_ON(root != vcpu->arch.mmu->root.hpa);
2251 iterator->shadow_addr
2252 = vcpu->arch.mmu->pae_root[(addr >> 30) & 3];
2253 iterator->shadow_addr &= SPTE_BASE_ADDR_MASK;
2255 if (!iterator->shadow_addr)
2256 iterator->level = 0;
2260 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
2261 struct kvm_vcpu *vcpu, u64 addr)
2263 shadow_walk_init_using_root(iterator, vcpu, vcpu->arch.mmu->root.hpa,
2267 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
2269 if (iterator->level < PG_LEVEL_4K)
2272 iterator->index = SPTE_INDEX(iterator->addr, iterator->level);
2273 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
2277 static void __shadow_walk_next(struct kvm_shadow_walk_iterator *iterator,
2280 if (!is_shadow_present_pte(spte) || is_last_spte(spte, iterator->level)) {
2281 iterator->level = 0;
2285 iterator->shadow_addr = spte & SPTE_BASE_ADDR_MASK;
2289 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
2291 __shadow_walk_next(iterator, *iterator->sptep);
2294 static void __link_shadow_page(struct kvm *kvm,
2295 struct kvm_mmu_memory_cache *cache, u64 *sptep,
2296 struct kvm_mmu_page *sp, bool flush)
2300 BUILD_BUG_ON(VMX_EPT_WRITABLE_MASK != PT_WRITABLE_MASK);
2303 * If an SPTE is present already, it must be a leaf and therefore
2304 * a large one. Drop it, and flush the TLB if needed, before
2307 if (is_shadow_present_pte(*sptep))
2308 drop_large_spte(kvm, sptep, flush);
2310 spte = make_nonleaf_spte(sp->spt, sp_ad_disabled(sp));
2312 mmu_spte_set(sptep, spte);
2314 mmu_page_add_parent_pte(cache, sp, sptep);
2316 if (sp->unsync_children || sp->unsync)
2320 static void link_shadow_page(struct kvm_vcpu *vcpu, u64 *sptep,
2321 struct kvm_mmu_page *sp)
2323 __link_shadow_page(vcpu->kvm, &vcpu->arch.mmu_pte_list_desc_cache, sptep, sp, true);
2326 static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2327 unsigned direct_access)
2329 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
2330 struct kvm_mmu_page *child;
2333 * For the direct sp, if the guest pte's dirty bit
2334 * changed form clean to dirty, it will corrupt the
2335 * sp's access: allow writable in the read-only sp,
2336 * so we should update the spte at this point to get
2337 * a new sp with the correct access.
2339 child = to_shadow_page(*sptep & SPTE_BASE_ADDR_MASK);
2340 if (child->role.access == direct_access)
2343 drop_parent_pte(child, sptep);
2344 kvm_flush_remote_tlbs_with_address(vcpu->kvm, child->gfn, 1);
2348 /* Returns the number of zapped non-leaf child shadow pages. */
2349 static int mmu_page_zap_pte(struct kvm *kvm, struct kvm_mmu_page *sp,
2350 u64 *spte, struct list_head *invalid_list)
2353 struct kvm_mmu_page *child;
2356 if (is_shadow_present_pte(pte)) {
2357 if (is_last_spte(pte, sp->role.level)) {
2358 drop_spte(kvm, spte);
2360 child = to_shadow_page(pte & SPTE_BASE_ADDR_MASK);
2361 drop_parent_pte(child, spte);
2364 * Recursively zap nested TDP SPs, parentless SPs are
2365 * unlikely to be used again in the near future. This
2366 * avoids retaining a large number of stale nested SPs.
2368 if (tdp_enabled && invalid_list &&
2369 child->role.guest_mode && !child->parent_ptes.val)
2370 return kvm_mmu_prepare_zap_page(kvm, child,
2373 } else if (is_mmio_spte(pte)) {
2374 mmu_spte_clear_no_track(spte);
2379 static int kvm_mmu_page_unlink_children(struct kvm *kvm,
2380 struct kvm_mmu_page *sp,
2381 struct list_head *invalid_list)
2386 for (i = 0; i < SPTE_ENT_PER_PAGE; ++i)
2387 zapped += mmu_page_zap_pte(kvm, sp, sp->spt + i, invalid_list);
2392 static void kvm_mmu_unlink_parents(struct kvm_mmu_page *sp)
2395 struct rmap_iterator iter;
2397 while ((sptep = rmap_get_first(&sp->parent_ptes, &iter)))
2398 drop_parent_pte(sp, sptep);
2401 static int mmu_zap_unsync_children(struct kvm *kvm,
2402 struct kvm_mmu_page *parent,
2403 struct list_head *invalid_list)
2406 struct mmu_page_path parents;
2407 struct kvm_mmu_pages pages;
2409 if (parent->role.level == PG_LEVEL_4K)
2412 while (mmu_unsync_walk(parent, &pages)) {
2413 struct kvm_mmu_page *sp;
2415 for_each_sp(pages, sp, parents, i) {
2416 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
2417 mmu_pages_clear_parents(&parents);
2425 static bool __kvm_mmu_prepare_zap_page(struct kvm *kvm,
2426 struct kvm_mmu_page *sp,
2427 struct list_head *invalid_list,
2430 bool list_unstable, zapped_root = false;
2432 trace_kvm_mmu_prepare_zap_page(sp);
2433 ++kvm->stat.mmu_shadow_zapped;
2434 *nr_zapped = mmu_zap_unsync_children(kvm, sp, invalid_list);
2435 *nr_zapped += kvm_mmu_page_unlink_children(kvm, sp, invalid_list);
2436 kvm_mmu_unlink_parents(sp);
2438 /* Zapping children means active_mmu_pages has become unstable. */
2439 list_unstable = *nr_zapped;
2441 if (!sp->role.invalid && sp_has_gptes(sp))
2442 unaccount_shadowed(kvm, sp);
2445 kvm_unlink_unsync_page(kvm, sp);
2446 if (!sp->root_count) {
2451 * Already invalid pages (previously active roots) are not on
2452 * the active page list. See list_del() in the "else" case of
2455 if (sp->role.invalid)
2456 list_add(&sp->link, invalid_list);
2458 list_move(&sp->link, invalid_list);
2459 kvm_mod_used_mmu_pages(kvm, -1);
2462 * Remove the active root from the active page list, the root
2463 * will be explicitly freed when the root_count hits zero.
2465 list_del(&sp->link);
2468 * Obsolete pages cannot be used on any vCPUs, see the comment
2469 * in kvm_mmu_zap_all_fast(). Note, is_obsolete_sp() also
2470 * treats invalid shadow pages as being obsolete.
2472 zapped_root = !is_obsolete_sp(kvm, sp);
2475 if (sp->lpage_disallowed)
2476 unaccount_huge_nx_page(kvm, sp);
2478 sp->role.invalid = 1;
2481 * Make the request to free obsolete roots after marking the root
2482 * invalid, otherwise other vCPUs may not see it as invalid.
2485 kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_FREE_OBSOLETE_ROOTS);
2486 return list_unstable;
2489 static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
2490 struct list_head *invalid_list)
2494 __kvm_mmu_prepare_zap_page(kvm, sp, invalid_list, &nr_zapped);
2498 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
2499 struct list_head *invalid_list)
2501 struct kvm_mmu_page *sp, *nsp;
2503 if (list_empty(invalid_list))
2507 * We need to make sure everyone sees our modifications to
2508 * the page tables and see changes to vcpu->mode here. The barrier
2509 * in the kvm_flush_remote_tlbs() achieves this. This pairs
2510 * with vcpu_enter_guest and walk_shadow_page_lockless_begin/end.
2512 * In addition, kvm_flush_remote_tlbs waits for all vcpus to exit
2513 * guest mode and/or lockless shadow page table walks.
2515 kvm_flush_remote_tlbs(kvm);
2517 list_for_each_entry_safe(sp, nsp, invalid_list, link) {
2518 WARN_ON(!sp->role.invalid || sp->root_count);
2519 kvm_mmu_free_shadow_page(sp);
2523 static unsigned long kvm_mmu_zap_oldest_mmu_pages(struct kvm *kvm,
2524 unsigned long nr_to_zap)
2526 unsigned long total_zapped = 0;
2527 struct kvm_mmu_page *sp, *tmp;
2528 LIST_HEAD(invalid_list);
2532 if (list_empty(&kvm->arch.active_mmu_pages))
2536 list_for_each_entry_safe_reverse(sp, tmp, &kvm->arch.active_mmu_pages, link) {
2538 * Don't zap active root pages, the page itself can't be freed
2539 * and zapping it will just force vCPUs to realloc and reload.
2544 unstable = __kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list,
2546 total_zapped += nr_zapped;
2547 if (total_zapped >= nr_to_zap)
2554 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2556 kvm->stat.mmu_recycled += total_zapped;
2557 return total_zapped;
2560 static inline unsigned long kvm_mmu_available_pages(struct kvm *kvm)
2562 if (kvm->arch.n_max_mmu_pages > kvm->arch.n_used_mmu_pages)
2563 return kvm->arch.n_max_mmu_pages -
2564 kvm->arch.n_used_mmu_pages;
2569 static int make_mmu_pages_available(struct kvm_vcpu *vcpu)
2571 unsigned long avail = kvm_mmu_available_pages(vcpu->kvm);
2573 if (likely(avail >= KVM_MIN_FREE_MMU_PAGES))
2576 kvm_mmu_zap_oldest_mmu_pages(vcpu->kvm, KVM_REFILL_PAGES - avail);
2579 * Note, this check is intentionally soft, it only guarantees that one
2580 * page is available, while the caller may end up allocating as many as
2581 * four pages, e.g. for PAE roots or for 5-level paging. Temporarily
2582 * exceeding the (arbitrary by default) limit will not harm the host,
2583 * being too aggressive may unnecessarily kill the guest, and getting an
2584 * exact count is far more trouble than it's worth, especially in the
2587 if (!kvm_mmu_available_pages(vcpu->kvm))
2593 * Changing the number of mmu pages allocated to the vm
2594 * Note: if goal_nr_mmu_pages is too small, you will get dead lock
2596 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned long goal_nr_mmu_pages)
2598 write_lock(&kvm->mmu_lock);
2600 if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) {
2601 kvm_mmu_zap_oldest_mmu_pages(kvm, kvm->arch.n_used_mmu_pages -
2604 goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages;
2607 kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages;
2609 write_unlock(&kvm->mmu_lock);
2612 int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
2614 struct kvm_mmu_page *sp;
2615 LIST_HEAD(invalid_list);
2618 pgprintk("%s: looking for gfn %llx\n", __func__, gfn);
2620 write_lock(&kvm->mmu_lock);
2621 for_each_gfn_valid_sp_with_gptes(kvm, sp, gfn) {
2622 pgprintk("%s: gfn %llx role %x\n", __func__, gfn,
2625 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
2627 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2628 write_unlock(&kvm->mmu_lock);
2633 static int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2638 if (vcpu->arch.mmu->root_role.direct)
2641 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2643 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2648 static void kvm_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
2650 trace_kvm_mmu_unsync_page(sp);
2651 ++kvm->stat.mmu_unsync;
2654 kvm_mmu_mark_parents_unsync(sp);
2658 * Attempt to unsync any shadow pages that can be reached by the specified gfn,
2659 * KVM is creating a writable mapping for said gfn. Returns 0 if all pages
2660 * were marked unsync (or if there is no shadow page), -EPERM if the SPTE must
2661 * be write-protected.
2663 int mmu_try_to_unsync_pages(struct kvm *kvm, const struct kvm_memory_slot *slot,
2664 gfn_t gfn, bool can_unsync, bool prefetch)
2666 struct kvm_mmu_page *sp;
2667 bool locked = false;
2670 * Force write-protection if the page is being tracked. Note, the page
2671 * track machinery is used to write-protect upper-level shadow pages,
2672 * i.e. this guards the role.level == 4K assertion below!
2674 if (kvm_slot_page_track_is_active(kvm, slot, gfn, KVM_PAGE_TRACK_WRITE))
2678 * The page is not write-tracked, mark existing shadow pages unsync
2679 * unless KVM is synchronizing an unsync SP (can_unsync = false). In
2680 * that case, KVM must complete emulation of the guest TLB flush before
2681 * allowing shadow pages to become unsync (writable by the guest).
2683 for_each_gfn_valid_sp_with_gptes(kvm, sp, gfn) {
2694 * TDP MMU page faults require an additional spinlock as they
2695 * run with mmu_lock held for read, not write, and the unsync
2696 * logic is not thread safe. Take the spinklock regardless of
2697 * the MMU type to avoid extra conditionals/parameters, there's
2698 * no meaningful penalty if mmu_lock is held for write.
2702 spin_lock(&kvm->arch.mmu_unsync_pages_lock);
2705 * Recheck after taking the spinlock, a different vCPU
2706 * may have since marked the page unsync. A false
2707 * positive on the unprotected check above is not
2708 * possible as clearing sp->unsync _must_ hold mmu_lock
2709 * for write, i.e. unsync cannot transition from 0->1
2710 * while this CPU holds mmu_lock for read (or write).
2712 if (READ_ONCE(sp->unsync))
2716 WARN_ON(sp->role.level != PG_LEVEL_4K);
2717 kvm_unsync_page(kvm, sp);
2720 spin_unlock(&kvm->arch.mmu_unsync_pages_lock);
2723 * We need to ensure that the marking of unsync pages is visible
2724 * before the SPTE is updated to allow writes because
2725 * kvm_mmu_sync_roots() checks the unsync flags without holding
2726 * the MMU lock and so can race with this. If the SPTE was updated
2727 * before the page had been marked as unsync-ed, something like the
2728 * following could happen:
2731 * ---------------------------------------------------------------------
2732 * 1.2 Host updates SPTE
2734 * 2.1 Guest writes a GPTE for GVA X.
2735 * (GPTE being in the guest page table shadowed
2736 * by the SP from CPU 1.)
2737 * This reads SPTE during the page table walk.
2738 * Since SPTE.W is read as 1, there is no
2741 * 2.2 Guest issues TLB flush.
2742 * That causes a VM Exit.
2744 * 2.3 Walking of unsync pages sees sp->unsync is
2745 * false and skips the page.
2747 * 2.4 Guest accesses GVA X.
2748 * Since the mapping in the SP was not updated,
2749 * so the old mapping for GVA X incorrectly
2753 * (sp->unsync = true)
2755 * The write barrier below ensures that 1.1 happens before 1.2 and thus
2756 * the situation in 2.4 does not arise. It pairs with the read barrier
2757 * in is_unsync_root(), placed between 2.1's load of SPTE.W and 2.3.
2764 static int mmu_set_spte(struct kvm_vcpu *vcpu, struct kvm_memory_slot *slot,
2765 u64 *sptep, unsigned int pte_access, gfn_t gfn,
2766 kvm_pfn_t pfn, struct kvm_page_fault *fault)
2768 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
2769 int level = sp->role.level;
2770 int was_rmapped = 0;
2771 int ret = RET_PF_FIXED;
2776 /* Prefetching always gets a writable pfn. */
2777 bool host_writable = !fault || fault->map_writable;
2778 bool prefetch = !fault || fault->prefetch;
2779 bool write_fault = fault && fault->write;
2781 pgprintk("%s: spte %llx write_fault %d gfn %llx\n", __func__,
2782 *sptep, write_fault, gfn);
2784 if (unlikely(is_noslot_pfn(pfn))) {
2785 vcpu->stat.pf_mmio_spte_created++;
2786 mark_mmio_spte(vcpu, sptep, gfn, pte_access);
2787 return RET_PF_EMULATE;
2790 if (is_shadow_present_pte(*sptep)) {
2792 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
2793 * the parent of the now unreachable PTE.
2795 if (level > PG_LEVEL_4K && !is_large_pte(*sptep)) {
2796 struct kvm_mmu_page *child;
2799 child = to_shadow_page(pte & SPTE_BASE_ADDR_MASK);
2800 drop_parent_pte(child, sptep);
2802 } else if (pfn != spte_to_pfn(*sptep)) {
2803 pgprintk("hfn old %llx new %llx\n",
2804 spte_to_pfn(*sptep), pfn);
2805 drop_spte(vcpu->kvm, sptep);
2811 wrprot = make_spte(vcpu, sp, slot, pte_access, gfn, pfn, *sptep, prefetch,
2812 true, host_writable, &spte);
2814 if (*sptep == spte) {
2815 ret = RET_PF_SPURIOUS;
2817 flush |= mmu_spte_update(sptep, spte);
2818 trace_kvm_mmu_set_spte(level, gfn, sptep);
2823 ret = RET_PF_EMULATE;
2827 kvm_flush_remote_tlbs_with_address(vcpu->kvm, gfn,
2828 KVM_PAGES_PER_HPAGE(level));
2830 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
2833 WARN_ON_ONCE(ret == RET_PF_SPURIOUS);
2834 rmap_add(vcpu, slot, sptep, gfn, pte_access);
2836 /* Already rmapped but the pte_access bits may have changed. */
2837 kvm_mmu_page_set_access(sp, spte_index(sptep), pte_access);
2843 static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu,
2844 struct kvm_mmu_page *sp,
2845 u64 *start, u64 *end)
2847 struct page *pages[PTE_PREFETCH_NUM];
2848 struct kvm_memory_slot *slot;
2849 unsigned int access = sp->role.access;
2853 gfn = kvm_mmu_page_get_gfn(sp, spte_index(start));
2854 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, access & ACC_WRITE_MASK);
2858 ret = gfn_to_page_many_atomic(slot, gfn, pages, end - start);
2862 for (i = 0; i < ret; i++, gfn++, start++) {
2863 mmu_set_spte(vcpu, slot, start, access, gfn,
2864 page_to_pfn(pages[i]), NULL);
2871 static void __direct_pte_prefetch(struct kvm_vcpu *vcpu,
2872 struct kvm_mmu_page *sp, u64 *sptep)
2874 u64 *spte, *start = NULL;
2877 WARN_ON(!sp->role.direct);
2879 i = spte_index(sptep) & ~(PTE_PREFETCH_NUM - 1);
2882 for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
2883 if (is_shadow_present_pte(*spte) || spte == sptep) {
2886 if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0)
2893 direct_pte_prefetch_many(vcpu, sp, start, spte);
2896 static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
2898 struct kvm_mmu_page *sp;
2900 sp = sptep_to_sp(sptep);
2903 * Without accessed bits, there's no way to distinguish between
2904 * actually accessed translations and prefetched, so disable pte
2905 * prefetch if accessed bits aren't available.
2907 if (sp_ad_disabled(sp))
2910 if (sp->role.level > PG_LEVEL_4K)
2914 * If addresses are being invalidated, skip prefetching to avoid
2915 * accidentally prefetching those addresses.
2917 if (unlikely(vcpu->kvm->mmu_notifier_count))
2920 __direct_pte_prefetch(vcpu, sp, sptep);
2924 * Lookup the mapping level for @gfn in the current mm.
2926 * WARNING! Use of host_pfn_mapping_level() requires the caller and the end
2927 * consumer to be tied into KVM's handlers for MMU notifier events!
2929 * There are several ways to safely use this helper:
2931 * - Check mmu_notifier_retry_hva() after grabbing the mapping level, before
2932 * consuming it. In this case, mmu_lock doesn't need to be held during the
2933 * lookup, but it does need to be held while checking the MMU notifier.
2935 * - Hold mmu_lock AND ensure there is no in-progress MMU notifier invalidation
2936 * event for the hva. This can be done by explicit checking the MMU notifier
2937 * or by ensuring that KVM already has a valid mapping that covers the hva.
2939 * - Do not use the result to install new mappings, e.g. use the host mapping
2940 * level only to decide whether or not to zap an entry. In this case, it's
2941 * not required to hold mmu_lock (though it's highly likely the caller will
2942 * want to hold mmu_lock anyways, e.g. to modify SPTEs).
2944 * Note! The lookup can still race with modifications to host page tables, but
2945 * the above "rules" ensure KVM will not _consume_ the result of the walk if a
2946 * race with the primary MMU occurs.
2948 static int host_pfn_mapping_level(struct kvm *kvm, gfn_t gfn,
2949 const struct kvm_memory_slot *slot)
2951 int level = PG_LEVEL_4K;
2953 unsigned long flags;
2960 * Note, using the already-retrieved memslot and __gfn_to_hva_memslot()
2961 * is not solely for performance, it's also necessary to avoid the
2962 * "writable" check in __gfn_to_hva_many(), which will always fail on
2963 * read-only memslots due to gfn_to_hva() assuming writes. Earlier
2964 * page fault steps have already verified the guest isn't writing a
2965 * read-only memslot.
2967 hva = __gfn_to_hva_memslot(slot, gfn);
2970 * Disable IRQs to prevent concurrent tear down of host page tables,
2971 * e.g. if the primary MMU promotes a P*D to a huge page and then frees
2972 * the original page table.
2974 local_irq_save(flags);
2977 * Read each entry once. As above, a non-leaf entry can be promoted to
2978 * a huge page _during_ this walk. Re-reading the entry could send the
2979 * walk into the weeks, e.g. p*d_large() returns false (sees the old
2980 * value) and then p*d_offset() walks into the target huge page instead
2981 * of the old page table (sees the new value).
2983 pgd = READ_ONCE(*pgd_offset(kvm->mm, hva));
2987 p4d = READ_ONCE(*p4d_offset(&pgd, hva));
2988 if (p4d_none(p4d) || !p4d_present(p4d))
2991 pud = READ_ONCE(*pud_offset(&p4d, hva));
2992 if (pud_none(pud) || !pud_present(pud))
2995 if (pud_large(pud)) {
2996 level = PG_LEVEL_1G;
3000 pmd = READ_ONCE(*pmd_offset(&pud, hva));
3001 if (pmd_none(pmd) || !pmd_present(pmd))
3005 level = PG_LEVEL_2M;
3008 local_irq_restore(flags);
3012 int kvm_mmu_max_mapping_level(struct kvm *kvm,
3013 const struct kvm_memory_slot *slot, gfn_t gfn,
3016 struct kvm_lpage_info *linfo;
3019 max_level = min(max_level, max_huge_page_level);
3020 for ( ; max_level > PG_LEVEL_4K; max_level--) {
3021 linfo = lpage_info_slot(gfn, slot, max_level);
3022 if (!linfo->disallow_lpage)
3026 if (max_level == PG_LEVEL_4K)
3029 host_level = host_pfn_mapping_level(kvm, gfn, slot);
3030 return min(host_level, max_level);
3033 void kvm_mmu_hugepage_adjust(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
3035 struct kvm_memory_slot *slot = fault->slot;
3038 fault->huge_page_disallowed = fault->exec && fault->nx_huge_page_workaround_enabled;
3040 if (unlikely(fault->max_level == PG_LEVEL_4K))
3043 if (is_error_noslot_pfn(fault->pfn))
3046 if (kvm_slot_dirty_track_enabled(slot))
3050 * Enforce the iTLB multihit workaround after capturing the requested
3051 * level, which will be used to do precise, accurate accounting.
3053 fault->req_level = kvm_mmu_max_mapping_level(vcpu->kvm, slot,
3054 fault->gfn, fault->max_level);
3055 if (fault->req_level == PG_LEVEL_4K || fault->huge_page_disallowed)
3059 * mmu_notifier_retry() was successful and mmu_lock is held, so
3060 * the pmd can't be split from under us.
3062 fault->goal_level = fault->req_level;
3063 mask = KVM_PAGES_PER_HPAGE(fault->goal_level) - 1;
3064 VM_BUG_ON((fault->gfn & mask) != (fault->pfn & mask));
3065 fault->pfn &= ~mask;
3068 void disallowed_hugepage_adjust(struct kvm_page_fault *fault, u64 spte, int cur_level)
3070 if (cur_level > PG_LEVEL_4K &&
3071 cur_level == fault->goal_level &&
3072 is_shadow_present_pte(spte) &&
3073 !is_large_pte(spte)) {
3075 * A small SPTE exists for this pfn, but FNAME(fetch)
3076 * and __direct_map would like to create a large PTE
3077 * instead: just force them to go down another level,
3078 * patching back for them into pfn the next 9 bits of
3081 u64 page_mask = KVM_PAGES_PER_HPAGE(cur_level) -
3082 KVM_PAGES_PER_HPAGE(cur_level - 1);
3083 fault->pfn |= fault->gfn & page_mask;
3084 fault->goal_level--;
3088 static int __direct_map(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
3090 struct kvm_shadow_walk_iterator it;
3091 struct kvm_mmu_page *sp;
3093 gfn_t base_gfn = fault->gfn;
3095 kvm_mmu_hugepage_adjust(vcpu, fault);
3097 trace_kvm_mmu_spte_requested(fault);
3098 for_each_shadow_entry(vcpu, fault->addr, it) {
3100 * We cannot overwrite existing page tables with an NX
3101 * large page, as the leaf could be executable.
3103 if (fault->nx_huge_page_workaround_enabled)
3104 disallowed_hugepage_adjust(fault, *it.sptep, it.level);
3106 base_gfn = fault->gfn & ~(KVM_PAGES_PER_HPAGE(it.level) - 1);
3107 if (it.level == fault->goal_level)
3110 sp = kvm_mmu_get_child_sp(vcpu, it.sptep, base_gfn, true, ACC_ALL);
3111 if (sp == ERR_PTR(-EEXIST))
3114 link_shadow_page(vcpu, it.sptep, sp);
3115 if (fault->is_tdp && fault->huge_page_disallowed &&
3116 fault->req_level >= it.level)
3117 account_huge_nx_page(vcpu->kvm, sp);
3120 if (WARN_ON_ONCE(it.level != fault->goal_level))
3123 ret = mmu_set_spte(vcpu, fault->slot, it.sptep, ACC_ALL,
3124 base_gfn, fault->pfn, fault);
3125 if (ret == RET_PF_SPURIOUS)
3128 direct_pte_prefetch(vcpu, it.sptep);
3132 static void kvm_send_hwpoison_signal(unsigned long address, struct task_struct *tsk)
3134 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, PAGE_SHIFT, tsk);
3137 static int kvm_handle_bad_page(struct kvm_vcpu *vcpu, gfn_t gfn, kvm_pfn_t pfn)
3140 * Do not cache the mmio info caused by writing the readonly gfn
3141 * into the spte otherwise read access on readonly gfn also can
3142 * caused mmio page fault and treat it as mmio access.
3144 if (pfn == KVM_PFN_ERR_RO_FAULT)
3145 return RET_PF_EMULATE;
3147 if (pfn == KVM_PFN_ERR_HWPOISON) {
3148 kvm_send_hwpoison_signal(kvm_vcpu_gfn_to_hva(vcpu, gfn), current);
3149 return RET_PF_RETRY;
3155 static int handle_abnormal_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault,
3156 unsigned int access)
3158 /* The pfn is invalid, report the error! */
3159 if (unlikely(is_error_pfn(fault->pfn)))
3160 return kvm_handle_bad_page(vcpu, fault->gfn, fault->pfn);
3162 if (unlikely(!fault->slot)) {
3163 gva_t gva = fault->is_tdp ? 0 : fault->addr;
3165 vcpu_cache_mmio_info(vcpu, gva, fault->gfn,
3166 access & shadow_mmio_access_mask);
3168 * If MMIO caching is disabled, emulate immediately without
3169 * touching the shadow page tables as attempting to install an
3170 * MMIO SPTE will just be an expensive nop. Do not cache MMIO
3171 * whose gfn is greater than host.MAXPHYADDR, any guest that
3172 * generates such gfns is running nested and is being tricked
3173 * by L0 userspace (you can observe gfn > L1.MAXPHYADDR if
3174 * and only if L1's MAXPHYADDR is inaccurate with respect to
3177 if (unlikely(!enable_mmio_caching) ||
3178 unlikely(fault->gfn > kvm_mmu_max_gfn()))
3179 return RET_PF_EMULATE;
3182 return RET_PF_CONTINUE;
3185 static bool page_fault_can_be_fast(struct kvm_page_fault *fault)
3188 * Page faults with reserved bits set, i.e. faults on MMIO SPTEs, only
3189 * reach the common page fault handler if the SPTE has an invalid MMIO
3190 * generation number. Refreshing the MMIO generation needs to go down
3191 * the slow path. Note, EPT Misconfigs do NOT set the PRESENT flag!
3197 * #PF can be fast if:
3199 * 1. The shadow page table entry is not present and A/D bits are
3200 * disabled _by KVM_, which could mean that the fault is potentially
3201 * caused by access tracking (if enabled). If A/D bits are enabled
3202 * by KVM, but disabled by L1 for L2, KVM is forced to disable A/D
3203 * bits for L2 and employ access tracking, but the fast page fault
3204 * mechanism only supports direct MMUs.
3205 * 2. The shadow page table entry is present, the access is a write,
3206 * and no reserved bits are set (MMIO SPTEs cannot be "fixed"), i.e.
3207 * the fault was caused by a write-protection violation. If the
3208 * SPTE is MMU-writable (determined later), the fault can be fixed
3209 * by setting the Writable bit, which can be done out of mmu_lock.
3211 if (!fault->present)
3212 return !kvm_ad_enabled();
3215 * Note, instruction fetches and writes are mutually exclusive, ignore
3218 return fault->write;
3222 * Returns true if the SPTE was fixed successfully. Otherwise,
3223 * someone else modified the SPTE from its original value.
3226 fast_pf_fix_direct_spte(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault,
3227 u64 *sptep, u64 old_spte, u64 new_spte)
3230 * Theoretically we could also set dirty bit (and flush TLB) here in
3231 * order to eliminate unnecessary PML logging. See comments in
3232 * set_spte. But fast_page_fault is very unlikely to happen with PML
3233 * enabled, so we do not do this. This might result in the same GPA
3234 * to be logged in PML buffer again when the write really happens, and
3235 * eventually to be called by mark_page_dirty twice. But it's also no
3236 * harm. This also avoids the TLB flush needed after setting dirty bit
3237 * so non-PML cases won't be impacted.
3239 * Compare with set_spte where instead shadow_dirty_mask is set.
3241 if (!try_cmpxchg64(sptep, &old_spte, new_spte))
3244 if (is_writable_pte(new_spte) && !is_writable_pte(old_spte))
3245 mark_page_dirty_in_slot(vcpu->kvm, fault->slot, fault->gfn);
3250 static bool is_access_allowed(struct kvm_page_fault *fault, u64 spte)
3253 return is_executable_pte(spte);
3256 return is_writable_pte(spte);
3258 /* Fault was on Read access */
3259 return spte & PT_PRESENT_MASK;
3263 * Returns the last level spte pointer of the shadow page walk for the given
3264 * gpa, and sets *spte to the spte value. This spte may be non-preset. If no
3265 * walk could be performed, returns NULL and *spte does not contain valid data.
3268 * - Must be called between walk_shadow_page_lockless_{begin,end}.
3269 * - The returned sptep must not be used after walk_shadow_page_lockless_end.
3271 static u64 *fast_pf_get_last_sptep(struct kvm_vcpu *vcpu, gpa_t gpa, u64 *spte)
3273 struct kvm_shadow_walk_iterator iterator;
3277 for_each_shadow_entry_lockless(vcpu, gpa, iterator, old_spte) {
3278 sptep = iterator.sptep;
3286 * Returns one of RET_PF_INVALID, RET_PF_FIXED or RET_PF_SPURIOUS.
3288 static int fast_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
3290 struct kvm_mmu_page *sp;
3291 int ret = RET_PF_INVALID;
3294 uint retry_count = 0;
3296 if (!page_fault_can_be_fast(fault))
3299 walk_shadow_page_lockless_begin(vcpu);
3304 if (is_tdp_mmu(vcpu->arch.mmu))
3305 sptep = kvm_tdp_mmu_fast_pf_get_last_sptep(vcpu, fault->addr, &spte);
3307 sptep = fast_pf_get_last_sptep(vcpu, fault->addr, &spte);
3309 if (!is_shadow_present_pte(spte))
3312 sp = sptep_to_sp(sptep);
3313 if (!is_last_spte(spte, sp->role.level))
3317 * Check whether the memory access that caused the fault would
3318 * still cause it if it were to be performed right now. If not,
3319 * then this is a spurious fault caused by TLB lazily flushed,
3320 * or some other CPU has already fixed the PTE after the
3321 * current CPU took the fault.
3323 * Need not check the access of upper level table entries since
3324 * they are always ACC_ALL.
3326 if (is_access_allowed(fault, spte)) {
3327 ret = RET_PF_SPURIOUS;
3334 * KVM only supports fixing page faults outside of MMU lock for
3335 * direct MMUs, nested MMUs are always indirect, and KVM always
3336 * uses A/D bits for non-nested MMUs. Thus, if A/D bits are
3337 * enabled, the SPTE can't be an access-tracked SPTE.
3339 if (unlikely(!kvm_ad_enabled()) && is_access_track_spte(spte))
3340 new_spte = restore_acc_track_spte(new_spte);
3343 * To keep things simple, only SPTEs that are MMU-writable can
3344 * be made fully writable outside of mmu_lock, e.g. only SPTEs
3345 * that were write-protected for dirty-logging or access
3346 * tracking are handled here. Don't bother checking if the
3347 * SPTE is writable to prioritize running with A/D bits enabled.
3348 * The is_access_allowed() check above handles the common case
3349 * of the fault being spurious, and the SPTE is known to be
3350 * shadow-present, i.e. except for access tracking restoration
3351 * making the new SPTE writable, the check is wasteful.
3353 if (fault->write && is_mmu_writable_spte(spte)) {
3354 new_spte |= PT_WRITABLE_MASK;
3357 * Do not fix write-permission on the large spte when
3358 * dirty logging is enabled. Since we only dirty the
3359 * first page into the dirty-bitmap in
3360 * fast_pf_fix_direct_spte(), other pages are missed
3361 * if its slot has dirty logging enabled.
3363 * Instead, we let the slow page fault path create a
3364 * normal spte to fix the access.
3366 if (sp->role.level > PG_LEVEL_4K &&
3367 kvm_slot_dirty_track_enabled(fault->slot))
3371 /* Verify that the fault can be handled in the fast path */
3372 if (new_spte == spte ||
3373 !is_access_allowed(fault, new_spte))
3377 * Currently, fast page fault only works for direct mapping
3378 * since the gfn is not stable for indirect shadow page. See
3379 * Documentation/virt/kvm/locking.rst to get more detail.
3381 if (fast_pf_fix_direct_spte(vcpu, fault, sptep, spte, new_spte)) {
3386 if (++retry_count > 4) {
3387 printk_once(KERN_WARNING
3388 "kvm: Fast #PF retrying more than 4 times.\n");
3394 trace_fast_page_fault(vcpu, fault, sptep, spte, ret);
3395 walk_shadow_page_lockless_end(vcpu);
3397 if (ret != RET_PF_INVALID)
3398 vcpu->stat.pf_fast++;
3403 static void mmu_free_root_page(struct kvm *kvm, hpa_t *root_hpa,
3404 struct list_head *invalid_list)
3406 struct kvm_mmu_page *sp;
3408 if (!VALID_PAGE(*root_hpa))
3411 sp = to_shadow_page(*root_hpa & SPTE_BASE_ADDR_MASK);
3415 if (is_tdp_mmu_page(sp))
3416 kvm_tdp_mmu_put_root(kvm, sp, false);
3417 else if (!--sp->root_count && sp->role.invalid)
3418 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
3420 *root_hpa = INVALID_PAGE;
3423 /* roots_to_free must be some combination of the KVM_MMU_ROOT_* flags */
3424 void kvm_mmu_free_roots(struct kvm *kvm, struct kvm_mmu *mmu,
3425 ulong roots_to_free)
3428 LIST_HEAD(invalid_list);
3429 bool free_active_root;
3431 BUILD_BUG_ON(KVM_MMU_NUM_PREV_ROOTS >= BITS_PER_LONG);
3433 /* Before acquiring the MMU lock, see if we need to do any real work. */
3434 free_active_root = (roots_to_free & KVM_MMU_ROOT_CURRENT)
3435 && VALID_PAGE(mmu->root.hpa);
3437 if (!free_active_root) {
3438 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3439 if ((roots_to_free & KVM_MMU_ROOT_PREVIOUS(i)) &&
3440 VALID_PAGE(mmu->prev_roots[i].hpa))
3443 if (i == KVM_MMU_NUM_PREV_ROOTS)
3447 write_lock(&kvm->mmu_lock);
3449 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3450 if (roots_to_free & KVM_MMU_ROOT_PREVIOUS(i))
3451 mmu_free_root_page(kvm, &mmu->prev_roots[i].hpa,
3454 if (free_active_root) {
3455 if (to_shadow_page(mmu->root.hpa)) {
3456 mmu_free_root_page(kvm, &mmu->root.hpa, &invalid_list);
3457 } else if (mmu->pae_root) {
3458 for (i = 0; i < 4; ++i) {
3459 if (!IS_VALID_PAE_ROOT(mmu->pae_root[i]))
3462 mmu_free_root_page(kvm, &mmu->pae_root[i],
3464 mmu->pae_root[i] = INVALID_PAE_ROOT;
3467 mmu->root.hpa = INVALID_PAGE;
3471 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3472 write_unlock(&kvm->mmu_lock);
3474 EXPORT_SYMBOL_GPL(kvm_mmu_free_roots);
3476 void kvm_mmu_free_guest_mode_roots(struct kvm *kvm, struct kvm_mmu *mmu)
3478 unsigned long roots_to_free = 0;
3483 * This should not be called while L2 is active, L2 can't invalidate
3484 * _only_ its own roots, e.g. INVVPID unconditionally exits.
3486 WARN_ON_ONCE(mmu->root_role.guest_mode);
3488 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
3489 root_hpa = mmu->prev_roots[i].hpa;
3490 if (!VALID_PAGE(root_hpa))
3493 if (!to_shadow_page(root_hpa) ||
3494 to_shadow_page(root_hpa)->role.guest_mode)
3495 roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i);
3498 kvm_mmu_free_roots(kvm, mmu, roots_to_free);
3500 EXPORT_SYMBOL_GPL(kvm_mmu_free_guest_mode_roots);
3503 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
3507 if (!kvm_vcpu_is_visible_gfn(vcpu, root_gfn)) {
3508 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
3515 static hpa_t mmu_alloc_root(struct kvm_vcpu *vcpu, gfn_t gfn, int quadrant,
3518 union kvm_mmu_page_role role = vcpu->arch.mmu->root_role;
3519 struct kvm_mmu_page *sp;
3522 role.quadrant = quadrant;
3524 WARN_ON_ONCE(quadrant && !role.has_4_byte_gpte);
3525 WARN_ON_ONCE(role.direct && role.has_4_byte_gpte);
3527 sp = kvm_mmu_get_shadow_page(vcpu, gfn, role);
3530 return __pa(sp->spt);
3533 static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
3535 struct kvm_mmu *mmu = vcpu->arch.mmu;
3536 u8 shadow_root_level = mmu->root_role.level;
3541 write_lock(&vcpu->kvm->mmu_lock);
3542 r = make_mmu_pages_available(vcpu);
3546 if (is_tdp_mmu_enabled(vcpu->kvm)) {
3547 root = kvm_tdp_mmu_get_vcpu_root_hpa(vcpu);
3548 mmu->root.hpa = root;
3549 } else if (shadow_root_level >= PT64_ROOT_4LEVEL) {
3550 root = mmu_alloc_root(vcpu, 0, 0, shadow_root_level);
3551 mmu->root.hpa = root;
3552 } else if (shadow_root_level == PT32E_ROOT_LEVEL) {
3553 if (WARN_ON_ONCE(!mmu->pae_root)) {
3558 for (i = 0; i < 4; ++i) {
3559 WARN_ON_ONCE(IS_VALID_PAE_ROOT(mmu->pae_root[i]));
3561 root = mmu_alloc_root(vcpu, i << (30 - PAGE_SHIFT), 0,
3563 mmu->pae_root[i] = root | PT_PRESENT_MASK |
3566 mmu->root.hpa = __pa(mmu->pae_root);
3568 WARN_ONCE(1, "Bad TDP root level = %d\n", shadow_root_level);
3573 /* root.pgd is ignored for direct MMUs. */
3576 write_unlock(&vcpu->kvm->mmu_lock);
3580 static int mmu_first_shadow_root_alloc(struct kvm *kvm)
3582 struct kvm_memslots *slots;
3583 struct kvm_memory_slot *slot;
3587 * Check if this is the first shadow root being allocated before
3590 if (kvm_shadow_root_allocated(kvm))
3593 mutex_lock(&kvm->slots_arch_lock);
3595 /* Recheck, under the lock, whether this is the first shadow root. */
3596 if (kvm_shadow_root_allocated(kvm))
3600 * Check if anything actually needs to be allocated, e.g. all metadata
3601 * will be allocated upfront if TDP is disabled.
3603 if (kvm_memslots_have_rmaps(kvm) &&
3604 kvm_page_track_write_tracking_enabled(kvm))
3607 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
3608 slots = __kvm_memslots(kvm, i);
3609 kvm_for_each_memslot(slot, bkt, slots) {
3611 * Both of these functions are no-ops if the target is
3612 * already allocated, so unconditionally calling both
3613 * is safe. Intentionally do NOT free allocations on
3614 * failure to avoid having to track which allocations
3615 * were made now versus when the memslot was created.
3616 * The metadata is guaranteed to be freed when the slot
3617 * is freed, and will be kept/used if userspace retries
3618 * KVM_RUN instead of killing the VM.
3620 r = memslot_rmap_alloc(slot, slot->npages);
3623 r = kvm_page_track_write_tracking_alloc(slot);
3630 * Ensure that shadow_root_allocated becomes true strictly after
3631 * all the related pointers are set.
3634 smp_store_release(&kvm->arch.shadow_root_allocated, true);
3637 mutex_unlock(&kvm->slots_arch_lock);
3641 static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
3643 struct kvm_mmu *mmu = vcpu->arch.mmu;
3644 u64 pdptrs[4], pm_mask;
3645 gfn_t root_gfn, root_pgd;
3649 root_pgd = mmu->get_guest_pgd(vcpu);
3650 root_gfn = root_pgd >> PAGE_SHIFT;
3652 if (mmu_check_root(vcpu, root_gfn))
3656 * On SVM, reading PDPTRs might access guest memory, which might fault
3657 * and thus might sleep. Grab the PDPTRs before acquiring mmu_lock.
3659 if (mmu->cpu_role.base.level == PT32E_ROOT_LEVEL) {
3660 for (i = 0; i < 4; ++i) {
3661 pdptrs[i] = mmu->get_pdptr(vcpu, i);
3662 if (!(pdptrs[i] & PT_PRESENT_MASK))
3665 if (mmu_check_root(vcpu, pdptrs[i] >> PAGE_SHIFT))
3670 r = mmu_first_shadow_root_alloc(vcpu->kvm);
3674 write_lock(&vcpu->kvm->mmu_lock);
3675 r = make_mmu_pages_available(vcpu);
3680 * Do we shadow a long mode page table? If so we need to
3681 * write-protect the guests page table root.
3683 if (mmu->cpu_role.base.level >= PT64_ROOT_4LEVEL) {
3684 root = mmu_alloc_root(vcpu, root_gfn, 0,
3685 mmu->root_role.level);
3686 mmu->root.hpa = root;
3690 if (WARN_ON_ONCE(!mmu->pae_root)) {
3696 * We shadow a 32 bit page table. This may be a legacy 2-level
3697 * or a PAE 3-level page table. In either case we need to be aware that
3698 * the shadow page table may be a PAE or a long mode page table.
3700 pm_mask = PT_PRESENT_MASK | shadow_me_value;
3701 if (mmu->root_role.level >= PT64_ROOT_4LEVEL) {
3702 pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK;
3704 if (WARN_ON_ONCE(!mmu->pml4_root)) {
3708 mmu->pml4_root[0] = __pa(mmu->pae_root) | pm_mask;
3710 if (mmu->root_role.level == PT64_ROOT_5LEVEL) {
3711 if (WARN_ON_ONCE(!mmu->pml5_root)) {
3715 mmu->pml5_root[0] = __pa(mmu->pml4_root) | pm_mask;
3719 for (i = 0; i < 4; ++i) {
3720 WARN_ON_ONCE(IS_VALID_PAE_ROOT(mmu->pae_root[i]));
3722 if (mmu->cpu_role.base.level == PT32E_ROOT_LEVEL) {
3723 if (!(pdptrs[i] & PT_PRESENT_MASK)) {
3724 mmu->pae_root[i] = INVALID_PAE_ROOT;
3727 root_gfn = pdptrs[i] >> PAGE_SHIFT;
3731 * If shadowing 32-bit non-PAE page tables, each PAE page
3732 * directory maps one quarter of the guest's non-PAE page
3733 * directory. Othwerise each PAE page direct shadows one guest
3734 * PAE page directory so that quadrant should be 0.
3736 quadrant = (mmu->cpu_role.base.level == PT32_ROOT_LEVEL) ? i : 0;
3738 root = mmu_alloc_root(vcpu, root_gfn, quadrant, PT32_ROOT_LEVEL);
3739 mmu->pae_root[i] = root | pm_mask;
3742 if (mmu->root_role.level == PT64_ROOT_5LEVEL)
3743 mmu->root.hpa = __pa(mmu->pml5_root);
3744 else if (mmu->root_role.level == PT64_ROOT_4LEVEL)
3745 mmu->root.hpa = __pa(mmu->pml4_root);
3747 mmu->root.hpa = __pa(mmu->pae_root);
3750 mmu->root.pgd = root_pgd;
3752 write_unlock(&vcpu->kvm->mmu_lock);
3757 static int mmu_alloc_special_roots(struct kvm_vcpu *vcpu)
3759 struct kvm_mmu *mmu = vcpu->arch.mmu;
3760 bool need_pml5 = mmu->root_role.level > PT64_ROOT_4LEVEL;
3761 u64 *pml5_root = NULL;
3762 u64 *pml4_root = NULL;
3766 * When shadowing 32-bit or PAE NPT with 64-bit NPT, the PML4 and PDP
3767 * tables are allocated and initialized at root creation as there is no
3768 * equivalent level in the guest's NPT to shadow. Allocate the tables
3769 * on demand, as running a 32-bit L1 VMM on 64-bit KVM is very rare.
3771 if (mmu->root_role.direct ||
3772 mmu->cpu_role.base.level >= PT64_ROOT_4LEVEL ||
3773 mmu->root_role.level < PT64_ROOT_4LEVEL)
3777 * NPT, the only paging mode that uses this horror, uses a fixed number
3778 * of levels for the shadow page tables, e.g. all MMUs are 4-level or
3779 * all MMus are 5-level. Thus, this can safely require that pml5_root
3780 * is allocated if the other roots are valid and pml5 is needed, as any
3781 * prior MMU would also have required pml5.
3783 if (mmu->pae_root && mmu->pml4_root && (!need_pml5 || mmu->pml5_root))
3787 * The special roots should always be allocated in concert. Yell and
3788 * bail if KVM ends up in a state where only one of the roots is valid.
3790 if (WARN_ON_ONCE(!tdp_enabled || mmu->pae_root || mmu->pml4_root ||
3791 (need_pml5 && mmu->pml5_root)))
3795 * Unlike 32-bit NPT, the PDP table doesn't need to be in low mem, and
3796 * doesn't need to be decrypted.
3798 pae_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT);
3802 #ifdef CONFIG_X86_64
3803 pml4_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT);
3808 pml5_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT);
3814 mmu->pae_root = pae_root;
3815 mmu->pml4_root = pml4_root;
3816 mmu->pml5_root = pml5_root;
3820 #ifdef CONFIG_X86_64
3822 free_page((unsigned long)pml4_root);
3824 free_page((unsigned long)pae_root);
3829 static bool is_unsync_root(hpa_t root)
3831 struct kvm_mmu_page *sp;
3833 if (!VALID_PAGE(root))
3837 * The read barrier orders the CPU's read of SPTE.W during the page table
3838 * walk before the reads of sp->unsync/sp->unsync_children here.
3840 * Even if another CPU was marking the SP as unsync-ed simultaneously,
3841 * any guest page table changes are not guaranteed to be visible anyway
3842 * until this VCPU issues a TLB flush strictly after those changes are
3843 * made. We only need to ensure that the other CPU sets these flags
3844 * before any actual changes to the page tables are made. The comments
3845 * in mmu_try_to_unsync_pages() describe what could go wrong if this
3846 * requirement isn't satisfied.
3849 sp = to_shadow_page(root);
3852 * PAE roots (somewhat arbitrarily) aren't backed by shadow pages, the
3853 * PDPTEs for a given PAE root need to be synchronized individually.
3855 if (WARN_ON_ONCE(!sp))
3858 if (sp->unsync || sp->unsync_children)
3864 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
3867 struct kvm_mmu_page *sp;
3869 if (vcpu->arch.mmu->root_role.direct)
3872 if (!VALID_PAGE(vcpu->arch.mmu->root.hpa))
3875 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
3877 if (vcpu->arch.mmu->cpu_role.base.level >= PT64_ROOT_4LEVEL) {
3878 hpa_t root = vcpu->arch.mmu->root.hpa;
3879 sp = to_shadow_page(root);
3881 if (!is_unsync_root(root))
3884 write_lock(&vcpu->kvm->mmu_lock);
3885 mmu_sync_children(vcpu, sp, true);
3886 write_unlock(&vcpu->kvm->mmu_lock);
3890 write_lock(&vcpu->kvm->mmu_lock);
3892 for (i = 0; i < 4; ++i) {
3893 hpa_t root = vcpu->arch.mmu->pae_root[i];
3895 if (IS_VALID_PAE_ROOT(root)) {
3896 root &= SPTE_BASE_ADDR_MASK;
3897 sp = to_shadow_page(root);
3898 mmu_sync_children(vcpu, sp, true);
3902 write_unlock(&vcpu->kvm->mmu_lock);
3905 void kvm_mmu_sync_prev_roots(struct kvm_vcpu *vcpu)
3907 unsigned long roots_to_free = 0;
3910 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3911 if (is_unsync_root(vcpu->arch.mmu->prev_roots[i].hpa))
3912 roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i);
3914 /* sync prev_roots by simply freeing them */
3915 kvm_mmu_free_roots(vcpu->kvm, vcpu->arch.mmu, roots_to_free);
3918 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
3919 gpa_t vaddr, u64 access,
3920 struct x86_exception *exception)
3923 exception->error_code = 0;
3924 return kvm_translate_gpa(vcpu, mmu, vaddr, access, exception);
3927 static bool mmio_info_in_cache(struct kvm_vcpu *vcpu, u64 addr, bool direct)
3930 * A nested guest cannot use the MMIO cache if it is using nested
3931 * page tables, because cr2 is a nGPA while the cache stores GPAs.
3933 if (mmu_is_nested(vcpu))
3937 return vcpu_match_mmio_gpa(vcpu, addr);
3939 return vcpu_match_mmio_gva(vcpu, addr);
3943 * Return the level of the lowest level SPTE added to sptes.
3944 * That SPTE may be non-present.
3946 * Must be called between walk_shadow_page_lockless_{begin,end}.
3948 static int get_walk(struct kvm_vcpu *vcpu, u64 addr, u64 *sptes, int *root_level)
3950 struct kvm_shadow_walk_iterator iterator;
3954 for (shadow_walk_init(&iterator, vcpu, addr),
3955 *root_level = iterator.level;
3956 shadow_walk_okay(&iterator);
3957 __shadow_walk_next(&iterator, spte)) {
3958 leaf = iterator.level;
3959 spte = mmu_spte_get_lockless(iterator.sptep);
3967 /* return true if reserved bit(s) are detected on a valid, non-MMIO SPTE. */
3968 static bool get_mmio_spte(struct kvm_vcpu *vcpu, u64 addr, u64 *sptep)
3970 u64 sptes[PT64_ROOT_MAX_LEVEL + 1];
3971 struct rsvd_bits_validate *rsvd_check;
3972 int root, leaf, level;
3973 bool reserved = false;
3975 walk_shadow_page_lockless_begin(vcpu);
3977 if (is_tdp_mmu(vcpu->arch.mmu))
3978 leaf = kvm_tdp_mmu_get_walk(vcpu, addr, sptes, &root);
3980 leaf = get_walk(vcpu, addr, sptes, &root);
3982 walk_shadow_page_lockless_end(vcpu);
3984 if (unlikely(leaf < 0)) {
3989 *sptep = sptes[leaf];
3992 * Skip reserved bits checks on the terminal leaf if it's not a valid
3993 * SPTE. Note, this also (intentionally) skips MMIO SPTEs, which, by
3994 * design, always have reserved bits set. The purpose of the checks is
3995 * to detect reserved bits on non-MMIO SPTEs. i.e. buggy SPTEs.
3997 if (!is_shadow_present_pte(sptes[leaf]))
4000 rsvd_check = &vcpu->arch.mmu->shadow_zero_check;
4002 for (level = root; level >= leaf; level--)
4003 reserved |= is_rsvd_spte(rsvd_check, sptes[level], level);
4006 pr_err("%s: reserved bits set on MMU-present spte, addr 0x%llx, hierarchy:\n",
4008 for (level = root; level >= leaf; level--)
4009 pr_err("------ spte = 0x%llx level = %d, rsvd bits = 0x%llx",
4010 sptes[level], level,
4011 get_rsvd_bits(rsvd_check, sptes[level], level));
4017 static int handle_mmio_page_fault(struct kvm_vcpu *vcpu, u64 addr, bool direct)
4022 if (mmio_info_in_cache(vcpu, addr, direct))
4023 return RET_PF_EMULATE;
4025 reserved = get_mmio_spte(vcpu, addr, &spte);
4026 if (WARN_ON(reserved))
4029 if (is_mmio_spte(spte)) {
4030 gfn_t gfn = get_mmio_spte_gfn(spte);
4031 unsigned int access = get_mmio_spte_access(spte);
4033 if (!check_mmio_spte(vcpu, spte))
4034 return RET_PF_INVALID;
4039 trace_handle_mmio_page_fault(addr, gfn, access);
4040 vcpu_cache_mmio_info(vcpu, addr, gfn, access);
4041 return RET_PF_EMULATE;
4045 * If the page table is zapped by other cpus, let CPU fault again on
4048 return RET_PF_RETRY;
4051 static bool page_fault_handle_page_track(struct kvm_vcpu *vcpu,
4052 struct kvm_page_fault *fault)
4054 if (unlikely(fault->rsvd))
4057 if (!fault->present || !fault->write)
4061 * guest is writing the page which is write tracked which can
4062 * not be fixed by page fault handler.
4064 if (kvm_slot_page_track_is_active(vcpu->kvm, fault->slot, fault->gfn, KVM_PAGE_TRACK_WRITE))
4070 static void shadow_page_table_clear_flood(struct kvm_vcpu *vcpu, gva_t addr)
4072 struct kvm_shadow_walk_iterator iterator;
4075 walk_shadow_page_lockless_begin(vcpu);
4076 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte)
4077 clear_sp_write_flooding_count(iterator.sptep);
4078 walk_shadow_page_lockless_end(vcpu);
4081 static u32 alloc_apf_token(struct kvm_vcpu *vcpu)
4083 /* make sure the token value is not 0 */
4084 u32 id = vcpu->arch.apf.id;
4087 vcpu->arch.apf.id = 1;
4089 return (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id;
4092 static bool kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
4095 struct kvm_arch_async_pf arch;
4097 arch.token = alloc_apf_token(vcpu);
4099 arch.direct_map = vcpu->arch.mmu->root_role.direct;
4100 arch.cr3 = vcpu->arch.mmu->get_guest_pgd(vcpu);
4102 return kvm_setup_async_pf(vcpu, cr2_or_gpa,
4103 kvm_vcpu_gfn_to_hva(vcpu, gfn), &arch);
4106 void kvm_arch_async_page_ready(struct kvm_vcpu *vcpu, struct kvm_async_pf *work)
4110 if ((vcpu->arch.mmu->root_role.direct != work->arch.direct_map) ||
4114 r = kvm_mmu_reload(vcpu);
4118 if (!vcpu->arch.mmu->root_role.direct &&
4119 work->arch.cr3 != vcpu->arch.mmu->get_guest_pgd(vcpu))
4122 kvm_mmu_do_page_fault(vcpu, work->cr2_or_gpa, 0, true);
4125 static int kvm_faultin_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
4127 struct kvm_memory_slot *slot = fault->slot;
4131 * Retry the page fault if the gfn hit a memslot that is being deleted
4132 * or moved. This ensures any existing SPTEs for the old memslot will
4133 * be zapped before KVM inserts a new MMIO SPTE for the gfn.
4135 if (slot && (slot->flags & KVM_MEMSLOT_INVALID))
4136 return RET_PF_RETRY;
4138 if (!kvm_is_visible_memslot(slot)) {
4139 /* Don't expose private memslots to L2. */
4140 if (is_guest_mode(vcpu)) {
4142 fault->pfn = KVM_PFN_NOSLOT;
4143 fault->map_writable = false;
4144 return RET_PF_CONTINUE;
4147 * If the APIC access page exists but is disabled, go directly
4148 * to emulation without caching the MMIO access or creating a
4149 * MMIO SPTE. That way the cache doesn't need to be purged
4150 * when the AVIC is re-enabled.
4152 if (slot && slot->id == APIC_ACCESS_PAGE_PRIVATE_MEMSLOT &&
4153 !kvm_apicv_activated(vcpu->kvm))
4154 return RET_PF_EMULATE;
4158 fault->pfn = __gfn_to_pfn_memslot(slot, fault->gfn, false, &async,
4159 fault->write, &fault->map_writable,
4162 return RET_PF_CONTINUE; /* *pfn has correct page already */
4164 if (!fault->prefetch && kvm_can_do_async_pf(vcpu)) {
4165 trace_kvm_try_async_get_page(fault->addr, fault->gfn);
4166 if (kvm_find_async_pf_gfn(vcpu, fault->gfn)) {
4167 trace_kvm_async_pf_doublefault(fault->addr, fault->gfn);
4168 kvm_make_request(KVM_REQ_APF_HALT, vcpu);
4169 return RET_PF_RETRY;
4170 } else if (kvm_arch_setup_async_pf(vcpu, fault->addr, fault->gfn)) {
4171 return RET_PF_RETRY;
4175 fault->pfn = __gfn_to_pfn_memslot(slot, fault->gfn, false, NULL,
4176 fault->write, &fault->map_writable,
4178 return RET_PF_CONTINUE;
4182 * Returns true if the page fault is stale and needs to be retried, i.e. if the
4183 * root was invalidated by a memslot update or a relevant mmu_notifier fired.
4185 static bool is_page_fault_stale(struct kvm_vcpu *vcpu,
4186 struct kvm_page_fault *fault, int mmu_seq)
4188 struct kvm_mmu_page *sp = to_shadow_page(vcpu->arch.mmu->root.hpa);
4190 /* Special roots, e.g. pae_root, are not backed by shadow pages. */
4191 if (sp && is_obsolete_sp(vcpu->kvm, sp))
4195 * Roots without an associated shadow page are considered invalid if
4196 * there is a pending request to free obsolete roots. The request is
4197 * only a hint that the current root _may_ be obsolete and needs to be
4198 * reloaded, e.g. if the guest frees a PGD that KVM is tracking as a
4199 * previous root, then __kvm_mmu_prepare_zap_page() signals all vCPUs
4200 * to reload even if no vCPU is actively using the root.
4202 if (!sp && kvm_test_request(KVM_REQ_MMU_FREE_OBSOLETE_ROOTS, vcpu))
4205 return fault->slot &&
4206 mmu_notifier_retry_hva(vcpu->kvm, mmu_seq, fault->hva);
4209 static int direct_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
4211 bool is_tdp_mmu_fault = is_tdp_mmu(vcpu->arch.mmu);
4213 unsigned long mmu_seq;
4216 fault->gfn = fault->addr >> PAGE_SHIFT;
4217 fault->slot = kvm_vcpu_gfn_to_memslot(vcpu, fault->gfn);
4219 if (page_fault_handle_page_track(vcpu, fault))
4220 return RET_PF_EMULATE;
4222 r = fast_page_fault(vcpu, fault);
4223 if (r != RET_PF_INVALID)
4226 r = mmu_topup_memory_caches(vcpu, false);
4230 mmu_seq = vcpu->kvm->mmu_notifier_seq;
4233 r = kvm_faultin_pfn(vcpu, fault);
4234 if (r != RET_PF_CONTINUE)
4237 r = handle_abnormal_pfn(vcpu, fault, ACC_ALL);
4238 if (r != RET_PF_CONTINUE)
4243 if (is_tdp_mmu_fault)
4244 read_lock(&vcpu->kvm->mmu_lock);
4246 write_lock(&vcpu->kvm->mmu_lock);
4248 if (is_page_fault_stale(vcpu, fault, mmu_seq))
4251 r = make_mmu_pages_available(vcpu);
4255 if (is_tdp_mmu_fault)
4256 r = kvm_tdp_mmu_map(vcpu, fault);
4258 r = __direct_map(vcpu, fault);
4261 if (is_tdp_mmu_fault)
4262 read_unlock(&vcpu->kvm->mmu_lock);
4264 write_unlock(&vcpu->kvm->mmu_lock);
4265 kvm_release_pfn_clean(fault->pfn);
4269 static int nonpaging_page_fault(struct kvm_vcpu *vcpu,
4270 struct kvm_page_fault *fault)
4272 pgprintk("%s: gva %lx error %x\n", __func__, fault->addr, fault->error_code);
4274 /* This path builds a PAE pagetable, we can map 2mb pages at maximum. */
4275 fault->max_level = PG_LEVEL_2M;
4276 return direct_page_fault(vcpu, fault);
4279 int kvm_handle_page_fault(struct kvm_vcpu *vcpu, u64 error_code,
4280 u64 fault_address, char *insn, int insn_len)
4283 u32 flags = vcpu->arch.apf.host_apf_flags;
4285 #ifndef CONFIG_X86_64
4286 /* A 64-bit CR2 should be impossible on 32-bit KVM. */
4287 if (WARN_ON_ONCE(fault_address >> 32))
4291 vcpu->arch.l1tf_flush_l1d = true;
4293 trace_kvm_page_fault(fault_address, error_code);
4295 if (kvm_event_needs_reinjection(vcpu))
4296 kvm_mmu_unprotect_page_virt(vcpu, fault_address);
4297 r = kvm_mmu_page_fault(vcpu, fault_address, error_code, insn,
4299 } else if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
4300 vcpu->arch.apf.host_apf_flags = 0;
4301 local_irq_disable();
4302 kvm_async_pf_task_wait_schedule(fault_address);
4305 WARN_ONCE(1, "Unexpected host async PF flags: %x\n", flags);
4310 EXPORT_SYMBOL_GPL(kvm_handle_page_fault);
4312 int kvm_tdp_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
4315 * If the guest's MTRRs may be used to compute the "real" memtype,
4316 * restrict the mapping level to ensure KVM uses a consistent memtype
4317 * across the entire mapping. If the host MTRRs are ignored by TDP
4318 * (shadow_memtype_mask is non-zero), and the VM has non-coherent DMA
4319 * (DMA doesn't snoop CPU caches), KVM's ABI is to honor the memtype
4320 * from the guest's MTRRs so that guest accesses to memory that is
4321 * DMA'd aren't cached against the guest's wishes.
4323 * Note, KVM may still ultimately ignore guest MTRRs for certain PFNs,
4324 * e.g. KVM will force UC memtype for host MMIO.
4326 if (shadow_memtype_mask && kvm_arch_has_noncoherent_dma(vcpu->kvm)) {
4327 for ( ; fault->max_level > PG_LEVEL_4K; --fault->max_level) {
4328 int page_num = KVM_PAGES_PER_HPAGE(fault->max_level);
4329 gfn_t base = (fault->addr >> PAGE_SHIFT) & ~(page_num - 1);
4331 if (kvm_mtrr_check_gfn_range_consistency(vcpu, base, page_num))
4336 return direct_page_fault(vcpu, fault);
4339 static void nonpaging_init_context(struct kvm_mmu *context)
4341 context->page_fault = nonpaging_page_fault;
4342 context->gva_to_gpa = nonpaging_gva_to_gpa;
4343 context->sync_page = nonpaging_sync_page;
4344 context->invlpg = NULL;
4347 static inline bool is_root_usable(struct kvm_mmu_root_info *root, gpa_t pgd,
4348 union kvm_mmu_page_role role)
4350 return (role.direct || pgd == root->pgd) &&
4351 VALID_PAGE(root->hpa) &&
4352 role.word == to_shadow_page(root->hpa)->role.word;
4356 * Find out if a previously cached root matching the new pgd/role is available,
4357 * and insert the current root as the MRU in the cache.
4358 * If a matching root is found, it is assigned to kvm_mmu->root and
4360 * If no match is found, kvm_mmu->root is left invalid, the LRU root is
4361 * evicted to make room for the current root, and false is returned.
4363 static bool cached_root_find_and_keep_current(struct kvm *kvm, struct kvm_mmu *mmu,
4365 union kvm_mmu_page_role new_role)
4369 if (is_root_usable(&mmu->root, new_pgd, new_role))
4372 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
4374 * The swaps end up rotating the cache like this:
4375 * C 0 1 2 3 (on entry to the function)
4379 * 3 C 0 1 2 (on exit from the loop)
4381 swap(mmu->root, mmu->prev_roots[i]);
4382 if (is_root_usable(&mmu->root, new_pgd, new_role))
4386 kvm_mmu_free_roots(kvm, mmu, KVM_MMU_ROOT_CURRENT);
4391 * Find out if a previously cached root matching the new pgd/role is available.
4392 * On entry, mmu->root is invalid.
4393 * If a matching root is found, it is assigned to kvm_mmu->root, the LRU entry
4394 * of the cache becomes invalid, and true is returned.
4395 * If no match is found, kvm_mmu->root is left invalid and false is returned.
4397 static bool cached_root_find_without_current(struct kvm *kvm, struct kvm_mmu *mmu,
4399 union kvm_mmu_page_role new_role)
4403 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
4404 if (is_root_usable(&mmu->prev_roots[i], new_pgd, new_role))
4410 swap(mmu->root, mmu->prev_roots[i]);
4411 /* Bubble up the remaining roots. */
4412 for (; i < KVM_MMU_NUM_PREV_ROOTS - 1; i++)
4413 mmu->prev_roots[i] = mmu->prev_roots[i + 1];
4414 mmu->prev_roots[i].hpa = INVALID_PAGE;
4418 static bool fast_pgd_switch(struct kvm *kvm, struct kvm_mmu *mmu,
4419 gpa_t new_pgd, union kvm_mmu_page_role new_role)
4422 * For now, limit the caching to 64-bit hosts+VMs in order to avoid
4423 * having to deal with PDPTEs. We may add support for 32-bit hosts/VMs
4424 * later if necessary.
4426 if (VALID_PAGE(mmu->root.hpa) && !to_shadow_page(mmu->root.hpa))
4427 kvm_mmu_free_roots(kvm, mmu, KVM_MMU_ROOT_CURRENT);
4429 if (VALID_PAGE(mmu->root.hpa))
4430 return cached_root_find_and_keep_current(kvm, mmu, new_pgd, new_role);
4432 return cached_root_find_without_current(kvm, mmu, new_pgd, new_role);
4435 void kvm_mmu_new_pgd(struct kvm_vcpu *vcpu, gpa_t new_pgd)
4437 struct kvm_mmu *mmu = vcpu->arch.mmu;
4438 union kvm_mmu_page_role new_role = mmu->root_role;
4440 if (!fast_pgd_switch(vcpu->kvm, mmu, new_pgd, new_role)) {
4441 /* kvm_mmu_ensure_valid_pgd will set up a new root. */
4446 * It's possible that the cached previous root page is obsolete because
4447 * of a change in the MMU generation number. However, changing the
4448 * generation number is accompanied by KVM_REQ_MMU_FREE_OBSOLETE_ROOTS,
4449 * which will free the root set here and allocate a new one.
4451 kvm_make_request(KVM_REQ_LOAD_MMU_PGD, vcpu);
4453 if (force_flush_and_sync_on_reuse) {
4454 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
4455 kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
4459 * The last MMIO access's GVA and GPA are cached in the VCPU. When
4460 * switching to a new CR3, that GVA->GPA mapping may no longer be
4461 * valid. So clear any cached MMIO info even when we don't need to sync
4462 * the shadow page tables.
4464 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
4467 * If this is a direct root page, it doesn't have a write flooding
4468 * count. Otherwise, clear the write flooding count.
4470 if (!new_role.direct)
4471 __clear_sp_write_flooding_count(
4472 to_shadow_page(vcpu->arch.mmu->root.hpa));
4474 EXPORT_SYMBOL_GPL(kvm_mmu_new_pgd);
4476 static unsigned long get_cr3(struct kvm_vcpu *vcpu)
4478 return kvm_read_cr3(vcpu);
4481 static bool sync_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn,
4482 unsigned int access)
4484 if (unlikely(is_mmio_spte(*sptep))) {
4485 if (gfn != get_mmio_spte_gfn(*sptep)) {
4486 mmu_spte_clear_no_track(sptep);
4490 mark_mmio_spte(vcpu, sptep, gfn, access);
4497 #define PTTYPE_EPT 18 /* arbitrary */
4498 #define PTTYPE PTTYPE_EPT
4499 #include "paging_tmpl.h"
4503 #include "paging_tmpl.h"
4507 #include "paging_tmpl.h"
4511 __reset_rsvds_bits_mask(struct rsvd_bits_validate *rsvd_check,
4512 u64 pa_bits_rsvd, int level, bool nx, bool gbpages,
4515 u64 gbpages_bit_rsvd = 0;
4516 u64 nonleaf_bit8_rsvd = 0;
4519 rsvd_check->bad_mt_xwr = 0;
4522 gbpages_bit_rsvd = rsvd_bits(7, 7);
4524 if (level == PT32E_ROOT_LEVEL)
4525 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 62);
4527 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 51);
4529 /* Note, NX doesn't exist in PDPTEs, this is handled below. */
4531 high_bits_rsvd |= rsvd_bits(63, 63);
4534 * Non-leaf PML4Es and PDPEs reserve bit 8 (which would be the G bit for
4535 * leaf entries) on AMD CPUs only.
4538 nonleaf_bit8_rsvd = rsvd_bits(8, 8);
4541 case PT32_ROOT_LEVEL:
4542 /* no rsvd bits for 2 level 4K page table entries */
4543 rsvd_check->rsvd_bits_mask[0][1] = 0;
4544 rsvd_check->rsvd_bits_mask[0][0] = 0;
4545 rsvd_check->rsvd_bits_mask[1][0] =
4546 rsvd_check->rsvd_bits_mask[0][0];
4549 rsvd_check->rsvd_bits_mask[1][1] = 0;
4553 if (is_cpuid_PSE36())
4554 /* 36bits PSE 4MB page */
4555 rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
4557 /* 32 bits PSE 4MB page */
4558 rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
4560 case PT32E_ROOT_LEVEL:
4561 rsvd_check->rsvd_bits_mask[0][2] = rsvd_bits(63, 63) |
4564 rsvd_bits(1, 2); /* PDPTE */
4565 rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd; /* PDE */
4566 rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd; /* PTE */
4567 rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd |
4568 rsvd_bits(13, 20); /* large page */
4569 rsvd_check->rsvd_bits_mask[1][0] =
4570 rsvd_check->rsvd_bits_mask[0][0];
4572 case PT64_ROOT_5LEVEL:
4573 rsvd_check->rsvd_bits_mask[0][4] = high_bits_rsvd |
4576 rsvd_check->rsvd_bits_mask[1][4] =
4577 rsvd_check->rsvd_bits_mask[0][4];
4579 case PT64_ROOT_4LEVEL:
4580 rsvd_check->rsvd_bits_mask[0][3] = high_bits_rsvd |
4583 rsvd_check->rsvd_bits_mask[0][2] = high_bits_rsvd |
4585 rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd;
4586 rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd;
4587 rsvd_check->rsvd_bits_mask[1][3] =
4588 rsvd_check->rsvd_bits_mask[0][3];
4589 rsvd_check->rsvd_bits_mask[1][2] = high_bits_rsvd |
4592 rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd |
4593 rsvd_bits(13, 20); /* large page */
4594 rsvd_check->rsvd_bits_mask[1][0] =
4595 rsvd_check->rsvd_bits_mask[0][0];
4600 static bool guest_can_use_gbpages(struct kvm_vcpu *vcpu)
4603 * If TDP is enabled, let the guest use GBPAGES if they're supported in
4604 * hardware. The hardware page walker doesn't let KVM disable GBPAGES,
4605 * i.e. won't treat them as reserved, and KVM doesn't redo the GVA->GPA
4606 * walk for performance and complexity reasons. Not to mention KVM
4607 * _can't_ solve the problem because GVA->GPA walks aren't visible to
4608 * KVM once a TDP translation is installed. Mimic hardware behavior so
4609 * that KVM's is at least consistent, i.e. doesn't randomly inject #PF.
4611 return tdp_enabled ? boot_cpu_has(X86_FEATURE_GBPAGES) :
4612 guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES);
4615 static void reset_guest_rsvds_bits_mask(struct kvm_vcpu *vcpu,
4616 struct kvm_mmu *context)
4618 __reset_rsvds_bits_mask(&context->guest_rsvd_check,
4619 vcpu->arch.reserved_gpa_bits,
4620 context->cpu_role.base.level, is_efer_nx(context),
4621 guest_can_use_gbpages(vcpu),
4622 is_cr4_pse(context),
4623 guest_cpuid_is_amd_or_hygon(vcpu));
4627 __reset_rsvds_bits_mask_ept(struct rsvd_bits_validate *rsvd_check,
4628 u64 pa_bits_rsvd, bool execonly, int huge_page_level)
4630 u64 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 51);
4631 u64 large_1g_rsvd = 0, large_2m_rsvd = 0;
4634 if (huge_page_level < PG_LEVEL_1G)
4635 large_1g_rsvd = rsvd_bits(7, 7);
4636 if (huge_page_level < PG_LEVEL_2M)
4637 large_2m_rsvd = rsvd_bits(7, 7);
4639 rsvd_check->rsvd_bits_mask[0][4] = high_bits_rsvd | rsvd_bits(3, 7);
4640 rsvd_check->rsvd_bits_mask[0][3] = high_bits_rsvd | rsvd_bits(3, 7);
4641 rsvd_check->rsvd_bits_mask[0][2] = high_bits_rsvd | rsvd_bits(3, 6) | large_1g_rsvd;
4642 rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd | rsvd_bits(3, 6) | large_2m_rsvd;
4643 rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd;
4646 rsvd_check->rsvd_bits_mask[1][4] = rsvd_check->rsvd_bits_mask[0][4];
4647 rsvd_check->rsvd_bits_mask[1][3] = rsvd_check->rsvd_bits_mask[0][3];
4648 rsvd_check->rsvd_bits_mask[1][2] = high_bits_rsvd | rsvd_bits(12, 29) | large_1g_rsvd;
4649 rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd | rsvd_bits(12, 20) | large_2m_rsvd;
4650 rsvd_check->rsvd_bits_mask[1][0] = rsvd_check->rsvd_bits_mask[0][0];
4652 bad_mt_xwr = 0xFFull << (2 * 8); /* bits 3..5 must not be 2 */
4653 bad_mt_xwr |= 0xFFull << (3 * 8); /* bits 3..5 must not be 3 */
4654 bad_mt_xwr |= 0xFFull << (7 * 8); /* bits 3..5 must not be 7 */
4655 bad_mt_xwr |= REPEAT_BYTE(1ull << 2); /* bits 0..2 must not be 010 */
4656 bad_mt_xwr |= REPEAT_BYTE(1ull << 6); /* bits 0..2 must not be 110 */
4658 /* bits 0..2 must not be 100 unless VMX capabilities allow it */
4659 bad_mt_xwr |= REPEAT_BYTE(1ull << 4);
4661 rsvd_check->bad_mt_xwr = bad_mt_xwr;
4664 static void reset_rsvds_bits_mask_ept(struct kvm_vcpu *vcpu,
4665 struct kvm_mmu *context, bool execonly, int huge_page_level)
4667 __reset_rsvds_bits_mask_ept(&context->guest_rsvd_check,
4668 vcpu->arch.reserved_gpa_bits, execonly,
4672 static inline u64 reserved_hpa_bits(void)
4674 return rsvd_bits(shadow_phys_bits, 63);
4678 * the page table on host is the shadow page table for the page
4679 * table in guest or amd nested guest, its mmu features completely
4680 * follow the features in guest.
4682 static void reset_shadow_zero_bits_mask(struct kvm_vcpu *vcpu,
4683 struct kvm_mmu *context)
4685 /* @amd adds a check on bit of SPTEs, which KVM shouldn't use anyways. */
4687 /* KVM doesn't use 2-level page tables for the shadow MMU. */
4688 bool is_pse = false;
4689 struct rsvd_bits_validate *shadow_zero_check;
4692 WARN_ON_ONCE(context->root_role.level < PT32E_ROOT_LEVEL);
4694 shadow_zero_check = &context->shadow_zero_check;
4695 __reset_rsvds_bits_mask(shadow_zero_check, reserved_hpa_bits(),
4696 context->root_role.level,
4697 context->root_role.efer_nx,
4698 guest_can_use_gbpages(vcpu), is_pse, is_amd);
4700 if (!shadow_me_mask)
4703 for (i = context->root_role.level; --i >= 0;) {
4705 * So far shadow_me_value is a constant during KVM's life
4706 * time. Bits in shadow_me_value are allowed to be set.
4707 * Bits in shadow_me_mask but not in shadow_me_value are
4708 * not allowed to be set.
4710 shadow_zero_check->rsvd_bits_mask[0][i] |= shadow_me_mask;
4711 shadow_zero_check->rsvd_bits_mask[1][i] |= shadow_me_mask;
4712 shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_value;
4713 shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_value;
4718 static inline bool boot_cpu_is_amd(void)
4720 WARN_ON_ONCE(!tdp_enabled);
4721 return shadow_x_mask == 0;
4725 * the direct page table on host, use as much mmu features as
4726 * possible, however, kvm currently does not do execution-protection.
4729 reset_tdp_shadow_zero_bits_mask(struct kvm_mmu *context)
4731 struct rsvd_bits_validate *shadow_zero_check;
4734 shadow_zero_check = &context->shadow_zero_check;
4736 if (boot_cpu_is_amd())
4737 __reset_rsvds_bits_mask(shadow_zero_check, reserved_hpa_bits(),
4738 context->root_role.level, false,
4739 boot_cpu_has(X86_FEATURE_GBPAGES),
4742 __reset_rsvds_bits_mask_ept(shadow_zero_check,
4743 reserved_hpa_bits(), false,
4744 max_huge_page_level);
4746 if (!shadow_me_mask)
4749 for (i = context->root_role.level; --i >= 0;) {
4750 shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask;
4751 shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask;
4756 * as the comments in reset_shadow_zero_bits_mask() except it
4757 * is the shadow page table for intel nested guest.
4760 reset_ept_shadow_zero_bits_mask(struct kvm_mmu *context, bool execonly)
4762 __reset_rsvds_bits_mask_ept(&context->shadow_zero_check,
4763 reserved_hpa_bits(), execonly,
4764 max_huge_page_level);
4767 #define BYTE_MASK(access) \
4768 ((1 & (access) ? 2 : 0) | \
4769 (2 & (access) ? 4 : 0) | \
4770 (3 & (access) ? 8 : 0) | \
4771 (4 & (access) ? 16 : 0) | \
4772 (5 & (access) ? 32 : 0) | \
4773 (6 & (access) ? 64 : 0) | \
4774 (7 & (access) ? 128 : 0))
4777 static void update_permission_bitmask(struct kvm_mmu *mmu, bool ept)
4781 const u8 x = BYTE_MASK(ACC_EXEC_MASK);
4782 const u8 w = BYTE_MASK(ACC_WRITE_MASK);
4783 const u8 u = BYTE_MASK(ACC_USER_MASK);
4785 bool cr4_smep = is_cr4_smep(mmu);
4786 bool cr4_smap = is_cr4_smap(mmu);
4787 bool cr0_wp = is_cr0_wp(mmu);
4788 bool efer_nx = is_efer_nx(mmu);
4790 for (byte = 0; byte < ARRAY_SIZE(mmu->permissions); ++byte) {
4791 unsigned pfec = byte << 1;
4794 * Each "*f" variable has a 1 bit for each UWX value
4795 * that causes a fault with the given PFEC.
4798 /* Faults from writes to non-writable pages */
4799 u8 wf = (pfec & PFERR_WRITE_MASK) ? (u8)~w : 0;
4800 /* Faults from user mode accesses to supervisor pages */
4801 u8 uf = (pfec & PFERR_USER_MASK) ? (u8)~u : 0;
4802 /* Faults from fetches of non-executable pages*/
4803 u8 ff = (pfec & PFERR_FETCH_MASK) ? (u8)~x : 0;
4804 /* Faults from kernel mode fetches of user pages */
4806 /* Faults from kernel mode accesses of user pages */
4810 /* Faults from kernel mode accesses to user pages */
4811 u8 kf = (pfec & PFERR_USER_MASK) ? 0 : u;
4813 /* Not really needed: !nx will cause pte.nx to fault */
4817 /* Allow supervisor writes if !cr0.wp */
4819 wf = (pfec & PFERR_USER_MASK) ? wf : 0;
4821 /* Disallow supervisor fetches of user code if cr4.smep */
4823 smepf = (pfec & PFERR_FETCH_MASK) ? kf : 0;
4826 * SMAP:kernel-mode data accesses from user-mode
4827 * mappings should fault. A fault is considered
4828 * as a SMAP violation if all of the following
4829 * conditions are true:
4830 * - X86_CR4_SMAP is set in CR4
4831 * - A user page is accessed
4832 * - The access is not a fetch
4833 * - The access is supervisor mode
4834 * - If implicit supervisor access or X86_EFLAGS_AC is clear
4836 * Here, we cover the first four conditions.
4837 * The fifth is computed dynamically in permission_fault();
4838 * PFERR_RSVD_MASK bit will be set in PFEC if the access is
4839 * *not* subject to SMAP restrictions.
4842 smapf = (pfec & (PFERR_RSVD_MASK|PFERR_FETCH_MASK)) ? 0 : kf;
4845 mmu->permissions[byte] = ff | uf | wf | smepf | smapf;
4850 * PKU is an additional mechanism by which the paging controls access to
4851 * user-mode addresses based on the value in the PKRU register. Protection
4852 * key violations are reported through a bit in the page fault error code.
4853 * Unlike other bits of the error code, the PK bit is not known at the
4854 * call site of e.g. gva_to_gpa; it must be computed directly in
4855 * permission_fault based on two bits of PKRU, on some machine state (CR4,
4856 * CR0, EFER, CPL), and on other bits of the error code and the page tables.
4858 * In particular the following conditions come from the error code, the
4859 * page tables and the machine state:
4860 * - PK is always zero unless CR4.PKE=1 and EFER.LMA=1
4861 * - PK is always zero if RSVD=1 (reserved bit set) or F=1 (instruction fetch)
4862 * - PK is always zero if U=0 in the page tables
4863 * - PKRU.WD is ignored if CR0.WP=0 and the access is a supervisor access.
4865 * The PKRU bitmask caches the result of these four conditions. The error
4866 * code (minus the P bit) and the page table's U bit form an index into the
4867 * PKRU bitmask. Two bits of the PKRU bitmask are then extracted and ANDed
4868 * with the two bits of the PKRU register corresponding to the protection key.
4869 * For the first three conditions above the bits will be 00, thus masking
4870 * away both AD and WD. For all reads or if the last condition holds, WD
4871 * only will be masked away.
4873 static void update_pkru_bitmask(struct kvm_mmu *mmu)
4880 if (!is_cr4_pke(mmu))
4883 wp = is_cr0_wp(mmu);
4885 for (bit = 0; bit < ARRAY_SIZE(mmu->permissions); ++bit) {
4886 unsigned pfec, pkey_bits;
4887 bool check_pkey, check_write, ff, uf, wf, pte_user;
4890 ff = pfec & PFERR_FETCH_MASK;
4891 uf = pfec & PFERR_USER_MASK;
4892 wf = pfec & PFERR_WRITE_MASK;
4894 /* PFEC.RSVD is replaced by ACC_USER_MASK. */
4895 pte_user = pfec & PFERR_RSVD_MASK;
4898 * Only need to check the access which is not an
4899 * instruction fetch and is to a user page.
4901 check_pkey = (!ff && pte_user);
4903 * write access is controlled by PKRU if it is a
4904 * user access or CR0.WP = 1.
4906 check_write = check_pkey && wf && (uf || wp);
4908 /* PKRU.AD stops both read and write access. */
4909 pkey_bits = !!check_pkey;
4910 /* PKRU.WD stops write access. */
4911 pkey_bits |= (!!check_write) << 1;
4913 mmu->pkru_mask |= (pkey_bits & 3) << pfec;
4917 static void reset_guest_paging_metadata(struct kvm_vcpu *vcpu,
4918 struct kvm_mmu *mmu)
4920 if (!is_cr0_pg(mmu))
4923 reset_guest_rsvds_bits_mask(vcpu, mmu);
4924 update_permission_bitmask(mmu, false);
4925 update_pkru_bitmask(mmu);
4928 static void paging64_init_context(struct kvm_mmu *context)
4930 context->page_fault = paging64_page_fault;
4931 context->gva_to_gpa = paging64_gva_to_gpa;
4932 context->sync_page = paging64_sync_page;
4933 context->invlpg = paging64_invlpg;
4936 static void paging32_init_context(struct kvm_mmu *context)
4938 context->page_fault = paging32_page_fault;
4939 context->gva_to_gpa = paging32_gva_to_gpa;
4940 context->sync_page = paging32_sync_page;
4941 context->invlpg = paging32_invlpg;
4944 static union kvm_cpu_role
4945 kvm_calc_cpu_role(struct kvm_vcpu *vcpu, const struct kvm_mmu_role_regs *regs)
4947 union kvm_cpu_role role = {0};
4949 role.base.access = ACC_ALL;
4950 role.base.smm = is_smm(vcpu);
4951 role.base.guest_mode = is_guest_mode(vcpu);
4954 if (!____is_cr0_pg(regs)) {
4955 role.base.direct = 1;
4959 role.base.efer_nx = ____is_efer_nx(regs);
4960 role.base.cr0_wp = ____is_cr0_wp(regs);
4961 role.base.smep_andnot_wp = ____is_cr4_smep(regs) && !____is_cr0_wp(regs);
4962 role.base.smap_andnot_wp = ____is_cr4_smap(regs) && !____is_cr0_wp(regs);
4963 role.base.has_4_byte_gpte = !____is_cr4_pae(regs);
4965 if (____is_efer_lma(regs))
4966 role.base.level = ____is_cr4_la57(regs) ? PT64_ROOT_5LEVEL
4968 else if (____is_cr4_pae(regs))
4969 role.base.level = PT32E_ROOT_LEVEL;
4971 role.base.level = PT32_ROOT_LEVEL;
4973 role.ext.cr4_smep = ____is_cr4_smep(regs);
4974 role.ext.cr4_smap = ____is_cr4_smap(regs);
4975 role.ext.cr4_pse = ____is_cr4_pse(regs);
4977 /* PKEY and LA57 are active iff long mode is active. */
4978 role.ext.cr4_pke = ____is_efer_lma(regs) && ____is_cr4_pke(regs);
4979 role.ext.cr4_la57 = ____is_efer_lma(regs) && ____is_cr4_la57(regs);
4980 role.ext.efer_lma = ____is_efer_lma(regs);
4984 static inline int kvm_mmu_get_tdp_level(struct kvm_vcpu *vcpu)
4986 /* tdp_root_level is architecture forced level, use it if nonzero */
4988 return tdp_root_level;
4990 /* Use 5-level TDP if and only if it's useful/necessary. */
4991 if (max_tdp_level == 5 && cpuid_maxphyaddr(vcpu) <= 48)
4994 return max_tdp_level;
4997 static union kvm_mmu_page_role
4998 kvm_calc_tdp_mmu_root_page_role(struct kvm_vcpu *vcpu,
4999 union kvm_cpu_role cpu_role)
5001 union kvm_mmu_page_role role = {0};
5003 role.access = ACC_ALL;
5005 role.efer_nx = true;
5006 role.smm = cpu_role.base.smm;
5007 role.guest_mode = cpu_role.base.guest_mode;
5008 role.ad_disabled = !kvm_ad_enabled();
5009 role.level = kvm_mmu_get_tdp_level(vcpu);
5011 role.has_4_byte_gpte = false;
5016 static void init_kvm_tdp_mmu(struct kvm_vcpu *vcpu,
5017 union kvm_cpu_role cpu_role)
5019 struct kvm_mmu *context = &vcpu->arch.root_mmu;
5020 union kvm_mmu_page_role root_role = kvm_calc_tdp_mmu_root_page_role(vcpu, cpu_role);
5022 if (cpu_role.as_u64 == context->cpu_role.as_u64 &&
5023 root_role.word == context->root_role.word)
5026 context->cpu_role.as_u64 = cpu_role.as_u64;
5027 context->root_role.word = root_role.word;
5028 context->page_fault = kvm_tdp_page_fault;
5029 context->sync_page = nonpaging_sync_page;
5030 context->invlpg = NULL;
5031 context->get_guest_pgd = get_cr3;
5032 context->get_pdptr = kvm_pdptr_read;
5033 context->inject_page_fault = kvm_inject_page_fault;
5035 if (!is_cr0_pg(context))
5036 context->gva_to_gpa = nonpaging_gva_to_gpa;
5037 else if (is_cr4_pae(context))
5038 context->gva_to_gpa = paging64_gva_to_gpa;
5040 context->gva_to_gpa = paging32_gva_to_gpa;
5042 reset_guest_paging_metadata(vcpu, context);
5043 reset_tdp_shadow_zero_bits_mask(context);
5046 static void shadow_mmu_init_context(struct kvm_vcpu *vcpu, struct kvm_mmu *context,
5047 union kvm_cpu_role cpu_role,
5048 union kvm_mmu_page_role root_role)
5050 if (cpu_role.as_u64 == context->cpu_role.as_u64 &&
5051 root_role.word == context->root_role.word)
5054 context->cpu_role.as_u64 = cpu_role.as_u64;
5055 context->root_role.word = root_role.word;
5057 if (!is_cr0_pg(context))
5058 nonpaging_init_context(context);
5059 else if (is_cr4_pae(context))
5060 paging64_init_context(context);
5062 paging32_init_context(context);
5064 reset_guest_paging_metadata(vcpu, context);
5065 reset_shadow_zero_bits_mask(vcpu, context);
5068 static void kvm_init_shadow_mmu(struct kvm_vcpu *vcpu,
5069 union kvm_cpu_role cpu_role)
5071 struct kvm_mmu *context = &vcpu->arch.root_mmu;
5072 union kvm_mmu_page_role root_role;
5074 root_role = cpu_role.base;
5076 /* KVM uses PAE paging whenever the guest isn't using 64-bit paging. */
5077 root_role.level = max_t(u32, root_role.level, PT32E_ROOT_LEVEL);
5080 * KVM forces EFER.NX=1 when TDP is disabled, reflect it in the MMU role.
5081 * KVM uses NX when TDP is disabled to handle a variety of scenarios,
5082 * notably for huge SPTEs if iTLB multi-hit mitigation is enabled and
5083 * to generate correct permissions for CR0.WP=0/CR4.SMEP=1/EFER.NX=0.
5084 * The iTLB multi-hit workaround can be toggled at any time, so assume
5085 * NX can be used by any non-nested shadow MMU to avoid having to reset
5088 root_role.efer_nx = true;
5090 shadow_mmu_init_context(vcpu, context, cpu_role, root_role);
5093 void kvm_init_shadow_npt_mmu(struct kvm_vcpu *vcpu, unsigned long cr0,
5094 unsigned long cr4, u64 efer, gpa_t nested_cr3)
5096 struct kvm_mmu *context = &vcpu->arch.guest_mmu;
5097 struct kvm_mmu_role_regs regs = {
5099 .cr4 = cr4 & ~X86_CR4_PKE,
5102 union kvm_cpu_role cpu_role = kvm_calc_cpu_role(vcpu, ®s);
5103 union kvm_mmu_page_role root_role;
5105 /* NPT requires CR0.PG=1. */
5106 WARN_ON_ONCE(cpu_role.base.direct);
5108 root_role = cpu_role.base;
5109 root_role.level = kvm_mmu_get_tdp_level(vcpu);
5110 if (root_role.level == PT64_ROOT_5LEVEL &&
5111 cpu_role.base.level == PT64_ROOT_4LEVEL)
5112 root_role.passthrough = 1;
5114 shadow_mmu_init_context(vcpu, context, cpu_role, root_role);
5115 kvm_mmu_new_pgd(vcpu, nested_cr3);
5117 EXPORT_SYMBOL_GPL(kvm_init_shadow_npt_mmu);
5119 static union kvm_cpu_role
5120 kvm_calc_shadow_ept_root_page_role(struct kvm_vcpu *vcpu, bool accessed_dirty,
5121 bool execonly, u8 level)
5123 union kvm_cpu_role role = {0};
5126 * KVM does not support SMM transfer monitors, and consequently does not
5127 * support the "entry to SMM" control either. role.base.smm is always 0.
5129 WARN_ON_ONCE(is_smm(vcpu));
5130 role.base.level = level;
5131 role.base.has_4_byte_gpte = false;
5132 role.base.direct = false;
5133 role.base.ad_disabled = !accessed_dirty;
5134 role.base.guest_mode = true;
5135 role.base.access = ACC_ALL;
5138 role.ext.execonly = execonly;
5144 void kvm_init_shadow_ept_mmu(struct kvm_vcpu *vcpu, bool execonly,
5145 int huge_page_level, bool accessed_dirty,
5148 struct kvm_mmu *context = &vcpu->arch.guest_mmu;
5149 u8 level = vmx_eptp_page_walk_level(new_eptp);
5150 union kvm_cpu_role new_mode =
5151 kvm_calc_shadow_ept_root_page_role(vcpu, accessed_dirty,
5154 if (new_mode.as_u64 != context->cpu_role.as_u64) {
5155 /* EPT, and thus nested EPT, does not consume CR0, CR4, nor EFER. */
5156 context->cpu_role.as_u64 = new_mode.as_u64;
5157 context->root_role.word = new_mode.base.word;
5159 context->page_fault = ept_page_fault;
5160 context->gva_to_gpa = ept_gva_to_gpa;
5161 context->sync_page = ept_sync_page;
5162 context->invlpg = ept_invlpg;
5164 update_permission_bitmask(context, true);
5165 context->pkru_mask = 0;
5166 reset_rsvds_bits_mask_ept(vcpu, context, execonly, huge_page_level);
5167 reset_ept_shadow_zero_bits_mask(context, execonly);
5170 kvm_mmu_new_pgd(vcpu, new_eptp);
5172 EXPORT_SYMBOL_GPL(kvm_init_shadow_ept_mmu);
5174 static void init_kvm_softmmu(struct kvm_vcpu *vcpu,
5175 union kvm_cpu_role cpu_role)
5177 struct kvm_mmu *context = &vcpu->arch.root_mmu;
5179 kvm_init_shadow_mmu(vcpu, cpu_role);
5181 context->get_guest_pgd = get_cr3;
5182 context->get_pdptr = kvm_pdptr_read;
5183 context->inject_page_fault = kvm_inject_page_fault;
5186 static void init_kvm_nested_mmu(struct kvm_vcpu *vcpu,
5187 union kvm_cpu_role new_mode)
5189 struct kvm_mmu *g_context = &vcpu->arch.nested_mmu;
5191 if (new_mode.as_u64 == g_context->cpu_role.as_u64)
5194 g_context->cpu_role.as_u64 = new_mode.as_u64;
5195 g_context->get_guest_pgd = get_cr3;
5196 g_context->get_pdptr = kvm_pdptr_read;
5197 g_context->inject_page_fault = kvm_inject_page_fault;
5200 * L2 page tables are never shadowed, so there is no need to sync
5203 g_context->invlpg = NULL;
5206 * Note that arch.mmu->gva_to_gpa translates l2_gpa to l1_gpa using
5207 * L1's nested page tables (e.g. EPT12). The nested translation
5208 * of l2_gva to l1_gpa is done by arch.nested_mmu.gva_to_gpa using
5209 * L2's page tables as the first level of translation and L1's
5210 * nested page tables as the second level of translation. Basically
5211 * the gva_to_gpa functions between mmu and nested_mmu are swapped.
5213 if (!is_paging(vcpu))
5214 g_context->gva_to_gpa = nonpaging_gva_to_gpa;
5215 else if (is_long_mode(vcpu))
5216 g_context->gva_to_gpa = paging64_gva_to_gpa;
5217 else if (is_pae(vcpu))
5218 g_context->gva_to_gpa = paging64_gva_to_gpa;
5220 g_context->gva_to_gpa = paging32_gva_to_gpa;
5222 reset_guest_paging_metadata(vcpu, g_context);
5225 void kvm_init_mmu(struct kvm_vcpu *vcpu)
5227 struct kvm_mmu_role_regs regs = vcpu_to_role_regs(vcpu);
5228 union kvm_cpu_role cpu_role = kvm_calc_cpu_role(vcpu, ®s);
5230 if (mmu_is_nested(vcpu))
5231 init_kvm_nested_mmu(vcpu, cpu_role);
5232 else if (tdp_enabled)
5233 init_kvm_tdp_mmu(vcpu, cpu_role);
5235 init_kvm_softmmu(vcpu, cpu_role);
5237 EXPORT_SYMBOL_GPL(kvm_init_mmu);
5239 void kvm_mmu_after_set_cpuid(struct kvm_vcpu *vcpu)
5242 * Invalidate all MMU roles to force them to reinitialize as CPUID
5243 * information is factored into reserved bit calculations.
5245 * Correctly handling multiple vCPU models with respect to paging and
5246 * physical address properties) in a single VM would require tracking
5247 * all relevant CPUID information in kvm_mmu_page_role. That is very
5248 * undesirable as it would increase the memory requirements for
5249 * gfn_track (see struct kvm_mmu_page_role comments). For now that
5250 * problem is swept under the rug; KVM's CPUID API is horrific and
5251 * it's all but impossible to solve it without introducing a new API.
5253 vcpu->arch.root_mmu.root_role.word = 0;
5254 vcpu->arch.guest_mmu.root_role.word = 0;
5255 vcpu->arch.nested_mmu.root_role.word = 0;
5256 vcpu->arch.root_mmu.cpu_role.ext.valid = 0;
5257 vcpu->arch.guest_mmu.cpu_role.ext.valid = 0;
5258 vcpu->arch.nested_mmu.cpu_role.ext.valid = 0;
5259 kvm_mmu_reset_context(vcpu);
5262 * Changing guest CPUID after KVM_RUN is forbidden, see the comment in
5263 * kvm_arch_vcpu_ioctl().
5265 KVM_BUG_ON(vcpu->arch.last_vmentry_cpu != -1, vcpu->kvm);
5268 void kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
5270 kvm_mmu_unload(vcpu);
5273 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
5275 int kvm_mmu_load(struct kvm_vcpu *vcpu)
5279 r = mmu_topup_memory_caches(vcpu, !vcpu->arch.mmu->root_role.direct);
5282 r = mmu_alloc_special_roots(vcpu);
5285 if (vcpu->arch.mmu->root_role.direct)
5286 r = mmu_alloc_direct_roots(vcpu);
5288 r = mmu_alloc_shadow_roots(vcpu);
5292 kvm_mmu_sync_roots(vcpu);
5294 kvm_mmu_load_pgd(vcpu);
5297 * Flush any TLB entries for the new root, the provenance of the root
5298 * is unknown. Even if KVM ensures there are no stale TLB entries
5299 * for a freed root, in theory another hypervisor could have left
5300 * stale entries. Flushing on alloc also allows KVM to skip the TLB
5301 * flush when freeing a root (see kvm_tdp_mmu_put_root()).
5303 static_call(kvm_x86_flush_tlb_current)(vcpu);
5308 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
5310 struct kvm *kvm = vcpu->kvm;
5312 kvm_mmu_free_roots(kvm, &vcpu->arch.root_mmu, KVM_MMU_ROOTS_ALL);
5313 WARN_ON(VALID_PAGE(vcpu->arch.root_mmu.root.hpa));
5314 kvm_mmu_free_roots(kvm, &vcpu->arch.guest_mmu, KVM_MMU_ROOTS_ALL);
5315 WARN_ON(VALID_PAGE(vcpu->arch.guest_mmu.root.hpa));
5316 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
5319 static bool is_obsolete_root(struct kvm *kvm, hpa_t root_hpa)
5321 struct kvm_mmu_page *sp;
5323 if (!VALID_PAGE(root_hpa))
5327 * When freeing obsolete roots, treat roots as obsolete if they don't
5328 * have an associated shadow page. This does mean KVM will get false
5329 * positives and free roots that don't strictly need to be freed, but
5330 * such false positives are relatively rare:
5332 * (a) only PAE paging and nested NPT has roots without shadow pages
5333 * (b) remote reloads due to a memslot update obsoletes _all_ roots
5334 * (c) KVM doesn't track previous roots for PAE paging, and the guest
5335 * is unlikely to zap an in-use PGD.
5337 sp = to_shadow_page(root_hpa);
5338 return !sp || is_obsolete_sp(kvm, sp);
5341 static void __kvm_mmu_free_obsolete_roots(struct kvm *kvm, struct kvm_mmu *mmu)
5343 unsigned long roots_to_free = 0;
5346 if (is_obsolete_root(kvm, mmu->root.hpa))
5347 roots_to_free |= KVM_MMU_ROOT_CURRENT;
5349 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
5350 if (is_obsolete_root(kvm, mmu->prev_roots[i].hpa))
5351 roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i);
5355 kvm_mmu_free_roots(kvm, mmu, roots_to_free);
5358 void kvm_mmu_free_obsolete_roots(struct kvm_vcpu *vcpu)
5360 __kvm_mmu_free_obsolete_roots(vcpu->kvm, &vcpu->arch.root_mmu);
5361 __kvm_mmu_free_obsolete_roots(vcpu->kvm, &vcpu->arch.guest_mmu);
5364 static bool need_remote_flush(u64 old, u64 new)
5366 if (!is_shadow_present_pte(old))
5368 if (!is_shadow_present_pte(new))
5370 if ((old ^ new) & SPTE_BASE_ADDR_MASK)
5372 old ^= shadow_nx_mask;
5373 new ^= shadow_nx_mask;
5374 return (old & ~new & SPTE_PERM_MASK) != 0;
5377 static u64 mmu_pte_write_fetch_gpte(struct kvm_vcpu *vcpu, gpa_t *gpa,
5384 * Assume that the pte write on a page table of the same type
5385 * as the current vcpu paging mode since we update the sptes only
5386 * when they have the same mode.
5388 if (is_pae(vcpu) && *bytes == 4) {
5389 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
5394 if (*bytes == 4 || *bytes == 8) {
5395 r = kvm_vcpu_read_guest_atomic(vcpu, *gpa, &gentry, *bytes);
5404 * If we're seeing too many writes to a page, it may no longer be a page table,
5405 * or we may be forking, in which case it is better to unmap the page.
5407 static bool detect_write_flooding(struct kvm_mmu_page *sp)
5410 * Skip write-flooding detected for the sp whose level is 1, because
5411 * it can become unsync, then the guest page is not write-protected.
5413 if (sp->role.level == PG_LEVEL_4K)
5416 atomic_inc(&sp->write_flooding_count);
5417 return atomic_read(&sp->write_flooding_count) >= 3;
5421 * Misaligned accesses are too much trouble to fix up; also, they usually
5422 * indicate a page is not used as a page table.
5424 static bool detect_write_misaligned(struct kvm_mmu_page *sp, gpa_t gpa,
5427 unsigned offset, pte_size, misaligned;
5429 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
5430 gpa, bytes, sp->role.word);
5432 offset = offset_in_page(gpa);
5433 pte_size = sp->role.has_4_byte_gpte ? 4 : 8;
5436 * Sometimes, the OS only writes the last one bytes to update status
5437 * bits, for example, in linux, andb instruction is used in clear_bit().
5439 if (!(offset & (pte_size - 1)) && bytes == 1)
5442 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
5443 misaligned |= bytes < 4;
5448 static u64 *get_written_sptes(struct kvm_mmu_page *sp, gpa_t gpa, int *nspte)
5450 unsigned page_offset, quadrant;
5454 page_offset = offset_in_page(gpa);
5455 level = sp->role.level;
5457 if (sp->role.has_4_byte_gpte) {
5458 page_offset <<= 1; /* 32->64 */
5460 * A 32-bit pde maps 4MB while the shadow pdes map
5461 * only 2MB. So we need to double the offset again
5462 * and zap two pdes instead of one.
5464 if (level == PT32_ROOT_LEVEL) {
5465 page_offset &= ~7; /* kill rounding error */
5469 quadrant = page_offset >> PAGE_SHIFT;
5470 page_offset &= ~PAGE_MASK;
5471 if (quadrant != sp->role.quadrant)
5475 spte = &sp->spt[page_offset / sizeof(*spte)];
5479 static void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
5480 const u8 *new, int bytes,
5481 struct kvm_page_track_notifier_node *node)
5483 gfn_t gfn = gpa >> PAGE_SHIFT;
5484 struct kvm_mmu_page *sp;
5485 LIST_HEAD(invalid_list);
5486 u64 entry, gentry, *spte;
5491 * If we don't have indirect shadow pages, it means no page is
5492 * write-protected, so we can exit simply.
5494 if (!READ_ONCE(vcpu->kvm->arch.indirect_shadow_pages))
5497 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
5499 write_lock(&vcpu->kvm->mmu_lock);
5501 gentry = mmu_pte_write_fetch_gpte(vcpu, &gpa, &bytes);
5503 ++vcpu->kvm->stat.mmu_pte_write;
5505 for_each_gfn_valid_sp_with_gptes(vcpu->kvm, sp, gfn) {
5506 if (detect_write_misaligned(sp, gpa, bytes) ||
5507 detect_write_flooding(sp)) {
5508 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
5509 ++vcpu->kvm->stat.mmu_flooded;
5513 spte = get_written_sptes(sp, gpa, &npte);
5519 mmu_page_zap_pte(vcpu->kvm, sp, spte, NULL);
5520 if (gentry && sp->role.level != PG_LEVEL_4K)
5521 ++vcpu->kvm->stat.mmu_pde_zapped;
5522 if (need_remote_flush(entry, *spte))
5527 kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, flush);
5528 write_unlock(&vcpu->kvm->mmu_lock);
5531 int noinline kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, u64 error_code,
5532 void *insn, int insn_len)
5534 int r, emulation_type = EMULTYPE_PF;
5535 bool direct = vcpu->arch.mmu->root_role.direct;
5537 if (WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root.hpa)))
5538 return RET_PF_RETRY;
5541 if (unlikely(error_code & PFERR_RSVD_MASK)) {
5542 r = handle_mmio_page_fault(vcpu, cr2_or_gpa, direct);
5543 if (r == RET_PF_EMULATE)
5547 if (r == RET_PF_INVALID) {
5548 r = kvm_mmu_do_page_fault(vcpu, cr2_or_gpa,
5549 lower_32_bits(error_code), false);
5550 if (KVM_BUG_ON(r == RET_PF_INVALID, vcpu->kvm))
5556 if (r != RET_PF_EMULATE)
5560 * Before emulating the instruction, check if the error code
5561 * was due to a RO violation while translating the guest page.
5562 * This can occur when using nested virtualization with nested
5563 * paging in both guests. If true, we simply unprotect the page
5564 * and resume the guest.
5566 if (vcpu->arch.mmu->root_role.direct &&
5567 (error_code & PFERR_NESTED_GUEST_PAGE) == PFERR_NESTED_GUEST_PAGE) {
5568 kvm_mmu_unprotect_page(vcpu->kvm, gpa_to_gfn(cr2_or_gpa));
5573 * vcpu->arch.mmu.page_fault returned RET_PF_EMULATE, but we can still
5574 * optimistically try to just unprotect the page and let the processor
5575 * re-execute the instruction that caused the page fault. Do not allow
5576 * retrying MMIO emulation, as it's not only pointless but could also
5577 * cause us to enter an infinite loop because the processor will keep
5578 * faulting on the non-existent MMIO address. Retrying an instruction
5579 * from a nested guest is also pointless and dangerous as we are only
5580 * explicitly shadowing L1's page tables, i.e. unprotecting something
5581 * for L1 isn't going to magically fix whatever issue cause L2 to fail.
5583 if (!mmio_info_in_cache(vcpu, cr2_or_gpa, direct) && !is_guest_mode(vcpu))
5584 emulation_type |= EMULTYPE_ALLOW_RETRY_PF;
5586 return x86_emulate_instruction(vcpu, cr2_or_gpa, emulation_type, insn,
5589 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
5591 void kvm_mmu_invalidate_gva(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
5592 gva_t gva, hpa_t root_hpa)
5596 /* It's actually a GPA for vcpu->arch.guest_mmu. */
5597 if (mmu != &vcpu->arch.guest_mmu) {
5598 /* INVLPG on a non-canonical address is a NOP according to the SDM. */
5599 if (is_noncanonical_address(gva, vcpu))
5602 static_call(kvm_x86_flush_tlb_gva)(vcpu, gva);
5608 if (root_hpa == INVALID_PAGE) {
5609 mmu->invlpg(vcpu, gva, mmu->root.hpa);
5612 * INVLPG is required to invalidate any global mappings for the VA,
5613 * irrespective of PCID. Since it would take us roughly similar amount
5614 * of work to determine whether any of the prev_root mappings of the VA
5615 * is marked global, or to just sync it blindly, so we might as well
5616 * just always sync it.
5618 * Mappings not reachable via the current cr3 or the prev_roots will be
5619 * synced when switching to that cr3, so nothing needs to be done here
5622 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5623 if (VALID_PAGE(mmu->prev_roots[i].hpa))
5624 mmu->invlpg(vcpu, gva, mmu->prev_roots[i].hpa);
5626 mmu->invlpg(vcpu, gva, root_hpa);
5630 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
5632 kvm_mmu_invalidate_gva(vcpu, vcpu->arch.walk_mmu, gva, INVALID_PAGE);
5633 ++vcpu->stat.invlpg;
5635 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
5638 void kvm_mmu_invpcid_gva(struct kvm_vcpu *vcpu, gva_t gva, unsigned long pcid)
5640 struct kvm_mmu *mmu = vcpu->arch.mmu;
5641 bool tlb_flush = false;
5644 if (pcid == kvm_get_active_pcid(vcpu)) {
5646 mmu->invlpg(vcpu, gva, mmu->root.hpa);
5650 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
5651 if (VALID_PAGE(mmu->prev_roots[i].hpa) &&
5652 pcid == kvm_get_pcid(vcpu, mmu->prev_roots[i].pgd)) {
5654 mmu->invlpg(vcpu, gva, mmu->prev_roots[i].hpa);
5660 static_call(kvm_x86_flush_tlb_gva)(vcpu, gva);
5662 ++vcpu->stat.invlpg;
5665 * Mappings not reachable via the current cr3 or the prev_roots will be
5666 * synced when switching to that cr3, so nothing needs to be done here
5671 void kvm_configure_mmu(bool enable_tdp, int tdp_forced_root_level,
5672 int tdp_max_root_level, int tdp_huge_page_level)
5674 tdp_enabled = enable_tdp;
5675 tdp_root_level = tdp_forced_root_level;
5676 max_tdp_level = tdp_max_root_level;
5679 * max_huge_page_level reflects KVM's MMU capabilities irrespective
5680 * of kernel support, e.g. KVM may be capable of using 1GB pages when
5681 * the kernel is not. But, KVM never creates a page size greater than
5682 * what is used by the kernel for any given HVA, i.e. the kernel's
5683 * capabilities are ultimately consulted by kvm_mmu_hugepage_adjust().
5686 max_huge_page_level = tdp_huge_page_level;
5687 else if (boot_cpu_has(X86_FEATURE_GBPAGES))
5688 max_huge_page_level = PG_LEVEL_1G;
5690 max_huge_page_level = PG_LEVEL_2M;
5692 EXPORT_SYMBOL_GPL(kvm_configure_mmu);
5694 /* The return value indicates if tlb flush on all vcpus is needed. */
5695 typedef bool (*slot_level_handler) (struct kvm *kvm,
5696 struct kvm_rmap_head *rmap_head,
5697 const struct kvm_memory_slot *slot);
5699 /* The caller should hold mmu-lock before calling this function. */
5700 static __always_inline bool
5701 slot_handle_level_range(struct kvm *kvm, const struct kvm_memory_slot *memslot,
5702 slot_level_handler fn, int start_level, int end_level,
5703 gfn_t start_gfn, gfn_t end_gfn, bool flush_on_yield,
5706 struct slot_rmap_walk_iterator iterator;
5708 for_each_slot_rmap_range(memslot, start_level, end_level, start_gfn,
5709 end_gfn, &iterator) {
5711 flush |= fn(kvm, iterator.rmap, memslot);
5713 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) {
5714 if (flush && flush_on_yield) {
5715 kvm_flush_remote_tlbs_with_address(kvm,
5717 iterator.gfn - start_gfn + 1);
5720 cond_resched_rwlock_write(&kvm->mmu_lock);
5727 static __always_inline bool
5728 slot_handle_level(struct kvm *kvm, const struct kvm_memory_slot *memslot,
5729 slot_level_handler fn, int start_level, int end_level,
5730 bool flush_on_yield)
5732 return slot_handle_level_range(kvm, memslot, fn, start_level,
5733 end_level, memslot->base_gfn,
5734 memslot->base_gfn + memslot->npages - 1,
5735 flush_on_yield, false);
5738 static __always_inline bool
5739 slot_handle_level_4k(struct kvm *kvm, const struct kvm_memory_slot *memslot,
5740 slot_level_handler fn, bool flush_on_yield)
5742 return slot_handle_level(kvm, memslot, fn, PG_LEVEL_4K,
5743 PG_LEVEL_4K, flush_on_yield);
5746 static void free_mmu_pages(struct kvm_mmu *mmu)
5748 if (!tdp_enabled && mmu->pae_root)
5749 set_memory_encrypted((unsigned long)mmu->pae_root, 1);
5750 free_page((unsigned long)mmu->pae_root);
5751 free_page((unsigned long)mmu->pml4_root);
5752 free_page((unsigned long)mmu->pml5_root);
5755 static int __kvm_mmu_create(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
5760 mmu->root.hpa = INVALID_PAGE;
5762 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5763 mmu->prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
5765 /* vcpu->arch.guest_mmu isn't used when !tdp_enabled. */
5766 if (!tdp_enabled && mmu == &vcpu->arch.guest_mmu)
5770 * When using PAE paging, the four PDPTEs are treated as 'root' pages,
5771 * while the PDP table is a per-vCPU construct that's allocated at MMU
5772 * creation. When emulating 32-bit mode, cr3 is only 32 bits even on
5773 * x86_64. Therefore we need to allocate the PDP table in the first
5774 * 4GB of memory, which happens to fit the DMA32 zone. TDP paging
5775 * generally doesn't use PAE paging and can skip allocating the PDP
5776 * table. The main exception, handled here, is SVM's 32-bit NPT. The
5777 * other exception is for shadowing L1's 32-bit or PAE NPT on 64-bit
5778 * KVM; that horror is handled on-demand by mmu_alloc_special_roots().
5780 if (tdp_enabled && kvm_mmu_get_tdp_level(vcpu) > PT32E_ROOT_LEVEL)
5783 page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_DMA32);
5787 mmu->pae_root = page_address(page);
5790 * CR3 is only 32 bits when PAE paging is used, thus it's impossible to
5791 * get the CPU to treat the PDPTEs as encrypted. Decrypt the page so
5792 * that KVM's writes and the CPU's reads get along. Note, this is
5793 * only necessary when using shadow paging, as 64-bit NPT can get at
5794 * the C-bit even when shadowing 32-bit NPT, and SME isn't supported
5795 * by 32-bit kernels (when KVM itself uses 32-bit NPT).
5798 set_memory_decrypted((unsigned long)mmu->pae_root, 1);
5800 WARN_ON_ONCE(shadow_me_value);
5802 for (i = 0; i < 4; ++i)
5803 mmu->pae_root[i] = INVALID_PAE_ROOT;
5808 int kvm_mmu_create(struct kvm_vcpu *vcpu)
5812 vcpu->arch.mmu_pte_list_desc_cache.kmem_cache = pte_list_desc_cache;
5813 vcpu->arch.mmu_pte_list_desc_cache.gfp_zero = __GFP_ZERO;
5815 vcpu->arch.mmu_page_header_cache.kmem_cache = mmu_page_header_cache;
5816 vcpu->arch.mmu_page_header_cache.gfp_zero = __GFP_ZERO;
5818 vcpu->arch.mmu_shadow_page_cache.gfp_zero = __GFP_ZERO;
5820 vcpu->arch.mmu = &vcpu->arch.root_mmu;
5821 vcpu->arch.walk_mmu = &vcpu->arch.root_mmu;
5823 ret = __kvm_mmu_create(vcpu, &vcpu->arch.guest_mmu);
5827 ret = __kvm_mmu_create(vcpu, &vcpu->arch.root_mmu);
5829 goto fail_allocate_root;
5833 free_mmu_pages(&vcpu->arch.guest_mmu);
5837 #define BATCH_ZAP_PAGES 10
5838 static void kvm_zap_obsolete_pages(struct kvm *kvm)
5840 struct kvm_mmu_page *sp, *node;
5841 int nr_zapped, batch = 0;
5845 list_for_each_entry_safe_reverse(sp, node,
5846 &kvm->arch.active_mmu_pages, link) {
5848 * No obsolete valid page exists before a newly created page
5849 * since active_mmu_pages is a FIFO list.
5851 if (!is_obsolete_sp(kvm, sp))
5855 * Invalid pages should never land back on the list of active
5856 * pages. Skip the bogus page, otherwise we'll get stuck in an
5857 * infinite loop if the page gets put back on the list (again).
5859 if (WARN_ON(sp->role.invalid))
5863 * No need to flush the TLB since we're only zapping shadow
5864 * pages with an obsolete generation number and all vCPUS have
5865 * loaded a new root, i.e. the shadow pages being zapped cannot
5866 * be in active use by the guest.
5868 if (batch >= BATCH_ZAP_PAGES &&
5869 cond_resched_rwlock_write(&kvm->mmu_lock)) {
5874 unstable = __kvm_mmu_prepare_zap_page(kvm, sp,
5875 &kvm->arch.zapped_obsolete_pages, &nr_zapped);
5883 * Kick all vCPUs (via remote TLB flush) before freeing the page tables
5884 * to ensure KVM is not in the middle of a lockless shadow page table
5885 * walk, which may reference the pages. The remote TLB flush itself is
5886 * not required and is simply a convenient way to kick vCPUs as needed.
5887 * KVM performs a local TLB flush when allocating a new root (see
5888 * kvm_mmu_load()), and the reload in the caller ensure no vCPUs are
5889 * running with an obsolete MMU.
5891 kvm_mmu_commit_zap_page(kvm, &kvm->arch.zapped_obsolete_pages);
5895 * Fast invalidate all shadow pages and use lock-break technique
5896 * to zap obsolete pages.
5898 * It's required when memslot is being deleted or VM is being
5899 * destroyed, in these cases, we should ensure that KVM MMU does
5900 * not use any resource of the being-deleted slot or all slots
5901 * after calling the function.
5903 static void kvm_mmu_zap_all_fast(struct kvm *kvm)
5905 lockdep_assert_held(&kvm->slots_lock);
5907 write_lock(&kvm->mmu_lock);
5908 trace_kvm_mmu_zap_all_fast(kvm);
5911 * Toggle mmu_valid_gen between '0' and '1'. Because slots_lock is
5912 * held for the entire duration of zapping obsolete pages, it's
5913 * impossible for there to be multiple invalid generations associated
5914 * with *valid* shadow pages at any given time, i.e. there is exactly
5915 * one valid generation and (at most) one invalid generation.
5917 kvm->arch.mmu_valid_gen = kvm->arch.mmu_valid_gen ? 0 : 1;
5920 * In order to ensure all vCPUs drop their soon-to-be invalid roots,
5921 * invalidating TDP MMU roots must be done while holding mmu_lock for
5922 * write and in the same critical section as making the reload request,
5923 * e.g. before kvm_zap_obsolete_pages() could drop mmu_lock and yield.
5925 if (is_tdp_mmu_enabled(kvm))
5926 kvm_tdp_mmu_invalidate_all_roots(kvm);
5929 * Notify all vcpus to reload its shadow page table and flush TLB.
5930 * Then all vcpus will switch to new shadow page table with the new
5933 * Note: we need to do this under the protection of mmu_lock,
5934 * otherwise, vcpu would purge shadow page but miss tlb flush.
5936 kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_FREE_OBSOLETE_ROOTS);
5938 kvm_zap_obsolete_pages(kvm);
5940 write_unlock(&kvm->mmu_lock);
5943 * Zap the invalidated TDP MMU roots, all SPTEs must be dropped before
5944 * returning to the caller, e.g. if the zap is in response to a memslot
5945 * deletion, mmu_notifier callbacks will be unable to reach the SPTEs
5946 * associated with the deleted memslot once the update completes, and
5947 * Deferring the zap until the final reference to the root is put would
5948 * lead to use-after-free.
5950 if (is_tdp_mmu_enabled(kvm))
5951 kvm_tdp_mmu_zap_invalidated_roots(kvm);
5954 static bool kvm_has_zapped_obsolete_pages(struct kvm *kvm)
5956 return unlikely(!list_empty_careful(&kvm->arch.zapped_obsolete_pages));
5959 static void kvm_mmu_invalidate_zap_pages_in_memslot(struct kvm *kvm,
5960 struct kvm_memory_slot *slot,
5961 struct kvm_page_track_notifier_node *node)
5963 kvm_mmu_zap_all_fast(kvm);
5966 int kvm_mmu_init_vm(struct kvm *kvm)
5968 struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
5971 INIT_LIST_HEAD(&kvm->arch.active_mmu_pages);
5972 INIT_LIST_HEAD(&kvm->arch.zapped_obsolete_pages);
5973 INIT_LIST_HEAD(&kvm->arch.lpage_disallowed_mmu_pages);
5974 spin_lock_init(&kvm->arch.mmu_unsync_pages_lock);
5976 r = kvm_mmu_init_tdp_mmu(kvm);
5980 node->track_write = kvm_mmu_pte_write;
5981 node->track_flush_slot = kvm_mmu_invalidate_zap_pages_in_memslot;
5982 kvm_page_track_register_notifier(kvm, node);
5984 kvm->arch.split_page_header_cache.kmem_cache = mmu_page_header_cache;
5985 kvm->arch.split_page_header_cache.gfp_zero = __GFP_ZERO;
5987 kvm->arch.split_shadow_page_cache.gfp_zero = __GFP_ZERO;
5989 kvm->arch.split_desc_cache.kmem_cache = pte_list_desc_cache;
5990 kvm->arch.split_desc_cache.gfp_zero = __GFP_ZERO;
5995 static void mmu_free_vm_memory_caches(struct kvm *kvm)
5997 kvm_mmu_free_memory_cache(&kvm->arch.split_desc_cache);
5998 kvm_mmu_free_memory_cache(&kvm->arch.split_page_header_cache);
5999 kvm_mmu_free_memory_cache(&kvm->arch.split_shadow_page_cache);
6002 void kvm_mmu_uninit_vm(struct kvm *kvm)
6004 struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
6006 kvm_page_track_unregister_notifier(kvm, node);
6008 kvm_mmu_uninit_tdp_mmu(kvm);
6010 mmu_free_vm_memory_caches(kvm);
6013 static bool kvm_rmap_zap_gfn_range(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end)
6015 const struct kvm_memory_slot *memslot;
6016 struct kvm_memslots *slots;
6017 struct kvm_memslot_iter iter;
6022 if (!kvm_memslots_have_rmaps(kvm))
6025 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
6026 slots = __kvm_memslots(kvm, i);
6028 kvm_for_each_memslot_in_gfn_range(&iter, slots, gfn_start, gfn_end) {
6029 memslot = iter.slot;
6030 start = max(gfn_start, memslot->base_gfn);
6031 end = min(gfn_end, memslot->base_gfn + memslot->npages);
6032 if (WARN_ON_ONCE(start >= end))
6035 flush = slot_handle_level_range(kvm, memslot, __kvm_zap_rmap,
6036 PG_LEVEL_4K, KVM_MAX_HUGEPAGE_LEVEL,
6037 start, end - 1, true, flush);
6045 * Invalidate (zap) SPTEs that cover GFNs from gfn_start and up to gfn_end
6046 * (not including it)
6048 void kvm_zap_gfn_range(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end)
6053 if (WARN_ON_ONCE(gfn_end <= gfn_start))
6056 write_lock(&kvm->mmu_lock);
6058 kvm_inc_notifier_count(kvm, gfn_start, gfn_end);
6060 flush = kvm_rmap_zap_gfn_range(kvm, gfn_start, gfn_end);
6062 if (is_tdp_mmu_enabled(kvm)) {
6063 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
6064 flush = kvm_tdp_mmu_zap_leafs(kvm, i, gfn_start,
6065 gfn_end, true, flush);
6069 kvm_flush_remote_tlbs_with_address(kvm, gfn_start,
6070 gfn_end - gfn_start);
6072 kvm_dec_notifier_count(kvm, gfn_start, gfn_end);
6074 write_unlock(&kvm->mmu_lock);
6077 static bool slot_rmap_write_protect(struct kvm *kvm,
6078 struct kvm_rmap_head *rmap_head,
6079 const struct kvm_memory_slot *slot)
6081 return rmap_write_protect(rmap_head, false);
6084 void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
6085 const struct kvm_memory_slot *memslot,
6090 if (kvm_memslots_have_rmaps(kvm)) {
6091 write_lock(&kvm->mmu_lock);
6092 flush = slot_handle_level(kvm, memslot, slot_rmap_write_protect,
6093 start_level, KVM_MAX_HUGEPAGE_LEVEL,
6095 write_unlock(&kvm->mmu_lock);
6098 if (is_tdp_mmu_enabled(kvm)) {
6099 read_lock(&kvm->mmu_lock);
6100 flush |= kvm_tdp_mmu_wrprot_slot(kvm, memslot, start_level);
6101 read_unlock(&kvm->mmu_lock);
6105 * Flush TLBs if any SPTEs had to be write-protected to ensure that
6106 * guest writes are reflected in the dirty bitmap before the memslot
6107 * update completes, i.e. before enabling dirty logging is visible to
6110 * Perform the TLB flush outside the mmu_lock to reduce the amount of
6111 * time the lock is held. However, this does mean that another CPU can
6112 * now grab mmu_lock and encounter a write-protected SPTE while CPUs
6113 * still have a writable mapping for the associated GFN in their TLB.
6115 * This is safe but requires KVM to be careful when making decisions
6116 * based on the write-protection status of an SPTE. Specifically, KVM
6117 * also write-protects SPTEs to monitor changes to guest page tables
6118 * during shadow paging, and must guarantee no CPUs can write to those
6119 * page before the lock is dropped. As mentioned in the previous
6120 * paragraph, a write-protected SPTE is no guarantee that CPU cannot
6121 * perform writes. So to determine if a TLB flush is truly required, KVM
6122 * will clear a separate software-only bit (MMU-writable) and skip the
6123 * flush if-and-only-if this bit was already clear.
6125 * See is_writable_pte() for more details.
6128 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
6131 static inline bool need_topup(struct kvm_mmu_memory_cache *cache, int min)
6133 return kvm_mmu_memory_cache_nr_free_objects(cache) < min;
6136 static bool need_topup_split_caches_or_resched(struct kvm *kvm)
6138 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock))
6142 * In the worst case, SPLIT_DESC_CACHE_MIN_NR_OBJECTS descriptors are needed
6143 * to split a single huge page. Calculating how many are actually needed
6144 * is possible but not worth the complexity.
6146 return need_topup(&kvm->arch.split_desc_cache, SPLIT_DESC_CACHE_MIN_NR_OBJECTS) ||
6147 need_topup(&kvm->arch.split_page_header_cache, 1) ||
6148 need_topup(&kvm->arch.split_shadow_page_cache, 1);
6151 static int topup_split_caches(struct kvm *kvm)
6154 * Allocating rmap list entries when splitting huge pages for nested
6155 * MMUs is uncommon as KVM needs to use a list if and only if there is
6156 * more than one rmap entry for a gfn, i.e. requires an L1 gfn to be
6157 * aliased by multiple L2 gfns and/or from multiple nested roots with
6158 * different roles. Aliasing gfns when using TDP is atypical for VMMs;
6159 * a few gfns are often aliased during boot, e.g. when remapping BIOS,
6160 * but aliasing rarely occurs post-boot or for many gfns. If there is
6161 * only one rmap entry, rmap->val points directly at that one entry and
6162 * doesn't need to allocate a list. Buffer the cache by the default
6163 * capacity so that KVM doesn't have to drop mmu_lock to topup if KVM
6164 * encounters an aliased gfn or two.
6166 const int capacity = SPLIT_DESC_CACHE_MIN_NR_OBJECTS +
6167 KVM_ARCH_NR_OBJS_PER_MEMORY_CACHE;
6170 lockdep_assert_held(&kvm->slots_lock);
6172 r = __kvm_mmu_topup_memory_cache(&kvm->arch.split_desc_cache, capacity,
6173 SPLIT_DESC_CACHE_MIN_NR_OBJECTS);
6177 r = kvm_mmu_topup_memory_cache(&kvm->arch.split_page_header_cache, 1);
6181 return kvm_mmu_topup_memory_cache(&kvm->arch.split_shadow_page_cache, 1);
6184 static struct kvm_mmu_page *shadow_mmu_get_sp_for_split(struct kvm *kvm, u64 *huge_sptep)
6186 struct kvm_mmu_page *huge_sp = sptep_to_sp(huge_sptep);
6187 struct shadow_page_caches caches = {};
6188 union kvm_mmu_page_role role;
6189 unsigned int access;
6192 gfn = kvm_mmu_page_get_gfn(huge_sp, spte_index(huge_sptep));
6193 access = kvm_mmu_page_get_access(huge_sp, spte_index(huge_sptep));
6196 * Note, huge page splitting always uses direct shadow pages, regardless
6197 * of whether the huge page itself is mapped by a direct or indirect
6198 * shadow page, since the huge page region itself is being directly
6199 * mapped with smaller pages.
6201 role = kvm_mmu_child_role(huge_sptep, /*direct=*/true, access);
6203 /* Direct SPs do not require a shadowed_info_cache. */
6204 caches.page_header_cache = &kvm->arch.split_page_header_cache;
6205 caches.shadow_page_cache = &kvm->arch.split_shadow_page_cache;
6207 /* Safe to pass NULL for vCPU since requesting a direct SP. */
6208 return __kvm_mmu_get_shadow_page(kvm, NULL, &caches, gfn, role);
6211 static void shadow_mmu_split_huge_page(struct kvm *kvm,
6212 const struct kvm_memory_slot *slot,
6216 struct kvm_mmu_memory_cache *cache = &kvm->arch.split_desc_cache;
6217 u64 huge_spte = READ_ONCE(*huge_sptep);
6218 struct kvm_mmu_page *sp;
6224 sp = shadow_mmu_get_sp_for_split(kvm, huge_sptep);
6226 for (index = 0; index < SPTE_ENT_PER_PAGE; index++) {
6227 sptep = &sp->spt[index];
6228 gfn = kvm_mmu_page_get_gfn(sp, index);
6231 * The SP may already have populated SPTEs, e.g. if this huge
6232 * page is aliased by multiple sptes with the same access
6233 * permissions. These entries are guaranteed to map the same
6234 * gfn-to-pfn translation since the SP is direct, so no need to
6237 * However, if a given SPTE points to a lower level page table,
6238 * that lower level page table may only be partially populated.
6239 * Installing such SPTEs would effectively unmap a potion of the
6240 * huge page. Unmapping guest memory always requires a TLB flush
6241 * since a subsequent operation on the unmapped regions would
6242 * fail to detect the need to flush.
6244 if (is_shadow_present_pte(*sptep)) {
6245 flush |= !is_last_spte(*sptep, sp->role.level);
6249 spte = make_huge_page_split_spte(kvm, huge_spte, sp->role, index);
6250 mmu_spte_set(sptep, spte);
6251 __rmap_add(kvm, cache, slot, sptep, gfn, sp->role.access);
6254 __link_shadow_page(kvm, cache, huge_sptep, sp, flush);
6257 static int shadow_mmu_try_split_huge_page(struct kvm *kvm,
6258 const struct kvm_memory_slot *slot,
6261 struct kvm_mmu_page *huge_sp = sptep_to_sp(huge_sptep);
6266 /* Grab information for the tracepoint before dropping the MMU lock. */
6267 gfn = kvm_mmu_page_get_gfn(huge_sp, spte_index(huge_sptep));
6268 level = huge_sp->role.level;
6271 if (kvm_mmu_available_pages(kvm) <= KVM_MIN_FREE_MMU_PAGES) {
6276 if (need_topup_split_caches_or_resched(kvm)) {
6277 write_unlock(&kvm->mmu_lock);
6280 * If the topup succeeds, return -EAGAIN to indicate that the
6281 * rmap iterator should be restarted because the MMU lock was
6284 r = topup_split_caches(kvm) ?: -EAGAIN;
6285 write_lock(&kvm->mmu_lock);
6289 shadow_mmu_split_huge_page(kvm, slot, huge_sptep);
6292 trace_kvm_mmu_split_huge_page(gfn, spte, level, r);
6296 static bool shadow_mmu_try_split_huge_pages(struct kvm *kvm,
6297 struct kvm_rmap_head *rmap_head,
6298 const struct kvm_memory_slot *slot)
6300 struct rmap_iterator iter;
6301 struct kvm_mmu_page *sp;
6306 for_each_rmap_spte(rmap_head, &iter, huge_sptep) {
6307 sp = sptep_to_sp(huge_sptep);
6309 /* TDP MMU is enabled, so rmap only contains nested MMU SPs. */
6310 if (WARN_ON_ONCE(!sp->role.guest_mode))
6313 /* The rmaps should never contain non-leaf SPTEs. */
6314 if (WARN_ON_ONCE(!is_large_pte(*huge_sptep)))
6317 /* SPs with level >PG_LEVEL_4K should never by unsync. */
6318 if (WARN_ON_ONCE(sp->unsync))
6321 /* Don't bother splitting huge pages on invalid SPs. */
6322 if (sp->role.invalid)
6325 r = shadow_mmu_try_split_huge_page(kvm, slot, huge_sptep);
6328 * The split succeeded or needs to be retried because the MMU
6329 * lock was dropped. Either way, restart the iterator to get it
6330 * back into a consistent state.
6332 if (!r || r == -EAGAIN)
6335 /* The split failed and shouldn't be retried (e.g. -ENOMEM). */
6342 static void kvm_shadow_mmu_try_split_huge_pages(struct kvm *kvm,
6343 const struct kvm_memory_slot *slot,
6344 gfn_t start, gfn_t end,
6350 * Split huge pages starting with KVM_MAX_HUGEPAGE_LEVEL and working
6351 * down to the target level. This ensures pages are recursively split
6352 * all the way to the target level. There's no need to split pages
6353 * already at the target level.
6355 for (level = KVM_MAX_HUGEPAGE_LEVEL; level > target_level; level--) {
6356 slot_handle_level_range(kvm, slot, shadow_mmu_try_split_huge_pages,
6357 level, level, start, end - 1, true, false);
6361 /* Must be called with the mmu_lock held in write-mode. */
6362 void kvm_mmu_try_split_huge_pages(struct kvm *kvm,
6363 const struct kvm_memory_slot *memslot,
6367 if (!is_tdp_mmu_enabled(kvm))
6370 if (kvm_memslots_have_rmaps(kvm))
6371 kvm_shadow_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level);
6373 kvm_tdp_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level, false);
6376 * A TLB flush is unnecessary at this point for the same resons as in
6377 * kvm_mmu_slot_try_split_huge_pages().
6381 void kvm_mmu_slot_try_split_huge_pages(struct kvm *kvm,
6382 const struct kvm_memory_slot *memslot,
6385 u64 start = memslot->base_gfn;
6386 u64 end = start + memslot->npages;
6388 if (!is_tdp_mmu_enabled(kvm))
6391 if (kvm_memslots_have_rmaps(kvm)) {
6392 write_lock(&kvm->mmu_lock);
6393 kvm_shadow_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level);
6394 write_unlock(&kvm->mmu_lock);
6397 read_lock(&kvm->mmu_lock);
6398 kvm_tdp_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level, true);
6399 read_unlock(&kvm->mmu_lock);
6402 * No TLB flush is necessary here. KVM will flush TLBs after
6403 * write-protecting and/or clearing dirty on the newly split SPTEs to
6404 * ensure that guest writes are reflected in the dirty log before the
6405 * ioctl to enable dirty logging on this memslot completes. Since the
6406 * split SPTEs retain the write and dirty bits of the huge SPTE, it is
6407 * safe for KVM to decide if a TLB flush is necessary based on the split
6412 static bool kvm_mmu_zap_collapsible_spte(struct kvm *kvm,
6413 struct kvm_rmap_head *rmap_head,
6414 const struct kvm_memory_slot *slot)
6417 struct rmap_iterator iter;
6418 int need_tlb_flush = 0;
6420 struct kvm_mmu_page *sp;
6423 for_each_rmap_spte(rmap_head, &iter, sptep) {
6424 sp = sptep_to_sp(sptep);
6425 pfn = spte_to_pfn(*sptep);
6428 * We cannot do huge page mapping for indirect shadow pages,
6429 * which are found on the last rmap (level = 1) when not using
6430 * tdp; such shadow pages are synced with the page table in
6431 * the guest, and the guest page table is using 4K page size
6432 * mapping if the indirect sp has level = 1.
6434 if (sp->role.direct &&
6435 sp->role.level < kvm_mmu_max_mapping_level(kvm, slot, sp->gfn,
6437 kvm_zap_one_rmap_spte(kvm, rmap_head, sptep);
6439 if (kvm_available_flush_tlb_with_range())
6440 kvm_flush_remote_tlbs_with_address(kvm, sp->gfn,
6441 KVM_PAGES_PER_HPAGE(sp->role.level));
6449 return need_tlb_flush;
6452 static void kvm_rmap_zap_collapsible_sptes(struct kvm *kvm,
6453 const struct kvm_memory_slot *slot)
6456 * Note, use KVM_MAX_HUGEPAGE_LEVEL - 1 since there's no need to zap
6457 * pages that are already mapped at the maximum hugepage level.
6459 if (slot_handle_level(kvm, slot, kvm_mmu_zap_collapsible_spte,
6460 PG_LEVEL_4K, KVM_MAX_HUGEPAGE_LEVEL - 1, true))
6461 kvm_arch_flush_remote_tlbs_memslot(kvm, slot);
6464 void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm,
6465 const struct kvm_memory_slot *slot)
6467 if (kvm_memslots_have_rmaps(kvm)) {
6468 write_lock(&kvm->mmu_lock);
6469 kvm_rmap_zap_collapsible_sptes(kvm, slot);
6470 write_unlock(&kvm->mmu_lock);
6473 if (is_tdp_mmu_enabled(kvm)) {
6474 read_lock(&kvm->mmu_lock);
6475 kvm_tdp_mmu_zap_collapsible_sptes(kvm, slot);
6476 read_unlock(&kvm->mmu_lock);
6480 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm,
6481 const struct kvm_memory_slot *memslot)
6484 * All current use cases for flushing the TLBs for a specific memslot
6485 * related to dirty logging, and many do the TLB flush out of mmu_lock.
6486 * The interaction between the various operations on memslot must be
6487 * serialized by slots_locks to ensure the TLB flush from one operation
6488 * is observed by any other operation on the same memslot.
6490 lockdep_assert_held(&kvm->slots_lock);
6491 kvm_flush_remote_tlbs_with_address(kvm, memslot->base_gfn,
6495 void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm,
6496 const struct kvm_memory_slot *memslot)
6500 if (kvm_memslots_have_rmaps(kvm)) {
6501 write_lock(&kvm->mmu_lock);
6503 * Clear dirty bits only on 4k SPTEs since the legacy MMU only
6504 * support dirty logging at a 4k granularity.
6506 flush = slot_handle_level_4k(kvm, memslot, __rmap_clear_dirty, false);
6507 write_unlock(&kvm->mmu_lock);
6510 if (is_tdp_mmu_enabled(kvm)) {
6511 read_lock(&kvm->mmu_lock);
6512 flush |= kvm_tdp_mmu_clear_dirty_slot(kvm, memslot);
6513 read_unlock(&kvm->mmu_lock);
6517 * It's also safe to flush TLBs out of mmu lock here as currently this
6518 * function is only used for dirty logging, in which case flushing TLB
6519 * out of mmu lock also guarantees no dirty pages will be lost in
6523 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
6526 void kvm_mmu_zap_all(struct kvm *kvm)
6528 struct kvm_mmu_page *sp, *node;
6529 LIST_HEAD(invalid_list);
6532 write_lock(&kvm->mmu_lock);
6534 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link) {
6535 if (WARN_ON(sp->role.invalid))
6537 if (__kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list, &ign))
6539 if (cond_resched_rwlock_write(&kvm->mmu_lock))
6543 kvm_mmu_commit_zap_page(kvm, &invalid_list);
6545 if (is_tdp_mmu_enabled(kvm))
6546 kvm_tdp_mmu_zap_all(kvm);
6548 write_unlock(&kvm->mmu_lock);
6551 void kvm_mmu_invalidate_mmio_sptes(struct kvm *kvm, u64 gen)
6553 WARN_ON(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS);
6555 gen &= MMIO_SPTE_GEN_MASK;
6558 * Generation numbers are incremented in multiples of the number of
6559 * address spaces in order to provide unique generations across all
6560 * address spaces. Strip what is effectively the address space
6561 * modifier prior to checking for a wrap of the MMIO generation so
6562 * that a wrap in any address space is detected.
6564 gen &= ~((u64)KVM_ADDRESS_SPACE_NUM - 1);
6567 * The very rare case: if the MMIO generation number has wrapped,
6568 * zap all shadow pages.
6570 if (unlikely(gen == 0)) {
6571 kvm_debug_ratelimited("kvm: zapping shadow pages for mmio generation wraparound\n");
6572 kvm_mmu_zap_all_fast(kvm);
6576 static unsigned long
6577 mmu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
6580 int nr_to_scan = sc->nr_to_scan;
6581 unsigned long freed = 0;
6583 mutex_lock(&kvm_lock);
6585 list_for_each_entry(kvm, &vm_list, vm_list) {
6587 LIST_HEAD(invalid_list);
6590 * Never scan more than sc->nr_to_scan VM instances.
6591 * Will not hit this condition practically since we do not try
6592 * to shrink more than one VM and it is very unlikely to see
6593 * !n_used_mmu_pages so many times.
6598 * n_used_mmu_pages is accessed without holding kvm->mmu_lock
6599 * here. We may skip a VM instance errorneosly, but we do not
6600 * want to shrink a VM that only started to populate its MMU
6603 if (!kvm->arch.n_used_mmu_pages &&
6604 !kvm_has_zapped_obsolete_pages(kvm))
6607 idx = srcu_read_lock(&kvm->srcu);
6608 write_lock(&kvm->mmu_lock);
6610 if (kvm_has_zapped_obsolete_pages(kvm)) {
6611 kvm_mmu_commit_zap_page(kvm,
6612 &kvm->arch.zapped_obsolete_pages);
6616 freed = kvm_mmu_zap_oldest_mmu_pages(kvm, sc->nr_to_scan);
6619 write_unlock(&kvm->mmu_lock);
6620 srcu_read_unlock(&kvm->srcu, idx);
6623 * unfair on small ones
6624 * per-vm shrinkers cry out
6625 * sadness comes quickly
6627 list_move_tail(&kvm->vm_list, &vm_list);
6631 mutex_unlock(&kvm_lock);
6635 static unsigned long
6636 mmu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
6638 return percpu_counter_read_positive(&kvm_total_used_mmu_pages);
6641 static struct shrinker mmu_shrinker = {
6642 .count_objects = mmu_shrink_count,
6643 .scan_objects = mmu_shrink_scan,
6644 .seeks = DEFAULT_SEEKS * 10,
6647 static void mmu_destroy_caches(void)
6649 kmem_cache_destroy(pte_list_desc_cache);
6650 kmem_cache_destroy(mmu_page_header_cache);
6653 static bool get_nx_auto_mode(void)
6655 /* Return true when CPU has the bug, and mitigations are ON */
6656 return boot_cpu_has_bug(X86_BUG_ITLB_MULTIHIT) && !cpu_mitigations_off();
6659 static void __set_nx_huge_pages(bool val)
6661 nx_huge_pages = itlb_multihit_kvm_mitigation = val;
6664 static int set_nx_huge_pages(const char *val, const struct kernel_param *kp)
6666 bool old_val = nx_huge_pages;
6669 /* In "auto" mode deploy workaround only if CPU has the bug. */
6670 if (sysfs_streq(val, "off"))
6672 else if (sysfs_streq(val, "force"))
6674 else if (sysfs_streq(val, "auto"))
6675 new_val = get_nx_auto_mode();
6676 else if (strtobool(val, &new_val) < 0)
6679 __set_nx_huge_pages(new_val);
6681 if (new_val != old_val) {
6684 mutex_lock(&kvm_lock);
6686 list_for_each_entry(kvm, &vm_list, vm_list) {
6687 mutex_lock(&kvm->slots_lock);
6688 kvm_mmu_zap_all_fast(kvm);
6689 mutex_unlock(&kvm->slots_lock);
6691 wake_up_process(kvm->arch.nx_lpage_recovery_thread);
6693 mutex_unlock(&kvm_lock);
6700 * nx_huge_pages needs to be resolved to true/false when kvm.ko is loaded, as
6701 * its default value of -1 is technically undefined behavior for a boolean.
6703 void kvm_mmu_x86_module_init(void)
6705 if (nx_huge_pages == -1)
6706 __set_nx_huge_pages(get_nx_auto_mode());
6710 * The bulk of the MMU initialization is deferred until the vendor module is
6711 * loaded as many of the masks/values may be modified by VMX or SVM, i.e. need
6712 * to be reset when a potentially different vendor module is loaded.
6714 int kvm_mmu_vendor_module_init(void)
6719 * MMU roles use union aliasing which is, generally speaking, an
6720 * undefined behavior. However, we supposedly know how compilers behave
6721 * and the current status quo is unlikely to change. Guardians below are
6722 * supposed to let us know if the assumption becomes false.
6724 BUILD_BUG_ON(sizeof(union kvm_mmu_page_role) != sizeof(u32));
6725 BUILD_BUG_ON(sizeof(union kvm_mmu_extended_role) != sizeof(u32));
6726 BUILD_BUG_ON(sizeof(union kvm_cpu_role) != sizeof(u64));
6728 kvm_mmu_reset_all_pte_masks();
6730 pte_list_desc_cache = kmem_cache_create("pte_list_desc",
6731 sizeof(struct pte_list_desc),
6732 0, SLAB_ACCOUNT, NULL);
6733 if (!pte_list_desc_cache)
6736 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
6737 sizeof(struct kvm_mmu_page),
6738 0, SLAB_ACCOUNT, NULL);
6739 if (!mmu_page_header_cache)
6742 if (percpu_counter_init(&kvm_total_used_mmu_pages, 0, GFP_KERNEL))
6745 ret = register_shrinker(&mmu_shrinker);
6752 mmu_destroy_caches();
6756 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
6758 kvm_mmu_unload(vcpu);
6759 free_mmu_pages(&vcpu->arch.root_mmu);
6760 free_mmu_pages(&vcpu->arch.guest_mmu);
6761 mmu_free_memory_caches(vcpu);
6764 void kvm_mmu_vendor_module_exit(void)
6766 mmu_destroy_caches();
6767 percpu_counter_destroy(&kvm_total_used_mmu_pages);
6768 unregister_shrinker(&mmu_shrinker);
6772 * Calculate the effective recovery period, accounting for '0' meaning "let KVM
6773 * select a halving time of 1 hour". Returns true if recovery is enabled.
6775 static bool calc_nx_huge_pages_recovery_period(uint *period)
6778 * Use READ_ONCE to get the params, this may be called outside of the
6779 * param setters, e.g. by the kthread to compute its next timeout.
6781 bool enabled = READ_ONCE(nx_huge_pages);
6782 uint ratio = READ_ONCE(nx_huge_pages_recovery_ratio);
6784 if (!enabled || !ratio)
6787 *period = READ_ONCE(nx_huge_pages_recovery_period_ms);
6789 /* Make sure the period is not less than one second. */
6790 ratio = min(ratio, 3600u);
6791 *period = 60 * 60 * 1000 / ratio;
6796 static int set_nx_huge_pages_recovery_param(const char *val, const struct kernel_param *kp)
6798 bool was_recovery_enabled, is_recovery_enabled;
6799 uint old_period, new_period;
6802 was_recovery_enabled = calc_nx_huge_pages_recovery_period(&old_period);
6804 err = param_set_uint(val, kp);
6808 is_recovery_enabled = calc_nx_huge_pages_recovery_period(&new_period);
6810 if (is_recovery_enabled &&
6811 (!was_recovery_enabled || old_period > new_period)) {
6814 mutex_lock(&kvm_lock);
6816 list_for_each_entry(kvm, &vm_list, vm_list)
6817 wake_up_process(kvm->arch.nx_lpage_recovery_thread);
6819 mutex_unlock(&kvm_lock);
6825 static void kvm_recover_nx_lpages(struct kvm *kvm)
6827 unsigned long nx_lpage_splits = kvm->stat.nx_lpage_splits;
6829 struct kvm_mmu_page *sp;
6831 LIST_HEAD(invalid_list);
6835 rcu_idx = srcu_read_lock(&kvm->srcu);
6836 write_lock(&kvm->mmu_lock);
6839 * Zapping TDP MMU shadow pages, including the remote TLB flush, must
6840 * be done under RCU protection, because the pages are freed via RCU
6845 ratio = READ_ONCE(nx_huge_pages_recovery_ratio);
6846 to_zap = ratio ? DIV_ROUND_UP(nx_lpage_splits, ratio) : 0;
6847 for ( ; to_zap; --to_zap) {
6848 if (list_empty(&kvm->arch.lpage_disallowed_mmu_pages))
6852 * We use a separate list instead of just using active_mmu_pages
6853 * because the number of lpage_disallowed pages is expected to
6854 * be relatively small compared to the total.
6856 sp = list_first_entry(&kvm->arch.lpage_disallowed_mmu_pages,
6857 struct kvm_mmu_page,
6858 lpage_disallowed_link);
6859 WARN_ON_ONCE(!sp->lpage_disallowed);
6860 if (is_tdp_mmu_page(sp)) {
6861 flush |= kvm_tdp_mmu_zap_sp(kvm, sp);
6863 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
6864 WARN_ON_ONCE(sp->lpage_disallowed);
6867 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) {
6868 kvm_mmu_remote_flush_or_zap(kvm, &invalid_list, flush);
6871 cond_resched_rwlock_write(&kvm->mmu_lock);
6877 kvm_mmu_remote_flush_or_zap(kvm, &invalid_list, flush);
6881 write_unlock(&kvm->mmu_lock);
6882 srcu_read_unlock(&kvm->srcu, rcu_idx);
6885 static long get_nx_lpage_recovery_timeout(u64 start_time)
6890 enabled = calc_nx_huge_pages_recovery_period(&period);
6892 return enabled ? start_time + msecs_to_jiffies(period) - get_jiffies_64()
6893 : MAX_SCHEDULE_TIMEOUT;
6896 static int kvm_nx_lpage_recovery_worker(struct kvm *kvm, uintptr_t data)
6899 long remaining_time;
6902 start_time = get_jiffies_64();
6903 remaining_time = get_nx_lpage_recovery_timeout(start_time);
6905 set_current_state(TASK_INTERRUPTIBLE);
6906 while (!kthread_should_stop() && remaining_time > 0) {
6907 schedule_timeout(remaining_time);
6908 remaining_time = get_nx_lpage_recovery_timeout(start_time);
6909 set_current_state(TASK_INTERRUPTIBLE);
6912 set_current_state(TASK_RUNNING);
6914 if (kthread_should_stop())
6917 kvm_recover_nx_lpages(kvm);
6921 int kvm_mmu_post_init_vm(struct kvm *kvm)
6925 err = kvm_vm_create_worker_thread(kvm, kvm_nx_lpage_recovery_worker, 0,
6926 "kvm-nx-lpage-recovery",
6927 &kvm->arch.nx_lpage_recovery_thread);
6929 kthread_unpark(kvm->arch.nx_lpage_recovery_thread);
6934 void kvm_mmu_pre_destroy_vm(struct kvm *kvm)
6936 if (kvm->arch.nx_lpage_recovery_thread)
6937 kthread_stop(kvm->arch.nx_lpage_recovery_thread);