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>
58 extern bool itlb_multihit_kvm_mitigation;
60 int __read_mostly nx_huge_pages = -1;
61 static uint __read_mostly nx_huge_pages_recovery_period_ms;
62 #ifdef CONFIG_PREEMPT_RT
63 /* Recovery can cause latency spikes, disable it for PREEMPT_RT. */
64 static uint __read_mostly nx_huge_pages_recovery_ratio = 0;
66 static uint __read_mostly nx_huge_pages_recovery_ratio = 60;
69 static int set_nx_huge_pages(const char *val, const struct kernel_param *kp);
70 static int set_nx_huge_pages_recovery_param(const char *val, const struct kernel_param *kp);
72 static const struct kernel_param_ops nx_huge_pages_ops = {
73 .set = set_nx_huge_pages,
74 .get = param_get_bool,
77 static const struct kernel_param_ops nx_huge_pages_recovery_param_ops = {
78 .set = set_nx_huge_pages_recovery_param,
79 .get = param_get_uint,
82 module_param_cb(nx_huge_pages, &nx_huge_pages_ops, &nx_huge_pages, 0644);
83 __MODULE_PARM_TYPE(nx_huge_pages, "bool");
84 module_param_cb(nx_huge_pages_recovery_ratio, &nx_huge_pages_recovery_param_ops,
85 &nx_huge_pages_recovery_ratio, 0644);
86 __MODULE_PARM_TYPE(nx_huge_pages_recovery_ratio, "uint");
87 module_param_cb(nx_huge_pages_recovery_period_ms, &nx_huge_pages_recovery_param_ops,
88 &nx_huge_pages_recovery_period_ms, 0644);
89 __MODULE_PARM_TYPE(nx_huge_pages_recovery_period_ms, "uint");
91 static bool __read_mostly force_flush_and_sync_on_reuse;
92 module_param_named(flush_on_reuse, force_flush_and_sync_on_reuse, bool, 0644);
95 * When setting this variable to true it enables Two-Dimensional-Paging
96 * where the hardware walks 2 page tables:
97 * 1. the guest-virtual to guest-physical
98 * 2. while doing 1. it walks guest-physical to host-physical
99 * If the hardware supports that we don't need to do shadow paging.
101 bool tdp_enabled = false;
103 static int max_huge_page_level __read_mostly;
104 static int tdp_root_level __read_mostly;
105 static int max_tdp_level __read_mostly;
109 module_param(dbg, bool, 0644);
112 #define PTE_PREFETCH_NUM 8
114 #define PT32_LEVEL_BITS 10
116 #define PT32_LEVEL_SHIFT(level) \
117 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
119 #define PT32_LVL_OFFSET_MASK(level) \
120 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
121 * PT32_LEVEL_BITS))) - 1))
123 #define PT32_INDEX(address, level)\
124 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
127 #define PT32_BASE_ADDR_MASK PAGE_MASK
128 #define PT32_DIR_BASE_ADDR_MASK \
129 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
130 #define PT32_LVL_ADDR_MASK(level) \
131 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
132 * PT32_LEVEL_BITS))) - 1))
134 #include <trace/events/kvm.h>
136 /* make pte_list_desc fit well in cache lines */
137 #define PTE_LIST_EXT 14
140 * Slight optimization of cacheline layout, by putting `more' and `spte_count'
141 * at the start; then accessing it will only use one single cacheline for
142 * either full (entries==PTE_LIST_EXT) case or entries<=6.
144 struct pte_list_desc {
145 struct pte_list_desc *more;
147 * Stores number of entries stored in the pte_list_desc. No need to be
148 * u64 but just for easier alignment. When PTE_LIST_EXT, means full.
151 u64 *sptes[PTE_LIST_EXT];
154 struct kvm_shadow_walk_iterator {
162 #define for_each_shadow_entry_using_root(_vcpu, _root, _addr, _walker) \
163 for (shadow_walk_init_using_root(&(_walker), (_vcpu), \
165 shadow_walk_okay(&(_walker)); \
166 shadow_walk_next(&(_walker)))
168 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
169 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
170 shadow_walk_okay(&(_walker)); \
171 shadow_walk_next(&(_walker)))
173 #define for_each_shadow_entry_lockless(_vcpu, _addr, _walker, spte) \
174 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
175 shadow_walk_okay(&(_walker)) && \
176 ({ spte = mmu_spte_get_lockless(_walker.sptep); 1; }); \
177 __shadow_walk_next(&(_walker), spte))
179 static struct kmem_cache *pte_list_desc_cache;
180 struct kmem_cache *mmu_page_header_cache;
181 static struct percpu_counter kvm_total_used_mmu_pages;
183 static void mmu_spte_set(u64 *sptep, u64 spte);
185 struct kvm_mmu_role_regs {
186 const unsigned long cr0;
187 const unsigned long cr4;
191 #define CREATE_TRACE_POINTS
192 #include "mmutrace.h"
195 * Yes, lot's of underscores. They're a hint that you probably shouldn't be
196 * reading from the role_regs. Once the mmu_role is constructed, it becomes
197 * the single source of truth for the MMU's state.
199 #define BUILD_MMU_ROLE_REGS_ACCESSOR(reg, name, flag) \
200 static inline bool __maybe_unused \
201 ____is_##reg##_##name(const struct kvm_mmu_role_regs *regs) \
203 return !!(regs->reg & flag); \
205 BUILD_MMU_ROLE_REGS_ACCESSOR(cr0, pg, X86_CR0_PG);
206 BUILD_MMU_ROLE_REGS_ACCESSOR(cr0, wp, X86_CR0_WP);
207 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pse, X86_CR4_PSE);
208 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pae, X86_CR4_PAE);
209 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, smep, X86_CR4_SMEP);
210 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, smap, X86_CR4_SMAP);
211 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pke, X86_CR4_PKE);
212 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, la57, X86_CR4_LA57);
213 BUILD_MMU_ROLE_REGS_ACCESSOR(efer, nx, EFER_NX);
214 BUILD_MMU_ROLE_REGS_ACCESSOR(efer, lma, EFER_LMA);
217 * The MMU itself (with a valid role) is the single source of truth for the
218 * MMU. Do not use the regs used to build the MMU/role, nor the vCPU. The
219 * regs don't account for dependencies, e.g. clearing CR4 bits if CR0.PG=1,
220 * and the vCPU may be incorrect/irrelevant.
222 #define BUILD_MMU_ROLE_ACCESSOR(base_or_ext, reg, name) \
223 static inline bool __maybe_unused is_##reg##_##name(struct kvm_mmu *mmu) \
225 return !!(mmu->mmu_role. base_or_ext . reg##_##name); \
227 BUILD_MMU_ROLE_ACCESSOR(ext, cr0, pg);
228 BUILD_MMU_ROLE_ACCESSOR(base, cr0, wp);
229 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, pse);
230 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, pae);
231 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, smep);
232 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, smap);
233 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, pke);
234 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, la57);
235 BUILD_MMU_ROLE_ACCESSOR(base, efer, nx);
237 static struct kvm_mmu_role_regs vcpu_to_role_regs(struct kvm_vcpu *vcpu)
239 struct kvm_mmu_role_regs regs = {
240 .cr0 = kvm_read_cr0_bits(vcpu, KVM_MMU_CR0_ROLE_BITS),
241 .cr4 = kvm_read_cr4_bits(vcpu, KVM_MMU_CR4_ROLE_BITS),
242 .efer = vcpu->arch.efer,
248 static int role_regs_to_root_level(const struct kvm_mmu_role_regs *regs)
250 if (!____is_cr0_pg(regs))
252 else if (____is_efer_lma(regs))
253 return ____is_cr4_la57(regs) ? PT64_ROOT_5LEVEL :
255 else if (____is_cr4_pae(regs))
256 return PT32E_ROOT_LEVEL;
258 return PT32_ROOT_LEVEL;
261 static inline bool kvm_available_flush_tlb_with_range(void)
263 return kvm_x86_ops.tlb_remote_flush_with_range;
266 static void kvm_flush_remote_tlbs_with_range(struct kvm *kvm,
267 struct kvm_tlb_range *range)
271 if (range && kvm_x86_ops.tlb_remote_flush_with_range)
272 ret = static_call(kvm_x86_tlb_remote_flush_with_range)(kvm, range);
275 kvm_flush_remote_tlbs(kvm);
278 void kvm_flush_remote_tlbs_with_address(struct kvm *kvm,
279 u64 start_gfn, u64 pages)
281 struct kvm_tlb_range range;
283 range.start_gfn = start_gfn;
286 kvm_flush_remote_tlbs_with_range(kvm, &range);
289 static void mark_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, u64 gfn,
292 u64 spte = make_mmio_spte(vcpu, gfn, access);
294 trace_mark_mmio_spte(sptep, gfn, spte);
295 mmu_spte_set(sptep, spte);
298 static gfn_t get_mmio_spte_gfn(u64 spte)
300 u64 gpa = spte & shadow_nonpresent_or_rsvd_lower_gfn_mask;
302 gpa |= (spte >> SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)
303 & shadow_nonpresent_or_rsvd_mask;
305 return gpa >> PAGE_SHIFT;
308 static unsigned get_mmio_spte_access(u64 spte)
310 return spte & shadow_mmio_access_mask;
313 static bool check_mmio_spte(struct kvm_vcpu *vcpu, u64 spte)
315 u64 kvm_gen, spte_gen, gen;
317 gen = kvm_vcpu_memslots(vcpu)->generation;
318 if (unlikely(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS))
321 kvm_gen = gen & MMIO_SPTE_GEN_MASK;
322 spte_gen = get_mmio_spte_generation(spte);
324 trace_check_mmio_spte(spte, kvm_gen, spte_gen);
325 return likely(kvm_gen == spte_gen);
328 static int is_cpuid_PSE36(void)
333 static gfn_t pse36_gfn_delta(u32 gpte)
335 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
337 return (gpte & PT32_DIR_PSE36_MASK) << shift;
341 static void __set_spte(u64 *sptep, u64 spte)
343 WRITE_ONCE(*sptep, spte);
346 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
348 WRITE_ONCE(*sptep, spte);
351 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
353 return xchg(sptep, spte);
356 static u64 __get_spte_lockless(u64 *sptep)
358 return READ_ONCE(*sptep);
369 static void count_spte_clear(u64 *sptep, u64 spte)
371 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
373 if (is_shadow_present_pte(spte))
376 /* Ensure the spte is completely set before we increase the count */
378 sp->clear_spte_count++;
381 static void __set_spte(u64 *sptep, u64 spte)
383 union split_spte *ssptep, sspte;
385 ssptep = (union split_spte *)sptep;
386 sspte = (union split_spte)spte;
388 ssptep->spte_high = sspte.spte_high;
391 * If we map the spte from nonpresent to present, We should store
392 * the high bits firstly, then set present bit, so cpu can not
393 * fetch this spte while we are setting the spte.
397 WRITE_ONCE(ssptep->spte_low, sspte.spte_low);
400 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
402 union split_spte *ssptep, sspte;
404 ssptep = (union split_spte *)sptep;
405 sspte = (union split_spte)spte;
407 WRITE_ONCE(ssptep->spte_low, sspte.spte_low);
410 * If we map the spte from present to nonpresent, we should clear
411 * present bit firstly to avoid vcpu fetch the old high bits.
415 ssptep->spte_high = sspte.spte_high;
416 count_spte_clear(sptep, spte);
419 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
421 union split_spte *ssptep, sspte, orig;
423 ssptep = (union split_spte *)sptep;
424 sspte = (union split_spte)spte;
426 /* xchg acts as a barrier before the setting of the high bits */
427 orig.spte_low = xchg(&ssptep->spte_low, sspte.spte_low);
428 orig.spte_high = ssptep->spte_high;
429 ssptep->spte_high = sspte.spte_high;
430 count_spte_clear(sptep, spte);
436 * The idea using the light way get the spte on x86_32 guest is from
437 * gup_get_pte (mm/gup.c).
439 * An spte tlb flush may be pending, because kvm_set_pte_rmapp
440 * coalesces them and we are running out of the MMU lock. Therefore
441 * we need to protect against in-progress updates of the spte.
443 * Reading the spte while an update is in progress may get the old value
444 * for the high part of the spte. The race is fine for a present->non-present
445 * change (because the high part of the spte is ignored for non-present spte),
446 * but for a present->present change we must reread the spte.
448 * All such changes are done in two steps (present->non-present and
449 * non-present->present), hence it is enough to count the number of
450 * present->non-present updates: if it changed while reading the spte,
451 * we might have hit the race. This is done using clear_spte_count.
453 static u64 __get_spte_lockless(u64 *sptep)
455 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
456 union split_spte spte, *orig = (union split_spte *)sptep;
460 count = sp->clear_spte_count;
463 spte.spte_low = orig->spte_low;
466 spte.spte_high = orig->spte_high;
469 if (unlikely(spte.spte_low != orig->spte_low ||
470 count != sp->clear_spte_count))
477 static bool spte_has_volatile_bits(u64 spte)
479 if (!is_shadow_present_pte(spte))
483 * Always atomically update spte if it can be updated
484 * out of mmu-lock, it can ensure dirty bit is not lost,
485 * also, it can help us to get a stable is_writable_pte()
486 * to ensure tlb flush is not missed.
488 if (spte_can_locklessly_be_made_writable(spte) ||
489 is_access_track_spte(spte))
492 if (spte_ad_enabled(spte)) {
493 if ((spte & shadow_accessed_mask) == 0 ||
494 (is_writable_pte(spte) && (spte & shadow_dirty_mask) == 0))
501 /* Rules for using mmu_spte_set:
502 * Set the sptep from nonpresent to present.
503 * Note: the sptep being assigned *must* be either not present
504 * or in a state where the hardware will not attempt to update
507 static void mmu_spte_set(u64 *sptep, u64 new_spte)
509 WARN_ON(is_shadow_present_pte(*sptep));
510 __set_spte(sptep, new_spte);
514 * Update the SPTE (excluding the PFN), but do not track changes in its
515 * accessed/dirty status.
517 static u64 mmu_spte_update_no_track(u64 *sptep, u64 new_spte)
519 u64 old_spte = *sptep;
521 WARN_ON(!is_shadow_present_pte(new_spte));
522 check_spte_writable_invariants(new_spte);
524 if (!is_shadow_present_pte(old_spte)) {
525 mmu_spte_set(sptep, new_spte);
529 if (!spte_has_volatile_bits(old_spte))
530 __update_clear_spte_fast(sptep, new_spte);
532 old_spte = __update_clear_spte_slow(sptep, new_spte);
534 WARN_ON(spte_to_pfn(old_spte) != spte_to_pfn(new_spte));
539 /* Rules for using mmu_spte_update:
540 * Update the state bits, it means the mapped pfn is not changed.
542 * Whenever an MMU-writable SPTE is overwritten with a read-only SPTE, remote
543 * TLBs must be flushed. Otherwise rmap_write_protect will find a read-only
544 * spte, even though the writable spte might be cached on a CPU's TLB.
546 * Returns true if the TLB needs to be flushed
548 static bool mmu_spte_update(u64 *sptep, u64 new_spte)
551 u64 old_spte = mmu_spte_update_no_track(sptep, new_spte);
553 if (!is_shadow_present_pte(old_spte))
557 * For the spte updated out of mmu-lock is safe, since
558 * we always atomically update it, see the comments in
559 * spte_has_volatile_bits().
561 if (spte_can_locklessly_be_made_writable(old_spte) &&
562 !is_writable_pte(new_spte))
566 * Flush TLB when accessed/dirty states are changed in the page tables,
567 * to guarantee consistency between TLB and page tables.
570 if (is_accessed_spte(old_spte) && !is_accessed_spte(new_spte)) {
572 kvm_set_pfn_accessed(spte_to_pfn(old_spte));
575 if (is_dirty_spte(old_spte) && !is_dirty_spte(new_spte)) {
577 kvm_set_pfn_dirty(spte_to_pfn(old_spte));
584 * Rules for using mmu_spte_clear_track_bits:
585 * It sets the sptep from present to nonpresent, and track the
586 * state bits, it is used to clear the last level sptep.
587 * Returns the old PTE.
589 static int mmu_spte_clear_track_bits(struct kvm *kvm, u64 *sptep)
592 u64 old_spte = *sptep;
593 int level = sptep_to_sp(sptep)->role.level;
595 if (!spte_has_volatile_bits(old_spte))
596 __update_clear_spte_fast(sptep, 0ull);
598 old_spte = __update_clear_spte_slow(sptep, 0ull);
600 if (!is_shadow_present_pte(old_spte))
603 kvm_update_page_stats(kvm, level, -1);
605 pfn = spte_to_pfn(old_spte);
608 * KVM does not hold the refcount of the page used by
609 * kvm mmu, before reclaiming the page, we should
610 * unmap it from mmu first.
612 WARN_ON(!kvm_is_reserved_pfn(pfn) && !page_count(pfn_to_page(pfn)));
614 if (is_accessed_spte(old_spte))
615 kvm_set_pfn_accessed(pfn);
617 if (is_dirty_spte(old_spte))
618 kvm_set_pfn_dirty(pfn);
624 * Rules for using mmu_spte_clear_no_track:
625 * Directly clear spte without caring the state bits of sptep,
626 * it is used to set the upper level spte.
628 static void mmu_spte_clear_no_track(u64 *sptep)
630 __update_clear_spte_fast(sptep, 0ull);
633 static u64 mmu_spte_get_lockless(u64 *sptep)
635 return __get_spte_lockless(sptep);
638 /* Returns the Accessed status of the PTE and resets it at the same time. */
639 static bool mmu_spte_age(u64 *sptep)
641 u64 spte = mmu_spte_get_lockless(sptep);
643 if (!is_accessed_spte(spte))
646 if (spte_ad_enabled(spte)) {
647 clear_bit((ffs(shadow_accessed_mask) - 1),
648 (unsigned long *)sptep);
651 * Capture the dirty status of the page, so that it doesn't get
652 * lost when the SPTE is marked for access tracking.
654 if (is_writable_pte(spte))
655 kvm_set_pfn_dirty(spte_to_pfn(spte));
657 spte = mark_spte_for_access_track(spte);
658 mmu_spte_update_no_track(sptep, spte);
664 static void walk_shadow_page_lockless_begin(struct kvm_vcpu *vcpu)
666 if (is_tdp_mmu(vcpu->arch.mmu)) {
667 kvm_tdp_mmu_walk_lockless_begin();
670 * Prevent page table teardown by making any free-er wait during
671 * kvm_flush_remote_tlbs() IPI to all active vcpus.
676 * Make sure a following spte read is not reordered ahead of the write
679 smp_store_mb(vcpu->mode, READING_SHADOW_PAGE_TABLES);
683 static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
685 if (is_tdp_mmu(vcpu->arch.mmu)) {
686 kvm_tdp_mmu_walk_lockless_end();
689 * Make sure the write to vcpu->mode is not reordered in front of
690 * reads to sptes. If it does, kvm_mmu_commit_zap_page() can see us
691 * OUTSIDE_GUEST_MODE and proceed to free the shadow page table.
693 smp_store_release(&vcpu->mode, OUTSIDE_GUEST_MODE);
698 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu, bool maybe_indirect)
702 /* 1 rmap, 1 parent PTE per level, and the prefetched rmaps. */
703 r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
704 1 + PT64_ROOT_MAX_LEVEL + PTE_PREFETCH_NUM);
707 r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_shadow_page_cache,
708 PT64_ROOT_MAX_LEVEL);
711 if (maybe_indirect) {
712 r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_gfn_array_cache,
713 PT64_ROOT_MAX_LEVEL);
717 return kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
718 PT64_ROOT_MAX_LEVEL);
721 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
723 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache);
724 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_shadow_page_cache);
725 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_gfn_array_cache);
726 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
729 static struct pte_list_desc *mmu_alloc_pte_list_desc(struct kvm_vcpu *vcpu)
731 return kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_list_desc_cache);
734 static void mmu_free_pte_list_desc(struct pte_list_desc *pte_list_desc)
736 kmem_cache_free(pte_list_desc_cache, pte_list_desc);
739 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
741 if (!sp->role.direct)
742 return sp->gfns[index];
744 return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
747 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
749 if (!sp->role.direct) {
750 sp->gfns[index] = gfn;
754 if (WARN_ON(gfn != kvm_mmu_page_get_gfn(sp, index)))
755 pr_err_ratelimited("gfn mismatch under direct page %llx "
756 "(expected %llx, got %llx)\n",
758 kvm_mmu_page_get_gfn(sp, index), gfn);
762 * Return the pointer to the large page information for a given gfn,
763 * handling slots that are not large page aligned.
765 static struct kvm_lpage_info *lpage_info_slot(gfn_t gfn,
766 const struct kvm_memory_slot *slot, int level)
770 idx = gfn_to_index(gfn, slot->base_gfn, level);
771 return &slot->arch.lpage_info[level - 2][idx];
774 static void update_gfn_disallow_lpage_count(const struct kvm_memory_slot *slot,
775 gfn_t gfn, int count)
777 struct kvm_lpage_info *linfo;
780 for (i = PG_LEVEL_2M; i <= KVM_MAX_HUGEPAGE_LEVEL; ++i) {
781 linfo = lpage_info_slot(gfn, slot, i);
782 linfo->disallow_lpage += count;
783 WARN_ON(linfo->disallow_lpage < 0);
787 void kvm_mmu_gfn_disallow_lpage(const struct kvm_memory_slot *slot, gfn_t gfn)
789 update_gfn_disallow_lpage_count(slot, gfn, 1);
792 void kvm_mmu_gfn_allow_lpage(const struct kvm_memory_slot *slot, gfn_t gfn)
794 update_gfn_disallow_lpage_count(slot, gfn, -1);
797 static void account_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
799 struct kvm_memslots *slots;
800 struct kvm_memory_slot *slot;
803 kvm->arch.indirect_shadow_pages++;
805 slots = kvm_memslots_for_spte_role(kvm, sp->role);
806 slot = __gfn_to_memslot(slots, gfn);
808 /* the non-leaf shadow pages are keeping readonly. */
809 if (sp->role.level > PG_LEVEL_4K)
810 return kvm_slot_page_track_add_page(kvm, slot, gfn,
811 KVM_PAGE_TRACK_WRITE);
813 kvm_mmu_gfn_disallow_lpage(slot, gfn);
816 void account_huge_nx_page(struct kvm *kvm, struct kvm_mmu_page *sp)
818 if (sp->lpage_disallowed)
821 ++kvm->stat.nx_lpage_splits;
822 list_add_tail(&sp->lpage_disallowed_link,
823 &kvm->arch.lpage_disallowed_mmu_pages);
824 sp->lpage_disallowed = true;
827 static void unaccount_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
829 struct kvm_memslots *slots;
830 struct kvm_memory_slot *slot;
833 kvm->arch.indirect_shadow_pages--;
835 slots = kvm_memslots_for_spte_role(kvm, sp->role);
836 slot = __gfn_to_memslot(slots, gfn);
837 if (sp->role.level > PG_LEVEL_4K)
838 return kvm_slot_page_track_remove_page(kvm, slot, gfn,
839 KVM_PAGE_TRACK_WRITE);
841 kvm_mmu_gfn_allow_lpage(slot, gfn);
844 void unaccount_huge_nx_page(struct kvm *kvm, struct kvm_mmu_page *sp)
846 --kvm->stat.nx_lpage_splits;
847 sp->lpage_disallowed = false;
848 list_del(&sp->lpage_disallowed_link);
851 static struct kvm_memory_slot *
852 gfn_to_memslot_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t gfn,
855 struct kvm_memory_slot *slot;
857 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
858 if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
860 if (no_dirty_log && kvm_slot_dirty_track_enabled(slot))
867 * About rmap_head encoding:
869 * If the bit zero of rmap_head->val is clear, then it points to the only spte
870 * in this rmap chain. Otherwise, (rmap_head->val & ~1) points to a struct
871 * pte_list_desc containing more mappings.
875 * Returns the number of pointers in the rmap chain, not counting the new one.
877 static int pte_list_add(struct kvm_vcpu *vcpu, u64 *spte,
878 struct kvm_rmap_head *rmap_head)
880 struct pte_list_desc *desc;
883 if (!rmap_head->val) {
884 rmap_printk("%p %llx 0->1\n", spte, *spte);
885 rmap_head->val = (unsigned long)spte;
886 } else if (!(rmap_head->val & 1)) {
887 rmap_printk("%p %llx 1->many\n", spte, *spte);
888 desc = mmu_alloc_pte_list_desc(vcpu);
889 desc->sptes[0] = (u64 *)rmap_head->val;
890 desc->sptes[1] = spte;
891 desc->spte_count = 2;
892 rmap_head->val = (unsigned long)desc | 1;
895 rmap_printk("%p %llx many->many\n", spte, *spte);
896 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
897 while (desc->spte_count == PTE_LIST_EXT) {
898 count += PTE_LIST_EXT;
900 desc->more = mmu_alloc_pte_list_desc(vcpu);
902 desc->spte_count = 0;
907 count += desc->spte_count;
908 desc->sptes[desc->spte_count++] = spte;
914 pte_list_desc_remove_entry(struct kvm_rmap_head *rmap_head,
915 struct pte_list_desc *desc, int i,
916 struct pte_list_desc *prev_desc)
918 int j = desc->spte_count - 1;
920 desc->sptes[i] = desc->sptes[j];
921 desc->sptes[j] = NULL;
923 if (desc->spte_count)
925 if (!prev_desc && !desc->more)
929 prev_desc->more = desc->more;
931 rmap_head->val = (unsigned long)desc->more | 1;
932 mmu_free_pte_list_desc(desc);
935 static void __pte_list_remove(u64 *spte, struct kvm_rmap_head *rmap_head)
937 struct pte_list_desc *desc;
938 struct pte_list_desc *prev_desc;
941 if (!rmap_head->val) {
942 pr_err("%s: %p 0->BUG\n", __func__, spte);
944 } else if (!(rmap_head->val & 1)) {
945 rmap_printk("%p 1->0\n", spte);
946 if ((u64 *)rmap_head->val != spte) {
947 pr_err("%s: %p 1->BUG\n", __func__, spte);
952 rmap_printk("%p many->many\n", spte);
953 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
956 for (i = 0; i < desc->spte_count; ++i) {
957 if (desc->sptes[i] == spte) {
958 pte_list_desc_remove_entry(rmap_head,
966 pr_err("%s: %p many->many\n", __func__, spte);
971 static void pte_list_remove(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
974 mmu_spte_clear_track_bits(kvm, sptep);
975 __pte_list_remove(sptep, rmap_head);
978 /* Return true if rmap existed, false otherwise */
979 static bool pte_list_destroy(struct kvm *kvm, struct kvm_rmap_head *rmap_head)
981 struct pte_list_desc *desc, *next;
987 if (!(rmap_head->val & 1)) {
988 mmu_spte_clear_track_bits(kvm, (u64 *)rmap_head->val);
992 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
994 for (; desc; desc = next) {
995 for (i = 0; i < desc->spte_count; i++)
996 mmu_spte_clear_track_bits(kvm, desc->sptes[i]);
998 mmu_free_pte_list_desc(desc);
1001 /* rmap_head is meaningless now, remember to reset it */
1006 unsigned int pte_list_count(struct kvm_rmap_head *rmap_head)
1008 struct pte_list_desc *desc;
1009 unsigned int count = 0;
1011 if (!rmap_head->val)
1013 else if (!(rmap_head->val & 1))
1016 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
1019 count += desc->spte_count;
1026 static struct kvm_rmap_head *gfn_to_rmap(gfn_t gfn, int level,
1027 const struct kvm_memory_slot *slot)
1031 idx = gfn_to_index(gfn, slot->base_gfn, level);
1032 return &slot->arch.rmap[level - PG_LEVEL_4K][idx];
1035 static bool rmap_can_add(struct kvm_vcpu *vcpu)
1037 struct kvm_mmu_memory_cache *mc;
1039 mc = &vcpu->arch.mmu_pte_list_desc_cache;
1040 return kvm_mmu_memory_cache_nr_free_objects(mc);
1043 static void rmap_remove(struct kvm *kvm, u64 *spte)
1045 struct kvm_memslots *slots;
1046 struct kvm_memory_slot *slot;
1047 struct kvm_mmu_page *sp;
1049 struct kvm_rmap_head *rmap_head;
1051 sp = sptep_to_sp(spte);
1052 gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
1055 * Unlike rmap_add, rmap_remove does not run in the context of a vCPU
1056 * so we have to determine which memslots to use based on context
1057 * information in sp->role.
1059 slots = kvm_memslots_for_spte_role(kvm, sp->role);
1061 slot = __gfn_to_memslot(slots, gfn);
1062 rmap_head = gfn_to_rmap(gfn, sp->role.level, slot);
1064 __pte_list_remove(spte, rmap_head);
1068 * Used by the following functions to iterate through the sptes linked by a
1069 * rmap. All fields are private and not assumed to be used outside.
1071 struct rmap_iterator {
1072 /* private fields */
1073 struct pte_list_desc *desc; /* holds the sptep if not NULL */
1074 int pos; /* index of the sptep */
1078 * Iteration must be started by this function. This should also be used after
1079 * removing/dropping sptes from the rmap link because in such cases the
1080 * information in the iterator may not be valid.
1082 * Returns sptep if found, NULL otherwise.
1084 static u64 *rmap_get_first(struct kvm_rmap_head *rmap_head,
1085 struct rmap_iterator *iter)
1089 if (!rmap_head->val)
1092 if (!(rmap_head->val & 1)) {
1094 sptep = (u64 *)rmap_head->val;
1098 iter->desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
1100 sptep = iter->desc->sptes[iter->pos];
1102 BUG_ON(!is_shadow_present_pte(*sptep));
1107 * Must be used with a valid iterator: e.g. after rmap_get_first().
1109 * Returns sptep if found, NULL otherwise.
1111 static u64 *rmap_get_next(struct rmap_iterator *iter)
1116 if (iter->pos < PTE_LIST_EXT - 1) {
1118 sptep = iter->desc->sptes[iter->pos];
1123 iter->desc = iter->desc->more;
1127 /* desc->sptes[0] cannot be NULL */
1128 sptep = iter->desc->sptes[iter->pos];
1135 BUG_ON(!is_shadow_present_pte(*sptep));
1139 #define for_each_rmap_spte(_rmap_head_, _iter_, _spte_) \
1140 for (_spte_ = rmap_get_first(_rmap_head_, _iter_); \
1141 _spte_; _spte_ = rmap_get_next(_iter_))
1143 static void drop_spte(struct kvm *kvm, u64 *sptep)
1145 u64 old_spte = mmu_spte_clear_track_bits(kvm, sptep);
1147 if (is_shadow_present_pte(old_spte))
1148 rmap_remove(kvm, sptep);
1152 static bool __drop_large_spte(struct kvm *kvm, u64 *sptep)
1154 if (is_large_pte(*sptep)) {
1155 WARN_ON(sptep_to_sp(sptep)->role.level == PG_LEVEL_4K);
1156 drop_spte(kvm, sptep);
1163 static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1165 if (__drop_large_spte(vcpu->kvm, sptep)) {
1166 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
1168 kvm_flush_remote_tlbs_with_address(vcpu->kvm, sp->gfn,
1169 KVM_PAGES_PER_HPAGE(sp->role.level));
1174 * Write-protect on the specified @sptep, @pt_protect indicates whether
1175 * spte write-protection is caused by protecting shadow page table.
1177 * Note: write protection is difference between dirty logging and spte
1179 * - for dirty logging, the spte can be set to writable at anytime if
1180 * its dirty bitmap is properly set.
1181 * - for spte protection, the spte can be writable only after unsync-ing
1184 * Return true if tlb need be flushed.
1186 static bool spte_write_protect(u64 *sptep, bool pt_protect)
1190 if (!is_writable_pte(spte) &&
1191 !(pt_protect && spte_can_locklessly_be_made_writable(spte)))
1194 rmap_printk("spte %p %llx\n", sptep, *sptep);
1197 spte &= ~shadow_mmu_writable_mask;
1198 spte = spte & ~PT_WRITABLE_MASK;
1200 return mmu_spte_update(sptep, spte);
1203 static bool rmap_write_protect(struct kvm_rmap_head *rmap_head,
1207 struct rmap_iterator iter;
1210 for_each_rmap_spte(rmap_head, &iter, sptep)
1211 flush |= spte_write_protect(sptep, pt_protect);
1216 static bool spte_clear_dirty(u64 *sptep)
1220 rmap_printk("spte %p %llx\n", sptep, *sptep);
1222 MMU_WARN_ON(!spte_ad_enabled(spte));
1223 spte &= ~shadow_dirty_mask;
1224 return mmu_spte_update(sptep, spte);
1227 static bool spte_wrprot_for_clear_dirty(u64 *sptep)
1229 bool was_writable = test_and_clear_bit(PT_WRITABLE_SHIFT,
1230 (unsigned long *)sptep);
1231 if (was_writable && !spte_ad_enabled(*sptep))
1232 kvm_set_pfn_dirty(spte_to_pfn(*sptep));
1234 return was_writable;
1238 * Gets the GFN ready for another round of dirty logging by clearing the
1239 * - D bit on ad-enabled SPTEs, and
1240 * - W bit on ad-disabled SPTEs.
1241 * Returns true iff any D or W bits were cleared.
1243 static bool __rmap_clear_dirty(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1244 const struct kvm_memory_slot *slot)
1247 struct rmap_iterator iter;
1250 for_each_rmap_spte(rmap_head, &iter, sptep)
1251 if (spte_ad_need_write_protect(*sptep))
1252 flush |= spte_wrprot_for_clear_dirty(sptep);
1254 flush |= spte_clear_dirty(sptep);
1260 * kvm_mmu_write_protect_pt_masked - write protect selected PT level pages
1261 * @kvm: kvm instance
1262 * @slot: slot to protect
1263 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1264 * @mask: indicates which pages we should protect
1266 * Used when we do not need to care about huge page mappings.
1268 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1269 struct kvm_memory_slot *slot,
1270 gfn_t gfn_offset, unsigned long mask)
1272 struct kvm_rmap_head *rmap_head;
1274 if (is_tdp_mmu_enabled(kvm))
1275 kvm_tdp_mmu_clear_dirty_pt_masked(kvm, slot,
1276 slot->base_gfn + gfn_offset, mask, true);
1278 if (!kvm_memslots_have_rmaps(kvm))
1282 rmap_head = gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1284 rmap_write_protect(rmap_head, false);
1286 /* clear the first set bit */
1292 * kvm_mmu_clear_dirty_pt_masked - clear MMU D-bit for PT level pages, or write
1293 * protect the page if the D-bit isn't supported.
1294 * @kvm: kvm instance
1295 * @slot: slot to clear D-bit
1296 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1297 * @mask: indicates which pages we should clear D-bit
1299 * Used for PML to re-log the dirty GPAs after userspace querying dirty_bitmap.
1301 static void kvm_mmu_clear_dirty_pt_masked(struct kvm *kvm,
1302 struct kvm_memory_slot *slot,
1303 gfn_t gfn_offset, unsigned long mask)
1305 struct kvm_rmap_head *rmap_head;
1307 if (is_tdp_mmu_enabled(kvm))
1308 kvm_tdp_mmu_clear_dirty_pt_masked(kvm, slot,
1309 slot->base_gfn + gfn_offset, mask, false);
1311 if (!kvm_memslots_have_rmaps(kvm))
1315 rmap_head = gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1317 __rmap_clear_dirty(kvm, rmap_head, slot);
1319 /* clear the first set bit */
1325 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1328 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1329 * enable dirty logging for them.
1331 * We need to care about huge page mappings: e.g. during dirty logging we may
1332 * have such mappings.
1334 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1335 struct kvm_memory_slot *slot,
1336 gfn_t gfn_offset, unsigned long mask)
1339 * Huge pages are NOT write protected when we start dirty logging in
1340 * initially-all-set mode; must write protect them here so that they
1341 * are split to 4K on the first write.
1343 * The gfn_offset is guaranteed to be aligned to 64, but the base_gfn
1344 * of memslot has no such restriction, so the range can cross two large
1347 if (kvm_dirty_log_manual_protect_and_init_set(kvm)) {
1348 gfn_t start = slot->base_gfn + gfn_offset + __ffs(mask);
1349 gfn_t end = slot->base_gfn + gfn_offset + __fls(mask);
1351 if (READ_ONCE(eager_page_split))
1352 kvm_mmu_try_split_huge_pages(kvm, slot, start, end, PG_LEVEL_4K);
1354 kvm_mmu_slot_gfn_write_protect(kvm, slot, start, PG_LEVEL_2M);
1356 /* Cross two large pages? */
1357 if (ALIGN(start << PAGE_SHIFT, PMD_SIZE) !=
1358 ALIGN(end << PAGE_SHIFT, PMD_SIZE))
1359 kvm_mmu_slot_gfn_write_protect(kvm, slot, end,
1363 /* Now handle 4K PTEs. */
1364 if (kvm_x86_ops.cpu_dirty_log_size)
1365 kvm_mmu_clear_dirty_pt_masked(kvm, slot, gfn_offset, mask);
1367 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1370 int kvm_cpu_dirty_log_size(void)
1372 return kvm_x86_ops.cpu_dirty_log_size;
1375 bool kvm_mmu_slot_gfn_write_protect(struct kvm *kvm,
1376 struct kvm_memory_slot *slot, u64 gfn,
1379 struct kvm_rmap_head *rmap_head;
1381 bool write_protected = false;
1383 if (kvm_memslots_have_rmaps(kvm)) {
1384 for (i = min_level; i <= KVM_MAX_HUGEPAGE_LEVEL; ++i) {
1385 rmap_head = gfn_to_rmap(gfn, i, slot);
1386 write_protected |= rmap_write_protect(rmap_head, true);
1390 if (is_tdp_mmu_enabled(kvm))
1392 kvm_tdp_mmu_write_protect_gfn(kvm, slot, gfn, min_level);
1394 return write_protected;
1397 static bool kvm_vcpu_write_protect_gfn(struct kvm_vcpu *vcpu, u64 gfn)
1399 struct kvm_memory_slot *slot;
1401 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1402 return kvm_mmu_slot_gfn_write_protect(vcpu->kvm, slot, gfn, PG_LEVEL_4K);
1405 static bool kvm_zap_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1406 const struct kvm_memory_slot *slot)
1408 return pte_list_destroy(kvm, rmap_head);
1411 static bool kvm_unmap_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1412 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1415 return kvm_zap_rmapp(kvm, rmap_head, slot);
1418 static bool kvm_set_pte_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1419 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1423 struct rmap_iterator iter;
1424 bool need_flush = false;
1428 WARN_ON(pte_huge(pte));
1429 new_pfn = pte_pfn(pte);
1432 for_each_rmap_spte(rmap_head, &iter, sptep) {
1433 rmap_printk("spte %p %llx gfn %llx (%d)\n",
1434 sptep, *sptep, gfn, level);
1438 if (pte_write(pte)) {
1439 pte_list_remove(kvm, rmap_head, sptep);
1442 new_spte = kvm_mmu_changed_pte_notifier_make_spte(
1445 mmu_spte_clear_track_bits(kvm, sptep);
1446 mmu_spte_set(sptep, new_spte);
1450 if (need_flush && kvm_available_flush_tlb_with_range()) {
1451 kvm_flush_remote_tlbs_with_address(kvm, gfn, 1);
1458 struct slot_rmap_walk_iterator {
1460 const struct kvm_memory_slot *slot;
1466 /* output fields. */
1468 struct kvm_rmap_head *rmap;
1471 /* private field. */
1472 struct kvm_rmap_head *end_rmap;
1476 rmap_walk_init_level(struct slot_rmap_walk_iterator *iterator, int level)
1478 iterator->level = level;
1479 iterator->gfn = iterator->start_gfn;
1480 iterator->rmap = gfn_to_rmap(iterator->gfn, level, iterator->slot);
1481 iterator->end_rmap = gfn_to_rmap(iterator->end_gfn, level, iterator->slot);
1485 slot_rmap_walk_init(struct slot_rmap_walk_iterator *iterator,
1486 const struct kvm_memory_slot *slot, int start_level,
1487 int end_level, gfn_t start_gfn, gfn_t end_gfn)
1489 iterator->slot = slot;
1490 iterator->start_level = start_level;
1491 iterator->end_level = end_level;
1492 iterator->start_gfn = start_gfn;
1493 iterator->end_gfn = end_gfn;
1495 rmap_walk_init_level(iterator, iterator->start_level);
1498 static bool slot_rmap_walk_okay(struct slot_rmap_walk_iterator *iterator)
1500 return !!iterator->rmap;
1503 static void slot_rmap_walk_next(struct slot_rmap_walk_iterator *iterator)
1505 if (++iterator->rmap <= iterator->end_rmap) {
1506 iterator->gfn += (1UL << KVM_HPAGE_GFN_SHIFT(iterator->level));
1510 if (++iterator->level > iterator->end_level) {
1511 iterator->rmap = NULL;
1515 rmap_walk_init_level(iterator, iterator->level);
1518 #define for_each_slot_rmap_range(_slot_, _start_level_, _end_level_, \
1519 _start_gfn, _end_gfn, _iter_) \
1520 for (slot_rmap_walk_init(_iter_, _slot_, _start_level_, \
1521 _end_level_, _start_gfn, _end_gfn); \
1522 slot_rmap_walk_okay(_iter_); \
1523 slot_rmap_walk_next(_iter_))
1525 typedef bool (*rmap_handler_t)(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1526 struct kvm_memory_slot *slot, gfn_t gfn,
1527 int level, pte_t pte);
1529 static __always_inline bool kvm_handle_gfn_range(struct kvm *kvm,
1530 struct kvm_gfn_range *range,
1531 rmap_handler_t handler)
1533 struct slot_rmap_walk_iterator iterator;
1536 for_each_slot_rmap_range(range->slot, PG_LEVEL_4K, KVM_MAX_HUGEPAGE_LEVEL,
1537 range->start, range->end - 1, &iterator)
1538 ret |= handler(kvm, iterator.rmap, range->slot, iterator.gfn,
1539 iterator.level, range->pte);
1544 bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
1548 if (kvm_memslots_have_rmaps(kvm))
1549 flush = kvm_handle_gfn_range(kvm, range, kvm_unmap_rmapp);
1551 if (is_tdp_mmu_enabled(kvm))
1552 flush = kvm_tdp_mmu_unmap_gfn_range(kvm, range, flush);
1557 bool kvm_set_spte_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1561 if (kvm_memslots_have_rmaps(kvm))
1562 flush = kvm_handle_gfn_range(kvm, range, kvm_set_pte_rmapp);
1564 if (is_tdp_mmu_enabled(kvm))
1565 flush |= kvm_tdp_mmu_set_spte_gfn(kvm, range);
1570 static bool kvm_age_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1571 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1575 struct rmap_iterator iter;
1578 for_each_rmap_spte(rmap_head, &iter, sptep)
1579 young |= mmu_spte_age(sptep);
1584 static bool kvm_test_age_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1585 struct kvm_memory_slot *slot, gfn_t gfn,
1586 int level, pte_t unused)
1589 struct rmap_iterator iter;
1591 for_each_rmap_spte(rmap_head, &iter, sptep)
1592 if (is_accessed_spte(*sptep))
1597 #define RMAP_RECYCLE_THRESHOLD 1000
1599 static void rmap_add(struct kvm_vcpu *vcpu, struct kvm_memory_slot *slot,
1600 u64 *spte, gfn_t gfn)
1602 struct kvm_mmu_page *sp;
1603 struct kvm_rmap_head *rmap_head;
1606 sp = sptep_to_sp(spte);
1607 kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
1608 rmap_head = gfn_to_rmap(gfn, sp->role.level, slot);
1609 rmap_count = pte_list_add(vcpu, spte, rmap_head);
1611 if (rmap_count > RMAP_RECYCLE_THRESHOLD) {
1612 kvm_unmap_rmapp(vcpu->kvm, rmap_head, NULL, gfn, sp->role.level, __pte(0));
1613 kvm_flush_remote_tlbs_with_address(
1614 vcpu->kvm, sp->gfn, KVM_PAGES_PER_HPAGE(sp->role.level));
1618 bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1622 if (kvm_memslots_have_rmaps(kvm))
1623 young = kvm_handle_gfn_range(kvm, range, kvm_age_rmapp);
1625 if (is_tdp_mmu_enabled(kvm))
1626 young |= kvm_tdp_mmu_age_gfn_range(kvm, range);
1631 bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1635 if (kvm_memslots_have_rmaps(kvm))
1636 young = kvm_handle_gfn_range(kvm, range, kvm_test_age_rmapp);
1638 if (is_tdp_mmu_enabled(kvm))
1639 young |= kvm_tdp_mmu_test_age_gfn(kvm, range);
1645 static int is_empty_shadow_page(u64 *spt)
1650 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
1651 if (is_shadow_present_pte(*pos)) {
1652 printk(KERN_ERR "%s: %p %llx\n", __func__,
1661 * This value is the sum of all of the kvm instances's
1662 * kvm->arch.n_used_mmu_pages values. We need a global,
1663 * aggregate version in order to make the slab shrinker
1666 static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, long nr)
1668 kvm->arch.n_used_mmu_pages += nr;
1669 percpu_counter_add(&kvm_total_used_mmu_pages, nr);
1672 static void kvm_mmu_free_page(struct kvm_mmu_page *sp)
1674 MMU_WARN_ON(!is_empty_shadow_page(sp->spt));
1675 hlist_del(&sp->hash_link);
1676 list_del(&sp->link);
1677 free_page((unsigned long)sp->spt);
1678 if (!sp->role.direct)
1679 free_page((unsigned long)sp->gfns);
1680 kmem_cache_free(mmu_page_header_cache, sp);
1683 static unsigned kvm_page_table_hashfn(gfn_t gfn)
1685 return hash_64(gfn, KVM_MMU_HASH_SHIFT);
1688 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
1689 struct kvm_mmu_page *sp, u64 *parent_pte)
1694 pte_list_add(vcpu, parent_pte, &sp->parent_ptes);
1697 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
1700 __pte_list_remove(parent_pte, &sp->parent_ptes);
1703 static void drop_parent_pte(struct kvm_mmu_page *sp,
1706 mmu_page_remove_parent_pte(sp, parent_pte);
1707 mmu_spte_clear_no_track(parent_pte);
1710 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu, int direct)
1712 struct kvm_mmu_page *sp;
1714 sp = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache);
1715 sp->spt = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_shadow_page_cache);
1717 sp->gfns = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_gfn_array_cache);
1718 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
1721 * active_mmu_pages must be a FIFO list, as kvm_zap_obsolete_pages()
1722 * depends on valid pages being added to the head of the list. See
1723 * comments in kvm_zap_obsolete_pages().
1725 sp->mmu_valid_gen = vcpu->kvm->arch.mmu_valid_gen;
1726 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
1727 kvm_mod_used_mmu_pages(vcpu->kvm, +1);
1731 static void mark_unsync(u64 *spte);
1732 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1735 struct rmap_iterator iter;
1737 for_each_rmap_spte(&sp->parent_ptes, &iter, sptep) {
1742 static void mark_unsync(u64 *spte)
1744 struct kvm_mmu_page *sp;
1747 sp = sptep_to_sp(spte);
1748 index = spte - sp->spt;
1749 if (__test_and_set_bit(index, sp->unsync_child_bitmap))
1751 if (sp->unsync_children++)
1753 kvm_mmu_mark_parents_unsync(sp);
1756 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1757 struct kvm_mmu_page *sp)
1762 #define KVM_PAGE_ARRAY_NR 16
1764 struct kvm_mmu_pages {
1765 struct mmu_page_and_offset {
1766 struct kvm_mmu_page *sp;
1768 } page[KVM_PAGE_ARRAY_NR];
1772 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1778 for (i=0; i < pvec->nr; i++)
1779 if (pvec->page[i].sp == sp)
1782 pvec->page[pvec->nr].sp = sp;
1783 pvec->page[pvec->nr].idx = idx;
1785 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1788 static inline void clear_unsync_child_bit(struct kvm_mmu_page *sp, int idx)
1790 --sp->unsync_children;
1791 WARN_ON((int)sp->unsync_children < 0);
1792 __clear_bit(idx, sp->unsync_child_bitmap);
1795 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1796 struct kvm_mmu_pages *pvec)
1798 int i, ret, nr_unsync_leaf = 0;
1800 for_each_set_bit(i, sp->unsync_child_bitmap, 512) {
1801 struct kvm_mmu_page *child;
1802 u64 ent = sp->spt[i];
1804 if (!is_shadow_present_pte(ent) || is_large_pte(ent)) {
1805 clear_unsync_child_bit(sp, i);
1809 child = to_shadow_page(ent & PT64_BASE_ADDR_MASK);
1811 if (child->unsync_children) {
1812 if (mmu_pages_add(pvec, child, i))
1815 ret = __mmu_unsync_walk(child, pvec);
1817 clear_unsync_child_bit(sp, i);
1819 } else if (ret > 0) {
1820 nr_unsync_leaf += ret;
1823 } else if (child->unsync) {
1825 if (mmu_pages_add(pvec, child, i))
1828 clear_unsync_child_bit(sp, i);
1831 return nr_unsync_leaf;
1834 #define INVALID_INDEX (-1)
1836 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1837 struct kvm_mmu_pages *pvec)
1840 if (!sp->unsync_children)
1843 mmu_pages_add(pvec, sp, INVALID_INDEX);
1844 return __mmu_unsync_walk(sp, pvec);
1847 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1849 WARN_ON(!sp->unsync);
1850 trace_kvm_mmu_sync_page(sp);
1852 --kvm->stat.mmu_unsync;
1855 static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1856 struct list_head *invalid_list);
1857 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1858 struct list_head *invalid_list);
1860 #define for_each_valid_sp(_kvm, _sp, _list) \
1861 hlist_for_each_entry(_sp, _list, hash_link) \
1862 if (is_obsolete_sp((_kvm), (_sp))) { \
1865 #define for_each_gfn_indirect_valid_sp(_kvm, _sp, _gfn) \
1866 for_each_valid_sp(_kvm, _sp, \
1867 &(_kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(_gfn)]) \
1868 if ((_sp)->gfn != (_gfn) || (_sp)->role.direct) {} else
1870 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1871 struct list_head *invalid_list)
1873 int ret = vcpu->arch.mmu->sync_page(vcpu, sp);
1876 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1880 static bool kvm_mmu_remote_flush_or_zap(struct kvm *kvm,
1881 struct list_head *invalid_list,
1884 if (!remote_flush && list_empty(invalid_list))
1887 if (!list_empty(invalid_list))
1888 kvm_mmu_commit_zap_page(kvm, invalid_list);
1890 kvm_flush_remote_tlbs(kvm);
1894 static bool is_obsolete_sp(struct kvm *kvm, struct kvm_mmu_page *sp)
1896 if (sp->role.invalid)
1899 /* TDP MMU pages due not use the MMU generation. */
1900 return !sp->tdp_mmu_page &&
1901 unlikely(sp->mmu_valid_gen != kvm->arch.mmu_valid_gen);
1904 struct mmu_page_path {
1905 struct kvm_mmu_page *parent[PT64_ROOT_MAX_LEVEL];
1906 unsigned int idx[PT64_ROOT_MAX_LEVEL];
1909 #define for_each_sp(pvec, sp, parents, i) \
1910 for (i = mmu_pages_first(&pvec, &parents); \
1911 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1912 i = mmu_pages_next(&pvec, &parents, i))
1914 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1915 struct mmu_page_path *parents,
1920 for (n = i+1; n < pvec->nr; n++) {
1921 struct kvm_mmu_page *sp = pvec->page[n].sp;
1922 unsigned idx = pvec->page[n].idx;
1923 int level = sp->role.level;
1925 parents->idx[level-1] = idx;
1926 if (level == PG_LEVEL_4K)
1929 parents->parent[level-2] = sp;
1935 static int mmu_pages_first(struct kvm_mmu_pages *pvec,
1936 struct mmu_page_path *parents)
1938 struct kvm_mmu_page *sp;
1944 WARN_ON(pvec->page[0].idx != INVALID_INDEX);
1946 sp = pvec->page[0].sp;
1947 level = sp->role.level;
1948 WARN_ON(level == PG_LEVEL_4K);
1950 parents->parent[level-2] = sp;
1952 /* Also set up a sentinel. Further entries in pvec are all
1953 * children of sp, so this element is never overwritten.
1955 parents->parent[level-1] = NULL;
1956 return mmu_pages_next(pvec, parents, 0);
1959 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1961 struct kvm_mmu_page *sp;
1962 unsigned int level = 0;
1965 unsigned int idx = parents->idx[level];
1966 sp = parents->parent[level];
1970 WARN_ON(idx == INVALID_INDEX);
1971 clear_unsync_child_bit(sp, idx);
1973 } while (!sp->unsync_children);
1976 static int mmu_sync_children(struct kvm_vcpu *vcpu,
1977 struct kvm_mmu_page *parent, bool can_yield)
1980 struct kvm_mmu_page *sp;
1981 struct mmu_page_path parents;
1982 struct kvm_mmu_pages pages;
1983 LIST_HEAD(invalid_list);
1986 while (mmu_unsync_walk(parent, &pages)) {
1987 bool protected = false;
1989 for_each_sp(pages, sp, parents, i)
1990 protected |= kvm_vcpu_write_protect_gfn(vcpu, sp->gfn);
1993 kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, true);
1997 for_each_sp(pages, sp, parents, i) {
1998 kvm_unlink_unsync_page(vcpu->kvm, sp);
1999 flush |= kvm_sync_page(vcpu, sp, &invalid_list) > 0;
2000 mmu_pages_clear_parents(&parents);
2002 if (need_resched() || rwlock_needbreak(&vcpu->kvm->mmu_lock)) {
2003 kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, flush);
2005 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
2009 cond_resched_rwlock_write(&vcpu->kvm->mmu_lock);
2014 kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, flush);
2018 static void __clear_sp_write_flooding_count(struct kvm_mmu_page *sp)
2020 atomic_set(&sp->write_flooding_count, 0);
2023 static void clear_sp_write_flooding_count(u64 *spte)
2025 __clear_sp_write_flooding_count(sptep_to_sp(spte));
2028 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
2033 unsigned int access)
2035 bool direct_mmu = vcpu->arch.mmu->direct_map;
2036 union kvm_mmu_page_role role;
2037 struct hlist_head *sp_list;
2039 struct kvm_mmu_page *sp;
2042 LIST_HEAD(invalid_list);
2044 role = vcpu->arch.mmu->mmu_role.base;
2046 role.direct = direct;
2047 role.access = access;
2048 if (role.has_4_byte_gpte) {
2049 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
2050 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
2051 role.quadrant = quadrant;
2054 sp_list = &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)];
2055 for_each_valid_sp(vcpu->kvm, sp, sp_list) {
2056 if (sp->gfn != gfn) {
2061 if (sp->role.word != role.word) {
2063 * If the guest is creating an upper-level page, zap
2064 * unsync pages for the same gfn. While it's possible
2065 * the guest is using recursive page tables, in all
2066 * likelihood the guest has stopped using the unsync
2067 * page and is installing a completely unrelated page.
2068 * Unsync pages must not be left as is, because the new
2069 * upper-level page will be write-protected.
2071 if (level > PG_LEVEL_4K && sp->unsync)
2072 kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2078 goto trace_get_page;
2082 * The page is good, but is stale. kvm_sync_page does
2083 * get the latest guest state, but (unlike mmu_unsync_children)
2084 * it doesn't write-protect the page or mark it synchronized!
2085 * This way the validity of the mapping is ensured, but the
2086 * overhead of write protection is not incurred until the
2087 * guest invalidates the TLB mapping. This allows multiple
2088 * SPs for a single gfn to be unsync.
2090 * If the sync fails, the page is zapped. If so, break
2091 * in order to rebuild it.
2093 ret = kvm_sync_page(vcpu, sp, &invalid_list);
2097 WARN_ON(!list_empty(&invalid_list));
2099 kvm_flush_remote_tlbs(vcpu->kvm);
2102 __clear_sp_write_flooding_count(sp);
2105 trace_kvm_mmu_get_page(sp, false);
2109 ++vcpu->kvm->stat.mmu_cache_miss;
2111 sp = kvm_mmu_alloc_page(vcpu, direct);
2115 hlist_add_head(&sp->hash_link, sp_list);
2117 account_shadowed(vcpu->kvm, sp);
2118 if (level == PG_LEVEL_4K && kvm_vcpu_write_protect_gfn(vcpu, gfn))
2119 kvm_flush_remote_tlbs_with_address(vcpu->kvm, gfn, 1);
2121 trace_kvm_mmu_get_page(sp, true);
2123 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2125 if (collisions > vcpu->kvm->stat.max_mmu_page_hash_collisions)
2126 vcpu->kvm->stat.max_mmu_page_hash_collisions = collisions;
2130 static void shadow_walk_init_using_root(struct kvm_shadow_walk_iterator *iterator,
2131 struct kvm_vcpu *vcpu, hpa_t root,
2134 iterator->addr = addr;
2135 iterator->shadow_addr = root;
2136 iterator->level = vcpu->arch.mmu->shadow_root_level;
2138 if (iterator->level >= PT64_ROOT_4LEVEL &&
2139 vcpu->arch.mmu->root_level < PT64_ROOT_4LEVEL &&
2140 !vcpu->arch.mmu->direct_map)
2141 iterator->level = PT32E_ROOT_LEVEL;
2143 if (iterator->level == PT32E_ROOT_LEVEL) {
2145 * prev_root is currently only used for 64-bit hosts. So only
2146 * the active root_hpa is valid here.
2148 BUG_ON(root != vcpu->arch.mmu->root.hpa);
2150 iterator->shadow_addr
2151 = vcpu->arch.mmu->pae_root[(addr >> 30) & 3];
2152 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
2154 if (!iterator->shadow_addr)
2155 iterator->level = 0;
2159 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
2160 struct kvm_vcpu *vcpu, u64 addr)
2162 shadow_walk_init_using_root(iterator, vcpu, vcpu->arch.mmu->root.hpa,
2166 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
2168 if (iterator->level < PG_LEVEL_4K)
2171 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
2172 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
2176 static void __shadow_walk_next(struct kvm_shadow_walk_iterator *iterator,
2179 if (!is_shadow_present_pte(spte) || is_last_spte(spte, iterator->level)) {
2180 iterator->level = 0;
2184 iterator->shadow_addr = spte & PT64_BASE_ADDR_MASK;
2188 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
2190 __shadow_walk_next(iterator, *iterator->sptep);
2193 static void link_shadow_page(struct kvm_vcpu *vcpu, u64 *sptep,
2194 struct kvm_mmu_page *sp)
2198 BUILD_BUG_ON(VMX_EPT_WRITABLE_MASK != PT_WRITABLE_MASK);
2200 spte = make_nonleaf_spte(sp->spt, sp_ad_disabled(sp));
2202 mmu_spte_set(sptep, spte);
2204 mmu_page_add_parent_pte(vcpu, sp, sptep);
2206 if (sp->unsync_children || sp->unsync)
2210 static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2211 unsigned direct_access)
2213 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
2214 struct kvm_mmu_page *child;
2217 * For the direct sp, if the guest pte's dirty bit
2218 * changed form clean to dirty, it will corrupt the
2219 * sp's access: allow writable in the read-only sp,
2220 * so we should update the spte at this point to get
2221 * a new sp with the correct access.
2223 child = to_shadow_page(*sptep & PT64_BASE_ADDR_MASK);
2224 if (child->role.access == direct_access)
2227 drop_parent_pte(child, sptep);
2228 kvm_flush_remote_tlbs_with_address(vcpu->kvm, child->gfn, 1);
2232 /* Returns the number of zapped non-leaf child shadow pages. */
2233 static int mmu_page_zap_pte(struct kvm *kvm, struct kvm_mmu_page *sp,
2234 u64 *spte, struct list_head *invalid_list)
2237 struct kvm_mmu_page *child;
2240 if (is_shadow_present_pte(pte)) {
2241 if (is_last_spte(pte, sp->role.level)) {
2242 drop_spte(kvm, spte);
2244 child = to_shadow_page(pte & PT64_BASE_ADDR_MASK);
2245 drop_parent_pte(child, spte);
2248 * Recursively zap nested TDP SPs, parentless SPs are
2249 * unlikely to be used again in the near future. This
2250 * avoids retaining a large number of stale nested SPs.
2252 if (tdp_enabled && invalid_list &&
2253 child->role.guest_mode && !child->parent_ptes.val)
2254 return kvm_mmu_prepare_zap_page(kvm, child,
2257 } else if (is_mmio_spte(pte)) {
2258 mmu_spte_clear_no_track(spte);
2263 static int kvm_mmu_page_unlink_children(struct kvm *kvm,
2264 struct kvm_mmu_page *sp,
2265 struct list_head *invalid_list)
2270 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2271 zapped += mmu_page_zap_pte(kvm, sp, sp->spt + i, invalid_list);
2276 static void kvm_mmu_unlink_parents(struct kvm_mmu_page *sp)
2279 struct rmap_iterator iter;
2281 while ((sptep = rmap_get_first(&sp->parent_ptes, &iter)))
2282 drop_parent_pte(sp, sptep);
2285 static int mmu_zap_unsync_children(struct kvm *kvm,
2286 struct kvm_mmu_page *parent,
2287 struct list_head *invalid_list)
2290 struct mmu_page_path parents;
2291 struct kvm_mmu_pages pages;
2293 if (parent->role.level == PG_LEVEL_4K)
2296 while (mmu_unsync_walk(parent, &pages)) {
2297 struct kvm_mmu_page *sp;
2299 for_each_sp(pages, sp, parents, i) {
2300 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
2301 mmu_pages_clear_parents(&parents);
2309 static bool __kvm_mmu_prepare_zap_page(struct kvm *kvm,
2310 struct kvm_mmu_page *sp,
2311 struct list_head *invalid_list,
2314 bool list_unstable, zapped_root = false;
2316 trace_kvm_mmu_prepare_zap_page(sp);
2317 ++kvm->stat.mmu_shadow_zapped;
2318 *nr_zapped = mmu_zap_unsync_children(kvm, sp, invalid_list);
2319 *nr_zapped += kvm_mmu_page_unlink_children(kvm, sp, invalid_list);
2320 kvm_mmu_unlink_parents(sp);
2322 /* Zapping children means active_mmu_pages has become unstable. */
2323 list_unstable = *nr_zapped;
2325 if (!sp->role.invalid && !sp->role.direct)
2326 unaccount_shadowed(kvm, sp);
2329 kvm_unlink_unsync_page(kvm, sp);
2330 if (!sp->root_count) {
2335 * Already invalid pages (previously active roots) are not on
2336 * the active page list. See list_del() in the "else" case of
2339 if (sp->role.invalid)
2340 list_add(&sp->link, invalid_list);
2342 list_move(&sp->link, invalid_list);
2343 kvm_mod_used_mmu_pages(kvm, -1);
2346 * Remove the active root from the active page list, the root
2347 * will be explicitly freed when the root_count hits zero.
2349 list_del(&sp->link);
2352 * Obsolete pages cannot be used on any vCPUs, see the comment
2353 * in kvm_mmu_zap_all_fast(). Note, is_obsolete_sp() also
2354 * treats invalid shadow pages as being obsolete.
2356 zapped_root = !is_obsolete_sp(kvm, sp);
2359 if (sp->lpage_disallowed)
2360 unaccount_huge_nx_page(kvm, sp);
2362 sp->role.invalid = 1;
2365 * Make the request to free obsolete roots after marking the root
2366 * invalid, otherwise other vCPUs may not see it as invalid.
2369 kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_FREE_OBSOLETE_ROOTS);
2370 return list_unstable;
2373 static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
2374 struct list_head *invalid_list)
2378 __kvm_mmu_prepare_zap_page(kvm, sp, invalid_list, &nr_zapped);
2382 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
2383 struct list_head *invalid_list)
2385 struct kvm_mmu_page *sp, *nsp;
2387 if (list_empty(invalid_list))
2391 * We need to make sure everyone sees our modifications to
2392 * the page tables and see changes to vcpu->mode here. The barrier
2393 * in the kvm_flush_remote_tlbs() achieves this. This pairs
2394 * with vcpu_enter_guest and walk_shadow_page_lockless_begin/end.
2396 * In addition, kvm_flush_remote_tlbs waits for all vcpus to exit
2397 * guest mode and/or lockless shadow page table walks.
2399 kvm_flush_remote_tlbs(kvm);
2401 list_for_each_entry_safe(sp, nsp, invalid_list, link) {
2402 WARN_ON(!sp->role.invalid || sp->root_count);
2403 kvm_mmu_free_page(sp);
2407 static unsigned long kvm_mmu_zap_oldest_mmu_pages(struct kvm *kvm,
2408 unsigned long nr_to_zap)
2410 unsigned long total_zapped = 0;
2411 struct kvm_mmu_page *sp, *tmp;
2412 LIST_HEAD(invalid_list);
2416 if (list_empty(&kvm->arch.active_mmu_pages))
2420 list_for_each_entry_safe_reverse(sp, tmp, &kvm->arch.active_mmu_pages, link) {
2422 * Don't zap active root pages, the page itself can't be freed
2423 * and zapping it will just force vCPUs to realloc and reload.
2428 unstable = __kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list,
2430 total_zapped += nr_zapped;
2431 if (total_zapped >= nr_to_zap)
2438 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2440 kvm->stat.mmu_recycled += total_zapped;
2441 return total_zapped;
2444 static inline unsigned long kvm_mmu_available_pages(struct kvm *kvm)
2446 if (kvm->arch.n_max_mmu_pages > kvm->arch.n_used_mmu_pages)
2447 return kvm->arch.n_max_mmu_pages -
2448 kvm->arch.n_used_mmu_pages;
2453 static int make_mmu_pages_available(struct kvm_vcpu *vcpu)
2455 unsigned long avail = kvm_mmu_available_pages(vcpu->kvm);
2457 if (likely(avail >= KVM_MIN_FREE_MMU_PAGES))
2460 kvm_mmu_zap_oldest_mmu_pages(vcpu->kvm, KVM_REFILL_PAGES - avail);
2463 * Note, this check is intentionally soft, it only guarantees that one
2464 * page is available, while the caller may end up allocating as many as
2465 * four pages, e.g. for PAE roots or for 5-level paging. Temporarily
2466 * exceeding the (arbitrary by default) limit will not harm the host,
2467 * being too aggressive may unnecessarily kill the guest, and getting an
2468 * exact count is far more trouble than it's worth, especially in the
2471 if (!kvm_mmu_available_pages(vcpu->kvm))
2477 * Changing the number of mmu pages allocated to the vm
2478 * Note: if goal_nr_mmu_pages is too small, you will get dead lock
2480 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned long goal_nr_mmu_pages)
2482 write_lock(&kvm->mmu_lock);
2484 if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) {
2485 kvm_mmu_zap_oldest_mmu_pages(kvm, kvm->arch.n_used_mmu_pages -
2488 goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages;
2491 kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages;
2493 write_unlock(&kvm->mmu_lock);
2496 int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
2498 struct kvm_mmu_page *sp;
2499 LIST_HEAD(invalid_list);
2502 pgprintk("%s: looking for gfn %llx\n", __func__, gfn);
2504 write_lock(&kvm->mmu_lock);
2505 for_each_gfn_indirect_valid_sp(kvm, sp, gfn) {
2506 pgprintk("%s: gfn %llx role %x\n", __func__, gfn,
2509 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
2511 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2512 write_unlock(&kvm->mmu_lock);
2517 static int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2522 if (vcpu->arch.mmu->direct_map)
2525 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2527 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2532 static void kvm_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
2534 trace_kvm_mmu_unsync_page(sp);
2535 ++kvm->stat.mmu_unsync;
2538 kvm_mmu_mark_parents_unsync(sp);
2542 * Attempt to unsync any shadow pages that can be reached by the specified gfn,
2543 * KVM is creating a writable mapping for said gfn. Returns 0 if all pages
2544 * were marked unsync (or if there is no shadow page), -EPERM if the SPTE must
2545 * be write-protected.
2547 int mmu_try_to_unsync_pages(struct kvm *kvm, const struct kvm_memory_slot *slot,
2548 gfn_t gfn, bool can_unsync, bool prefetch)
2550 struct kvm_mmu_page *sp;
2551 bool locked = false;
2554 * Force write-protection if the page is being tracked. Note, the page
2555 * track machinery is used to write-protect upper-level shadow pages,
2556 * i.e. this guards the role.level == 4K assertion below!
2558 if (kvm_slot_page_track_is_active(kvm, slot, gfn, KVM_PAGE_TRACK_WRITE))
2562 * The page is not write-tracked, mark existing shadow pages unsync
2563 * unless KVM is synchronizing an unsync SP (can_unsync = false). In
2564 * that case, KVM must complete emulation of the guest TLB flush before
2565 * allowing shadow pages to become unsync (writable by the guest).
2567 for_each_gfn_indirect_valid_sp(kvm, sp, gfn) {
2578 * TDP MMU page faults require an additional spinlock as they
2579 * run with mmu_lock held for read, not write, and the unsync
2580 * logic is not thread safe. Take the spinklock regardless of
2581 * the MMU type to avoid extra conditionals/parameters, there's
2582 * no meaningful penalty if mmu_lock is held for write.
2586 spin_lock(&kvm->arch.mmu_unsync_pages_lock);
2589 * Recheck after taking the spinlock, a different vCPU
2590 * may have since marked the page unsync. A false
2591 * positive on the unprotected check above is not
2592 * possible as clearing sp->unsync _must_ hold mmu_lock
2593 * for write, i.e. unsync cannot transition from 0->1
2594 * while this CPU holds mmu_lock for read (or write).
2596 if (READ_ONCE(sp->unsync))
2600 WARN_ON(sp->role.level != PG_LEVEL_4K);
2601 kvm_unsync_page(kvm, sp);
2604 spin_unlock(&kvm->arch.mmu_unsync_pages_lock);
2607 * We need to ensure that the marking of unsync pages is visible
2608 * before the SPTE is updated to allow writes because
2609 * kvm_mmu_sync_roots() checks the unsync flags without holding
2610 * the MMU lock and so can race with this. If the SPTE was updated
2611 * before the page had been marked as unsync-ed, something like the
2612 * following could happen:
2615 * ---------------------------------------------------------------------
2616 * 1.2 Host updates SPTE
2618 * 2.1 Guest writes a GPTE for GVA X.
2619 * (GPTE being in the guest page table shadowed
2620 * by the SP from CPU 1.)
2621 * This reads SPTE during the page table walk.
2622 * Since SPTE.W is read as 1, there is no
2625 * 2.2 Guest issues TLB flush.
2626 * That causes a VM Exit.
2628 * 2.3 Walking of unsync pages sees sp->unsync is
2629 * false and skips the page.
2631 * 2.4 Guest accesses GVA X.
2632 * Since the mapping in the SP was not updated,
2633 * so the old mapping for GVA X incorrectly
2637 * (sp->unsync = true)
2639 * The write barrier below ensures that 1.1 happens before 1.2 and thus
2640 * the situation in 2.4 does not arise. It pairs with the read barrier
2641 * in is_unsync_root(), placed between 2.1's load of SPTE.W and 2.3.
2648 static int mmu_set_spte(struct kvm_vcpu *vcpu, struct kvm_memory_slot *slot,
2649 u64 *sptep, unsigned int pte_access, gfn_t gfn,
2650 kvm_pfn_t pfn, struct kvm_page_fault *fault)
2652 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
2653 int level = sp->role.level;
2654 int was_rmapped = 0;
2655 int ret = RET_PF_FIXED;
2660 /* Prefetching always gets a writable pfn. */
2661 bool host_writable = !fault || fault->map_writable;
2662 bool prefetch = !fault || fault->prefetch;
2663 bool write_fault = fault && fault->write;
2665 pgprintk("%s: spte %llx write_fault %d gfn %llx\n", __func__,
2666 *sptep, write_fault, gfn);
2668 if (unlikely(is_noslot_pfn(pfn))) {
2669 mark_mmio_spte(vcpu, sptep, gfn, pte_access);
2670 return RET_PF_EMULATE;
2673 if (is_shadow_present_pte(*sptep)) {
2675 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
2676 * the parent of the now unreachable PTE.
2678 if (level > PG_LEVEL_4K && !is_large_pte(*sptep)) {
2679 struct kvm_mmu_page *child;
2682 child = to_shadow_page(pte & PT64_BASE_ADDR_MASK);
2683 drop_parent_pte(child, sptep);
2685 } else if (pfn != spte_to_pfn(*sptep)) {
2686 pgprintk("hfn old %llx new %llx\n",
2687 spte_to_pfn(*sptep), pfn);
2688 drop_spte(vcpu->kvm, sptep);
2694 wrprot = make_spte(vcpu, sp, slot, pte_access, gfn, pfn, *sptep, prefetch,
2695 true, host_writable, &spte);
2697 if (*sptep == spte) {
2698 ret = RET_PF_SPURIOUS;
2700 flush |= mmu_spte_update(sptep, spte);
2701 trace_kvm_mmu_set_spte(level, gfn, sptep);
2706 ret = RET_PF_EMULATE;
2710 kvm_flush_remote_tlbs_with_address(vcpu->kvm, gfn,
2711 KVM_PAGES_PER_HPAGE(level));
2713 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
2716 WARN_ON_ONCE(ret == RET_PF_SPURIOUS);
2717 kvm_update_page_stats(vcpu->kvm, level, 1);
2718 rmap_add(vcpu, slot, sptep, gfn);
2724 static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu,
2725 struct kvm_mmu_page *sp,
2726 u64 *start, u64 *end)
2728 struct page *pages[PTE_PREFETCH_NUM];
2729 struct kvm_memory_slot *slot;
2730 unsigned int access = sp->role.access;
2734 gfn = kvm_mmu_page_get_gfn(sp, start - sp->spt);
2735 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, access & ACC_WRITE_MASK);
2739 ret = gfn_to_page_many_atomic(slot, gfn, pages, end - start);
2743 for (i = 0; i < ret; i++, gfn++, start++) {
2744 mmu_set_spte(vcpu, slot, start, access, gfn,
2745 page_to_pfn(pages[i]), NULL);
2752 static void __direct_pte_prefetch(struct kvm_vcpu *vcpu,
2753 struct kvm_mmu_page *sp, u64 *sptep)
2755 u64 *spte, *start = NULL;
2758 WARN_ON(!sp->role.direct);
2760 i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
2763 for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
2764 if (is_shadow_present_pte(*spte) || spte == sptep) {
2767 if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0)
2774 direct_pte_prefetch_many(vcpu, sp, start, spte);
2777 static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
2779 struct kvm_mmu_page *sp;
2781 sp = sptep_to_sp(sptep);
2784 * Without accessed bits, there's no way to distinguish between
2785 * actually accessed translations and prefetched, so disable pte
2786 * prefetch if accessed bits aren't available.
2788 if (sp_ad_disabled(sp))
2791 if (sp->role.level > PG_LEVEL_4K)
2795 * If addresses are being invalidated, skip prefetching to avoid
2796 * accidentally prefetching those addresses.
2798 if (unlikely(vcpu->kvm->mmu_notifier_count))
2801 __direct_pte_prefetch(vcpu, sp, sptep);
2804 static int host_pfn_mapping_level(struct kvm *kvm, gfn_t gfn, kvm_pfn_t pfn,
2805 const struct kvm_memory_slot *slot)
2808 unsigned long flags;
2809 int level = PG_LEVEL_4K;
2815 if (!PageCompound(pfn_to_page(pfn)) && !kvm_is_zone_device_pfn(pfn))
2819 * Note, using the already-retrieved memslot and __gfn_to_hva_memslot()
2820 * is not solely for performance, it's also necessary to avoid the
2821 * "writable" check in __gfn_to_hva_many(), which will always fail on
2822 * read-only memslots due to gfn_to_hva() assuming writes. Earlier
2823 * page fault steps have already verified the guest isn't writing a
2824 * read-only memslot.
2826 hva = __gfn_to_hva_memslot(slot, gfn);
2829 * Lookup the mapping level in the current mm. The information
2830 * may become stale soon, but it is safe to use as long as
2831 * 1) mmu_notifier_retry was checked after taking mmu_lock, and
2832 * 2) mmu_lock is taken now.
2834 * We still need to disable IRQs to prevent concurrent tear down
2837 local_irq_save(flags);
2839 pgd = READ_ONCE(*pgd_offset(kvm->mm, hva));
2843 p4d = READ_ONCE(*p4d_offset(&pgd, hva));
2844 if (p4d_none(p4d) || !p4d_present(p4d))
2847 pud = READ_ONCE(*pud_offset(&p4d, hva));
2848 if (pud_none(pud) || !pud_present(pud))
2851 if (pud_large(pud)) {
2852 level = PG_LEVEL_1G;
2856 pmd = READ_ONCE(*pmd_offset(&pud, hva));
2857 if (pmd_none(pmd) || !pmd_present(pmd))
2861 level = PG_LEVEL_2M;
2864 local_irq_restore(flags);
2868 int kvm_mmu_max_mapping_level(struct kvm *kvm,
2869 const struct kvm_memory_slot *slot, gfn_t gfn,
2870 kvm_pfn_t pfn, int max_level)
2872 struct kvm_lpage_info *linfo;
2875 max_level = min(max_level, max_huge_page_level);
2876 for ( ; max_level > PG_LEVEL_4K; max_level--) {
2877 linfo = lpage_info_slot(gfn, slot, max_level);
2878 if (!linfo->disallow_lpage)
2882 if (max_level == PG_LEVEL_4K)
2885 host_level = host_pfn_mapping_level(kvm, gfn, pfn, slot);
2886 return min(host_level, max_level);
2889 void kvm_mmu_hugepage_adjust(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
2891 struct kvm_memory_slot *slot = fault->slot;
2894 fault->huge_page_disallowed = fault->exec && fault->nx_huge_page_workaround_enabled;
2896 if (unlikely(fault->max_level == PG_LEVEL_4K))
2899 if (is_error_noslot_pfn(fault->pfn) || kvm_is_reserved_pfn(fault->pfn))
2902 if (kvm_slot_dirty_track_enabled(slot))
2906 * Enforce the iTLB multihit workaround after capturing the requested
2907 * level, which will be used to do precise, accurate accounting.
2909 fault->req_level = kvm_mmu_max_mapping_level(vcpu->kvm, slot,
2910 fault->gfn, fault->pfn,
2912 if (fault->req_level == PG_LEVEL_4K || fault->huge_page_disallowed)
2916 * mmu_notifier_retry() was successful and mmu_lock is held, so
2917 * the pmd can't be split from under us.
2919 fault->goal_level = fault->req_level;
2920 mask = KVM_PAGES_PER_HPAGE(fault->goal_level) - 1;
2921 VM_BUG_ON((fault->gfn & mask) != (fault->pfn & mask));
2922 fault->pfn &= ~mask;
2925 void disallowed_hugepage_adjust(struct kvm_page_fault *fault, u64 spte, int cur_level)
2927 if (cur_level > PG_LEVEL_4K &&
2928 cur_level == fault->goal_level &&
2929 is_shadow_present_pte(spte) &&
2930 !is_large_pte(spte)) {
2932 * A small SPTE exists for this pfn, but FNAME(fetch)
2933 * and __direct_map would like to create a large PTE
2934 * instead: just force them to go down another level,
2935 * patching back for them into pfn the next 9 bits of
2938 u64 page_mask = KVM_PAGES_PER_HPAGE(cur_level) -
2939 KVM_PAGES_PER_HPAGE(cur_level - 1);
2940 fault->pfn |= fault->gfn & page_mask;
2941 fault->goal_level--;
2945 static int __direct_map(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
2947 struct kvm_shadow_walk_iterator it;
2948 struct kvm_mmu_page *sp;
2950 gfn_t base_gfn = fault->gfn;
2952 kvm_mmu_hugepage_adjust(vcpu, fault);
2954 trace_kvm_mmu_spte_requested(fault);
2955 for_each_shadow_entry(vcpu, fault->addr, it) {
2957 * We cannot overwrite existing page tables with an NX
2958 * large page, as the leaf could be executable.
2960 if (fault->nx_huge_page_workaround_enabled)
2961 disallowed_hugepage_adjust(fault, *it.sptep, it.level);
2963 base_gfn = fault->gfn & ~(KVM_PAGES_PER_HPAGE(it.level) - 1);
2964 if (it.level == fault->goal_level)
2967 drop_large_spte(vcpu, it.sptep);
2968 if (is_shadow_present_pte(*it.sptep))
2971 sp = kvm_mmu_get_page(vcpu, base_gfn, it.addr,
2972 it.level - 1, true, ACC_ALL);
2974 link_shadow_page(vcpu, it.sptep, sp);
2975 if (fault->is_tdp && fault->huge_page_disallowed &&
2976 fault->req_level >= it.level)
2977 account_huge_nx_page(vcpu->kvm, sp);
2980 if (WARN_ON_ONCE(it.level != fault->goal_level))
2983 ret = mmu_set_spte(vcpu, fault->slot, it.sptep, ACC_ALL,
2984 base_gfn, fault->pfn, fault);
2985 if (ret == RET_PF_SPURIOUS)
2988 direct_pte_prefetch(vcpu, it.sptep);
2989 ++vcpu->stat.pf_fixed;
2993 static void kvm_send_hwpoison_signal(unsigned long address, struct task_struct *tsk)
2995 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, PAGE_SHIFT, tsk);
2998 static int kvm_handle_bad_page(struct kvm_vcpu *vcpu, gfn_t gfn, kvm_pfn_t pfn)
3001 * Do not cache the mmio info caused by writing the readonly gfn
3002 * into the spte otherwise read access on readonly gfn also can
3003 * caused mmio page fault and treat it as mmio access.
3005 if (pfn == KVM_PFN_ERR_RO_FAULT)
3006 return RET_PF_EMULATE;
3008 if (pfn == KVM_PFN_ERR_HWPOISON) {
3009 kvm_send_hwpoison_signal(kvm_vcpu_gfn_to_hva(vcpu, gfn), current);
3010 return RET_PF_RETRY;
3016 static bool handle_abnormal_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault,
3017 unsigned int access, int *ret_val)
3019 /* The pfn is invalid, report the error! */
3020 if (unlikely(is_error_pfn(fault->pfn))) {
3021 *ret_val = kvm_handle_bad_page(vcpu, fault->gfn, fault->pfn);
3025 if (unlikely(!fault->slot)) {
3026 gva_t gva = fault->is_tdp ? 0 : fault->addr;
3028 vcpu_cache_mmio_info(vcpu, gva, fault->gfn,
3029 access & shadow_mmio_access_mask);
3031 * If MMIO caching is disabled, emulate immediately without
3032 * touching the shadow page tables as attempting to install an
3033 * MMIO SPTE will just be an expensive nop. Do not cache MMIO
3034 * whose gfn is greater than host.MAXPHYADDR, any guest that
3035 * generates such gfns is running nested and is being tricked
3036 * by L0 userspace (you can observe gfn > L1.MAXPHYADDR if
3037 * and only if L1's MAXPHYADDR is inaccurate with respect to
3040 if (unlikely(!enable_mmio_caching) ||
3041 unlikely(fault->gfn > kvm_mmu_max_gfn())) {
3042 *ret_val = RET_PF_EMULATE;
3050 static bool page_fault_can_be_fast(struct kvm_page_fault *fault)
3053 * Do not fix the mmio spte with invalid generation number which
3054 * need to be updated by slow page fault path.
3059 /* See if the page fault is due to an NX violation */
3060 if (unlikely(fault->exec && fault->present))
3064 * #PF can be fast if:
3065 * 1. The shadow page table entry is not present, which could mean that
3066 * the fault is potentially caused by access tracking (if enabled).
3067 * 2. The shadow page table entry is present and the fault
3068 * is caused by write-protect, that means we just need change the W
3069 * bit of the spte which can be done out of mmu-lock.
3071 * However, if access tracking is disabled we know that a non-present
3072 * page must be a genuine page fault where we have to create a new SPTE.
3073 * So, if access tracking is disabled, we return true only for write
3074 * accesses to a present page.
3077 return shadow_acc_track_mask != 0 || (fault->write && fault->present);
3081 * Returns true if the SPTE was fixed successfully. Otherwise,
3082 * someone else modified the SPTE from its original value.
3085 fast_pf_fix_direct_spte(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault,
3086 u64 *sptep, u64 old_spte, u64 new_spte)
3089 * Theoretically we could also set dirty bit (and flush TLB) here in
3090 * order to eliminate unnecessary PML logging. See comments in
3091 * set_spte. But fast_page_fault is very unlikely to happen with PML
3092 * enabled, so we do not do this. This might result in the same GPA
3093 * to be logged in PML buffer again when the write really happens, and
3094 * eventually to be called by mark_page_dirty twice. But it's also no
3095 * harm. This also avoids the TLB flush needed after setting dirty bit
3096 * so non-PML cases won't be impacted.
3098 * Compare with set_spte where instead shadow_dirty_mask is set.
3100 if (cmpxchg64(sptep, old_spte, new_spte) != old_spte)
3103 if (is_writable_pte(new_spte) && !is_writable_pte(old_spte))
3104 mark_page_dirty_in_slot(vcpu->kvm, fault->slot, fault->gfn);
3109 static bool is_access_allowed(struct kvm_page_fault *fault, u64 spte)
3112 return is_executable_pte(spte);
3115 return is_writable_pte(spte);
3117 /* Fault was on Read access */
3118 return spte & PT_PRESENT_MASK;
3122 * Returns the last level spte pointer of the shadow page walk for the given
3123 * gpa, and sets *spte to the spte value. This spte may be non-preset. If no
3124 * walk could be performed, returns NULL and *spte does not contain valid data.
3127 * - Must be called between walk_shadow_page_lockless_{begin,end}.
3128 * - The returned sptep must not be used after walk_shadow_page_lockless_end.
3130 static u64 *fast_pf_get_last_sptep(struct kvm_vcpu *vcpu, gpa_t gpa, u64 *spte)
3132 struct kvm_shadow_walk_iterator iterator;
3136 for_each_shadow_entry_lockless(vcpu, gpa, iterator, old_spte) {
3137 sptep = iterator.sptep;
3145 * Returns one of RET_PF_INVALID, RET_PF_FIXED or RET_PF_SPURIOUS.
3147 static int fast_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
3149 struct kvm_mmu_page *sp;
3150 int ret = RET_PF_INVALID;
3153 uint retry_count = 0;
3155 if (!page_fault_can_be_fast(fault))
3158 walk_shadow_page_lockless_begin(vcpu);
3163 if (is_tdp_mmu(vcpu->arch.mmu))
3164 sptep = kvm_tdp_mmu_fast_pf_get_last_sptep(vcpu, fault->addr, &spte);
3166 sptep = fast_pf_get_last_sptep(vcpu, fault->addr, &spte);
3168 if (!is_shadow_present_pte(spte))
3171 sp = sptep_to_sp(sptep);
3172 if (!is_last_spte(spte, sp->role.level))
3176 * Check whether the memory access that caused the fault would
3177 * still cause it if it were to be performed right now. If not,
3178 * then this is a spurious fault caused by TLB lazily flushed,
3179 * or some other CPU has already fixed the PTE after the
3180 * current CPU took the fault.
3182 * Need not check the access of upper level table entries since
3183 * they are always ACC_ALL.
3185 if (is_access_allowed(fault, spte)) {
3186 ret = RET_PF_SPURIOUS;
3192 if (is_access_track_spte(spte))
3193 new_spte = restore_acc_track_spte(new_spte);
3196 * Currently, to simplify the code, write-protection can
3197 * be removed in the fast path only if the SPTE was
3198 * write-protected for dirty-logging or access tracking.
3201 spte_can_locklessly_be_made_writable(spte)) {
3202 new_spte |= PT_WRITABLE_MASK;
3205 * Do not fix write-permission on the large spte when
3206 * dirty logging is enabled. Since we only dirty the
3207 * first page into the dirty-bitmap in
3208 * fast_pf_fix_direct_spte(), other pages are missed
3209 * if its slot has dirty logging enabled.
3211 * Instead, we let the slow page fault path create a
3212 * normal spte to fix the access.
3214 if (sp->role.level > PG_LEVEL_4K &&
3215 kvm_slot_dirty_track_enabled(fault->slot))
3219 /* Verify that the fault can be handled in the fast path */
3220 if (new_spte == spte ||
3221 !is_access_allowed(fault, new_spte))
3225 * Currently, fast page fault only works for direct mapping
3226 * since the gfn is not stable for indirect shadow page. See
3227 * Documentation/virt/kvm/locking.rst to get more detail.
3229 if (fast_pf_fix_direct_spte(vcpu, fault, sptep, spte, new_spte)) {
3234 if (++retry_count > 4) {
3235 printk_once(KERN_WARNING
3236 "kvm: Fast #PF retrying more than 4 times.\n");
3242 trace_fast_page_fault(vcpu, fault, sptep, spte, ret);
3243 walk_shadow_page_lockless_end(vcpu);
3248 static void mmu_free_root_page(struct kvm *kvm, hpa_t *root_hpa,
3249 struct list_head *invalid_list)
3251 struct kvm_mmu_page *sp;
3253 if (!VALID_PAGE(*root_hpa))
3256 sp = to_shadow_page(*root_hpa & PT64_BASE_ADDR_MASK);
3260 if (is_tdp_mmu_page(sp))
3261 kvm_tdp_mmu_put_root(kvm, sp, false);
3262 else if (!--sp->root_count && sp->role.invalid)
3263 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
3265 *root_hpa = INVALID_PAGE;
3268 /* roots_to_free must be some combination of the KVM_MMU_ROOT_* flags */
3269 void kvm_mmu_free_roots(struct kvm *kvm, struct kvm_mmu *mmu,
3270 ulong roots_to_free)
3273 LIST_HEAD(invalid_list);
3274 bool free_active_root;
3276 BUILD_BUG_ON(KVM_MMU_NUM_PREV_ROOTS >= BITS_PER_LONG);
3278 /* Before acquiring the MMU lock, see if we need to do any real work. */
3279 free_active_root = (roots_to_free & KVM_MMU_ROOT_CURRENT)
3280 && VALID_PAGE(mmu->root.hpa);
3282 if (!free_active_root) {
3283 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3284 if ((roots_to_free & KVM_MMU_ROOT_PREVIOUS(i)) &&
3285 VALID_PAGE(mmu->prev_roots[i].hpa))
3288 if (i == KVM_MMU_NUM_PREV_ROOTS)
3292 write_lock(&kvm->mmu_lock);
3294 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3295 if (roots_to_free & KVM_MMU_ROOT_PREVIOUS(i))
3296 mmu_free_root_page(kvm, &mmu->prev_roots[i].hpa,
3299 if (free_active_root) {
3300 if (to_shadow_page(mmu->root.hpa)) {
3301 mmu_free_root_page(kvm, &mmu->root.hpa, &invalid_list);
3302 } else if (mmu->pae_root) {
3303 for (i = 0; i < 4; ++i) {
3304 if (!IS_VALID_PAE_ROOT(mmu->pae_root[i]))
3307 mmu_free_root_page(kvm, &mmu->pae_root[i],
3309 mmu->pae_root[i] = INVALID_PAE_ROOT;
3312 mmu->root.hpa = INVALID_PAGE;
3316 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3317 write_unlock(&kvm->mmu_lock);
3319 EXPORT_SYMBOL_GPL(kvm_mmu_free_roots);
3321 void kvm_mmu_free_guest_mode_roots(struct kvm *kvm, struct kvm_mmu *mmu)
3323 unsigned long roots_to_free = 0;
3328 * This should not be called while L2 is active, L2 can't invalidate
3329 * _only_ its own roots, e.g. INVVPID unconditionally exits.
3331 WARN_ON_ONCE(mmu->mmu_role.base.guest_mode);
3333 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
3334 root_hpa = mmu->prev_roots[i].hpa;
3335 if (!VALID_PAGE(root_hpa))
3338 if (!to_shadow_page(root_hpa) ||
3339 to_shadow_page(root_hpa)->role.guest_mode)
3340 roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i);
3343 kvm_mmu_free_roots(kvm, mmu, roots_to_free);
3345 EXPORT_SYMBOL_GPL(kvm_mmu_free_guest_mode_roots);
3348 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
3352 if (!kvm_vcpu_is_visible_gfn(vcpu, root_gfn)) {
3353 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
3360 static hpa_t mmu_alloc_root(struct kvm_vcpu *vcpu, gfn_t gfn, gva_t gva,
3361 u8 level, bool direct)
3363 struct kvm_mmu_page *sp;
3365 sp = kvm_mmu_get_page(vcpu, gfn, gva, level, direct, ACC_ALL);
3368 return __pa(sp->spt);
3371 static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
3373 struct kvm_mmu *mmu = vcpu->arch.mmu;
3374 u8 shadow_root_level = mmu->shadow_root_level;
3379 write_lock(&vcpu->kvm->mmu_lock);
3380 r = make_mmu_pages_available(vcpu);
3384 if (is_tdp_mmu_enabled(vcpu->kvm)) {
3385 root = kvm_tdp_mmu_get_vcpu_root_hpa(vcpu);
3386 mmu->root.hpa = root;
3387 } else if (shadow_root_level >= PT64_ROOT_4LEVEL) {
3388 root = mmu_alloc_root(vcpu, 0, 0, shadow_root_level, true);
3389 mmu->root.hpa = root;
3390 } else if (shadow_root_level == PT32E_ROOT_LEVEL) {
3391 if (WARN_ON_ONCE(!mmu->pae_root)) {
3396 for (i = 0; i < 4; ++i) {
3397 WARN_ON_ONCE(IS_VALID_PAE_ROOT(mmu->pae_root[i]));
3399 root = mmu_alloc_root(vcpu, i << (30 - PAGE_SHIFT),
3400 i << 30, PT32_ROOT_LEVEL, true);
3401 mmu->pae_root[i] = root | PT_PRESENT_MASK |
3404 mmu->root.hpa = __pa(mmu->pae_root);
3406 WARN_ONCE(1, "Bad TDP root level = %d\n", shadow_root_level);
3411 /* root.pgd is ignored for direct MMUs. */
3414 write_unlock(&vcpu->kvm->mmu_lock);
3418 static int mmu_first_shadow_root_alloc(struct kvm *kvm)
3420 struct kvm_memslots *slots;
3421 struct kvm_memory_slot *slot;
3425 * Check if this is the first shadow root being allocated before
3428 if (kvm_shadow_root_allocated(kvm))
3431 mutex_lock(&kvm->slots_arch_lock);
3433 /* Recheck, under the lock, whether this is the first shadow root. */
3434 if (kvm_shadow_root_allocated(kvm))
3438 * Check if anything actually needs to be allocated, e.g. all metadata
3439 * will be allocated upfront if TDP is disabled.
3441 if (kvm_memslots_have_rmaps(kvm) &&
3442 kvm_page_track_write_tracking_enabled(kvm))
3445 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
3446 slots = __kvm_memslots(kvm, i);
3447 kvm_for_each_memslot(slot, bkt, slots) {
3449 * Both of these functions are no-ops if the target is
3450 * already allocated, so unconditionally calling both
3451 * is safe. Intentionally do NOT free allocations on
3452 * failure to avoid having to track which allocations
3453 * were made now versus when the memslot was created.
3454 * The metadata is guaranteed to be freed when the slot
3455 * is freed, and will be kept/used if userspace retries
3456 * KVM_RUN instead of killing the VM.
3458 r = memslot_rmap_alloc(slot, slot->npages);
3461 r = kvm_page_track_write_tracking_alloc(slot);
3468 * Ensure that shadow_root_allocated becomes true strictly after
3469 * all the related pointers are set.
3472 smp_store_release(&kvm->arch.shadow_root_allocated, true);
3475 mutex_unlock(&kvm->slots_arch_lock);
3479 static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
3481 struct kvm_mmu *mmu = vcpu->arch.mmu;
3482 u64 pdptrs[4], pm_mask;
3483 gfn_t root_gfn, root_pgd;
3488 root_pgd = mmu->get_guest_pgd(vcpu);
3489 root_gfn = root_pgd >> PAGE_SHIFT;
3491 if (mmu_check_root(vcpu, root_gfn))
3495 * On SVM, reading PDPTRs might access guest memory, which might fault
3496 * and thus might sleep. Grab the PDPTRs before acquiring mmu_lock.
3498 if (mmu->root_level == PT32E_ROOT_LEVEL) {
3499 for (i = 0; i < 4; ++i) {
3500 pdptrs[i] = mmu->get_pdptr(vcpu, i);
3501 if (!(pdptrs[i] & PT_PRESENT_MASK))
3504 if (mmu_check_root(vcpu, pdptrs[i] >> PAGE_SHIFT))
3509 r = mmu_first_shadow_root_alloc(vcpu->kvm);
3513 write_lock(&vcpu->kvm->mmu_lock);
3514 r = make_mmu_pages_available(vcpu);
3519 * Do we shadow a long mode page table? If so we need to
3520 * write-protect the guests page table root.
3522 if (mmu->root_level >= PT64_ROOT_4LEVEL) {
3523 root = mmu_alloc_root(vcpu, root_gfn, 0,
3524 mmu->shadow_root_level, false);
3525 mmu->root.hpa = root;
3529 if (WARN_ON_ONCE(!mmu->pae_root)) {
3535 * We shadow a 32 bit page table. This may be a legacy 2-level
3536 * or a PAE 3-level page table. In either case we need to be aware that
3537 * the shadow page table may be a PAE or a long mode page table.
3539 pm_mask = PT_PRESENT_MASK | shadow_me_mask;
3540 if (mmu->shadow_root_level >= PT64_ROOT_4LEVEL) {
3541 pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK;
3543 if (WARN_ON_ONCE(!mmu->pml4_root)) {
3547 mmu->pml4_root[0] = __pa(mmu->pae_root) | pm_mask;
3549 if (mmu->shadow_root_level == PT64_ROOT_5LEVEL) {
3550 if (WARN_ON_ONCE(!mmu->pml5_root)) {
3554 mmu->pml5_root[0] = __pa(mmu->pml4_root) | pm_mask;
3558 for (i = 0; i < 4; ++i) {
3559 WARN_ON_ONCE(IS_VALID_PAE_ROOT(mmu->pae_root[i]));
3561 if (mmu->root_level == PT32E_ROOT_LEVEL) {
3562 if (!(pdptrs[i] & PT_PRESENT_MASK)) {
3563 mmu->pae_root[i] = INVALID_PAE_ROOT;
3566 root_gfn = pdptrs[i] >> PAGE_SHIFT;
3569 root = mmu_alloc_root(vcpu, root_gfn, i << 30,
3570 PT32_ROOT_LEVEL, false);
3571 mmu->pae_root[i] = root | pm_mask;
3574 if (mmu->shadow_root_level == PT64_ROOT_5LEVEL)
3575 mmu->root.hpa = __pa(mmu->pml5_root);
3576 else if (mmu->shadow_root_level == PT64_ROOT_4LEVEL)
3577 mmu->root.hpa = __pa(mmu->pml4_root);
3579 mmu->root.hpa = __pa(mmu->pae_root);
3582 mmu->root.pgd = root_pgd;
3584 write_unlock(&vcpu->kvm->mmu_lock);
3589 static int mmu_alloc_special_roots(struct kvm_vcpu *vcpu)
3591 struct kvm_mmu *mmu = vcpu->arch.mmu;
3592 bool need_pml5 = mmu->shadow_root_level > PT64_ROOT_4LEVEL;
3593 u64 *pml5_root = NULL;
3594 u64 *pml4_root = NULL;
3598 * When shadowing 32-bit or PAE NPT with 64-bit NPT, the PML4 and PDP
3599 * tables are allocated and initialized at root creation as there is no
3600 * equivalent level in the guest's NPT to shadow. Allocate the tables
3601 * on demand, as running a 32-bit L1 VMM on 64-bit KVM is very rare.
3603 if (mmu->direct_map || mmu->root_level >= PT64_ROOT_4LEVEL ||
3604 mmu->shadow_root_level < PT64_ROOT_4LEVEL)
3608 * NPT, the only paging mode that uses this horror, uses a fixed number
3609 * of levels for the shadow page tables, e.g. all MMUs are 4-level or
3610 * all MMus are 5-level. Thus, this can safely require that pml5_root
3611 * is allocated if the other roots are valid and pml5 is needed, as any
3612 * prior MMU would also have required pml5.
3614 if (mmu->pae_root && mmu->pml4_root && (!need_pml5 || mmu->pml5_root))
3618 * The special roots should always be allocated in concert. Yell and
3619 * bail if KVM ends up in a state where only one of the roots is valid.
3621 if (WARN_ON_ONCE(!tdp_enabled || mmu->pae_root || mmu->pml4_root ||
3622 (need_pml5 && mmu->pml5_root)))
3626 * Unlike 32-bit NPT, the PDP table doesn't need to be in low mem, and
3627 * doesn't need to be decrypted.
3629 pae_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT);
3633 #ifdef CONFIG_X86_64
3634 pml4_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT);
3639 pml5_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT);
3645 mmu->pae_root = pae_root;
3646 mmu->pml4_root = pml4_root;
3647 mmu->pml5_root = pml5_root;
3651 #ifdef CONFIG_X86_64
3653 free_page((unsigned long)pml4_root);
3655 free_page((unsigned long)pae_root);
3660 static bool is_unsync_root(hpa_t root)
3662 struct kvm_mmu_page *sp;
3664 if (!VALID_PAGE(root))
3668 * The read barrier orders the CPU's read of SPTE.W during the page table
3669 * walk before the reads of sp->unsync/sp->unsync_children here.
3671 * Even if another CPU was marking the SP as unsync-ed simultaneously,
3672 * any guest page table changes are not guaranteed to be visible anyway
3673 * until this VCPU issues a TLB flush strictly after those changes are
3674 * made. We only need to ensure that the other CPU sets these flags
3675 * before any actual changes to the page tables are made. The comments
3676 * in mmu_try_to_unsync_pages() describe what could go wrong if this
3677 * requirement isn't satisfied.
3680 sp = to_shadow_page(root);
3683 * PAE roots (somewhat arbitrarily) aren't backed by shadow pages, the
3684 * PDPTEs for a given PAE root need to be synchronized individually.
3686 if (WARN_ON_ONCE(!sp))
3689 if (sp->unsync || sp->unsync_children)
3695 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
3698 struct kvm_mmu_page *sp;
3700 if (vcpu->arch.mmu->direct_map)
3703 if (!VALID_PAGE(vcpu->arch.mmu->root.hpa))
3706 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
3708 if (vcpu->arch.mmu->root_level >= PT64_ROOT_4LEVEL) {
3709 hpa_t root = vcpu->arch.mmu->root.hpa;
3710 sp = to_shadow_page(root);
3712 if (!is_unsync_root(root))
3715 write_lock(&vcpu->kvm->mmu_lock);
3716 mmu_sync_children(vcpu, sp, true);
3717 write_unlock(&vcpu->kvm->mmu_lock);
3721 write_lock(&vcpu->kvm->mmu_lock);
3723 for (i = 0; i < 4; ++i) {
3724 hpa_t root = vcpu->arch.mmu->pae_root[i];
3726 if (IS_VALID_PAE_ROOT(root)) {
3727 root &= PT64_BASE_ADDR_MASK;
3728 sp = to_shadow_page(root);
3729 mmu_sync_children(vcpu, sp, true);
3733 write_unlock(&vcpu->kvm->mmu_lock);
3736 void kvm_mmu_sync_prev_roots(struct kvm_vcpu *vcpu)
3738 unsigned long roots_to_free = 0;
3741 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3742 if (is_unsync_root(vcpu->arch.mmu->prev_roots[i].hpa))
3743 roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i);
3745 /* sync prev_roots by simply freeing them */
3746 kvm_mmu_free_roots(vcpu->kvm, vcpu->arch.mmu, roots_to_free);
3749 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
3750 gpa_t vaddr, u64 access,
3751 struct x86_exception *exception)
3754 exception->error_code = 0;
3755 return kvm_translate_gpa(vcpu, mmu, vaddr, access, exception);
3758 static bool mmio_info_in_cache(struct kvm_vcpu *vcpu, u64 addr, bool direct)
3761 * A nested guest cannot use the MMIO cache if it is using nested
3762 * page tables, because cr2 is a nGPA while the cache stores GPAs.
3764 if (mmu_is_nested(vcpu))
3768 return vcpu_match_mmio_gpa(vcpu, addr);
3770 return vcpu_match_mmio_gva(vcpu, addr);
3774 * Return the level of the lowest level SPTE added to sptes.
3775 * That SPTE may be non-present.
3777 * Must be called between walk_shadow_page_lockless_{begin,end}.
3779 static int get_walk(struct kvm_vcpu *vcpu, u64 addr, u64 *sptes, int *root_level)
3781 struct kvm_shadow_walk_iterator iterator;
3785 for (shadow_walk_init(&iterator, vcpu, addr),
3786 *root_level = iterator.level;
3787 shadow_walk_okay(&iterator);
3788 __shadow_walk_next(&iterator, spte)) {
3789 leaf = iterator.level;
3790 spte = mmu_spte_get_lockless(iterator.sptep);
3798 /* return true if reserved bit(s) are detected on a valid, non-MMIO SPTE. */
3799 static bool get_mmio_spte(struct kvm_vcpu *vcpu, u64 addr, u64 *sptep)
3801 u64 sptes[PT64_ROOT_MAX_LEVEL + 1];
3802 struct rsvd_bits_validate *rsvd_check;
3803 int root, leaf, level;
3804 bool reserved = false;
3806 walk_shadow_page_lockless_begin(vcpu);
3808 if (is_tdp_mmu(vcpu->arch.mmu))
3809 leaf = kvm_tdp_mmu_get_walk(vcpu, addr, sptes, &root);
3811 leaf = get_walk(vcpu, addr, sptes, &root);
3813 walk_shadow_page_lockless_end(vcpu);
3815 if (unlikely(leaf < 0)) {
3820 *sptep = sptes[leaf];
3823 * Skip reserved bits checks on the terminal leaf if it's not a valid
3824 * SPTE. Note, this also (intentionally) skips MMIO SPTEs, which, by
3825 * design, always have reserved bits set. The purpose of the checks is
3826 * to detect reserved bits on non-MMIO SPTEs. i.e. buggy SPTEs.
3828 if (!is_shadow_present_pte(sptes[leaf]))
3831 rsvd_check = &vcpu->arch.mmu->shadow_zero_check;
3833 for (level = root; level >= leaf; level--)
3834 reserved |= is_rsvd_spte(rsvd_check, sptes[level], level);
3837 pr_err("%s: reserved bits set on MMU-present spte, addr 0x%llx, hierarchy:\n",
3839 for (level = root; level >= leaf; level--)
3840 pr_err("------ spte = 0x%llx level = %d, rsvd bits = 0x%llx",
3841 sptes[level], level,
3842 get_rsvd_bits(rsvd_check, sptes[level], level));
3848 static int handle_mmio_page_fault(struct kvm_vcpu *vcpu, u64 addr, bool direct)
3853 if (mmio_info_in_cache(vcpu, addr, direct))
3854 return RET_PF_EMULATE;
3856 reserved = get_mmio_spte(vcpu, addr, &spte);
3857 if (WARN_ON(reserved))
3860 if (is_mmio_spte(spte)) {
3861 gfn_t gfn = get_mmio_spte_gfn(spte);
3862 unsigned int access = get_mmio_spte_access(spte);
3864 if (!check_mmio_spte(vcpu, spte))
3865 return RET_PF_INVALID;
3870 trace_handle_mmio_page_fault(addr, gfn, access);
3871 vcpu_cache_mmio_info(vcpu, addr, gfn, access);
3872 return RET_PF_EMULATE;
3876 * If the page table is zapped by other cpus, let CPU fault again on
3879 return RET_PF_RETRY;
3882 static bool page_fault_handle_page_track(struct kvm_vcpu *vcpu,
3883 struct kvm_page_fault *fault)
3885 if (unlikely(fault->rsvd))
3888 if (!fault->present || !fault->write)
3892 * guest is writing the page which is write tracked which can
3893 * not be fixed by page fault handler.
3895 if (kvm_slot_page_track_is_active(vcpu->kvm, fault->slot, fault->gfn, KVM_PAGE_TRACK_WRITE))
3901 static void shadow_page_table_clear_flood(struct kvm_vcpu *vcpu, gva_t addr)
3903 struct kvm_shadow_walk_iterator iterator;
3906 walk_shadow_page_lockless_begin(vcpu);
3907 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte)
3908 clear_sp_write_flooding_count(iterator.sptep);
3909 walk_shadow_page_lockless_end(vcpu);
3912 static u32 alloc_apf_token(struct kvm_vcpu *vcpu)
3914 /* make sure the token value is not 0 */
3915 u32 id = vcpu->arch.apf.id;
3918 vcpu->arch.apf.id = 1;
3920 return (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id;
3923 static bool kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
3926 struct kvm_arch_async_pf arch;
3928 arch.token = alloc_apf_token(vcpu);
3930 arch.direct_map = vcpu->arch.mmu->direct_map;
3931 arch.cr3 = vcpu->arch.mmu->get_guest_pgd(vcpu);
3933 return kvm_setup_async_pf(vcpu, cr2_or_gpa,
3934 kvm_vcpu_gfn_to_hva(vcpu, gfn), &arch);
3937 static bool kvm_faultin_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault, int *r)
3939 struct kvm_memory_slot *slot = fault->slot;
3943 * Retry the page fault if the gfn hit a memslot that is being deleted
3944 * or moved. This ensures any existing SPTEs for the old memslot will
3945 * be zapped before KVM inserts a new MMIO SPTE for the gfn.
3947 if (slot && (slot->flags & KVM_MEMSLOT_INVALID))
3950 if (!kvm_is_visible_memslot(slot)) {
3951 /* Don't expose private memslots to L2. */
3952 if (is_guest_mode(vcpu)) {
3954 fault->pfn = KVM_PFN_NOSLOT;
3955 fault->map_writable = false;
3959 * If the APIC access page exists but is disabled, go directly
3960 * to emulation without caching the MMIO access or creating a
3961 * MMIO SPTE. That way the cache doesn't need to be purged
3962 * when the AVIC is re-enabled.
3964 if (slot && slot->id == APIC_ACCESS_PAGE_PRIVATE_MEMSLOT &&
3965 !kvm_apicv_activated(vcpu->kvm)) {
3966 *r = RET_PF_EMULATE;
3972 fault->pfn = __gfn_to_pfn_memslot(slot, fault->gfn, false, &async,
3973 fault->write, &fault->map_writable,
3976 return false; /* *pfn has correct page already */
3978 if (!fault->prefetch && kvm_can_do_async_pf(vcpu)) {
3979 trace_kvm_try_async_get_page(fault->addr, fault->gfn);
3980 if (kvm_find_async_pf_gfn(vcpu, fault->gfn)) {
3981 trace_kvm_async_pf_doublefault(fault->addr, fault->gfn);
3982 kvm_make_request(KVM_REQ_APF_HALT, vcpu);
3984 } else if (kvm_arch_setup_async_pf(vcpu, fault->addr, fault->gfn))
3988 fault->pfn = __gfn_to_pfn_memslot(slot, fault->gfn, false, NULL,
3989 fault->write, &fault->map_writable,
3999 * Returns true if the page fault is stale and needs to be retried, i.e. if the
4000 * root was invalidated by a memslot update or a relevant mmu_notifier fired.
4002 static bool is_page_fault_stale(struct kvm_vcpu *vcpu,
4003 struct kvm_page_fault *fault, int mmu_seq)
4005 struct kvm_mmu_page *sp = to_shadow_page(vcpu->arch.mmu->root.hpa);
4007 /* Special roots, e.g. pae_root, are not backed by shadow pages. */
4008 if (sp && is_obsolete_sp(vcpu->kvm, sp))
4012 * Roots without an associated shadow page are considered invalid if
4013 * there is a pending request to free obsolete roots. The request is
4014 * only a hint that the current root _may_ be obsolete and needs to be
4015 * reloaded, e.g. if the guest frees a PGD that KVM is tracking as a
4016 * previous root, then __kvm_mmu_prepare_zap_page() signals all vCPUs
4017 * to reload even if no vCPU is actively using the root.
4019 if (!sp && kvm_test_request(KVM_REQ_MMU_FREE_OBSOLETE_ROOTS, vcpu))
4022 return fault->slot &&
4023 mmu_notifier_retry_hva(vcpu->kvm, mmu_seq, fault->hva);
4026 static int direct_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
4028 bool is_tdp_mmu_fault = is_tdp_mmu(vcpu->arch.mmu);
4030 unsigned long mmu_seq;
4033 fault->gfn = fault->addr >> PAGE_SHIFT;
4034 fault->slot = kvm_vcpu_gfn_to_memslot(vcpu, fault->gfn);
4036 if (page_fault_handle_page_track(vcpu, fault))
4037 return RET_PF_EMULATE;
4039 r = fast_page_fault(vcpu, fault);
4040 if (r != RET_PF_INVALID)
4043 r = mmu_topup_memory_caches(vcpu, false);
4047 mmu_seq = vcpu->kvm->mmu_notifier_seq;
4050 if (kvm_faultin_pfn(vcpu, fault, &r))
4053 if (handle_abnormal_pfn(vcpu, fault, ACC_ALL, &r))
4058 if (is_tdp_mmu_fault)
4059 read_lock(&vcpu->kvm->mmu_lock);
4061 write_lock(&vcpu->kvm->mmu_lock);
4063 if (is_page_fault_stale(vcpu, fault, mmu_seq))
4066 r = make_mmu_pages_available(vcpu);
4070 if (is_tdp_mmu_fault)
4071 r = kvm_tdp_mmu_map(vcpu, fault);
4073 r = __direct_map(vcpu, fault);
4076 if (is_tdp_mmu_fault)
4077 read_unlock(&vcpu->kvm->mmu_lock);
4079 write_unlock(&vcpu->kvm->mmu_lock);
4080 kvm_release_pfn_clean(fault->pfn);
4084 static int nonpaging_page_fault(struct kvm_vcpu *vcpu,
4085 struct kvm_page_fault *fault)
4087 pgprintk("%s: gva %lx error %x\n", __func__, fault->addr, fault->error_code);
4089 /* This path builds a PAE pagetable, we can map 2mb pages at maximum. */
4090 fault->max_level = PG_LEVEL_2M;
4091 return direct_page_fault(vcpu, fault);
4094 int kvm_handle_page_fault(struct kvm_vcpu *vcpu, u64 error_code,
4095 u64 fault_address, char *insn, int insn_len)
4098 u32 flags = vcpu->arch.apf.host_apf_flags;
4100 #ifndef CONFIG_X86_64
4101 /* A 64-bit CR2 should be impossible on 32-bit KVM. */
4102 if (WARN_ON_ONCE(fault_address >> 32))
4106 vcpu->arch.l1tf_flush_l1d = true;
4108 trace_kvm_page_fault(fault_address, error_code);
4110 if (kvm_event_needs_reinjection(vcpu))
4111 kvm_mmu_unprotect_page_virt(vcpu, fault_address);
4112 r = kvm_mmu_page_fault(vcpu, fault_address, error_code, insn,
4114 } else if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
4115 vcpu->arch.apf.host_apf_flags = 0;
4116 local_irq_disable();
4117 kvm_async_pf_task_wait_schedule(fault_address);
4120 WARN_ONCE(1, "Unexpected host async PF flags: %x\n", flags);
4125 EXPORT_SYMBOL_GPL(kvm_handle_page_fault);
4127 int kvm_tdp_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
4129 while (fault->max_level > PG_LEVEL_4K) {
4130 int page_num = KVM_PAGES_PER_HPAGE(fault->max_level);
4131 gfn_t base = (fault->addr >> PAGE_SHIFT) & ~(page_num - 1);
4133 if (kvm_mtrr_check_gfn_range_consistency(vcpu, base, page_num))
4139 return direct_page_fault(vcpu, fault);
4142 static void nonpaging_init_context(struct kvm_mmu *context)
4144 context->page_fault = nonpaging_page_fault;
4145 context->gva_to_gpa = nonpaging_gva_to_gpa;
4146 context->sync_page = nonpaging_sync_page;
4147 context->invlpg = NULL;
4148 context->direct_map = true;
4151 static inline bool is_root_usable(struct kvm_mmu_root_info *root, gpa_t pgd,
4152 union kvm_mmu_page_role role)
4154 return (role.direct || pgd == root->pgd) &&
4155 VALID_PAGE(root->hpa) &&
4156 role.word == to_shadow_page(root->hpa)->role.word;
4160 * Find out if a previously cached root matching the new pgd/role is available,
4161 * and insert the current root as the MRU in the cache.
4162 * If a matching root is found, it is assigned to kvm_mmu->root and
4164 * If no match is found, kvm_mmu->root is left invalid, the LRU root is
4165 * evicted to make room for the current root, and false is returned.
4167 static bool cached_root_find_and_keep_current(struct kvm *kvm, struct kvm_mmu *mmu,
4169 union kvm_mmu_page_role new_role)
4173 if (is_root_usable(&mmu->root, new_pgd, new_role))
4176 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
4178 * The swaps end up rotating the cache like this:
4179 * C 0 1 2 3 (on entry to the function)
4183 * 3 C 0 1 2 (on exit from the loop)
4185 swap(mmu->root, mmu->prev_roots[i]);
4186 if (is_root_usable(&mmu->root, new_pgd, new_role))
4190 kvm_mmu_free_roots(kvm, mmu, KVM_MMU_ROOT_CURRENT);
4195 * Find out if a previously cached root matching the new pgd/role is available.
4196 * On entry, mmu->root is invalid.
4197 * If a matching root is found, it is assigned to kvm_mmu->root, the LRU entry
4198 * of the cache becomes invalid, and true is returned.
4199 * If no match is found, kvm_mmu->root is left invalid and false is returned.
4201 static bool cached_root_find_without_current(struct kvm *kvm, struct kvm_mmu *mmu,
4203 union kvm_mmu_page_role new_role)
4207 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
4208 if (is_root_usable(&mmu->prev_roots[i], new_pgd, new_role))
4214 swap(mmu->root, mmu->prev_roots[i]);
4215 /* Bubble up the remaining roots. */
4216 for (; i < KVM_MMU_NUM_PREV_ROOTS - 1; i++)
4217 mmu->prev_roots[i] = mmu->prev_roots[i + 1];
4218 mmu->prev_roots[i].hpa = INVALID_PAGE;
4222 static bool fast_pgd_switch(struct kvm *kvm, struct kvm_mmu *mmu,
4223 gpa_t new_pgd, union kvm_mmu_page_role new_role)
4226 * For now, limit the caching to 64-bit hosts+VMs in order to avoid
4227 * having to deal with PDPTEs. We may add support for 32-bit hosts/VMs
4228 * later if necessary.
4230 if (VALID_PAGE(mmu->root.hpa) && !to_shadow_page(mmu->root.hpa))
4231 kvm_mmu_free_roots(kvm, mmu, KVM_MMU_ROOT_CURRENT);
4233 if (VALID_PAGE(mmu->root.hpa))
4234 return cached_root_find_and_keep_current(kvm, mmu, new_pgd, new_role);
4236 return cached_root_find_without_current(kvm, mmu, new_pgd, new_role);
4239 void kvm_mmu_new_pgd(struct kvm_vcpu *vcpu, gpa_t new_pgd)
4241 struct kvm_mmu *mmu = vcpu->arch.mmu;
4242 union kvm_mmu_page_role new_role = mmu->mmu_role.base;
4244 if (!fast_pgd_switch(vcpu->kvm, mmu, new_pgd, new_role)) {
4245 /* kvm_mmu_ensure_valid_pgd will set up a new root. */
4250 * It's possible that the cached previous root page is obsolete because
4251 * of a change in the MMU generation number. However, changing the
4252 * generation number is accompanied by KVM_REQ_MMU_FREE_OBSOLETE_ROOTS,
4253 * which will free the root set here and allocate a new one.
4255 kvm_make_request(KVM_REQ_LOAD_MMU_PGD, vcpu);
4257 if (force_flush_and_sync_on_reuse) {
4258 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
4259 kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
4263 * The last MMIO access's GVA and GPA are cached in the VCPU. When
4264 * switching to a new CR3, that GVA->GPA mapping may no longer be
4265 * valid. So clear any cached MMIO info even when we don't need to sync
4266 * the shadow page tables.
4268 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
4271 * If this is a direct root page, it doesn't have a write flooding
4272 * count. Otherwise, clear the write flooding count.
4274 if (!new_role.direct)
4275 __clear_sp_write_flooding_count(
4276 to_shadow_page(vcpu->arch.mmu->root.hpa));
4278 EXPORT_SYMBOL_GPL(kvm_mmu_new_pgd);
4280 static unsigned long get_cr3(struct kvm_vcpu *vcpu)
4282 return kvm_read_cr3(vcpu);
4285 static bool sync_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn,
4286 unsigned int access)
4288 if (unlikely(is_mmio_spte(*sptep))) {
4289 if (gfn != get_mmio_spte_gfn(*sptep)) {
4290 mmu_spte_clear_no_track(sptep);
4294 mark_mmio_spte(vcpu, sptep, gfn, access);
4301 #define PTTYPE_EPT 18 /* arbitrary */
4302 #define PTTYPE PTTYPE_EPT
4303 #include "paging_tmpl.h"
4307 #include "paging_tmpl.h"
4311 #include "paging_tmpl.h"
4315 __reset_rsvds_bits_mask(struct rsvd_bits_validate *rsvd_check,
4316 u64 pa_bits_rsvd, int level, bool nx, bool gbpages,
4319 u64 gbpages_bit_rsvd = 0;
4320 u64 nonleaf_bit8_rsvd = 0;
4323 rsvd_check->bad_mt_xwr = 0;
4326 gbpages_bit_rsvd = rsvd_bits(7, 7);
4328 if (level == PT32E_ROOT_LEVEL)
4329 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 62);
4331 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 51);
4333 /* Note, NX doesn't exist in PDPTEs, this is handled below. */
4335 high_bits_rsvd |= rsvd_bits(63, 63);
4338 * Non-leaf PML4Es and PDPEs reserve bit 8 (which would be the G bit for
4339 * leaf entries) on AMD CPUs only.
4342 nonleaf_bit8_rsvd = rsvd_bits(8, 8);
4345 case PT32_ROOT_LEVEL:
4346 /* no rsvd bits for 2 level 4K page table entries */
4347 rsvd_check->rsvd_bits_mask[0][1] = 0;
4348 rsvd_check->rsvd_bits_mask[0][0] = 0;
4349 rsvd_check->rsvd_bits_mask[1][0] =
4350 rsvd_check->rsvd_bits_mask[0][0];
4353 rsvd_check->rsvd_bits_mask[1][1] = 0;
4357 if (is_cpuid_PSE36())
4358 /* 36bits PSE 4MB page */
4359 rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
4361 /* 32 bits PSE 4MB page */
4362 rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
4364 case PT32E_ROOT_LEVEL:
4365 rsvd_check->rsvd_bits_mask[0][2] = rsvd_bits(63, 63) |
4368 rsvd_bits(1, 2); /* PDPTE */
4369 rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd; /* PDE */
4370 rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd; /* PTE */
4371 rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd |
4372 rsvd_bits(13, 20); /* large page */
4373 rsvd_check->rsvd_bits_mask[1][0] =
4374 rsvd_check->rsvd_bits_mask[0][0];
4376 case PT64_ROOT_5LEVEL:
4377 rsvd_check->rsvd_bits_mask[0][4] = high_bits_rsvd |
4380 rsvd_check->rsvd_bits_mask[1][4] =
4381 rsvd_check->rsvd_bits_mask[0][4];
4383 case PT64_ROOT_4LEVEL:
4384 rsvd_check->rsvd_bits_mask[0][3] = high_bits_rsvd |
4387 rsvd_check->rsvd_bits_mask[0][2] = high_bits_rsvd |
4389 rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd;
4390 rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd;
4391 rsvd_check->rsvd_bits_mask[1][3] =
4392 rsvd_check->rsvd_bits_mask[0][3];
4393 rsvd_check->rsvd_bits_mask[1][2] = high_bits_rsvd |
4396 rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd |
4397 rsvd_bits(13, 20); /* large page */
4398 rsvd_check->rsvd_bits_mask[1][0] =
4399 rsvd_check->rsvd_bits_mask[0][0];
4404 static bool guest_can_use_gbpages(struct kvm_vcpu *vcpu)
4407 * If TDP is enabled, let the guest use GBPAGES if they're supported in
4408 * hardware. The hardware page walker doesn't let KVM disable GBPAGES,
4409 * i.e. won't treat them as reserved, and KVM doesn't redo the GVA->GPA
4410 * walk for performance and complexity reasons. Not to mention KVM
4411 * _can't_ solve the problem because GVA->GPA walks aren't visible to
4412 * KVM once a TDP translation is installed. Mimic hardware behavior so
4413 * that KVM's is at least consistent, i.e. doesn't randomly inject #PF.
4415 return tdp_enabled ? boot_cpu_has(X86_FEATURE_GBPAGES) :
4416 guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES);
4419 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
4420 struct kvm_mmu *context)
4422 __reset_rsvds_bits_mask(&context->guest_rsvd_check,
4423 vcpu->arch.reserved_gpa_bits,
4424 context->root_level, is_efer_nx(context),
4425 guest_can_use_gbpages(vcpu),
4426 is_cr4_pse(context),
4427 guest_cpuid_is_amd_or_hygon(vcpu));
4431 __reset_rsvds_bits_mask_ept(struct rsvd_bits_validate *rsvd_check,
4432 u64 pa_bits_rsvd, bool execonly, int huge_page_level)
4434 u64 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 51);
4435 u64 large_1g_rsvd = 0, large_2m_rsvd = 0;
4438 if (huge_page_level < PG_LEVEL_1G)
4439 large_1g_rsvd = rsvd_bits(7, 7);
4440 if (huge_page_level < PG_LEVEL_2M)
4441 large_2m_rsvd = rsvd_bits(7, 7);
4443 rsvd_check->rsvd_bits_mask[0][4] = high_bits_rsvd | rsvd_bits(3, 7);
4444 rsvd_check->rsvd_bits_mask[0][3] = high_bits_rsvd | rsvd_bits(3, 7);
4445 rsvd_check->rsvd_bits_mask[0][2] = high_bits_rsvd | rsvd_bits(3, 6) | large_1g_rsvd;
4446 rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd | rsvd_bits(3, 6) | large_2m_rsvd;
4447 rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd;
4450 rsvd_check->rsvd_bits_mask[1][4] = rsvd_check->rsvd_bits_mask[0][4];
4451 rsvd_check->rsvd_bits_mask[1][3] = rsvd_check->rsvd_bits_mask[0][3];
4452 rsvd_check->rsvd_bits_mask[1][2] = high_bits_rsvd | rsvd_bits(12, 29) | large_1g_rsvd;
4453 rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd | rsvd_bits(12, 20) | large_2m_rsvd;
4454 rsvd_check->rsvd_bits_mask[1][0] = rsvd_check->rsvd_bits_mask[0][0];
4456 bad_mt_xwr = 0xFFull << (2 * 8); /* bits 3..5 must not be 2 */
4457 bad_mt_xwr |= 0xFFull << (3 * 8); /* bits 3..5 must not be 3 */
4458 bad_mt_xwr |= 0xFFull << (7 * 8); /* bits 3..5 must not be 7 */
4459 bad_mt_xwr |= REPEAT_BYTE(1ull << 2); /* bits 0..2 must not be 010 */
4460 bad_mt_xwr |= REPEAT_BYTE(1ull << 6); /* bits 0..2 must not be 110 */
4462 /* bits 0..2 must not be 100 unless VMX capabilities allow it */
4463 bad_mt_xwr |= REPEAT_BYTE(1ull << 4);
4465 rsvd_check->bad_mt_xwr = bad_mt_xwr;
4468 static void reset_rsvds_bits_mask_ept(struct kvm_vcpu *vcpu,
4469 struct kvm_mmu *context, bool execonly, int huge_page_level)
4471 __reset_rsvds_bits_mask_ept(&context->guest_rsvd_check,
4472 vcpu->arch.reserved_gpa_bits, execonly,
4476 static inline u64 reserved_hpa_bits(void)
4478 return rsvd_bits(shadow_phys_bits, 63);
4482 * the page table on host is the shadow page table for the page
4483 * table in guest or amd nested guest, its mmu features completely
4484 * follow the features in guest.
4486 static void reset_shadow_zero_bits_mask(struct kvm_vcpu *vcpu,
4487 struct kvm_mmu *context)
4490 * KVM uses NX when TDP is disabled to handle a variety of scenarios,
4491 * notably for huge SPTEs if iTLB multi-hit mitigation is enabled and
4492 * to generate correct permissions for CR0.WP=0/CR4.SMEP=1/EFER.NX=0.
4493 * The iTLB multi-hit workaround can be toggled at any time, so assume
4494 * NX can be used by any non-nested shadow MMU to avoid having to reset
4495 * MMU contexts. Note, KVM forces EFER.NX=1 when TDP is disabled.
4497 bool uses_nx = is_efer_nx(context) || !tdp_enabled;
4499 /* @amd adds a check on bit of SPTEs, which KVM shouldn't use anyways. */
4501 /* KVM doesn't use 2-level page tables for the shadow MMU. */
4502 bool is_pse = false;
4503 struct rsvd_bits_validate *shadow_zero_check;
4506 WARN_ON_ONCE(context->shadow_root_level < PT32E_ROOT_LEVEL);
4508 shadow_zero_check = &context->shadow_zero_check;
4509 __reset_rsvds_bits_mask(shadow_zero_check, reserved_hpa_bits(),
4510 context->shadow_root_level, uses_nx,
4511 guest_can_use_gbpages(vcpu), is_pse, is_amd);
4513 if (!shadow_me_mask)
4516 for (i = context->shadow_root_level; --i >= 0;) {
4517 shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask;
4518 shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask;
4523 static inline bool boot_cpu_is_amd(void)
4525 WARN_ON_ONCE(!tdp_enabled);
4526 return shadow_x_mask == 0;
4530 * the direct page table on host, use as much mmu features as
4531 * possible, however, kvm currently does not do execution-protection.
4534 reset_tdp_shadow_zero_bits_mask(struct kvm_mmu *context)
4536 struct rsvd_bits_validate *shadow_zero_check;
4539 shadow_zero_check = &context->shadow_zero_check;
4541 if (boot_cpu_is_amd())
4542 __reset_rsvds_bits_mask(shadow_zero_check, reserved_hpa_bits(),
4543 context->shadow_root_level, false,
4544 boot_cpu_has(X86_FEATURE_GBPAGES),
4547 __reset_rsvds_bits_mask_ept(shadow_zero_check,
4548 reserved_hpa_bits(), false,
4549 max_huge_page_level);
4551 if (!shadow_me_mask)
4554 for (i = context->shadow_root_level; --i >= 0;) {
4555 shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask;
4556 shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask;
4561 * as the comments in reset_shadow_zero_bits_mask() except it
4562 * is the shadow page table for intel nested guest.
4565 reset_ept_shadow_zero_bits_mask(struct kvm_mmu *context, bool execonly)
4567 __reset_rsvds_bits_mask_ept(&context->shadow_zero_check,
4568 reserved_hpa_bits(), execonly,
4569 max_huge_page_level);
4572 #define BYTE_MASK(access) \
4573 ((1 & (access) ? 2 : 0) | \
4574 (2 & (access) ? 4 : 0) | \
4575 (3 & (access) ? 8 : 0) | \
4576 (4 & (access) ? 16 : 0) | \
4577 (5 & (access) ? 32 : 0) | \
4578 (6 & (access) ? 64 : 0) | \
4579 (7 & (access) ? 128 : 0))
4582 static void update_permission_bitmask(struct kvm_mmu *mmu, bool ept)
4586 const u8 x = BYTE_MASK(ACC_EXEC_MASK);
4587 const u8 w = BYTE_MASK(ACC_WRITE_MASK);
4588 const u8 u = BYTE_MASK(ACC_USER_MASK);
4590 bool cr4_smep = is_cr4_smep(mmu);
4591 bool cr4_smap = is_cr4_smap(mmu);
4592 bool cr0_wp = is_cr0_wp(mmu);
4593 bool efer_nx = is_efer_nx(mmu);
4595 for (byte = 0; byte < ARRAY_SIZE(mmu->permissions); ++byte) {
4596 unsigned pfec = byte << 1;
4599 * Each "*f" variable has a 1 bit for each UWX value
4600 * that causes a fault with the given PFEC.
4603 /* Faults from writes to non-writable pages */
4604 u8 wf = (pfec & PFERR_WRITE_MASK) ? (u8)~w : 0;
4605 /* Faults from user mode accesses to supervisor pages */
4606 u8 uf = (pfec & PFERR_USER_MASK) ? (u8)~u : 0;
4607 /* Faults from fetches of non-executable pages*/
4608 u8 ff = (pfec & PFERR_FETCH_MASK) ? (u8)~x : 0;
4609 /* Faults from kernel mode fetches of user pages */
4611 /* Faults from kernel mode accesses of user pages */
4615 /* Faults from kernel mode accesses to user pages */
4616 u8 kf = (pfec & PFERR_USER_MASK) ? 0 : u;
4618 /* Not really needed: !nx will cause pte.nx to fault */
4622 /* Allow supervisor writes if !cr0.wp */
4624 wf = (pfec & PFERR_USER_MASK) ? wf : 0;
4626 /* Disallow supervisor fetches of user code if cr4.smep */
4628 smepf = (pfec & PFERR_FETCH_MASK) ? kf : 0;
4631 * SMAP:kernel-mode data accesses from user-mode
4632 * mappings should fault. A fault is considered
4633 * as a SMAP violation if all of the following
4634 * conditions are true:
4635 * - X86_CR4_SMAP is set in CR4
4636 * - A user page is accessed
4637 * - The access is not a fetch
4638 * - The access is supervisor mode
4639 * - If implicit supervisor access or X86_EFLAGS_AC is clear
4641 * Here, we cover the first four conditions.
4642 * The fifth is computed dynamically in permission_fault();
4643 * PFERR_RSVD_MASK bit will be set in PFEC if the access is
4644 * *not* subject to SMAP restrictions.
4647 smapf = (pfec & (PFERR_RSVD_MASK|PFERR_FETCH_MASK)) ? 0 : kf;
4650 mmu->permissions[byte] = ff | uf | wf | smepf | smapf;
4655 * PKU is an additional mechanism by which the paging controls access to
4656 * user-mode addresses based on the value in the PKRU register. Protection
4657 * key violations are reported through a bit in the page fault error code.
4658 * Unlike other bits of the error code, the PK bit is not known at the
4659 * call site of e.g. gva_to_gpa; it must be computed directly in
4660 * permission_fault based on two bits of PKRU, on some machine state (CR4,
4661 * CR0, EFER, CPL), and on other bits of the error code and the page tables.
4663 * In particular the following conditions come from the error code, the
4664 * page tables and the machine state:
4665 * - PK is always zero unless CR4.PKE=1 and EFER.LMA=1
4666 * - PK is always zero if RSVD=1 (reserved bit set) or F=1 (instruction fetch)
4667 * - PK is always zero if U=0 in the page tables
4668 * - PKRU.WD is ignored if CR0.WP=0 and the access is a supervisor access.
4670 * The PKRU bitmask caches the result of these four conditions. The error
4671 * code (minus the P bit) and the page table's U bit form an index into the
4672 * PKRU bitmask. Two bits of the PKRU bitmask are then extracted and ANDed
4673 * with the two bits of the PKRU register corresponding to the protection key.
4674 * For the first three conditions above the bits will be 00, thus masking
4675 * away both AD and WD. For all reads or if the last condition holds, WD
4676 * only will be masked away.
4678 static void update_pkru_bitmask(struct kvm_mmu *mmu)
4685 if (!is_cr4_pke(mmu))
4688 wp = is_cr0_wp(mmu);
4690 for (bit = 0; bit < ARRAY_SIZE(mmu->permissions); ++bit) {
4691 unsigned pfec, pkey_bits;
4692 bool check_pkey, check_write, ff, uf, wf, pte_user;
4695 ff = pfec & PFERR_FETCH_MASK;
4696 uf = pfec & PFERR_USER_MASK;
4697 wf = pfec & PFERR_WRITE_MASK;
4699 /* PFEC.RSVD is replaced by ACC_USER_MASK. */
4700 pte_user = pfec & PFERR_RSVD_MASK;
4703 * Only need to check the access which is not an
4704 * instruction fetch and is to a user page.
4706 check_pkey = (!ff && pte_user);
4708 * write access is controlled by PKRU if it is a
4709 * user access or CR0.WP = 1.
4711 check_write = check_pkey && wf && (uf || wp);
4713 /* PKRU.AD stops both read and write access. */
4714 pkey_bits = !!check_pkey;
4715 /* PKRU.WD stops write access. */
4716 pkey_bits |= (!!check_write) << 1;
4718 mmu->pkru_mask |= (pkey_bits & 3) << pfec;
4722 static void reset_guest_paging_metadata(struct kvm_vcpu *vcpu,
4723 struct kvm_mmu *mmu)
4725 if (!is_cr0_pg(mmu))
4728 reset_rsvds_bits_mask(vcpu, mmu);
4729 update_permission_bitmask(mmu, false);
4730 update_pkru_bitmask(mmu);
4733 static void paging64_init_context(struct kvm_mmu *context)
4735 context->page_fault = paging64_page_fault;
4736 context->gva_to_gpa = paging64_gva_to_gpa;
4737 context->sync_page = paging64_sync_page;
4738 context->invlpg = paging64_invlpg;
4739 context->direct_map = false;
4742 static void paging32_init_context(struct kvm_mmu *context)
4744 context->page_fault = paging32_page_fault;
4745 context->gva_to_gpa = paging32_gva_to_gpa;
4746 context->sync_page = paging32_sync_page;
4747 context->invlpg = paging32_invlpg;
4748 context->direct_map = false;
4751 static union kvm_mmu_role kvm_calc_mmu_role_common(struct kvm_vcpu *vcpu,
4752 const struct kvm_mmu_role_regs *regs)
4754 union kvm_mmu_role role = {0};
4756 role.base.access = ACC_ALL;
4757 if (____is_cr0_pg(regs)) {
4758 role.ext.cr0_pg = 1;
4759 role.base.efer_nx = ____is_efer_nx(regs);
4760 role.base.cr0_wp = ____is_cr0_wp(regs);
4762 role.ext.cr4_pae = ____is_cr4_pae(regs);
4763 role.ext.cr4_smep = ____is_cr4_smep(regs);
4764 role.ext.cr4_smap = ____is_cr4_smap(regs);
4765 role.ext.cr4_pse = ____is_cr4_pse(regs);
4767 /* PKEY and LA57 are active iff long mode is active. */
4768 role.ext.cr4_pke = ____is_efer_lma(regs) && ____is_cr4_pke(regs);
4769 role.ext.cr4_la57 = ____is_efer_lma(regs) && ____is_cr4_la57(regs);
4770 role.ext.efer_lma = ____is_efer_lma(regs);
4772 role.base.smm = is_smm(vcpu);
4773 role.base.guest_mode = is_guest_mode(vcpu);
4779 static inline int kvm_mmu_get_tdp_level(struct kvm_vcpu *vcpu)
4781 /* tdp_root_level is architecture forced level, use it if nonzero */
4783 return tdp_root_level;
4785 /* Use 5-level TDP if and only if it's useful/necessary. */
4786 if (max_tdp_level == 5 && cpuid_maxphyaddr(vcpu) <= 48)
4789 return max_tdp_level;
4792 static union kvm_mmu_role
4793 kvm_calc_tdp_mmu_root_page_role(struct kvm_vcpu *vcpu,
4794 const struct kvm_mmu_role_regs *regs)
4796 union kvm_mmu_role role = kvm_calc_mmu_role_common(vcpu, regs);
4798 role.base.ad_disabled = (shadow_accessed_mask == 0);
4799 role.base.level = kvm_mmu_get_tdp_level(vcpu);
4800 role.base.direct = true;
4801 role.base.has_4_byte_gpte = false;
4806 static void init_kvm_tdp_mmu(struct kvm_vcpu *vcpu,
4807 const struct kvm_mmu_role_regs *regs)
4809 struct kvm_mmu *context = &vcpu->arch.root_mmu;
4810 union kvm_mmu_role new_role =
4811 kvm_calc_tdp_mmu_root_page_role(vcpu, regs);
4813 if (new_role.as_u64 == context->mmu_role.as_u64)
4816 context->mmu_role.as_u64 = new_role.as_u64;
4817 context->page_fault = kvm_tdp_page_fault;
4818 context->sync_page = nonpaging_sync_page;
4819 context->invlpg = NULL;
4820 context->shadow_root_level = kvm_mmu_get_tdp_level(vcpu);
4821 context->direct_map = true;
4822 context->get_guest_pgd = get_cr3;
4823 context->get_pdptr = kvm_pdptr_read;
4824 context->inject_page_fault = kvm_inject_page_fault;
4825 context->root_level = role_regs_to_root_level(regs);
4827 if (!is_cr0_pg(context))
4828 context->gva_to_gpa = nonpaging_gva_to_gpa;
4829 else if (is_cr4_pae(context))
4830 context->gva_to_gpa = paging64_gva_to_gpa;
4832 context->gva_to_gpa = paging32_gva_to_gpa;
4834 reset_guest_paging_metadata(vcpu, context);
4835 reset_tdp_shadow_zero_bits_mask(context);
4838 static union kvm_mmu_role
4839 kvm_calc_shadow_root_page_role_common(struct kvm_vcpu *vcpu,
4840 const struct kvm_mmu_role_regs *regs)
4842 union kvm_mmu_role role = kvm_calc_mmu_role_common(vcpu, regs);
4844 role.base.smep_andnot_wp = role.ext.cr4_smep && !____is_cr0_wp(regs);
4845 role.base.smap_andnot_wp = role.ext.cr4_smap && !____is_cr0_wp(regs);
4846 role.base.has_4_byte_gpte = ____is_cr0_pg(regs) && !____is_cr4_pae(regs);
4851 static union kvm_mmu_role
4852 kvm_calc_shadow_mmu_root_page_role(struct kvm_vcpu *vcpu,
4853 const struct kvm_mmu_role_regs *regs)
4855 union kvm_mmu_role role =
4856 kvm_calc_shadow_root_page_role_common(vcpu, regs);
4858 role.base.direct = !____is_cr0_pg(regs);
4860 if (!____is_efer_lma(regs))
4861 role.base.level = PT32E_ROOT_LEVEL;
4862 else if (____is_cr4_la57(regs))
4863 role.base.level = PT64_ROOT_5LEVEL;
4865 role.base.level = PT64_ROOT_4LEVEL;
4870 static void shadow_mmu_init_context(struct kvm_vcpu *vcpu, struct kvm_mmu *context,
4871 const struct kvm_mmu_role_regs *regs,
4872 union kvm_mmu_role new_role)
4874 if (new_role.as_u64 == context->mmu_role.as_u64)
4877 context->mmu_role.as_u64 = new_role.as_u64;
4879 if (!is_cr0_pg(context))
4880 nonpaging_init_context(context);
4881 else if (is_cr4_pae(context))
4882 paging64_init_context(context);
4884 paging32_init_context(context);
4885 context->root_level = role_regs_to_root_level(regs);
4887 reset_guest_paging_metadata(vcpu, context);
4888 context->shadow_root_level = new_role.base.level;
4890 reset_shadow_zero_bits_mask(vcpu, context);
4893 static void kvm_init_shadow_mmu(struct kvm_vcpu *vcpu,
4894 const struct kvm_mmu_role_regs *regs)
4896 struct kvm_mmu *context = &vcpu->arch.root_mmu;
4897 union kvm_mmu_role new_role =
4898 kvm_calc_shadow_mmu_root_page_role(vcpu, regs);
4900 shadow_mmu_init_context(vcpu, context, regs, new_role);
4903 static union kvm_mmu_role
4904 kvm_calc_shadow_npt_root_page_role(struct kvm_vcpu *vcpu,
4905 const struct kvm_mmu_role_regs *regs)
4907 union kvm_mmu_role role =
4908 kvm_calc_shadow_root_page_role_common(vcpu, regs);
4910 role.base.direct = false;
4911 role.base.level = kvm_mmu_get_tdp_level(vcpu);
4916 void kvm_init_shadow_npt_mmu(struct kvm_vcpu *vcpu, unsigned long cr0,
4917 unsigned long cr4, u64 efer, gpa_t nested_cr3)
4919 struct kvm_mmu *context = &vcpu->arch.guest_mmu;
4920 struct kvm_mmu_role_regs regs = {
4922 .cr4 = cr4 & ~X86_CR4_PKE,
4925 union kvm_mmu_role new_role;
4927 new_role = kvm_calc_shadow_npt_root_page_role(vcpu, ®s);
4929 shadow_mmu_init_context(vcpu, context, ®s, new_role);
4930 kvm_mmu_new_pgd(vcpu, nested_cr3);
4932 EXPORT_SYMBOL_GPL(kvm_init_shadow_npt_mmu);
4934 static union kvm_mmu_role
4935 kvm_calc_shadow_ept_root_page_role(struct kvm_vcpu *vcpu, bool accessed_dirty,
4936 bool execonly, u8 level)
4938 union kvm_mmu_role role = {0};
4941 * KVM does not support SMM transfer monitors, and consequently does not
4942 * support the "entry to SMM" control either. role.base.smm is always 0.
4944 WARN_ON_ONCE(is_smm(vcpu));
4945 role.base.level = level;
4946 role.base.has_4_byte_gpte = false;
4947 role.base.direct = false;
4948 role.base.ad_disabled = !accessed_dirty;
4949 role.base.guest_mode = true;
4950 role.base.access = ACC_ALL;
4952 /* EPT, and thus nested EPT, does not consume CR0, CR4, nor EFER. */
4954 role.ext.execonly = execonly;
4960 void kvm_init_shadow_ept_mmu(struct kvm_vcpu *vcpu, bool execonly,
4961 int huge_page_level, bool accessed_dirty,
4964 struct kvm_mmu *context = &vcpu->arch.guest_mmu;
4965 u8 level = vmx_eptp_page_walk_level(new_eptp);
4966 union kvm_mmu_role new_role =
4967 kvm_calc_shadow_ept_root_page_role(vcpu, accessed_dirty,
4970 if (new_role.as_u64 != context->mmu_role.as_u64) {
4971 context->mmu_role.as_u64 = new_role.as_u64;
4973 context->shadow_root_level = level;
4975 context->ept_ad = accessed_dirty;
4976 context->page_fault = ept_page_fault;
4977 context->gva_to_gpa = ept_gva_to_gpa;
4978 context->sync_page = ept_sync_page;
4979 context->invlpg = ept_invlpg;
4980 context->root_level = level;
4981 context->direct_map = false;
4982 update_permission_bitmask(context, true);
4983 context->pkru_mask = 0;
4984 reset_rsvds_bits_mask_ept(vcpu, context, execonly, huge_page_level);
4985 reset_ept_shadow_zero_bits_mask(context, execonly);
4988 kvm_mmu_new_pgd(vcpu, new_eptp);
4990 EXPORT_SYMBOL_GPL(kvm_init_shadow_ept_mmu);
4992 static void init_kvm_softmmu(struct kvm_vcpu *vcpu,
4993 const struct kvm_mmu_role_regs *regs)
4995 struct kvm_mmu *context = &vcpu->arch.root_mmu;
4997 kvm_init_shadow_mmu(vcpu, regs);
4999 context->get_guest_pgd = get_cr3;
5000 context->get_pdptr = kvm_pdptr_read;
5001 context->inject_page_fault = kvm_inject_page_fault;
5004 static union kvm_mmu_role
5005 kvm_calc_nested_mmu_role(struct kvm_vcpu *vcpu, const struct kvm_mmu_role_regs *regs)
5007 union kvm_mmu_role role;
5009 role = kvm_calc_shadow_root_page_role_common(vcpu, regs);
5012 * Nested MMUs are used only for walking L2's gva->gpa, they never have
5013 * shadow pages of their own and so "direct" has no meaning. Set it
5014 * to "true" to try to detect bogus usage of the nested MMU.
5016 role.base.direct = true;
5017 role.base.level = role_regs_to_root_level(regs);
5021 static void init_kvm_nested_mmu(struct kvm_vcpu *vcpu,
5022 const struct kvm_mmu_role_regs *regs)
5024 union kvm_mmu_role new_role = kvm_calc_nested_mmu_role(vcpu, regs);
5025 struct kvm_mmu *g_context = &vcpu->arch.nested_mmu;
5027 if (new_role.as_u64 == g_context->mmu_role.as_u64)
5030 g_context->mmu_role.as_u64 = new_role.as_u64;
5031 g_context->get_guest_pgd = get_cr3;
5032 g_context->get_pdptr = kvm_pdptr_read;
5033 g_context->inject_page_fault = kvm_inject_page_fault;
5034 g_context->root_level = new_role.base.level;
5037 * L2 page tables are never shadowed, so there is no need to sync
5040 g_context->invlpg = NULL;
5043 * Note that arch.mmu->gva_to_gpa translates l2_gpa to l1_gpa using
5044 * L1's nested page tables (e.g. EPT12). The nested translation
5045 * of l2_gva to l1_gpa is done by arch.nested_mmu.gva_to_gpa using
5046 * L2's page tables as the first level of translation and L1's
5047 * nested page tables as the second level of translation. Basically
5048 * the gva_to_gpa functions between mmu and nested_mmu are swapped.
5050 if (!is_paging(vcpu))
5051 g_context->gva_to_gpa = nonpaging_gva_to_gpa;
5052 else if (is_long_mode(vcpu))
5053 g_context->gva_to_gpa = paging64_gva_to_gpa;
5054 else if (is_pae(vcpu))
5055 g_context->gva_to_gpa = paging64_gva_to_gpa;
5057 g_context->gva_to_gpa = paging32_gva_to_gpa;
5059 reset_guest_paging_metadata(vcpu, g_context);
5062 void kvm_init_mmu(struct kvm_vcpu *vcpu)
5064 struct kvm_mmu_role_regs regs = vcpu_to_role_regs(vcpu);
5066 if (mmu_is_nested(vcpu))
5067 init_kvm_nested_mmu(vcpu, ®s);
5068 else if (tdp_enabled)
5069 init_kvm_tdp_mmu(vcpu, ®s);
5071 init_kvm_softmmu(vcpu, ®s);
5073 EXPORT_SYMBOL_GPL(kvm_init_mmu);
5075 void kvm_mmu_after_set_cpuid(struct kvm_vcpu *vcpu)
5078 * Invalidate all MMU roles to force them to reinitialize as CPUID
5079 * information is factored into reserved bit calculations.
5081 * Correctly handling multiple vCPU models with respect to paging and
5082 * physical address properties) in a single VM would require tracking
5083 * all relevant CPUID information in kvm_mmu_page_role. That is very
5084 * undesirable as it would increase the memory requirements for
5085 * gfn_track (see struct kvm_mmu_page_role comments). For now that
5086 * problem is swept under the rug; KVM's CPUID API is horrific and
5087 * it's all but impossible to solve it without introducing a new API.
5089 vcpu->arch.root_mmu.mmu_role.ext.valid = 0;
5090 vcpu->arch.guest_mmu.mmu_role.ext.valid = 0;
5091 vcpu->arch.nested_mmu.mmu_role.ext.valid = 0;
5092 kvm_mmu_reset_context(vcpu);
5095 * Changing guest CPUID after KVM_RUN is forbidden, see the comment in
5096 * kvm_arch_vcpu_ioctl().
5098 KVM_BUG_ON(vcpu->arch.last_vmentry_cpu != -1, vcpu->kvm);
5101 void kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
5103 kvm_mmu_unload(vcpu);
5106 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
5108 int kvm_mmu_load(struct kvm_vcpu *vcpu)
5112 r = mmu_topup_memory_caches(vcpu, !vcpu->arch.mmu->direct_map);
5115 r = mmu_alloc_special_roots(vcpu);
5118 if (vcpu->arch.mmu->direct_map)
5119 r = mmu_alloc_direct_roots(vcpu);
5121 r = mmu_alloc_shadow_roots(vcpu);
5125 kvm_mmu_sync_roots(vcpu);
5127 kvm_mmu_load_pgd(vcpu);
5130 * Flush any TLB entries for the new root, the provenance of the root
5131 * is unknown. Even if KVM ensures there are no stale TLB entries
5132 * for a freed root, in theory another hypervisor could have left
5133 * stale entries. Flushing on alloc also allows KVM to skip the TLB
5134 * flush when freeing a root (see kvm_tdp_mmu_put_root()).
5136 static_call(kvm_x86_flush_tlb_current)(vcpu);
5141 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
5143 struct kvm *kvm = vcpu->kvm;
5145 kvm_mmu_free_roots(kvm, &vcpu->arch.root_mmu, KVM_MMU_ROOTS_ALL);
5146 WARN_ON(VALID_PAGE(vcpu->arch.root_mmu.root.hpa));
5147 kvm_mmu_free_roots(kvm, &vcpu->arch.guest_mmu, KVM_MMU_ROOTS_ALL);
5148 WARN_ON(VALID_PAGE(vcpu->arch.guest_mmu.root.hpa));
5149 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
5152 static bool is_obsolete_root(struct kvm *kvm, hpa_t root_hpa)
5154 struct kvm_mmu_page *sp;
5156 if (!VALID_PAGE(root_hpa))
5160 * When freeing obsolete roots, treat roots as obsolete if they don't
5161 * have an associated shadow page. This does mean KVM will get false
5162 * positives and free roots that don't strictly need to be freed, but
5163 * such false positives are relatively rare:
5165 * (a) only PAE paging and nested NPT has roots without shadow pages
5166 * (b) remote reloads due to a memslot update obsoletes _all_ roots
5167 * (c) KVM doesn't track previous roots for PAE paging, and the guest
5168 * is unlikely to zap an in-use PGD.
5170 sp = to_shadow_page(root_hpa);
5171 return !sp || is_obsolete_sp(kvm, sp);
5174 static void __kvm_mmu_free_obsolete_roots(struct kvm *kvm, struct kvm_mmu *mmu)
5176 unsigned long roots_to_free = 0;
5179 if (is_obsolete_root(kvm, mmu->root.hpa))
5180 roots_to_free |= KVM_MMU_ROOT_CURRENT;
5182 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
5183 if (is_obsolete_root(kvm, mmu->root.hpa))
5184 roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i);
5188 kvm_mmu_free_roots(kvm, mmu, roots_to_free);
5191 void kvm_mmu_free_obsolete_roots(struct kvm_vcpu *vcpu)
5193 __kvm_mmu_free_obsolete_roots(vcpu->kvm, &vcpu->arch.root_mmu);
5194 __kvm_mmu_free_obsolete_roots(vcpu->kvm, &vcpu->arch.guest_mmu);
5197 static bool need_remote_flush(u64 old, u64 new)
5199 if (!is_shadow_present_pte(old))
5201 if (!is_shadow_present_pte(new))
5203 if ((old ^ new) & PT64_BASE_ADDR_MASK)
5205 old ^= shadow_nx_mask;
5206 new ^= shadow_nx_mask;
5207 return (old & ~new & PT64_PERM_MASK) != 0;
5210 static u64 mmu_pte_write_fetch_gpte(struct kvm_vcpu *vcpu, gpa_t *gpa,
5217 * Assume that the pte write on a page table of the same type
5218 * as the current vcpu paging mode since we update the sptes only
5219 * when they have the same mode.
5221 if (is_pae(vcpu) && *bytes == 4) {
5222 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
5227 if (*bytes == 4 || *bytes == 8) {
5228 r = kvm_vcpu_read_guest_atomic(vcpu, *gpa, &gentry, *bytes);
5237 * If we're seeing too many writes to a page, it may no longer be a page table,
5238 * or we may be forking, in which case it is better to unmap the page.
5240 static bool detect_write_flooding(struct kvm_mmu_page *sp)
5243 * Skip write-flooding detected for the sp whose level is 1, because
5244 * it can become unsync, then the guest page is not write-protected.
5246 if (sp->role.level == PG_LEVEL_4K)
5249 atomic_inc(&sp->write_flooding_count);
5250 return atomic_read(&sp->write_flooding_count) >= 3;
5254 * Misaligned accesses are too much trouble to fix up; also, they usually
5255 * indicate a page is not used as a page table.
5257 static bool detect_write_misaligned(struct kvm_mmu_page *sp, gpa_t gpa,
5260 unsigned offset, pte_size, misaligned;
5262 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
5263 gpa, bytes, sp->role.word);
5265 offset = offset_in_page(gpa);
5266 pte_size = sp->role.has_4_byte_gpte ? 4 : 8;
5269 * Sometimes, the OS only writes the last one bytes to update status
5270 * bits, for example, in linux, andb instruction is used in clear_bit().
5272 if (!(offset & (pte_size - 1)) && bytes == 1)
5275 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
5276 misaligned |= bytes < 4;
5281 static u64 *get_written_sptes(struct kvm_mmu_page *sp, gpa_t gpa, int *nspte)
5283 unsigned page_offset, quadrant;
5287 page_offset = offset_in_page(gpa);
5288 level = sp->role.level;
5290 if (sp->role.has_4_byte_gpte) {
5291 page_offset <<= 1; /* 32->64 */
5293 * A 32-bit pde maps 4MB while the shadow pdes map
5294 * only 2MB. So we need to double the offset again
5295 * and zap two pdes instead of one.
5297 if (level == PT32_ROOT_LEVEL) {
5298 page_offset &= ~7; /* kill rounding error */
5302 quadrant = page_offset >> PAGE_SHIFT;
5303 page_offset &= ~PAGE_MASK;
5304 if (quadrant != sp->role.quadrant)
5308 spte = &sp->spt[page_offset / sizeof(*spte)];
5312 static void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
5313 const u8 *new, int bytes,
5314 struct kvm_page_track_notifier_node *node)
5316 gfn_t gfn = gpa >> PAGE_SHIFT;
5317 struct kvm_mmu_page *sp;
5318 LIST_HEAD(invalid_list);
5319 u64 entry, gentry, *spte;
5324 * If we don't have indirect shadow pages, it means no page is
5325 * write-protected, so we can exit simply.
5327 if (!READ_ONCE(vcpu->kvm->arch.indirect_shadow_pages))
5330 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
5333 * No need to care whether allocation memory is successful
5334 * or not since pte prefetch is skipped if it does not have
5335 * enough objects in the cache.
5337 mmu_topup_memory_caches(vcpu, true);
5339 write_lock(&vcpu->kvm->mmu_lock);
5341 gentry = mmu_pte_write_fetch_gpte(vcpu, &gpa, &bytes);
5343 ++vcpu->kvm->stat.mmu_pte_write;
5345 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn) {
5346 if (detect_write_misaligned(sp, gpa, bytes) ||
5347 detect_write_flooding(sp)) {
5348 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
5349 ++vcpu->kvm->stat.mmu_flooded;
5353 spte = get_written_sptes(sp, gpa, &npte);
5359 mmu_page_zap_pte(vcpu->kvm, sp, spte, NULL);
5360 if (gentry && sp->role.level != PG_LEVEL_4K)
5361 ++vcpu->kvm->stat.mmu_pde_zapped;
5362 if (need_remote_flush(entry, *spte))
5367 kvm_mmu_remote_flush_or_zap(vcpu->kvm, &invalid_list, flush);
5368 write_unlock(&vcpu->kvm->mmu_lock);
5371 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, u64 error_code,
5372 void *insn, int insn_len)
5374 int r, emulation_type = EMULTYPE_PF;
5375 bool direct = vcpu->arch.mmu->direct_map;
5377 if (WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root.hpa)))
5378 return RET_PF_RETRY;
5381 if (unlikely(error_code & PFERR_RSVD_MASK)) {
5382 r = handle_mmio_page_fault(vcpu, cr2_or_gpa, direct);
5383 if (r == RET_PF_EMULATE)
5387 if (r == RET_PF_INVALID) {
5388 r = kvm_mmu_do_page_fault(vcpu, cr2_or_gpa,
5389 lower_32_bits(error_code), false);
5390 if (KVM_BUG_ON(r == RET_PF_INVALID, vcpu->kvm))
5396 if (r != RET_PF_EMULATE)
5400 * Before emulating the instruction, check if the error code
5401 * was due to a RO violation while translating the guest page.
5402 * This can occur when using nested virtualization with nested
5403 * paging in both guests. If true, we simply unprotect the page
5404 * and resume the guest.
5406 if (vcpu->arch.mmu->direct_map &&
5407 (error_code & PFERR_NESTED_GUEST_PAGE) == PFERR_NESTED_GUEST_PAGE) {
5408 kvm_mmu_unprotect_page(vcpu->kvm, gpa_to_gfn(cr2_or_gpa));
5413 * vcpu->arch.mmu.page_fault returned RET_PF_EMULATE, but we can still
5414 * optimistically try to just unprotect the page and let the processor
5415 * re-execute the instruction that caused the page fault. Do not allow
5416 * retrying MMIO emulation, as it's not only pointless but could also
5417 * cause us to enter an infinite loop because the processor will keep
5418 * faulting on the non-existent MMIO address. Retrying an instruction
5419 * from a nested guest is also pointless and dangerous as we are only
5420 * explicitly shadowing L1's page tables, i.e. unprotecting something
5421 * for L1 isn't going to magically fix whatever issue cause L2 to fail.
5423 if (!mmio_info_in_cache(vcpu, cr2_or_gpa, direct) && !is_guest_mode(vcpu))
5424 emulation_type |= EMULTYPE_ALLOW_RETRY_PF;
5426 return x86_emulate_instruction(vcpu, cr2_or_gpa, emulation_type, insn,
5429 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
5431 void kvm_mmu_invalidate_gva(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
5432 gva_t gva, hpa_t root_hpa)
5436 /* It's actually a GPA for vcpu->arch.guest_mmu. */
5437 if (mmu != &vcpu->arch.guest_mmu) {
5438 /* INVLPG on a non-canonical address is a NOP according to the SDM. */
5439 if (is_noncanonical_address(gva, vcpu))
5442 static_call(kvm_x86_flush_tlb_gva)(vcpu, gva);
5448 if (root_hpa == INVALID_PAGE) {
5449 mmu->invlpg(vcpu, gva, mmu->root.hpa);
5452 * INVLPG is required to invalidate any global mappings for the VA,
5453 * irrespective of PCID. Since it would take us roughly similar amount
5454 * of work to determine whether any of the prev_root mappings of the VA
5455 * is marked global, or to just sync it blindly, so we might as well
5456 * just always sync it.
5458 * Mappings not reachable via the current cr3 or the prev_roots will be
5459 * synced when switching to that cr3, so nothing needs to be done here
5462 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5463 if (VALID_PAGE(mmu->prev_roots[i].hpa))
5464 mmu->invlpg(vcpu, gva, mmu->prev_roots[i].hpa);
5466 mmu->invlpg(vcpu, gva, root_hpa);
5470 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
5472 kvm_mmu_invalidate_gva(vcpu, vcpu->arch.walk_mmu, gva, INVALID_PAGE);
5473 ++vcpu->stat.invlpg;
5475 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
5478 void kvm_mmu_invpcid_gva(struct kvm_vcpu *vcpu, gva_t gva, unsigned long pcid)
5480 struct kvm_mmu *mmu = vcpu->arch.mmu;
5481 bool tlb_flush = false;
5484 if (pcid == kvm_get_active_pcid(vcpu)) {
5485 mmu->invlpg(vcpu, gva, mmu->root.hpa);
5489 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
5490 if (VALID_PAGE(mmu->prev_roots[i].hpa) &&
5491 pcid == kvm_get_pcid(vcpu, mmu->prev_roots[i].pgd)) {
5492 mmu->invlpg(vcpu, gva, mmu->prev_roots[i].hpa);
5498 static_call(kvm_x86_flush_tlb_gva)(vcpu, gva);
5500 ++vcpu->stat.invlpg;
5503 * Mappings not reachable via the current cr3 or the prev_roots will be
5504 * synced when switching to that cr3, so nothing needs to be done here
5509 void kvm_configure_mmu(bool enable_tdp, int tdp_forced_root_level,
5510 int tdp_max_root_level, int tdp_huge_page_level)
5512 tdp_enabled = enable_tdp;
5513 tdp_root_level = tdp_forced_root_level;
5514 max_tdp_level = tdp_max_root_level;
5517 * max_huge_page_level reflects KVM's MMU capabilities irrespective
5518 * of kernel support, e.g. KVM may be capable of using 1GB pages when
5519 * the kernel is not. But, KVM never creates a page size greater than
5520 * what is used by the kernel for any given HVA, i.e. the kernel's
5521 * capabilities are ultimately consulted by kvm_mmu_hugepage_adjust().
5524 max_huge_page_level = tdp_huge_page_level;
5525 else if (boot_cpu_has(X86_FEATURE_GBPAGES))
5526 max_huge_page_level = PG_LEVEL_1G;
5528 max_huge_page_level = PG_LEVEL_2M;
5530 EXPORT_SYMBOL_GPL(kvm_configure_mmu);
5532 /* The return value indicates if tlb flush on all vcpus is needed. */
5533 typedef bool (*slot_level_handler) (struct kvm *kvm,
5534 struct kvm_rmap_head *rmap_head,
5535 const struct kvm_memory_slot *slot);
5537 /* The caller should hold mmu-lock before calling this function. */
5538 static __always_inline bool
5539 slot_handle_level_range(struct kvm *kvm, const struct kvm_memory_slot *memslot,
5540 slot_level_handler fn, int start_level, int end_level,
5541 gfn_t start_gfn, gfn_t end_gfn, bool flush_on_yield,
5544 struct slot_rmap_walk_iterator iterator;
5546 for_each_slot_rmap_range(memslot, start_level, end_level, start_gfn,
5547 end_gfn, &iterator) {
5549 flush |= fn(kvm, iterator.rmap, memslot);
5551 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) {
5552 if (flush && flush_on_yield) {
5553 kvm_flush_remote_tlbs_with_address(kvm,
5555 iterator.gfn - start_gfn + 1);
5558 cond_resched_rwlock_write(&kvm->mmu_lock);
5565 static __always_inline bool
5566 slot_handle_level(struct kvm *kvm, const struct kvm_memory_slot *memslot,
5567 slot_level_handler fn, int start_level, int end_level,
5568 bool flush_on_yield)
5570 return slot_handle_level_range(kvm, memslot, fn, start_level,
5571 end_level, memslot->base_gfn,
5572 memslot->base_gfn + memslot->npages - 1,
5573 flush_on_yield, false);
5576 static __always_inline bool
5577 slot_handle_level_4k(struct kvm *kvm, const struct kvm_memory_slot *memslot,
5578 slot_level_handler fn, bool flush_on_yield)
5580 return slot_handle_level(kvm, memslot, fn, PG_LEVEL_4K,
5581 PG_LEVEL_4K, flush_on_yield);
5584 static void free_mmu_pages(struct kvm_mmu *mmu)
5586 if (!tdp_enabled && mmu->pae_root)
5587 set_memory_encrypted((unsigned long)mmu->pae_root, 1);
5588 free_page((unsigned long)mmu->pae_root);
5589 free_page((unsigned long)mmu->pml4_root);
5590 free_page((unsigned long)mmu->pml5_root);
5593 static int __kvm_mmu_create(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
5598 mmu->root.hpa = INVALID_PAGE;
5600 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5601 mmu->prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
5603 /* vcpu->arch.guest_mmu isn't used when !tdp_enabled. */
5604 if (!tdp_enabled && mmu == &vcpu->arch.guest_mmu)
5608 * When using PAE paging, the four PDPTEs are treated as 'root' pages,
5609 * while the PDP table is a per-vCPU construct that's allocated at MMU
5610 * creation. When emulating 32-bit mode, cr3 is only 32 bits even on
5611 * x86_64. Therefore we need to allocate the PDP table in the first
5612 * 4GB of memory, which happens to fit the DMA32 zone. TDP paging
5613 * generally doesn't use PAE paging and can skip allocating the PDP
5614 * table. The main exception, handled here, is SVM's 32-bit NPT. The
5615 * other exception is for shadowing L1's 32-bit or PAE NPT on 64-bit
5616 * KVM; that horror is handled on-demand by mmu_alloc_special_roots().
5618 if (tdp_enabled && kvm_mmu_get_tdp_level(vcpu) > PT32E_ROOT_LEVEL)
5621 page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_DMA32);
5625 mmu->pae_root = page_address(page);
5628 * CR3 is only 32 bits when PAE paging is used, thus it's impossible to
5629 * get the CPU to treat the PDPTEs as encrypted. Decrypt the page so
5630 * that KVM's writes and the CPU's reads get along. Note, this is
5631 * only necessary when using shadow paging, as 64-bit NPT can get at
5632 * the C-bit even when shadowing 32-bit NPT, and SME isn't supported
5633 * by 32-bit kernels (when KVM itself uses 32-bit NPT).
5636 set_memory_decrypted((unsigned long)mmu->pae_root, 1);
5638 WARN_ON_ONCE(shadow_me_mask);
5640 for (i = 0; i < 4; ++i)
5641 mmu->pae_root[i] = INVALID_PAE_ROOT;
5646 int kvm_mmu_create(struct kvm_vcpu *vcpu)
5650 vcpu->arch.mmu_pte_list_desc_cache.kmem_cache = pte_list_desc_cache;
5651 vcpu->arch.mmu_pte_list_desc_cache.gfp_zero = __GFP_ZERO;
5653 vcpu->arch.mmu_page_header_cache.kmem_cache = mmu_page_header_cache;
5654 vcpu->arch.mmu_page_header_cache.gfp_zero = __GFP_ZERO;
5656 vcpu->arch.mmu_shadow_page_cache.gfp_zero = __GFP_ZERO;
5658 vcpu->arch.mmu = &vcpu->arch.root_mmu;
5659 vcpu->arch.walk_mmu = &vcpu->arch.root_mmu;
5661 ret = __kvm_mmu_create(vcpu, &vcpu->arch.guest_mmu);
5665 ret = __kvm_mmu_create(vcpu, &vcpu->arch.root_mmu);
5667 goto fail_allocate_root;
5671 free_mmu_pages(&vcpu->arch.guest_mmu);
5675 #define BATCH_ZAP_PAGES 10
5676 static void kvm_zap_obsolete_pages(struct kvm *kvm)
5678 struct kvm_mmu_page *sp, *node;
5679 int nr_zapped, batch = 0;
5682 list_for_each_entry_safe_reverse(sp, node,
5683 &kvm->arch.active_mmu_pages, link) {
5685 * No obsolete valid page exists before a newly created page
5686 * since active_mmu_pages is a FIFO list.
5688 if (!is_obsolete_sp(kvm, sp))
5692 * Invalid pages should never land back on the list of active
5693 * pages. Skip the bogus page, otherwise we'll get stuck in an
5694 * infinite loop if the page gets put back on the list (again).
5696 if (WARN_ON(sp->role.invalid))
5700 * No need to flush the TLB since we're only zapping shadow
5701 * pages with an obsolete generation number and all vCPUS have
5702 * loaded a new root, i.e. the shadow pages being zapped cannot
5703 * be in active use by the guest.
5705 if (batch >= BATCH_ZAP_PAGES &&
5706 cond_resched_rwlock_write(&kvm->mmu_lock)) {
5711 if (__kvm_mmu_prepare_zap_page(kvm, sp,
5712 &kvm->arch.zapped_obsolete_pages, &nr_zapped)) {
5719 * Kick all vCPUs (via remote TLB flush) before freeing the page tables
5720 * to ensure KVM is not in the middle of a lockless shadow page table
5721 * walk, which may reference the pages. The remote TLB flush itself is
5722 * not required and is simply a convenient way to kick vCPUs as needed.
5723 * KVM performs a local TLB flush when allocating a new root (see
5724 * kvm_mmu_load()), and the reload in the caller ensure no vCPUs are
5725 * running with an obsolete MMU.
5727 kvm_mmu_commit_zap_page(kvm, &kvm->arch.zapped_obsolete_pages);
5731 * Fast invalidate all shadow pages and use lock-break technique
5732 * to zap obsolete pages.
5734 * It's required when memslot is being deleted or VM is being
5735 * destroyed, in these cases, we should ensure that KVM MMU does
5736 * not use any resource of the being-deleted slot or all slots
5737 * after calling the function.
5739 static void kvm_mmu_zap_all_fast(struct kvm *kvm)
5741 lockdep_assert_held(&kvm->slots_lock);
5743 write_lock(&kvm->mmu_lock);
5744 trace_kvm_mmu_zap_all_fast(kvm);
5747 * Toggle mmu_valid_gen between '0' and '1'. Because slots_lock is
5748 * held for the entire duration of zapping obsolete pages, it's
5749 * impossible for there to be multiple invalid generations associated
5750 * with *valid* shadow pages at any given time, i.e. there is exactly
5751 * one valid generation and (at most) one invalid generation.
5753 kvm->arch.mmu_valid_gen = kvm->arch.mmu_valid_gen ? 0 : 1;
5756 * In order to ensure all vCPUs drop their soon-to-be invalid roots,
5757 * invalidating TDP MMU roots must be done while holding mmu_lock for
5758 * write and in the same critical section as making the reload request,
5759 * e.g. before kvm_zap_obsolete_pages() could drop mmu_lock and yield.
5761 if (is_tdp_mmu_enabled(kvm))
5762 kvm_tdp_mmu_invalidate_all_roots(kvm);
5765 * Notify all vcpus to reload its shadow page table and flush TLB.
5766 * Then all vcpus will switch to new shadow page table with the new
5769 * Note: we need to do this under the protection of mmu_lock,
5770 * otherwise, vcpu would purge shadow page but miss tlb flush.
5772 kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_FREE_OBSOLETE_ROOTS);
5774 kvm_zap_obsolete_pages(kvm);
5776 write_unlock(&kvm->mmu_lock);
5779 * Zap the invalidated TDP MMU roots, all SPTEs must be dropped before
5780 * returning to the caller, e.g. if the zap is in response to a memslot
5781 * deletion, mmu_notifier callbacks will be unable to reach the SPTEs
5782 * associated with the deleted memslot once the update completes, and
5783 * Deferring the zap until the final reference to the root is put would
5784 * lead to use-after-free.
5786 if (is_tdp_mmu_enabled(kvm))
5787 kvm_tdp_mmu_zap_invalidated_roots(kvm);
5790 static bool kvm_has_zapped_obsolete_pages(struct kvm *kvm)
5792 return unlikely(!list_empty_careful(&kvm->arch.zapped_obsolete_pages));
5795 static void kvm_mmu_invalidate_zap_pages_in_memslot(struct kvm *kvm,
5796 struct kvm_memory_slot *slot,
5797 struct kvm_page_track_notifier_node *node)
5799 kvm_mmu_zap_all_fast(kvm);
5802 int kvm_mmu_init_vm(struct kvm *kvm)
5804 struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
5807 INIT_LIST_HEAD(&kvm->arch.active_mmu_pages);
5808 INIT_LIST_HEAD(&kvm->arch.zapped_obsolete_pages);
5809 INIT_LIST_HEAD(&kvm->arch.lpage_disallowed_mmu_pages);
5810 spin_lock_init(&kvm->arch.mmu_unsync_pages_lock);
5812 r = kvm_mmu_init_tdp_mmu(kvm);
5816 node->track_write = kvm_mmu_pte_write;
5817 node->track_flush_slot = kvm_mmu_invalidate_zap_pages_in_memslot;
5818 kvm_page_track_register_notifier(kvm, node);
5822 void kvm_mmu_uninit_vm(struct kvm *kvm)
5824 struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
5826 kvm_page_track_unregister_notifier(kvm, node);
5828 kvm_mmu_uninit_tdp_mmu(kvm);
5831 static bool __kvm_zap_rmaps(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end)
5833 const struct kvm_memory_slot *memslot;
5834 struct kvm_memslots *slots;
5835 struct kvm_memslot_iter iter;
5840 if (!kvm_memslots_have_rmaps(kvm))
5843 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
5844 slots = __kvm_memslots(kvm, i);
5846 kvm_for_each_memslot_in_gfn_range(&iter, slots, gfn_start, gfn_end) {
5847 memslot = iter.slot;
5848 start = max(gfn_start, memslot->base_gfn);
5849 end = min(gfn_end, memslot->base_gfn + memslot->npages);
5850 if (WARN_ON_ONCE(start >= end))
5853 flush = slot_handle_level_range(kvm, memslot, kvm_zap_rmapp,
5855 PG_LEVEL_4K, KVM_MAX_HUGEPAGE_LEVEL,
5856 start, end - 1, true, flush);
5864 * Invalidate (zap) SPTEs that cover GFNs from gfn_start and up to gfn_end
5865 * (not including it)
5867 void kvm_zap_gfn_range(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end)
5872 if (WARN_ON_ONCE(gfn_end <= gfn_start))
5875 write_lock(&kvm->mmu_lock);
5877 kvm_inc_notifier_count(kvm, gfn_start, gfn_end);
5879 flush = __kvm_zap_rmaps(kvm, gfn_start, gfn_end);
5881 if (is_tdp_mmu_enabled(kvm)) {
5882 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
5883 flush = kvm_tdp_mmu_zap_leafs(kvm, i, gfn_start,
5884 gfn_end, true, flush);
5888 kvm_flush_remote_tlbs_with_address(kvm, gfn_start,
5889 gfn_end - gfn_start);
5891 kvm_dec_notifier_count(kvm, gfn_start, gfn_end);
5893 write_unlock(&kvm->mmu_lock);
5896 static bool slot_rmap_write_protect(struct kvm *kvm,
5897 struct kvm_rmap_head *rmap_head,
5898 const struct kvm_memory_slot *slot)
5900 return rmap_write_protect(rmap_head, false);
5903 void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
5904 const struct kvm_memory_slot *memslot,
5909 if (kvm_memslots_have_rmaps(kvm)) {
5910 write_lock(&kvm->mmu_lock);
5911 flush = slot_handle_level(kvm, memslot, slot_rmap_write_protect,
5912 start_level, KVM_MAX_HUGEPAGE_LEVEL,
5914 write_unlock(&kvm->mmu_lock);
5917 if (is_tdp_mmu_enabled(kvm)) {
5918 read_lock(&kvm->mmu_lock);
5919 flush |= kvm_tdp_mmu_wrprot_slot(kvm, memslot, start_level);
5920 read_unlock(&kvm->mmu_lock);
5924 * Flush TLBs if any SPTEs had to be write-protected to ensure that
5925 * guest writes are reflected in the dirty bitmap before the memslot
5926 * update completes, i.e. before enabling dirty logging is visible to
5929 * Perform the TLB flush outside the mmu_lock to reduce the amount of
5930 * time the lock is held. However, this does mean that another CPU can
5931 * now grab mmu_lock and encounter a write-protected SPTE while CPUs
5932 * still have a writable mapping for the associated GFN in their TLB.
5934 * This is safe but requires KVM to be careful when making decisions
5935 * based on the write-protection status of an SPTE. Specifically, KVM
5936 * also write-protects SPTEs to monitor changes to guest page tables
5937 * during shadow paging, and must guarantee no CPUs can write to those
5938 * page before the lock is dropped. As mentioned in the previous
5939 * paragraph, a write-protected SPTE is no guarantee that CPU cannot
5940 * perform writes. So to determine if a TLB flush is truly required, KVM
5941 * will clear a separate software-only bit (MMU-writable) and skip the
5942 * flush if-and-only-if this bit was already clear.
5944 * See is_writable_pte() for more details.
5947 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
5950 /* Must be called with the mmu_lock held in write-mode. */
5951 void kvm_mmu_try_split_huge_pages(struct kvm *kvm,
5952 const struct kvm_memory_slot *memslot,
5956 if (is_tdp_mmu_enabled(kvm))
5957 kvm_tdp_mmu_try_split_huge_pages(kvm, memslot, start, end,
5958 target_level, false);
5961 * A TLB flush is unnecessary at this point for the same resons as in
5962 * kvm_mmu_slot_try_split_huge_pages().
5966 void kvm_mmu_slot_try_split_huge_pages(struct kvm *kvm,
5967 const struct kvm_memory_slot *memslot,
5970 u64 start = memslot->base_gfn;
5971 u64 end = start + memslot->npages;
5973 if (is_tdp_mmu_enabled(kvm)) {
5974 read_lock(&kvm->mmu_lock);
5975 kvm_tdp_mmu_try_split_huge_pages(kvm, memslot, start, end, target_level, true);
5976 read_unlock(&kvm->mmu_lock);
5980 * No TLB flush is necessary here. KVM will flush TLBs after
5981 * write-protecting and/or clearing dirty on the newly split SPTEs to
5982 * ensure that guest writes are reflected in the dirty log before the
5983 * ioctl to enable dirty logging on this memslot completes. Since the
5984 * split SPTEs retain the write and dirty bits of the huge SPTE, it is
5985 * safe for KVM to decide if a TLB flush is necessary based on the split
5990 static bool kvm_mmu_zap_collapsible_spte(struct kvm *kvm,
5991 struct kvm_rmap_head *rmap_head,
5992 const struct kvm_memory_slot *slot)
5995 struct rmap_iterator iter;
5996 int need_tlb_flush = 0;
5998 struct kvm_mmu_page *sp;
6001 for_each_rmap_spte(rmap_head, &iter, sptep) {
6002 sp = sptep_to_sp(sptep);
6003 pfn = spte_to_pfn(*sptep);
6006 * We cannot do huge page mapping for indirect shadow pages,
6007 * which are found on the last rmap (level = 1) when not using
6008 * tdp; such shadow pages are synced with the page table in
6009 * the guest, and the guest page table is using 4K page size
6010 * mapping if the indirect sp has level = 1.
6012 if (sp->role.direct && !kvm_is_reserved_pfn(pfn) &&
6013 sp->role.level < kvm_mmu_max_mapping_level(kvm, slot, sp->gfn,
6014 pfn, PG_LEVEL_NUM)) {
6015 pte_list_remove(kvm, rmap_head, sptep);
6017 if (kvm_available_flush_tlb_with_range())
6018 kvm_flush_remote_tlbs_with_address(kvm, sp->gfn,
6019 KVM_PAGES_PER_HPAGE(sp->role.level));
6027 return need_tlb_flush;
6030 void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm,
6031 const struct kvm_memory_slot *slot)
6033 if (kvm_memslots_have_rmaps(kvm)) {
6034 write_lock(&kvm->mmu_lock);
6036 * Zap only 4k SPTEs since the legacy MMU only supports dirty
6037 * logging at a 4k granularity and never creates collapsible
6038 * 2m SPTEs during dirty logging.
6040 if (slot_handle_level_4k(kvm, slot, kvm_mmu_zap_collapsible_spte, true))
6041 kvm_arch_flush_remote_tlbs_memslot(kvm, slot);
6042 write_unlock(&kvm->mmu_lock);
6045 if (is_tdp_mmu_enabled(kvm)) {
6046 read_lock(&kvm->mmu_lock);
6047 kvm_tdp_mmu_zap_collapsible_sptes(kvm, slot);
6048 read_unlock(&kvm->mmu_lock);
6052 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm,
6053 const struct kvm_memory_slot *memslot)
6056 * All current use cases for flushing the TLBs for a specific memslot
6057 * related to dirty logging, and many do the TLB flush out of mmu_lock.
6058 * The interaction between the various operations on memslot must be
6059 * serialized by slots_locks to ensure the TLB flush from one operation
6060 * is observed by any other operation on the same memslot.
6062 lockdep_assert_held(&kvm->slots_lock);
6063 kvm_flush_remote_tlbs_with_address(kvm, memslot->base_gfn,
6067 void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm,
6068 const struct kvm_memory_slot *memslot)
6072 if (kvm_memslots_have_rmaps(kvm)) {
6073 write_lock(&kvm->mmu_lock);
6075 * Clear dirty bits only on 4k SPTEs since the legacy MMU only
6076 * support dirty logging at a 4k granularity.
6078 flush = slot_handle_level_4k(kvm, memslot, __rmap_clear_dirty, false);
6079 write_unlock(&kvm->mmu_lock);
6082 if (is_tdp_mmu_enabled(kvm)) {
6083 read_lock(&kvm->mmu_lock);
6084 flush |= kvm_tdp_mmu_clear_dirty_slot(kvm, memslot);
6085 read_unlock(&kvm->mmu_lock);
6089 * It's also safe to flush TLBs out of mmu lock here as currently this
6090 * function is only used for dirty logging, in which case flushing TLB
6091 * out of mmu lock also guarantees no dirty pages will be lost in
6095 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
6098 void kvm_mmu_zap_all(struct kvm *kvm)
6100 struct kvm_mmu_page *sp, *node;
6101 LIST_HEAD(invalid_list);
6104 write_lock(&kvm->mmu_lock);
6106 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link) {
6107 if (WARN_ON(sp->role.invalid))
6109 if (__kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list, &ign))
6111 if (cond_resched_rwlock_write(&kvm->mmu_lock))
6115 kvm_mmu_commit_zap_page(kvm, &invalid_list);
6117 if (is_tdp_mmu_enabled(kvm))
6118 kvm_tdp_mmu_zap_all(kvm);
6120 write_unlock(&kvm->mmu_lock);
6123 void kvm_mmu_invalidate_mmio_sptes(struct kvm *kvm, u64 gen)
6125 WARN_ON(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS);
6127 gen &= MMIO_SPTE_GEN_MASK;
6130 * Generation numbers are incremented in multiples of the number of
6131 * address spaces in order to provide unique generations across all
6132 * address spaces. Strip what is effectively the address space
6133 * modifier prior to checking for a wrap of the MMIO generation so
6134 * that a wrap in any address space is detected.
6136 gen &= ~((u64)KVM_ADDRESS_SPACE_NUM - 1);
6139 * The very rare case: if the MMIO generation number has wrapped,
6140 * zap all shadow pages.
6142 if (unlikely(gen == 0)) {
6143 kvm_debug_ratelimited("kvm: zapping shadow pages for mmio generation wraparound\n");
6144 kvm_mmu_zap_all_fast(kvm);
6148 static unsigned long
6149 mmu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
6152 int nr_to_scan = sc->nr_to_scan;
6153 unsigned long freed = 0;
6155 mutex_lock(&kvm_lock);
6157 list_for_each_entry(kvm, &vm_list, vm_list) {
6159 LIST_HEAD(invalid_list);
6162 * Never scan more than sc->nr_to_scan VM instances.
6163 * Will not hit this condition practically since we do not try
6164 * to shrink more than one VM and it is very unlikely to see
6165 * !n_used_mmu_pages so many times.
6170 * n_used_mmu_pages is accessed without holding kvm->mmu_lock
6171 * here. We may skip a VM instance errorneosly, but we do not
6172 * want to shrink a VM that only started to populate its MMU
6175 if (!kvm->arch.n_used_mmu_pages &&
6176 !kvm_has_zapped_obsolete_pages(kvm))
6179 idx = srcu_read_lock(&kvm->srcu);
6180 write_lock(&kvm->mmu_lock);
6182 if (kvm_has_zapped_obsolete_pages(kvm)) {
6183 kvm_mmu_commit_zap_page(kvm,
6184 &kvm->arch.zapped_obsolete_pages);
6188 freed = kvm_mmu_zap_oldest_mmu_pages(kvm, sc->nr_to_scan);
6191 write_unlock(&kvm->mmu_lock);
6192 srcu_read_unlock(&kvm->srcu, idx);
6195 * unfair on small ones
6196 * per-vm shrinkers cry out
6197 * sadness comes quickly
6199 list_move_tail(&kvm->vm_list, &vm_list);
6203 mutex_unlock(&kvm_lock);
6207 static unsigned long
6208 mmu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
6210 return percpu_counter_read_positive(&kvm_total_used_mmu_pages);
6213 static struct shrinker mmu_shrinker = {
6214 .count_objects = mmu_shrink_count,
6215 .scan_objects = mmu_shrink_scan,
6216 .seeks = DEFAULT_SEEKS * 10,
6219 static void mmu_destroy_caches(void)
6221 kmem_cache_destroy(pte_list_desc_cache);
6222 kmem_cache_destroy(mmu_page_header_cache);
6225 static bool get_nx_auto_mode(void)
6227 /* Return true when CPU has the bug, and mitigations are ON */
6228 return boot_cpu_has_bug(X86_BUG_ITLB_MULTIHIT) && !cpu_mitigations_off();
6231 static void __set_nx_huge_pages(bool val)
6233 nx_huge_pages = itlb_multihit_kvm_mitigation = val;
6236 static int set_nx_huge_pages(const char *val, const struct kernel_param *kp)
6238 bool old_val = nx_huge_pages;
6241 /* In "auto" mode deploy workaround only if CPU has the bug. */
6242 if (sysfs_streq(val, "off"))
6244 else if (sysfs_streq(val, "force"))
6246 else if (sysfs_streq(val, "auto"))
6247 new_val = get_nx_auto_mode();
6248 else if (strtobool(val, &new_val) < 0)
6251 __set_nx_huge_pages(new_val);
6253 if (new_val != old_val) {
6256 mutex_lock(&kvm_lock);
6258 list_for_each_entry(kvm, &vm_list, vm_list) {
6259 mutex_lock(&kvm->slots_lock);
6260 kvm_mmu_zap_all_fast(kvm);
6261 mutex_unlock(&kvm->slots_lock);
6263 wake_up_process(kvm->arch.nx_lpage_recovery_thread);
6265 mutex_unlock(&kvm_lock);
6272 * nx_huge_pages needs to be resolved to true/false when kvm.ko is loaded, as
6273 * its default value of -1 is technically undefined behavior for a boolean.
6275 void kvm_mmu_x86_module_init(void)
6277 if (nx_huge_pages == -1)
6278 __set_nx_huge_pages(get_nx_auto_mode());
6282 * The bulk of the MMU initialization is deferred until the vendor module is
6283 * loaded as many of the masks/values may be modified by VMX or SVM, i.e. need
6284 * to be reset when a potentially different vendor module is loaded.
6286 int kvm_mmu_vendor_module_init(void)
6291 * MMU roles use union aliasing which is, generally speaking, an
6292 * undefined behavior. However, we supposedly know how compilers behave
6293 * and the current status quo is unlikely to change. Guardians below are
6294 * supposed to let us know if the assumption becomes false.
6296 BUILD_BUG_ON(sizeof(union kvm_mmu_page_role) != sizeof(u32));
6297 BUILD_BUG_ON(sizeof(union kvm_mmu_extended_role) != sizeof(u32));
6298 BUILD_BUG_ON(sizeof(union kvm_mmu_role) != sizeof(u64));
6300 kvm_mmu_reset_all_pte_masks();
6302 pte_list_desc_cache = kmem_cache_create("pte_list_desc",
6303 sizeof(struct pte_list_desc),
6304 0, SLAB_ACCOUNT, NULL);
6305 if (!pte_list_desc_cache)
6308 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
6309 sizeof(struct kvm_mmu_page),
6310 0, SLAB_ACCOUNT, NULL);
6311 if (!mmu_page_header_cache)
6314 if (percpu_counter_init(&kvm_total_used_mmu_pages, 0, GFP_KERNEL))
6317 ret = register_shrinker(&mmu_shrinker);
6324 mmu_destroy_caches();
6328 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
6330 kvm_mmu_unload(vcpu);
6331 free_mmu_pages(&vcpu->arch.root_mmu);
6332 free_mmu_pages(&vcpu->arch.guest_mmu);
6333 mmu_free_memory_caches(vcpu);
6336 void kvm_mmu_vendor_module_exit(void)
6338 mmu_destroy_caches();
6339 percpu_counter_destroy(&kvm_total_used_mmu_pages);
6340 unregister_shrinker(&mmu_shrinker);
6344 * Calculate the effective recovery period, accounting for '0' meaning "let KVM
6345 * select a halving time of 1 hour". Returns true if recovery is enabled.
6347 static bool calc_nx_huge_pages_recovery_period(uint *period)
6350 * Use READ_ONCE to get the params, this may be called outside of the
6351 * param setters, e.g. by the kthread to compute its next timeout.
6353 bool enabled = READ_ONCE(nx_huge_pages);
6354 uint ratio = READ_ONCE(nx_huge_pages_recovery_ratio);
6356 if (!enabled || !ratio)
6359 *period = READ_ONCE(nx_huge_pages_recovery_period_ms);
6361 /* Make sure the period is not less than one second. */
6362 ratio = min(ratio, 3600u);
6363 *period = 60 * 60 * 1000 / ratio;
6368 static int set_nx_huge_pages_recovery_param(const char *val, const struct kernel_param *kp)
6370 bool was_recovery_enabled, is_recovery_enabled;
6371 uint old_period, new_period;
6374 was_recovery_enabled = calc_nx_huge_pages_recovery_period(&old_period);
6376 err = param_set_uint(val, kp);
6380 is_recovery_enabled = calc_nx_huge_pages_recovery_period(&new_period);
6382 if (is_recovery_enabled &&
6383 (!was_recovery_enabled || old_period > new_period)) {
6386 mutex_lock(&kvm_lock);
6388 list_for_each_entry(kvm, &vm_list, vm_list)
6389 wake_up_process(kvm->arch.nx_lpage_recovery_thread);
6391 mutex_unlock(&kvm_lock);
6397 static void kvm_recover_nx_lpages(struct kvm *kvm)
6399 unsigned long nx_lpage_splits = kvm->stat.nx_lpage_splits;
6401 struct kvm_mmu_page *sp;
6403 LIST_HEAD(invalid_list);
6407 rcu_idx = srcu_read_lock(&kvm->srcu);
6408 write_lock(&kvm->mmu_lock);
6411 * Zapping TDP MMU shadow pages, including the remote TLB flush, must
6412 * be done under RCU protection, because the pages are freed via RCU
6417 ratio = READ_ONCE(nx_huge_pages_recovery_ratio);
6418 to_zap = ratio ? DIV_ROUND_UP(nx_lpage_splits, ratio) : 0;
6419 for ( ; to_zap; --to_zap) {
6420 if (list_empty(&kvm->arch.lpage_disallowed_mmu_pages))
6424 * We use a separate list instead of just using active_mmu_pages
6425 * because the number of lpage_disallowed pages is expected to
6426 * be relatively small compared to the total.
6428 sp = list_first_entry(&kvm->arch.lpage_disallowed_mmu_pages,
6429 struct kvm_mmu_page,
6430 lpage_disallowed_link);
6431 WARN_ON_ONCE(!sp->lpage_disallowed);
6432 if (is_tdp_mmu_page(sp)) {
6433 flush |= kvm_tdp_mmu_zap_sp(kvm, sp);
6435 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
6436 WARN_ON_ONCE(sp->lpage_disallowed);
6439 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) {
6440 kvm_mmu_remote_flush_or_zap(kvm, &invalid_list, flush);
6443 cond_resched_rwlock_write(&kvm->mmu_lock);
6449 kvm_mmu_remote_flush_or_zap(kvm, &invalid_list, flush);
6453 write_unlock(&kvm->mmu_lock);
6454 srcu_read_unlock(&kvm->srcu, rcu_idx);
6457 static long get_nx_lpage_recovery_timeout(u64 start_time)
6462 enabled = calc_nx_huge_pages_recovery_period(&period);
6464 return enabled ? start_time + msecs_to_jiffies(period) - get_jiffies_64()
6465 : MAX_SCHEDULE_TIMEOUT;
6468 static int kvm_nx_lpage_recovery_worker(struct kvm *kvm, uintptr_t data)
6471 long remaining_time;
6474 start_time = get_jiffies_64();
6475 remaining_time = get_nx_lpage_recovery_timeout(start_time);
6477 set_current_state(TASK_INTERRUPTIBLE);
6478 while (!kthread_should_stop() && remaining_time > 0) {
6479 schedule_timeout(remaining_time);
6480 remaining_time = get_nx_lpage_recovery_timeout(start_time);
6481 set_current_state(TASK_INTERRUPTIBLE);
6484 set_current_state(TASK_RUNNING);
6486 if (kthread_should_stop())
6489 kvm_recover_nx_lpages(kvm);
6493 int kvm_mmu_post_init_vm(struct kvm *kvm)
6497 err = kvm_vm_create_worker_thread(kvm, kvm_nx_lpage_recovery_worker, 0,
6498 "kvm-nx-lpage-recovery",
6499 &kvm->arch.nx_lpage_recovery_thread);
6501 kthread_unpark(kvm->arch.nx_lpage_recovery_thread);
6506 void kvm_mmu_pre_destroy_vm(struct kvm *kvm)
6508 if (kvm->arch.nx_lpage_recovery_thread)
6509 kthread_stop(kvm->arch.nx_lpage_recovery_thread);