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>
52 #include <asm/kvm_page_track.h>
55 extern bool itlb_multihit_kvm_mitigation;
57 static int __read_mostly nx_huge_pages = -1;
58 #ifdef CONFIG_PREEMPT_RT
59 /* Recovery can cause latency spikes, disable it for PREEMPT_RT. */
60 static uint __read_mostly nx_huge_pages_recovery_ratio = 0;
62 static uint __read_mostly nx_huge_pages_recovery_ratio = 60;
65 static int set_nx_huge_pages(const char *val, const struct kernel_param *kp);
66 static int set_nx_huge_pages_recovery_ratio(const char *val, const struct kernel_param *kp);
68 static const struct kernel_param_ops nx_huge_pages_ops = {
69 .set = set_nx_huge_pages,
70 .get = param_get_bool,
73 static const struct kernel_param_ops nx_huge_pages_recovery_ratio_ops = {
74 .set = set_nx_huge_pages_recovery_ratio,
75 .get = param_get_uint,
78 module_param_cb(nx_huge_pages, &nx_huge_pages_ops, &nx_huge_pages, 0644);
79 __MODULE_PARM_TYPE(nx_huge_pages, "bool");
80 module_param_cb(nx_huge_pages_recovery_ratio, &nx_huge_pages_recovery_ratio_ops,
81 &nx_huge_pages_recovery_ratio, 0644);
82 __MODULE_PARM_TYPE(nx_huge_pages_recovery_ratio, "uint");
84 static bool __read_mostly force_flush_and_sync_on_reuse;
85 module_param_named(flush_on_reuse, force_flush_and_sync_on_reuse, bool, 0644);
88 * When setting this variable to true it enables Two-Dimensional-Paging
89 * where the hardware walks 2 page tables:
90 * 1. the guest-virtual to guest-physical
91 * 2. while doing 1. it walks guest-physical to host-physical
92 * If the hardware supports that we don't need to do shadow paging.
94 bool tdp_enabled = false;
96 static int max_huge_page_level __read_mostly;
97 static int max_tdp_level __read_mostly;
100 AUDIT_PRE_PAGE_FAULT,
101 AUDIT_POST_PAGE_FAULT,
103 AUDIT_POST_PTE_WRITE,
110 module_param(dbg, bool, 0644);
113 #define PTE_PREFETCH_NUM 8
115 #define PT32_LEVEL_BITS 10
117 #define PT32_LEVEL_SHIFT(level) \
118 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
120 #define PT32_LVL_OFFSET_MASK(level) \
121 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
122 * PT32_LEVEL_BITS))) - 1))
124 #define PT32_INDEX(address, level)\
125 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
128 #define PT32_BASE_ADDR_MASK PAGE_MASK
129 #define PT32_DIR_BASE_ADDR_MASK \
130 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
131 #define PT32_LVL_ADDR_MASK(level) \
132 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
133 * PT32_LEVEL_BITS))) - 1))
135 #include <trace/events/kvm.h>
137 /* make pte_list_desc fit well in cache line */
138 #define PTE_LIST_EXT 3
140 struct pte_list_desc {
141 u64 *sptes[PTE_LIST_EXT];
142 struct pte_list_desc *more;
145 struct kvm_shadow_walk_iterator {
153 #define for_each_shadow_entry_using_root(_vcpu, _root, _addr, _walker) \
154 for (shadow_walk_init_using_root(&(_walker), (_vcpu), \
156 shadow_walk_okay(&(_walker)); \
157 shadow_walk_next(&(_walker)))
159 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
160 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
161 shadow_walk_okay(&(_walker)); \
162 shadow_walk_next(&(_walker)))
164 #define for_each_shadow_entry_lockless(_vcpu, _addr, _walker, spte) \
165 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
166 shadow_walk_okay(&(_walker)) && \
167 ({ spte = mmu_spte_get_lockless(_walker.sptep); 1; }); \
168 __shadow_walk_next(&(_walker), spte))
170 static struct kmem_cache *pte_list_desc_cache;
171 struct kmem_cache *mmu_page_header_cache;
172 static struct percpu_counter kvm_total_used_mmu_pages;
174 static void mmu_spte_set(u64 *sptep, u64 spte);
175 static union kvm_mmu_page_role
176 kvm_mmu_calc_root_page_role(struct kvm_vcpu *vcpu);
178 #define CREATE_TRACE_POINTS
179 #include "mmutrace.h"
182 static inline bool kvm_available_flush_tlb_with_range(void)
184 return kvm_x86_ops.tlb_remote_flush_with_range;
187 static void kvm_flush_remote_tlbs_with_range(struct kvm *kvm,
188 struct kvm_tlb_range *range)
192 if (range && kvm_x86_ops.tlb_remote_flush_with_range)
193 ret = static_call(kvm_x86_tlb_remote_flush_with_range)(kvm, range);
196 kvm_flush_remote_tlbs(kvm);
199 void kvm_flush_remote_tlbs_with_address(struct kvm *kvm,
200 u64 start_gfn, u64 pages)
202 struct kvm_tlb_range range;
204 range.start_gfn = start_gfn;
207 kvm_flush_remote_tlbs_with_range(kvm, &range);
210 bool is_nx_huge_page_enabled(void)
212 return READ_ONCE(nx_huge_pages);
215 static void mark_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, u64 gfn,
218 u64 mask = make_mmio_spte(vcpu, gfn, access);
220 trace_mark_mmio_spte(sptep, gfn, mask);
221 mmu_spte_set(sptep, mask);
224 static gfn_t get_mmio_spte_gfn(u64 spte)
226 u64 gpa = spte & shadow_nonpresent_or_rsvd_lower_gfn_mask;
228 gpa |= (spte >> SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)
229 & shadow_nonpresent_or_rsvd_mask;
231 return gpa >> PAGE_SHIFT;
234 static unsigned get_mmio_spte_access(u64 spte)
236 return spte & shadow_mmio_access_mask;
239 static bool set_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn,
240 kvm_pfn_t pfn, unsigned int access)
242 if (unlikely(is_noslot_pfn(pfn))) {
243 mark_mmio_spte(vcpu, sptep, gfn, access);
250 static bool check_mmio_spte(struct kvm_vcpu *vcpu, u64 spte)
252 u64 kvm_gen, spte_gen, gen;
254 gen = kvm_vcpu_memslots(vcpu)->generation;
255 if (unlikely(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS))
258 kvm_gen = gen & MMIO_SPTE_GEN_MASK;
259 spte_gen = get_mmio_spte_generation(spte);
261 trace_check_mmio_spte(spte, kvm_gen, spte_gen);
262 return likely(kvm_gen == spte_gen);
265 static gpa_t translate_gpa(struct kvm_vcpu *vcpu, gpa_t gpa, u32 access,
266 struct x86_exception *exception)
268 /* Check if guest physical address doesn't exceed guest maximum */
269 if (kvm_vcpu_is_illegal_gpa(vcpu, gpa)) {
270 exception->error_code |= PFERR_RSVD_MASK;
277 static int is_cpuid_PSE36(void)
282 static int is_nx(struct kvm_vcpu *vcpu)
284 return vcpu->arch.efer & EFER_NX;
287 static gfn_t pse36_gfn_delta(u32 gpte)
289 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
291 return (gpte & PT32_DIR_PSE36_MASK) << shift;
295 static void __set_spte(u64 *sptep, u64 spte)
297 WRITE_ONCE(*sptep, spte);
300 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
302 WRITE_ONCE(*sptep, spte);
305 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
307 return xchg(sptep, spte);
310 static u64 __get_spte_lockless(u64 *sptep)
312 return READ_ONCE(*sptep);
323 static void count_spte_clear(u64 *sptep, u64 spte)
325 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
327 if (is_shadow_present_pte(spte))
330 /* Ensure the spte is completely set before we increase the count */
332 sp->clear_spte_count++;
335 static void __set_spte(u64 *sptep, u64 spte)
337 union split_spte *ssptep, sspte;
339 ssptep = (union split_spte *)sptep;
340 sspte = (union split_spte)spte;
342 ssptep->spte_high = sspte.spte_high;
345 * If we map the spte from nonpresent to present, We should store
346 * the high bits firstly, then set present bit, so cpu can not
347 * fetch this spte while we are setting the spte.
351 WRITE_ONCE(ssptep->spte_low, sspte.spte_low);
354 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
356 union split_spte *ssptep, sspte;
358 ssptep = (union split_spte *)sptep;
359 sspte = (union split_spte)spte;
361 WRITE_ONCE(ssptep->spte_low, sspte.spte_low);
364 * If we map the spte from present to nonpresent, we should clear
365 * present bit firstly to avoid vcpu fetch the old high bits.
369 ssptep->spte_high = sspte.spte_high;
370 count_spte_clear(sptep, spte);
373 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
375 union split_spte *ssptep, sspte, orig;
377 ssptep = (union split_spte *)sptep;
378 sspte = (union split_spte)spte;
380 /* xchg acts as a barrier before the setting of the high bits */
381 orig.spte_low = xchg(&ssptep->spte_low, sspte.spte_low);
382 orig.spte_high = ssptep->spte_high;
383 ssptep->spte_high = sspte.spte_high;
384 count_spte_clear(sptep, spte);
390 * The idea using the light way get the spte on x86_32 guest is from
391 * gup_get_pte (mm/gup.c).
393 * An spte tlb flush may be pending, because kvm_set_pte_rmapp
394 * coalesces them and we are running out of the MMU lock. Therefore
395 * we need to protect against in-progress updates of the spte.
397 * Reading the spte while an update is in progress may get the old value
398 * for the high part of the spte. The race is fine for a present->non-present
399 * change (because the high part of the spte is ignored for non-present spte),
400 * but for a present->present change we must reread the spte.
402 * All such changes are done in two steps (present->non-present and
403 * non-present->present), hence it is enough to count the number of
404 * present->non-present updates: if it changed while reading the spte,
405 * we might have hit the race. This is done using clear_spte_count.
407 static u64 __get_spte_lockless(u64 *sptep)
409 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
410 union split_spte spte, *orig = (union split_spte *)sptep;
414 count = sp->clear_spte_count;
417 spte.spte_low = orig->spte_low;
420 spte.spte_high = orig->spte_high;
423 if (unlikely(spte.spte_low != orig->spte_low ||
424 count != sp->clear_spte_count))
431 static bool spte_has_volatile_bits(u64 spte)
433 if (!is_shadow_present_pte(spte))
437 * Always atomically update spte if it can be updated
438 * out of mmu-lock, it can ensure dirty bit is not lost,
439 * also, it can help us to get a stable is_writable_pte()
440 * to ensure tlb flush is not missed.
442 if (spte_can_locklessly_be_made_writable(spte) ||
443 is_access_track_spte(spte))
446 if (spte_ad_enabled(spte)) {
447 if ((spte & shadow_accessed_mask) == 0 ||
448 (is_writable_pte(spte) && (spte & shadow_dirty_mask) == 0))
455 /* Rules for using mmu_spte_set:
456 * Set the sptep from nonpresent to present.
457 * Note: the sptep being assigned *must* be either not present
458 * or in a state where the hardware will not attempt to update
461 static void mmu_spte_set(u64 *sptep, u64 new_spte)
463 WARN_ON(is_shadow_present_pte(*sptep));
464 __set_spte(sptep, new_spte);
468 * Update the SPTE (excluding the PFN), but do not track changes in its
469 * accessed/dirty status.
471 static u64 mmu_spte_update_no_track(u64 *sptep, u64 new_spte)
473 u64 old_spte = *sptep;
475 WARN_ON(!is_shadow_present_pte(new_spte));
477 if (!is_shadow_present_pte(old_spte)) {
478 mmu_spte_set(sptep, new_spte);
482 if (!spte_has_volatile_bits(old_spte))
483 __update_clear_spte_fast(sptep, new_spte);
485 old_spte = __update_clear_spte_slow(sptep, new_spte);
487 WARN_ON(spte_to_pfn(old_spte) != spte_to_pfn(new_spte));
492 /* Rules for using mmu_spte_update:
493 * Update the state bits, it means the mapped pfn is not changed.
495 * Whenever we overwrite a writable spte with a read-only one we
496 * should flush remote TLBs. Otherwise rmap_write_protect
497 * will find a read-only spte, even though the writable spte
498 * might be cached on a CPU's TLB, the return value indicates this
501 * Returns true if the TLB needs to be flushed
503 static bool mmu_spte_update(u64 *sptep, u64 new_spte)
506 u64 old_spte = mmu_spte_update_no_track(sptep, new_spte);
508 if (!is_shadow_present_pte(old_spte))
512 * For the spte updated out of mmu-lock is safe, since
513 * we always atomically update it, see the comments in
514 * spte_has_volatile_bits().
516 if (spte_can_locklessly_be_made_writable(old_spte) &&
517 !is_writable_pte(new_spte))
521 * Flush TLB when accessed/dirty states are changed in the page tables,
522 * to guarantee consistency between TLB and page tables.
525 if (is_accessed_spte(old_spte) && !is_accessed_spte(new_spte)) {
527 kvm_set_pfn_accessed(spte_to_pfn(old_spte));
530 if (is_dirty_spte(old_spte) && !is_dirty_spte(new_spte)) {
532 kvm_set_pfn_dirty(spte_to_pfn(old_spte));
539 * Rules for using mmu_spte_clear_track_bits:
540 * It sets the sptep from present to nonpresent, and track the
541 * state bits, it is used to clear the last level sptep.
542 * Returns non-zero if the PTE was previously valid.
544 static int mmu_spte_clear_track_bits(u64 *sptep)
547 u64 old_spte = *sptep;
549 if (!spte_has_volatile_bits(old_spte))
550 __update_clear_spte_fast(sptep, 0ull);
552 old_spte = __update_clear_spte_slow(sptep, 0ull);
554 if (!is_shadow_present_pte(old_spte))
557 pfn = spte_to_pfn(old_spte);
560 * KVM does not hold the refcount of the page used by
561 * kvm mmu, before reclaiming the page, we should
562 * unmap it from mmu first.
564 WARN_ON(!kvm_is_reserved_pfn(pfn) && !page_count(pfn_to_page(pfn)));
566 if (is_accessed_spte(old_spte))
567 kvm_set_pfn_accessed(pfn);
569 if (is_dirty_spte(old_spte))
570 kvm_set_pfn_dirty(pfn);
576 * Rules for using mmu_spte_clear_no_track:
577 * Directly clear spte without caring the state bits of sptep,
578 * it is used to set the upper level spte.
580 static void mmu_spte_clear_no_track(u64 *sptep)
582 __update_clear_spte_fast(sptep, 0ull);
585 static u64 mmu_spte_get_lockless(u64 *sptep)
587 return __get_spte_lockless(sptep);
590 /* Restore an acc-track PTE back to a regular PTE */
591 static u64 restore_acc_track_spte(u64 spte)
594 u64 saved_bits = (spte >> SHADOW_ACC_TRACK_SAVED_BITS_SHIFT)
595 & SHADOW_ACC_TRACK_SAVED_BITS_MASK;
597 WARN_ON_ONCE(spte_ad_enabled(spte));
598 WARN_ON_ONCE(!is_access_track_spte(spte));
600 new_spte &= ~shadow_acc_track_mask;
601 new_spte &= ~(SHADOW_ACC_TRACK_SAVED_BITS_MASK <<
602 SHADOW_ACC_TRACK_SAVED_BITS_SHIFT);
603 new_spte |= saved_bits;
608 /* Returns the Accessed status of the PTE and resets it at the same time. */
609 static bool mmu_spte_age(u64 *sptep)
611 u64 spte = mmu_spte_get_lockless(sptep);
613 if (!is_accessed_spte(spte))
616 if (spte_ad_enabled(spte)) {
617 clear_bit((ffs(shadow_accessed_mask) - 1),
618 (unsigned long *)sptep);
621 * Capture the dirty status of the page, so that it doesn't get
622 * lost when the SPTE is marked for access tracking.
624 if (is_writable_pte(spte))
625 kvm_set_pfn_dirty(spte_to_pfn(spte));
627 spte = mark_spte_for_access_track(spte);
628 mmu_spte_update_no_track(sptep, spte);
634 static void walk_shadow_page_lockless_begin(struct kvm_vcpu *vcpu)
637 * Prevent page table teardown by making any free-er wait during
638 * kvm_flush_remote_tlbs() IPI to all active vcpus.
643 * Make sure a following spte read is not reordered ahead of the write
646 smp_store_mb(vcpu->mode, READING_SHADOW_PAGE_TABLES);
649 static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
652 * Make sure the write to vcpu->mode is not reordered in front of
653 * reads to sptes. If it does, kvm_mmu_commit_zap_page() can see us
654 * OUTSIDE_GUEST_MODE and proceed to free the shadow page table.
656 smp_store_release(&vcpu->mode, OUTSIDE_GUEST_MODE);
660 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu, bool maybe_indirect)
664 /* 1 rmap, 1 parent PTE per level, and the prefetched rmaps. */
665 r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
666 1 + PT64_ROOT_MAX_LEVEL + PTE_PREFETCH_NUM);
669 r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_shadow_page_cache,
670 PT64_ROOT_MAX_LEVEL);
673 if (maybe_indirect) {
674 r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_gfn_array_cache,
675 PT64_ROOT_MAX_LEVEL);
679 return kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
680 PT64_ROOT_MAX_LEVEL);
683 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
685 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache);
686 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_shadow_page_cache);
687 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_gfn_array_cache);
688 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
691 static struct pte_list_desc *mmu_alloc_pte_list_desc(struct kvm_vcpu *vcpu)
693 return kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_list_desc_cache);
696 static void mmu_free_pte_list_desc(struct pte_list_desc *pte_list_desc)
698 kmem_cache_free(pte_list_desc_cache, pte_list_desc);
701 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
703 if (!sp->role.direct)
704 return sp->gfns[index];
706 return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
709 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
711 if (!sp->role.direct) {
712 sp->gfns[index] = gfn;
716 if (WARN_ON(gfn != kvm_mmu_page_get_gfn(sp, index)))
717 pr_err_ratelimited("gfn mismatch under direct page %llx "
718 "(expected %llx, got %llx)\n",
720 kvm_mmu_page_get_gfn(sp, index), gfn);
724 * Return the pointer to the large page information for a given gfn,
725 * handling slots that are not large page aligned.
727 static struct kvm_lpage_info *lpage_info_slot(gfn_t gfn,
728 struct kvm_memory_slot *slot,
733 idx = gfn_to_index(gfn, slot->base_gfn, level);
734 return &slot->arch.lpage_info[level - 2][idx];
737 static void update_gfn_disallow_lpage_count(struct kvm_memory_slot *slot,
738 gfn_t gfn, int count)
740 struct kvm_lpage_info *linfo;
743 for (i = PG_LEVEL_2M; i <= KVM_MAX_HUGEPAGE_LEVEL; ++i) {
744 linfo = lpage_info_slot(gfn, slot, i);
745 linfo->disallow_lpage += count;
746 WARN_ON(linfo->disallow_lpage < 0);
750 void kvm_mmu_gfn_disallow_lpage(struct kvm_memory_slot *slot, gfn_t gfn)
752 update_gfn_disallow_lpage_count(slot, gfn, 1);
755 void kvm_mmu_gfn_allow_lpage(struct kvm_memory_slot *slot, gfn_t gfn)
757 update_gfn_disallow_lpage_count(slot, gfn, -1);
760 static void account_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
762 struct kvm_memslots *slots;
763 struct kvm_memory_slot *slot;
766 kvm->arch.indirect_shadow_pages++;
768 slots = kvm_memslots_for_spte_role(kvm, sp->role);
769 slot = __gfn_to_memslot(slots, gfn);
771 /* the non-leaf shadow pages are keeping readonly. */
772 if (sp->role.level > PG_LEVEL_4K)
773 return kvm_slot_page_track_add_page(kvm, slot, gfn,
774 KVM_PAGE_TRACK_WRITE);
776 kvm_mmu_gfn_disallow_lpage(slot, gfn);
779 void account_huge_nx_page(struct kvm *kvm, struct kvm_mmu_page *sp)
781 if (sp->lpage_disallowed)
784 ++kvm->stat.nx_lpage_splits;
785 list_add_tail(&sp->lpage_disallowed_link,
786 &kvm->arch.lpage_disallowed_mmu_pages);
787 sp->lpage_disallowed = true;
790 static void unaccount_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
792 struct kvm_memslots *slots;
793 struct kvm_memory_slot *slot;
796 kvm->arch.indirect_shadow_pages--;
798 slots = kvm_memslots_for_spte_role(kvm, sp->role);
799 slot = __gfn_to_memslot(slots, gfn);
800 if (sp->role.level > PG_LEVEL_4K)
801 return kvm_slot_page_track_remove_page(kvm, slot, gfn,
802 KVM_PAGE_TRACK_WRITE);
804 kvm_mmu_gfn_allow_lpage(slot, gfn);
807 void unaccount_huge_nx_page(struct kvm *kvm, struct kvm_mmu_page *sp)
809 --kvm->stat.nx_lpage_splits;
810 sp->lpage_disallowed = false;
811 list_del(&sp->lpage_disallowed_link);
814 static struct kvm_memory_slot *
815 gfn_to_memslot_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t gfn,
818 struct kvm_memory_slot *slot;
820 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
821 if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
823 if (no_dirty_log && kvm_slot_dirty_track_enabled(slot))
830 * About rmap_head encoding:
832 * If the bit zero of rmap_head->val is clear, then it points to the only spte
833 * in this rmap chain. Otherwise, (rmap_head->val & ~1) points to a struct
834 * pte_list_desc containing more mappings.
838 * Returns the number of pointers in the rmap chain, not counting the new one.
840 static int pte_list_add(struct kvm_vcpu *vcpu, u64 *spte,
841 struct kvm_rmap_head *rmap_head)
843 struct pte_list_desc *desc;
846 if (!rmap_head->val) {
847 rmap_printk("%p %llx 0->1\n", spte, *spte);
848 rmap_head->val = (unsigned long)spte;
849 } else if (!(rmap_head->val & 1)) {
850 rmap_printk("%p %llx 1->many\n", spte, *spte);
851 desc = mmu_alloc_pte_list_desc(vcpu);
852 desc->sptes[0] = (u64 *)rmap_head->val;
853 desc->sptes[1] = spte;
854 rmap_head->val = (unsigned long)desc | 1;
857 rmap_printk("%p %llx many->many\n", spte, *spte);
858 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
859 while (desc->sptes[PTE_LIST_EXT-1]) {
860 count += PTE_LIST_EXT;
863 desc->more = mmu_alloc_pte_list_desc(vcpu);
869 for (i = 0; desc->sptes[i]; ++i)
871 desc->sptes[i] = spte;
877 pte_list_desc_remove_entry(struct kvm_rmap_head *rmap_head,
878 struct pte_list_desc *desc, int i,
879 struct pte_list_desc *prev_desc)
883 for (j = PTE_LIST_EXT - 1; !desc->sptes[j] && j > i; --j)
885 desc->sptes[i] = desc->sptes[j];
886 desc->sptes[j] = NULL;
889 if (!prev_desc && !desc->more)
893 prev_desc->more = desc->more;
895 rmap_head->val = (unsigned long)desc->more | 1;
896 mmu_free_pte_list_desc(desc);
899 static void __pte_list_remove(u64 *spte, struct kvm_rmap_head *rmap_head)
901 struct pte_list_desc *desc;
902 struct pte_list_desc *prev_desc;
905 if (!rmap_head->val) {
906 pr_err("%s: %p 0->BUG\n", __func__, spte);
908 } else if (!(rmap_head->val & 1)) {
909 rmap_printk("%p 1->0\n", spte);
910 if ((u64 *)rmap_head->val != spte) {
911 pr_err("%s: %p 1->BUG\n", __func__, spte);
916 rmap_printk("%p many->many\n", spte);
917 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
920 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i) {
921 if (desc->sptes[i] == spte) {
922 pte_list_desc_remove_entry(rmap_head,
930 pr_err("%s: %p many->many\n", __func__, spte);
935 static void pte_list_remove(struct kvm_rmap_head *rmap_head, u64 *sptep)
937 mmu_spte_clear_track_bits(sptep);
938 __pte_list_remove(sptep, rmap_head);
941 static struct kvm_rmap_head *__gfn_to_rmap(gfn_t gfn, int level,
942 struct kvm_memory_slot *slot)
946 idx = gfn_to_index(gfn, slot->base_gfn, level);
947 return &slot->arch.rmap[level - PG_LEVEL_4K][idx];
950 static struct kvm_rmap_head *gfn_to_rmap(struct kvm *kvm, gfn_t gfn,
951 struct kvm_mmu_page *sp)
953 struct kvm_memslots *slots;
954 struct kvm_memory_slot *slot;
956 slots = kvm_memslots_for_spte_role(kvm, sp->role);
957 slot = __gfn_to_memslot(slots, gfn);
958 return __gfn_to_rmap(gfn, sp->role.level, slot);
961 static bool rmap_can_add(struct kvm_vcpu *vcpu)
963 struct kvm_mmu_memory_cache *mc;
965 mc = &vcpu->arch.mmu_pte_list_desc_cache;
966 return kvm_mmu_memory_cache_nr_free_objects(mc);
969 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
971 struct kvm_mmu_page *sp;
972 struct kvm_rmap_head *rmap_head;
974 sp = sptep_to_sp(spte);
975 kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
976 rmap_head = gfn_to_rmap(vcpu->kvm, gfn, sp);
977 return pte_list_add(vcpu, spte, rmap_head);
980 static void rmap_remove(struct kvm *kvm, u64 *spte)
982 struct kvm_mmu_page *sp;
984 struct kvm_rmap_head *rmap_head;
986 sp = sptep_to_sp(spte);
987 gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
988 rmap_head = gfn_to_rmap(kvm, gfn, sp);
989 __pte_list_remove(spte, rmap_head);
993 * Used by the following functions to iterate through the sptes linked by a
994 * rmap. All fields are private and not assumed to be used outside.
996 struct rmap_iterator {
998 struct pte_list_desc *desc; /* holds the sptep if not NULL */
999 int pos; /* index of the sptep */
1003 * Iteration must be started by this function. This should also be used after
1004 * removing/dropping sptes from the rmap link because in such cases the
1005 * information in the iterator may not be valid.
1007 * Returns sptep if found, NULL otherwise.
1009 static u64 *rmap_get_first(struct kvm_rmap_head *rmap_head,
1010 struct rmap_iterator *iter)
1014 if (!rmap_head->val)
1017 if (!(rmap_head->val & 1)) {
1019 sptep = (u64 *)rmap_head->val;
1023 iter->desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
1025 sptep = iter->desc->sptes[iter->pos];
1027 BUG_ON(!is_shadow_present_pte(*sptep));
1032 * Must be used with a valid iterator: e.g. after rmap_get_first().
1034 * Returns sptep if found, NULL otherwise.
1036 static u64 *rmap_get_next(struct rmap_iterator *iter)
1041 if (iter->pos < PTE_LIST_EXT - 1) {
1043 sptep = iter->desc->sptes[iter->pos];
1048 iter->desc = iter->desc->more;
1052 /* desc->sptes[0] cannot be NULL */
1053 sptep = iter->desc->sptes[iter->pos];
1060 BUG_ON(!is_shadow_present_pte(*sptep));
1064 #define for_each_rmap_spte(_rmap_head_, _iter_, _spte_) \
1065 for (_spte_ = rmap_get_first(_rmap_head_, _iter_); \
1066 _spte_; _spte_ = rmap_get_next(_iter_))
1068 static void drop_spte(struct kvm *kvm, u64 *sptep)
1070 if (mmu_spte_clear_track_bits(sptep))
1071 rmap_remove(kvm, sptep);
1075 static bool __drop_large_spte(struct kvm *kvm, u64 *sptep)
1077 if (is_large_pte(*sptep)) {
1078 WARN_ON(sptep_to_sp(sptep)->role.level == PG_LEVEL_4K);
1079 drop_spte(kvm, sptep);
1087 static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1089 if (__drop_large_spte(vcpu->kvm, sptep)) {
1090 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
1092 kvm_flush_remote_tlbs_with_address(vcpu->kvm, sp->gfn,
1093 KVM_PAGES_PER_HPAGE(sp->role.level));
1098 * Write-protect on the specified @sptep, @pt_protect indicates whether
1099 * spte write-protection is caused by protecting shadow page table.
1101 * Note: write protection is difference between dirty logging and spte
1103 * - for dirty logging, the spte can be set to writable at anytime if
1104 * its dirty bitmap is properly set.
1105 * - for spte protection, the spte can be writable only after unsync-ing
1108 * Return true if tlb need be flushed.
1110 static bool spte_write_protect(u64 *sptep, bool pt_protect)
1114 if (!is_writable_pte(spte) &&
1115 !(pt_protect && spte_can_locklessly_be_made_writable(spte)))
1118 rmap_printk("spte %p %llx\n", sptep, *sptep);
1121 spte &= ~SPTE_MMU_WRITEABLE;
1122 spte = spte & ~PT_WRITABLE_MASK;
1124 return mmu_spte_update(sptep, spte);
1127 static bool __rmap_write_protect(struct kvm *kvm,
1128 struct kvm_rmap_head *rmap_head,
1132 struct rmap_iterator iter;
1135 for_each_rmap_spte(rmap_head, &iter, sptep)
1136 flush |= spte_write_protect(sptep, pt_protect);
1141 static bool spte_clear_dirty(u64 *sptep)
1145 rmap_printk("spte %p %llx\n", sptep, *sptep);
1147 MMU_WARN_ON(!spte_ad_enabled(spte));
1148 spte &= ~shadow_dirty_mask;
1149 return mmu_spte_update(sptep, spte);
1152 static bool spte_wrprot_for_clear_dirty(u64 *sptep)
1154 bool was_writable = test_and_clear_bit(PT_WRITABLE_SHIFT,
1155 (unsigned long *)sptep);
1156 if (was_writable && !spte_ad_enabled(*sptep))
1157 kvm_set_pfn_dirty(spte_to_pfn(*sptep));
1159 return was_writable;
1163 * Gets the GFN ready for another round of dirty logging by clearing the
1164 * - D bit on ad-enabled SPTEs, and
1165 * - W bit on ad-disabled SPTEs.
1166 * Returns true iff any D or W bits were cleared.
1168 static bool __rmap_clear_dirty(struct kvm *kvm, struct kvm_rmap_head *rmap_head)
1171 struct rmap_iterator iter;
1174 for_each_rmap_spte(rmap_head, &iter, sptep)
1175 if (spte_ad_need_write_protect(*sptep))
1176 flush |= spte_wrprot_for_clear_dirty(sptep);
1178 flush |= spte_clear_dirty(sptep);
1183 static bool spte_set_dirty(u64 *sptep)
1187 rmap_printk("spte %p %llx\n", sptep, *sptep);
1190 * Similar to the !kvm_x86_ops.slot_disable_log_dirty case,
1191 * do not bother adding back write access to pages marked
1192 * SPTE_AD_WRPROT_ONLY_MASK.
1194 spte |= shadow_dirty_mask;
1196 return mmu_spte_update(sptep, spte);
1199 static bool __rmap_set_dirty(struct kvm *kvm, struct kvm_rmap_head *rmap_head)
1202 struct rmap_iterator iter;
1205 for_each_rmap_spte(rmap_head, &iter, sptep)
1206 if (spte_ad_enabled(*sptep))
1207 flush |= spte_set_dirty(sptep);
1213 * kvm_mmu_write_protect_pt_masked - write protect selected PT level pages
1214 * @kvm: kvm instance
1215 * @slot: slot to protect
1216 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1217 * @mask: indicates which pages we should protect
1219 * Used when we do not need to care about huge page mappings: e.g. during dirty
1220 * logging we do not have any such mappings.
1222 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1223 struct kvm_memory_slot *slot,
1224 gfn_t gfn_offset, unsigned long mask)
1226 struct kvm_rmap_head *rmap_head;
1228 if (is_tdp_mmu_enabled(kvm))
1229 kvm_tdp_mmu_clear_dirty_pt_masked(kvm, slot,
1230 slot->base_gfn + gfn_offset, mask, true);
1232 rmap_head = __gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1234 __rmap_write_protect(kvm, rmap_head, false);
1236 /* clear the first set bit */
1242 * kvm_mmu_clear_dirty_pt_masked - clear MMU D-bit for PT level pages, or write
1243 * protect the page if the D-bit isn't supported.
1244 * @kvm: kvm instance
1245 * @slot: slot to clear D-bit
1246 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1247 * @mask: indicates which pages we should clear D-bit
1249 * Used for PML to re-log the dirty GPAs after userspace querying dirty_bitmap.
1251 void kvm_mmu_clear_dirty_pt_masked(struct kvm *kvm,
1252 struct kvm_memory_slot *slot,
1253 gfn_t gfn_offset, unsigned long mask)
1255 struct kvm_rmap_head *rmap_head;
1257 if (is_tdp_mmu_enabled(kvm))
1258 kvm_tdp_mmu_clear_dirty_pt_masked(kvm, slot,
1259 slot->base_gfn + gfn_offset, mask, false);
1261 rmap_head = __gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1263 __rmap_clear_dirty(kvm, rmap_head);
1265 /* clear the first set bit */
1269 EXPORT_SYMBOL_GPL(kvm_mmu_clear_dirty_pt_masked);
1272 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1275 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1276 * enable dirty logging for them.
1278 * Used when we do not need to care about huge page mappings: e.g. during dirty
1279 * logging we do not have any such mappings.
1281 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1282 struct kvm_memory_slot *slot,
1283 gfn_t gfn_offset, unsigned long mask)
1285 if (kvm_x86_ops.enable_log_dirty_pt_masked)
1286 static_call(kvm_x86_enable_log_dirty_pt_masked)(kvm, slot,
1290 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1293 int kvm_cpu_dirty_log_size(void)
1295 if (kvm_x86_ops.cpu_dirty_log_size)
1296 return static_call(kvm_x86_cpu_dirty_log_size)();
1301 bool kvm_mmu_slot_gfn_write_protect(struct kvm *kvm,
1302 struct kvm_memory_slot *slot, u64 gfn)
1304 struct kvm_rmap_head *rmap_head;
1306 bool write_protected = false;
1308 for (i = PG_LEVEL_4K; i <= KVM_MAX_HUGEPAGE_LEVEL; ++i) {
1309 rmap_head = __gfn_to_rmap(gfn, i, slot);
1310 write_protected |= __rmap_write_protect(kvm, rmap_head, true);
1313 if (is_tdp_mmu_enabled(kvm))
1315 kvm_tdp_mmu_write_protect_gfn(kvm, slot, gfn);
1317 return write_protected;
1320 static bool rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
1322 struct kvm_memory_slot *slot;
1324 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1325 return kvm_mmu_slot_gfn_write_protect(vcpu->kvm, slot, gfn);
1328 static bool kvm_zap_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head)
1331 struct rmap_iterator iter;
1334 while ((sptep = rmap_get_first(rmap_head, &iter))) {
1335 rmap_printk("spte %p %llx.\n", sptep, *sptep);
1337 pte_list_remove(rmap_head, sptep);
1344 static int kvm_unmap_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1345 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1348 return kvm_zap_rmapp(kvm, rmap_head);
1351 static int kvm_set_pte_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1352 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1356 struct rmap_iterator iter;
1359 pte_t *ptep = (pte_t *)data;
1362 WARN_ON(pte_huge(*ptep));
1363 new_pfn = pte_pfn(*ptep);
1366 for_each_rmap_spte(rmap_head, &iter, sptep) {
1367 rmap_printk("spte %p %llx gfn %llx (%d)\n",
1368 sptep, *sptep, gfn, level);
1372 if (pte_write(*ptep)) {
1373 pte_list_remove(rmap_head, sptep);
1376 new_spte = kvm_mmu_changed_pte_notifier_make_spte(
1379 mmu_spte_clear_track_bits(sptep);
1380 mmu_spte_set(sptep, new_spte);
1384 if (need_flush && kvm_available_flush_tlb_with_range()) {
1385 kvm_flush_remote_tlbs_with_address(kvm, gfn, 1);
1392 struct slot_rmap_walk_iterator {
1394 struct kvm_memory_slot *slot;
1400 /* output fields. */
1402 struct kvm_rmap_head *rmap;
1405 /* private field. */
1406 struct kvm_rmap_head *end_rmap;
1410 rmap_walk_init_level(struct slot_rmap_walk_iterator *iterator, int level)
1412 iterator->level = level;
1413 iterator->gfn = iterator->start_gfn;
1414 iterator->rmap = __gfn_to_rmap(iterator->gfn, level, iterator->slot);
1415 iterator->end_rmap = __gfn_to_rmap(iterator->end_gfn, level,
1420 slot_rmap_walk_init(struct slot_rmap_walk_iterator *iterator,
1421 struct kvm_memory_slot *slot, int start_level,
1422 int end_level, gfn_t start_gfn, gfn_t end_gfn)
1424 iterator->slot = slot;
1425 iterator->start_level = start_level;
1426 iterator->end_level = end_level;
1427 iterator->start_gfn = start_gfn;
1428 iterator->end_gfn = end_gfn;
1430 rmap_walk_init_level(iterator, iterator->start_level);
1433 static bool slot_rmap_walk_okay(struct slot_rmap_walk_iterator *iterator)
1435 return !!iterator->rmap;
1438 static void slot_rmap_walk_next(struct slot_rmap_walk_iterator *iterator)
1440 if (++iterator->rmap <= iterator->end_rmap) {
1441 iterator->gfn += (1UL << KVM_HPAGE_GFN_SHIFT(iterator->level));
1445 if (++iterator->level > iterator->end_level) {
1446 iterator->rmap = NULL;
1450 rmap_walk_init_level(iterator, iterator->level);
1453 #define for_each_slot_rmap_range(_slot_, _start_level_, _end_level_, \
1454 _start_gfn, _end_gfn, _iter_) \
1455 for (slot_rmap_walk_init(_iter_, _slot_, _start_level_, \
1456 _end_level_, _start_gfn, _end_gfn); \
1457 slot_rmap_walk_okay(_iter_); \
1458 slot_rmap_walk_next(_iter_))
1460 static __always_inline int
1461 kvm_handle_hva_range(struct kvm *kvm,
1462 unsigned long start,
1465 int (*handler)(struct kvm *kvm,
1466 struct kvm_rmap_head *rmap_head,
1467 struct kvm_memory_slot *slot,
1470 unsigned long data))
1472 struct kvm_memslots *slots;
1473 struct kvm_memory_slot *memslot;
1474 struct slot_rmap_walk_iterator iterator;
1478 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
1479 slots = __kvm_memslots(kvm, i);
1480 kvm_for_each_memslot(memslot, slots) {
1481 unsigned long hva_start, hva_end;
1482 gfn_t gfn_start, gfn_end;
1484 hva_start = max(start, memslot->userspace_addr);
1485 hva_end = min(end, memslot->userspace_addr +
1486 (memslot->npages << PAGE_SHIFT));
1487 if (hva_start >= hva_end)
1490 * {gfn(page) | page intersects with [hva_start, hva_end)} =
1491 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
1493 gfn_start = hva_to_gfn_memslot(hva_start, memslot);
1494 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
1496 for_each_slot_rmap_range(memslot, PG_LEVEL_4K,
1497 KVM_MAX_HUGEPAGE_LEVEL,
1498 gfn_start, gfn_end - 1,
1500 ret |= handler(kvm, iterator.rmap, memslot,
1501 iterator.gfn, iterator.level, data);
1508 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
1510 int (*handler)(struct kvm *kvm,
1511 struct kvm_rmap_head *rmap_head,
1512 struct kvm_memory_slot *slot,
1513 gfn_t gfn, int level,
1514 unsigned long data))
1516 return kvm_handle_hva_range(kvm, hva, hva + 1, data, handler);
1519 int kvm_unmap_hva_range(struct kvm *kvm, unsigned long start, unsigned long end,
1524 r = kvm_handle_hva_range(kvm, start, end, 0, kvm_unmap_rmapp);
1526 if (is_tdp_mmu_enabled(kvm))
1527 r |= kvm_tdp_mmu_zap_hva_range(kvm, start, end);
1532 int kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1536 r = kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
1538 if (is_tdp_mmu_enabled(kvm))
1539 r |= kvm_tdp_mmu_set_spte_hva(kvm, hva, &pte);
1544 static int kvm_age_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1545 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1549 struct rmap_iterator iter;
1552 for_each_rmap_spte(rmap_head, &iter, sptep)
1553 young |= mmu_spte_age(sptep);
1555 trace_kvm_age_page(gfn, level, slot, young);
1559 static int kvm_test_age_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1560 struct kvm_memory_slot *slot, gfn_t gfn,
1561 int level, unsigned long data)
1564 struct rmap_iterator iter;
1566 for_each_rmap_spte(rmap_head, &iter, sptep)
1567 if (is_accessed_spte(*sptep))
1572 #define RMAP_RECYCLE_THRESHOLD 1000
1574 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
1576 struct kvm_rmap_head *rmap_head;
1577 struct kvm_mmu_page *sp;
1579 sp = sptep_to_sp(spte);
1581 rmap_head = gfn_to_rmap(vcpu->kvm, gfn, sp);
1583 kvm_unmap_rmapp(vcpu->kvm, rmap_head, NULL, gfn, sp->role.level, 0);
1584 kvm_flush_remote_tlbs_with_address(vcpu->kvm, sp->gfn,
1585 KVM_PAGES_PER_HPAGE(sp->role.level));
1588 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1592 young = kvm_handle_hva_range(kvm, start, end, 0, kvm_age_rmapp);
1593 if (is_tdp_mmu_enabled(kvm))
1594 young |= kvm_tdp_mmu_age_hva_range(kvm, start, end);
1599 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1603 young = kvm_handle_hva(kvm, hva, 0, kvm_test_age_rmapp);
1604 if (is_tdp_mmu_enabled(kvm))
1605 young |= kvm_tdp_mmu_test_age_hva(kvm, hva);
1611 static int is_empty_shadow_page(u64 *spt)
1616 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
1617 if (is_shadow_present_pte(*pos)) {
1618 printk(KERN_ERR "%s: %p %llx\n", __func__,
1627 * This value is the sum of all of the kvm instances's
1628 * kvm->arch.n_used_mmu_pages values. We need a global,
1629 * aggregate version in order to make the slab shrinker
1632 static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, unsigned long nr)
1634 kvm->arch.n_used_mmu_pages += nr;
1635 percpu_counter_add(&kvm_total_used_mmu_pages, nr);
1638 static void kvm_mmu_free_page(struct kvm_mmu_page *sp)
1640 MMU_WARN_ON(!is_empty_shadow_page(sp->spt));
1641 hlist_del(&sp->hash_link);
1642 list_del(&sp->link);
1643 free_page((unsigned long)sp->spt);
1644 if (!sp->role.direct)
1645 free_page((unsigned long)sp->gfns);
1646 kmem_cache_free(mmu_page_header_cache, sp);
1649 static unsigned kvm_page_table_hashfn(gfn_t gfn)
1651 return hash_64(gfn, KVM_MMU_HASH_SHIFT);
1654 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
1655 struct kvm_mmu_page *sp, u64 *parent_pte)
1660 pte_list_add(vcpu, parent_pte, &sp->parent_ptes);
1663 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
1666 __pte_list_remove(parent_pte, &sp->parent_ptes);
1669 static void drop_parent_pte(struct kvm_mmu_page *sp,
1672 mmu_page_remove_parent_pte(sp, parent_pte);
1673 mmu_spte_clear_no_track(parent_pte);
1676 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu, int direct)
1678 struct kvm_mmu_page *sp;
1680 sp = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache);
1681 sp->spt = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_shadow_page_cache);
1683 sp->gfns = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_gfn_array_cache);
1684 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
1687 * active_mmu_pages must be a FIFO list, as kvm_zap_obsolete_pages()
1688 * depends on valid pages being added to the head of the list. See
1689 * comments in kvm_zap_obsolete_pages().
1691 sp->mmu_valid_gen = vcpu->kvm->arch.mmu_valid_gen;
1692 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
1693 kvm_mod_used_mmu_pages(vcpu->kvm, +1);
1697 static void mark_unsync(u64 *spte);
1698 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1701 struct rmap_iterator iter;
1703 for_each_rmap_spte(&sp->parent_ptes, &iter, sptep) {
1708 static void mark_unsync(u64 *spte)
1710 struct kvm_mmu_page *sp;
1713 sp = sptep_to_sp(spte);
1714 index = spte - sp->spt;
1715 if (__test_and_set_bit(index, sp->unsync_child_bitmap))
1717 if (sp->unsync_children++)
1719 kvm_mmu_mark_parents_unsync(sp);
1722 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1723 struct kvm_mmu_page *sp)
1728 #define KVM_PAGE_ARRAY_NR 16
1730 struct kvm_mmu_pages {
1731 struct mmu_page_and_offset {
1732 struct kvm_mmu_page *sp;
1734 } page[KVM_PAGE_ARRAY_NR];
1738 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1744 for (i=0; i < pvec->nr; i++)
1745 if (pvec->page[i].sp == sp)
1748 pvec->page[pvec->nr].sp = sp;
1749 pvec->page[pvec->nr].idx = idx;
1751 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1754 static inline void clear_unsync_child_bit(struct kvm_mmu_page *sp, int idx)
1756 --sp->unsync_children;
1757 WARN_ON((int)sp->unsync_children < 0);
1758 __clear_bit(idx, sp->unsync_child_bitmap);
1761 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1762 struct kvm_mmu_pages *pvec)
1764 int i, ret, nr_unsync_leaf = 0;
1766 for_each_set_bit(i, sp->unsync_child_bitmap, 512) {
1767 struct kvm_mmu_page *child;
1768 u64 ent = sp->spt[i];
1770 if (!is_shadow_present_pte(ent) || is_large_pte(ent)) {
1771 clear_unsync_child_bit(sp, i);
1775 child = to_shadow_page(ent & PT64_BASE_ADDR_MASK);
1777 if (child->unsync_children) {
1778 if (mmu_pages_add(pvec, child, i))
1781 ret = __mmu_unsync_walk(child, pvec);
1783 clear_unsync_child_bit(sp, i);
1785 } else if (ret > 0) {
1786 nr_unsync_leaf += ret;
1789 } else if (child->unsync) {
1791 if (mmu_pages_add(pvec, child, i))
1794 clear_unsync_child_bit(sp, i);
1797 return nr_unsync_leaf;
1800 #define INVALID_INDEX (-1)
1802 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1803 struct kvm_mmu_pages *pvec)
1806 if (!sp->unsync_children)
1809 mmu_pages_add(pvec, sp, INVALID_INDEX);
1810 return __mmu_unsync_walk(sp, pvec);
1813 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1815 WARN_ON(!sp->unsync);
1816 trace_kvm_mmu_sync_page(sp);
1818 --kvm->stat.mmu_unsync;
1821 static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1822 struct list_head *invalid_list);
1823 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1824 struct list_head *invalid_list);
1826 #define for_each_valid_sp(_kvm, _sp, _list) \
1827 hlist_for_each_entry(_sp, _list, hash_link) \
1828 if (is_obsolete_sp((_kvm), (_sp))) { \
1831 #define for_each_gfn_indirect_valid_sp(_kvm, _sp, _gfn) \
1832 for_each_valid_sp(_kvm, _sp, \
1833 &(_kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(_gfn)]) \
1834 if ((_sp)->gfn != (_gfn) || (_sp)->role.direct) {} else
1836 static inline bool is_ept_sp(struct kvm_mmu_page *sp)
1838 return sp->role.cr0_wp && sp->role.smap_andnot_wp;
1841 /* @sp->gfn should be write-protected at the call site */
1842 static bool __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1843 struct list_head *invalid_list)
1845 if ((!is_ept_sp(sp) && sp->role.gpte_is_8_bytes != !!is_pae(vcpu)) ||
1846 vcpu->arch.mmu->sync_page(vcpu, sp) == 0) {
1847 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1854 static bool kvm_mmu_remote_flush_or_zap(struct kvm *kvm,
1855 struct list_head *invalid_list,
1858 if (!remote_flush && list_empty(invalid_list))
1861 if (!list_empty(invalid_list))
1862 kvm_mmu_commit_zap_page(kvm, invalid_list);
1864 kvm_flush_remote_tlbs(kvm);
1868 static void kvm_mmu_flush_or_zap(struct kvm_vcpu *vcpu,
1869 struct list_head *invalid_list,
1870 bool remote_flush, bool local_flush)
1872 if (kvm_mmu_remote_flush_or_zap(vcpu->kvm, invalid_list, remote_flush))
1876 kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
1879 #ifdef CONFIG_KVM_MMU_AUDIT
1880 #include "mmu_audit.c"
1882 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, int point) { }
1883 static void mmu_audit_disable(void) { }
1886 static bool is_obsolete_sp(struct kvm *kvm, struct kvm_mmu_page *sp)
1888 return sp->role.invalid ||
1889 unlikely(sp->mmu_valid_gen != kvm->arch.mmu_valid_gen);
1892 static bool kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1893 struct list_head *invalid_list)
1895 kvm_unlink_unsync_page(vcpu->kvm, sp);
1896 return __kvm_sync_page(vcpu, sp, invalid_list);
1899 /* @gfn should be write-protected at the call site */
1900 static bool kvm_sync_pages(struct kvm_vcpu *vcpu, gfn_t gfn,
1901 struct list_head *invalid_list)
1903 struct kvm_mmu_page *s;
1906 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn) {
1910 WARN_ON(s->role.level != PG_LEVEL_4K);
1911 ret |= kvm_sync_page(vcpu, s, invalid_list);
1917 struct mmu_page_path {
1918 struct kvm_mmu_page *parent[PT64_ROOT_MAX_LEVEL];
1919 unsigned int idx[PT64_ROOT_MAX_LEVEL];
1922 #define for_each_sp(pvec, sp, parents, i) \
1923 for (i = mmu_pages_first(&pvec, &parents); \
1924 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1925 i = mmu_pages_next(&pvec, &parents, i))
1927 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1928 struct mmu_page_path *parents,
1933 for (n = i+1; n < pvec->nr; n++) {
1934 struct kvm_mmu_page *sp = pvec->page[n].sp;
1935 unsigned idx = pvec->page[n].idx;
1936 int level = sp->role.level;
1938 parents->idx[level-1] = idx;
1939 if (level == PG_LEVEL_4K)
1942 parents->parent[level-2] = sp;
1948 static int mmu_pages_first(struct kvm_mmu_pages *pvec,
1949 struct mmu_page_path *parents)
1951 struct kvm_mmu_page *sp;
1957 WARN_ON(pvec->page[0].idx != INVALID_INDEX);
1959 sp = pvec->page[0].sp;
1960 level = sp->role.level;
1961 WARN_ON(level == PG_LEVEL_4K);
1963 parents->parent[level-2] = sp;
1965 /* Also set up a sentinel. Further entries in pvec are all
1966 * children of sp, so this element is never overwritten.
1968 parents->parent[level-1] = NULL;
1969 return mmu_pages_next(pvec, parents, 0);
1972 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1974 struct kvm_mmu_page *sp;
1975 unsigned int level = 0;
1978 unsigned int idx = parents->idx[level];
1979 sp = parents->parent[level];
1983 WARN_ON(idx == INVALID_INDEX);
1984 clear_unsync_child_bit(sp, idx);
1986 } while (!sp->unsync_children);
1989 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1990 struct kvm_mmu_page *parent)
1993 struct kvm_mmu_page *sp;
1994 struct mmu_page_path parents;
1995 struct kvm_mmu_pages pages;
1996 LIST_HEAD(invalid_list);
1999 while (mmu_unsync_walk(parent, &pages)) {
2000 bool protected = false;
2002 for_each_sp(pages, sp, parents, i)
2003 protected |= rmap_write_protect(vcpu, sp->gfn);
2006 kvm_flush_remote_tlbs(vcpu->kvm);
2010 for_each_sp(pages, sp, parents, i) {
2011 flush |= kvm_sync_page(vcpu, sp, &invalid_list);
2012 mmu_pages_clear_parents(&parents);
2014 if (need_resched() || rwlock_needbreak(&vcpu->kvm->mmu_lock)) {
2015 kvm_mmu_flush_or_zap(vcpu, &invalid_list, false, flush);
2016 cond_resched_rwlock_write(&vcpu->kvm->mmu_lock);
2021 kvm_mmu_flush_or_zap(vcpu, &invalid_list, false, flush);
2024 static void __clear_sp_write_flooding_count(struct kvm_mmu_page *sp)
2026 atomic_set(&sp->write_flooding_count, 0);
2029 static void clear_sp_write_flooding_count(u64 *spte)
2031 __clear_sp_write_flooding_count(sptep_to_sp(spte));
2034 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
2039 unsigned int access)
2041 bool direct_mmu = vcpu->arch.mmu->direct_map;
2042 union kvm_mmu_page_role role;
2043 struct hlist_head *sp_list;
2045 struct kvm_mmu_page *sp;
2046 bool need_sync = false;
2049 LIST_HEAD(invalid_list);
2051 role = vcpu->arch.mmu->mmu_role.base;
2053 role.direct = direct;
2055 role.gpte_is_8_bytes = true;
2056 role.access = access;
2057 if (!direct_mmu && vcpu->arch.mmu->root_level <= PT32_ROOT_LEVEL) {
2058 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
2059 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
2060 role.quadrant = quadrant;
2063 sp_list = &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)];
2064 for_each_valid_sp(vcpu->kvm, sp, sp_list) {
2065 if (sp->gfn != gfn) {
2070 if (!need_sync && sp->unsync)
2073 if (sp->role.word != role.word)
2077 goto trace_get_page;
2080 /* The page is good, but __kvm_sync_page might still end
2081 * up zapping it. If so, break in order to rebuild it.
2083 if (!__kvm_sync_page(vcpu, sp, &invalid_list))
2086 WARN_ON(!list_empty(&invalid_list));
2087 kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
2090 if (sp->unsync_children)
2091 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
2093 __clear_sp_write_flooding_count(sp);
2096 trace_kvm_mmu_get_page(sp, false);
2100 ++vcpu->kvm->stat.mmu_cache_miss;
2102 sp = kvm_mmu_alloc_page(vcpu, direct);
2106 hlist_add_head(&sp->hash_link, sp_list);
2109 * we should do write protection before syncing pages
2110 * otherwise the content of the synced shadow page may
2111 * be inconsistent with guest page table.
2113 account_shadowed(vcpu->kvm, sp);
2114 if (level == PG_LEVEL_4K && rmap_write_protect(vcpu, gfn))
2115 kvm_flush_remote_tlbs_with_address(vcpu->kvm, gfn, 1);
2117 if (level > PG_LEVEL_4K && need_sync)
2118 flush |= kvm_sync_pages(vcpu, gfn, &invalid_list);
2120 trace_kvm_mmu_get_page(sp, true);
2122 kvm_mmu_flush_or_zap(vcpu, &invalid_list, false, flush);
2124 if (collisions > vcpu->kvm->stat.max_mmu_page_hash_collisions)
2125 vcpu->kvm->stat.max_mmu_page_hash_collisions = collisions;
2129 static void shadow_walk_init_using_root(struct kvm_shadow_walk_iterator *iterator,
2130 struct kvm_vcpu *vcpu, hpa_t root,
2133 iterator->addr = addr;
2134 iterator->shadow_addr = root;
2135 iterator->level = vcpu->arch.mmu->shadow_root_level;
2137 if (iterator->level == PT64_ROOT_4LEVEL &&
2138 vcpu->arch.mmu->root_level < PT64_ROOT_4LEVEL &&
2139 !vcpu->arch.mmu->direct_map)
2142 if (iterator->level == PT32E_ROOT_LEVEL) {
2144 * prev_root is currently only used for 64-bit hosts. So only
2145 * the active root_hpa is valid here.
2147 BUG_ON(root != vcpu->arch.mmu->root_hpa);
2149 iterator->shadow_addr
2150 = vcpu->arch.mmu->pae_root[(addr >> 30) & 3];
2151 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
2153 if (!iterator->shadow_addr)
2154 iterator->level = 0;
2158 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
2159 struct kvm_vcpu *vcpu, u64 addr)
2161 shadow_walk_init_using_root(iterator, vcpu, vcpu->arch.mmu->root_hpa,
2165 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
2167 if (iterator->level < PG_LEVEL_4K)
2170 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
2171 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
2175 static void __shadow_walk_next(struct kvm_shadow_walk_iterator *iterator,
2178 if (is_last_spte(spte, iterator->level)) {
2179 iterator->level = 0;
2183 iterator->shadow_addr = spte & PT64_BASE_ADDR_MASK;
2187 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
2189 __shadow_walk_next(iterator, *iterator->sptep);
2192 static void link_shadow_page(struct kvm_vcpu *vcpu, u64 *sptep,
2193 struct kvm_mmu_page *sp)
2197 BUILD_BUG_ON(VMX_EPT_WRITABLE_MASK != PT_WRITABLE_MASK);
2199 spte = make_nonleaf_spte(sp->spt, sp_ad_disabled(sp));
2201 mmu_spte_set(sptep, spte);
2203 mmu_page_add_parent_pte(vcpu, sp, sptep);
2205 if (sp->unsync_children || sp->unsync)
2209 static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2210 unsigned direct_access)
2212 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
2213 struct kvm_mmu_page *child;
2216 * For the direct sp, if the guest pte's dirty bit
2217 * changed form clean to dirty, it will corrupt the
2218 * sp's access: allow writable in the read-only sp,
2219 * so we should update the spte at this point to get
2220 * a new sp with the correct access.
2222 child = to_shadow_page(*sptep & PT64_BASE_ADDR_MASK);
2223 if (child->role.access == direct_access)
2226 drop_parent_pte(child, sptep);
2227 kvm_flush_remote_tlbs_with_address(vcpu->kvm, child->gfn, 1);
2231 /* Returns the number of zapped non-leaf child shadow pages. */
2232 static int mmu_page_zap_pte(struct kvm *kvm, struct kvm_mmu_page *sp,
2233 u64 *spte, struct list_head *invalid_list)
2236 struct kvm_mmu_page *child;
2239 if (is_shadow_present_pte(pte)) {
2240 if (is_last_spte(pte, sp->role.level)) {
2241 drop_spte(kvm, spte);
2242 if (is_large_pte(pte))
2245 child = to_shadow_page(pte & PT64_BASE_ADDR_MASK);
2246 drop_parent_pte(child, spte);
2249 * Recursively zap nested TDP SPs, parentless SPs are
2250 * unlikely to be used again in the near future. This
2251 * avoids retaining a large number of stale nested SPs.
2253 if (tdp_enabled && invalid_list &&
2254 child->role.guest_mode && !child->parent_ptes.val)
2255 return kvm_mmu_prepare_zap_page(kvm, child,
2258 } else if (is_mmio_spte(pte)) {
2259 mmu_spte_clear_no_track(spte);
2264 static int kvm_mmu_page_unlink_children(struct kvm *kvm,
2265 struct kvm_mmu_page *sp,
2266 struct list_head *invalid_list)
2271 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2272 zapped += mmu_page_zap_pte(kvm, sp, sp->spt + i, invalid_list);
2277 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
2280 struct rmap_iterator iter;
2282 while ((sptep = rmap_get_first(&sp->parent_ptes, &iter)))
2283 drop_parent_pte(sp, sptep);
2286 static int mmu_zap_unsync_children(struct kvm *kvm,
2287 struct kvm_mmu_page *parent,
2288 struct list_head *invalid_list)
2291 struct mmu_page_path parents;
2292 struct kvm_mmu_pages pages;
2294 if (parent->role.level == PG_LEVEL_4K)
2297 while (mmu_unsync_walk(parent, &pages)) {
2298 struct kvm_mmu_page *sp;
2300 for_each_sp(pages, sp, parents, i) {
2301 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
2302 mmu_pages_clear_parents(&parents);
2310 static bool __kvm_mmu_prepare_zap_page(struct kvm *kvm,
2311 struct kvm_mmu_page *sp,
2312 struct list_head *invalid_list,
2317 trace_kvm_mmu_prepare_zap_page(sp);
2318 ++kvm->stat.mmu_shadow_zapped;
2319 *nr_zapped = mmu_zap_unsync_children(kvm, sp, invalid_list);
2320 *nr_zapped += kvm_mmu_page_unlink_children(kvm, sp, invalid_list);
2321 kvm_mmu_unlink_parents(kvm, sp);
2323 /* Zapping children means active_mmu_pages has become unstable. */
2324 list_unstable = *nr_zapped;
2326 if (!sp->role.invalid && !sp->role.direct)
2327 unaccount_shadowed(kvm, sp);
2330 kvm_unlink_unsync_page(kvm, sp);
2331 if (!sp->root_count) {
2336 * Already invalid pages (previously active roots) are not on
2337 * the active page list. See list_del() in the "else" case of
2340 if (sp->role.invalid)
2341 list_add(&sp->link, invalid_list);
2343 list_move(&sp->link, invalid_list);
2344 kvm_mod_used_mmu_pages(kvm, -1);
2347 * Remove the active root from the active page list, the root
2348 * will be explicitly freed when the root_count hits zero.
2350 list_del(&sp->link);
2353 * Obsolete pages cannot be used on any vCPUs, see the comment
2354 * in kvm_mmu_zap_all_fast(). Note, is_obsolete_sp() also
2355 * treats invalid shadow pages as being obsolete.
2357 if (!is_obsolete_sp(kvm, sp))
2358 kvm_reload_remote_mmus(kvm);
2361 if (sp->lpage_disallowed)
2362 unaccount_huge_nx_page(kvm, sp);
2364 sp->role.invalid = 1;
2365 return list_unstable;
2368 static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
2369 struct list_head *invalid_list)
2373 __kvm_mmu_prepare_zap_page(kvm, sp, invalid_list, &nr_zapped);
2377 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
2378 struct list_head *invalid_list)
2380 struct kvm_mmu_page *sp, *nsp;
2382 if (list_empty(invalid_list))
2386 * We need to make sure everyone sees our modifications to
2387 * the page tables and see changes to vcpu->mode here. The barrier
2388 * in the kvm_flush_remote_tlbs() achieves this. This pairs
2389 * with vcpu_enter_guest and walk_shadow_page_lockless_begin/end.
2391 * In addition, kvm_flush_remote_tlbs waits for all vcpus to exit
2392 * guest mode and/or lockless shadow page table walks.
2394 kvm_flush_remote_tlbs(kvm);
2396 list_for_each_entry_safe(sp, nsp, invalid_list, link) {
2397 WARN_ON(!sp->role.invalid || sp->root_count);
2398 kvm_mmu_free_page(sp);
2402 static unsigned long kvm_mmu_zap_oldest_mmu_pages(struct kvm *kvm,
2403 unsigned long nr_to_zap)
2405 unsigned long total_zapped = 0;
2406 struct kvm_mmu_page *sp, *tmp;
2407 LIST_HEAD(invalid_list);
2411 if (list_empty(&kvm->arch.active_mmu_pages))
2415 list_for_each_entry_safe_reverse(sp, tmp, &kvm->arch.active_mmu_pages, link) {
2417 * Don't zap active root pages, the page itself can't be freed
2418 * and zapping it will just force vCPUs to realloc and reload.
2423 unstable = __kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list,
2425 total_zapped += nr_zapped;
2426 if (total_zapped >= nr_to_zap)
2433 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2435 kvm->stat.mmu_recycled += total_zapped;
2436 return total_zapped;
2439 static inline unsigned long kvm_mmu_available_pages(struct kvm *kvm)
2441 if (kvm->arch.n_max_mmu_pages > kvm->arch.n_used_mmu_pages)
2442 return kvm->arch.n_max_mmu_pages -
2443 kvm->arch.n_used_mmu_pages;
2448 static int make_mmu_pages_available(struct kvm_vcpu *vcpu)
2450 unsigned long avail = kvm_mmu_available_pages(vcpu->kvm);
2452 if (likely(avail >= KVM_MIN_FREE_MMU_PAGES))
2455 kvm_mmu_zap_oldest_mmu_pages(vcpu->kvm, KVM_REFILL_PAGES - avail);
2457 if (!kvm_mmu_available_pages(vcpu->kvm))
2463 * Changing the number of mmu pages allocated to the vm
2464 * Note: if goal_nr_mmu_pages is too small, you will get dead lock
2466 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned long goal_nr_mmu_pages)
2468 write_lock(&kvm->mmu_lock);
2470 if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) {
2471 kvm_mmu_zap_oldest_mmu_pages(kvm, kvm->arch.n_used_mmu_pages -
2474 goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages;
2477 kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages;
2479 write_unlock(&kvm->mmu_lock);
2482 int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
2484 struct kvm_mmu_page *sp;
2485 LIST_HEAD(invalid_list);
2488 pgprintk("%s: looking for gfn %llx\n", __func__, gfn);
2490 write_lock(&kvm->mmu_lock);
2491 for_each_gfn_indirect_valid_sp(kvm, sp, gfn) {
2492 pgprintk("%s: gfn %llx role %x\n", __func__, gfn,
2495 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
2497 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2498 write_unlock(&kvm->mmu_lock);
2502 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page);
2504 static void kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
2506 trace_kvm_mmu_unsync_page(sp);
2507 ++vcpu->kvm->stat.mmu_unsync;
2510 kvm_mmu_mark_parents_unsync(sp);
2513 bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
2516 struct kvm_mmu_page *sp;
2518 if (kvm_page_track_is_active(vcpu, gfn, KVM_PAGE_TRACK_WRITE))
2521 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn) {
2528 WARN_ON(sp->role.level != PG_LEVEL_4K);
2529 kvm_unsync_page(vcpu, sp);
2533 * We need to ensure that the marking of unsync pages is visible
2534 * before the SPTE is updated to allow writes because
2535 * kvm_mmu_sync_roots() checks the unsync flags without holding
2536 * the MMU lock and so can race with this. If the SPTE was updated
2537 * before the page had been marked as unsync-ed, something like the
2538 * following could happen:
2541 * ---------------------------------------------------------------------
2542 * 1.2 Host updates SPTE
2544 * 2.1 Guest writes a GPTE for GVA X.
2545 * (GPTE being in the guest page table shadowed
2546 * by the SP from CPU 1.)
2547 * This reads SPTE during the page table walk.
2548 * Since SPTE.W is read as 1, there is no
2551 * 2.2 Guest issues TLB flush.
2552 * That causes a VM Exit.
2554 * 2.3 kvm_mmu_sync_pages() reads sp->unsync.
2555 * Since it is false, so it just returns.
2557 * 2.4 Guest accesses GVA X.
2558 * Since the mapping in the SP was not updated,
2559 * so the old mapping for GVA X incorrectly
2563 * (sp->unsync = true)
2565 * The write barrier below ensures that 1.1 happens before 1.2 and thus
2566 * the situation in 2.4 does not arise. The implicit barrier in 2.2
2567 * pairs with this write barrier.
2574 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2575 unsigned int pte_access, int level,
2576 gfn_t gfn, kvm_pfn_t pfn, bool speculative,
2577 bool can_unsync, bool host_writable)
2580 struct kvm_mmu_page *sp;
2583 if (set_mmio_spte(vcpu, sptep, gfn, pfn, pte_access))
2586 sp = sptep_to_sp(sptep);
2588 ret = make_spte(vcpu, pte_access, level, gfn, pfn, *sptep, speculative,
2589 can_unsync, host_writable, sp_ad_disabled(sp), &spte);
2591 if (spte & PT_WRITABLE_MASK)
2592 kvm_vcpu_mark_page_dirty(vcpu, gfn);
2595 ret |= SET_SPTE_SPURIOUS;
2596 else if (mmu_spte_update(sptep, spte))
2597 ret |= SET_SPTE_NEED_REMOTE_TLB_FLUSH;
2601 static int mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2602 unsigned int pte_access, bool write_fault, int level,
2603 gfn_t gfn, kvm_pfn_t pfn, bool speculative,
2606 int was_rmapped = 0;
2609 int ret = RET_PF_FIXED;
2612 pgprintk("%s: spte %llx write_fault %d gfn %llx\n", __func__,
2613 *sptep, write_fault, gfn);
2615 if (is_shadow_present_pte(*sptep)) {
2617 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
2618 * the parent of the now unreachable PTE.
2620 if (level > PG_LEVEL_4K && !is_large_pte(*sptep)) {
2621 struct kvm_mmu_page *child;
2624 child = to_shadow_page(pte & PT64_BASE_ADDR_MASK);
2625 drop_parent_pte(child, sptep);
2627 } else if (pfn != spte_to_pfn(*sptep)) {
2628 pgprintk("hfn old %llx new %llx\n",
2629 spte_to_pfn(*sptep), pfn);
2630 drop_spte(vcpu->kvm, sptep);
2636 set_spte_ret = set_spte(vcpu, sptep, pte_access, level, gfn, pfn,
2637 speculative, true, host_writable);
2638 if (set_spte_ret & SET_SPTE_WRITE_PROTECTED_PT) {
2640 ret = RET_PF_EMULATE;
2641 kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
2644 if (set_spte_ret & SET_SPTE_NEED_REMOTE_TLB_FLUSH || flush)
2645 kvm_flush_remote_tlbs_with_address(vcpu->kvm, gfn,
2646 KVM_PAGES_PER_HPAGE(level));
2648 if (unlikely(is_mmio_spte(*sptep)))
2649 ret = RET_PF_EMULATE;
2652 * The fault is fully spurious if and only if the new SPTE and old SPTE
2653 * are identical, and emulation is not required.
2655 if ((set_spte_ret & SET_SPTE_SPURIOUS) && ret == RET_PF_FIXED) {
2656 WARN_ON_ONCE(!was_rmapped);
2657 return RET_PF_SPURIOUS;
2660 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
2661 trace_kvm_mmu_set_spte(level, gfn, sptep);
2662 if (!was_rmapped && is_large_pte(*sptep))
2663 ++vcpu->kvm->stat.lpages;
2665 if (is_shadow_present_pte(*sptep)) {
2667 rmap_count = rmap_add(vcpu, sptep, gfn);
2668 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
2669 rmap_recycle(vcpu, sptep, gfn);
2676 static kvm_pfn_t pte_prefetch_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn,
2679 struct kvm_memory_slot *slot;
2681 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, no_dirty_log);
2683 return KVM_PFN_ERR_FAULT;
2685 return gfn_to_pfn_memslot_atomic(slot, gfn);
2688 static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu,
2689 struct kvm_mmu_page *sp,
2690 u64 *start, u64 *end)
2692 struct page *pages[PTE_PREFETCH_NUM];
2693 struct kvm_memory_slot *slot;
2694 unsigned int access = sp->role.access;
2698 gfn = kvm_mmu_page_get_gfn(sp, start - sp->spt);
2699 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, access & ACC_WRITE_MASK);
2703 ret = gfn_to_page_many_atomic(slot, gfn, pages, end - start);
2707 for (i = 0; i < ret; i++, gfn++, start++) {
2708 mmu_set_spte(vcpu, start, access, false, sp->role.level, gfn,
2709 page_to_pfn(pages[i]), true, true);
2716 static void __direct_pte_prefetch(struct kvm_vcpu *vcpu,
2717 struct kvm_mmu_page *sp, u64 *sptep)
2719 u64 *spte, *start = NULL;
2722 WARN_ON(!sp->role.direct);
2724 i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
2727 for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
2728 if (is_shadow_present_pte(*spte) || spte == sptep) {
2731 if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0)
2739 static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
2741 struct kvm_mmu_page *sp;
2743 sp = sptep_to_sp(sptep);
2746 * Without accessed bits, there's no way to distinguish between
2747 * actually accessed translations and prefetched, so disable pte
2748 * prefetch if accessed bits aren't available.
2750 if (sp_ad_disabled(sp))
2753 if (sp->role.level > PG_LEVEL_4K)
2756 __direct_pte_prefetch(vcpu, sp, sptep);
2759 static int host_pfn_mapping_level(struct kvm_vcpu *vcpu, gfn_t gfn,
2760 kvm_pfn_t pfn, struct kvm_memory_slot *slot)
2766 if (!PageCompound(pfn_to_page(pfn)) && !kvm_is_zone_device_pfn(pfn))
2770 * Note, using the already-retrieved memslot and __gfn_to_hva_memslot()
2771 * is not solely for performance, it's also necessary to avoid the
2772 * "writable" check in __gfn_to_hva_many(), which will always fail on
2773 * read-only memslots due to gfn_to_hva() assuming writes. Earlier
2774 * page fault steps have already verified the guest isn't writing a
2775 * read-only memslot.
2777 hva = __gfn_to_hva_memslot(slot, gfn);
2779 pte = lookup_address_in_mm(vcpu->kvm->mm, hva, &level);
2786 int kvm_mmu_hugepage_adjust(struct kvm_vcpu *vcpu, gfn_t gfn,
2787 int max_level, kvm_pfn_t *pfnp,
2788 bool huge_page_disallowed, int *req_level)
2790 struct kvm_memory_slot *slot;
2791 struct kvm_lpage_info *linfo;
2792 kvm_pfn_t pfn = *pfnp;
2796 *req_level = PG_LEVEL_4K;
2798 if (unlikely(max_level == PG_LEVEL_4K))
2801 if (is_error_noslot_pfn(pfn) || kvm_is_reserved_pfn(pfn))
2804 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, true);
2808 max_level = min(max_level, max_huge_page_level);
2809 for ( ; max_level > PG_LEVEL_4K; max_level--) {
2810 linfo = lpage_info_slot(gfn, slot, max_level);
2811 if (!linfo->disallow_lpage)
2815 if (max_level == PG_LEVEL_4K)
2818 level = host_pfn_mapping_level(vcpu, gfn, pfn, slot);
2819 if (level == PG_LEVEL_4K)
2822 *req_level = level = min(level, max_level);
2825 * Enforce the iTLB multihit workaround after capturing the requested
2826 * level, which will be used to do precise, accurate accounting.
2828 if (huge_page_disallowed)
2832 * mmu_notifier_retry() was successful and mmu_lock is held, so
2833 * the pmd can't be split from under us.
2835 mask = KVM_PAGES_PER_HPAGE(level) - 1;
2836 VM_BUG_ON((gfn & mask) != (pfn & mask));
2837 *pfnp = pfn & ~mask;
2842 void disallowed_hugepage_adjust(u64 spte, gfn_t gfn, int cur_level,
2843 kvm_pfn_t *pfnp, int *goal_levelp)
2845 int level = *goal_levelp;
2847 if (cur_level == level && level > PG_LEVEL_4K &&
2848 is_shadow_present_pte(spte) &&
2849 !is_large_pte(spte)) {
2851 * A small SPTE exists for this pfn, but FNAME(fetch)
2852 * and __direct_map would like to create a large PTE
2853 * instead: just force them to go down another level,
2854 * patching back for them into pfn the next 9 bits of
2857 u64 page_mask = KVM_PAGES_PER_HPAGE(level) -
2858 KVM_PAGES_PER_HPAGE(level - 1);
2859 *pfnp |= gfn & page_mask;
2864 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
2865 int map_writable, int max_level, kvm_pfn_t pfn,
2866 bool prefault, bool is_tdp)
2868 bool nx_huge_page_workaround_enabled = is_nx_huge_page_enabled();
2869 bool write = error_code & PFERR_WRITE_MASK;
2870 bool exec = error_code & PFERR_FETCH_MASK;
2871 bool huge_page_disallowed = exec && nx_huge_page_workaround_enabled;
2872 struct kvm_shadow_walk_iterator it;
2873 struct kvm_mmu_page *sp;
2874 int level, req_level, ret;
2875 gfn_t gfn = gpa >> PAGE_SHIFT;
2876 gfn_t base_gfn = gfn;
2878 if (WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root_hpa)))
2879 return RET_PF_RETRY;
2881 level = kvm_mmu_hugepage_adjust(vcpu, gfn, max_level, &pfn,
2882 huge_page_disallowed, &req_level);
2884 trace_kvm_mmu_spte_requested(gpa, level, pfn);
2885 for_each_shadow_entry(vcpu, gpa, it) {
2887 * We cannot overwrite existing page tables with an NX
2888 * large page, as the leaf could be executable.
2890 if (nx_huge_page_workaround_enabled)
2891 disallowed_hugepage_adjust(*it.sptep, gfn, it.level,
2894 base_gfn = gfn & ~(KVM_PAGES_PER_HPAGE(it.level) - 1);
2895 if (it.level == level)
2898 drop_large_spte(vcpu, it.sptep);
2899 if (!is_shadow_present_pte(*it.sptep)) {
2900 sp = kvm_mmu_get_page(vcpu, base_gfn, it.addr,
2901 it.level - 1, true, ACC_ALL);
2903 link_shadow_page(vcpu, it.sptep, sp);
2904 if (is_tdp && huge_page_disallowed &&
2905 req_level >= it.level)
2906 account_huge_nx_page(vcpu->kvm, sp);
2910 ret = mmu_set_spte(vcpu, it.sptep, ACC_ALL,
2911 write, level, base_gfn, pfn, prefault,
2913 if (ret == RET_PF_SPURIOUS)
2916 direct_pte_prefetch(vcpu, it.sptep);
2917 ++vcpu->stat.pf_fixed;
2921 static void kvm_send_hwpoison_signal(unsigned long address, struct task_struct *tsk)
2923 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, PAGE_SHIFT, tsk);
2926 static int kvm_handle_bad_page(struct kvm_vcpu *vcpu, gfn_t gfn, kvm_pfn_t pfn)
2929 * Do not cache the mmio info caused by writing the readonly gfn
2930 * into the spte otherwise read access on readonly gfn also can
2931 * caused mmio page fault and treat it as mmio access.
2933 if (pfn == KVM_PFN_ERR_RO_FAULT)
2934 return RET_PF_EMULATE;
2936 if (pfn == KVM_PFN_ERR_HWPOISON) {
2937 kvm_send_hwpoison_signal(kvm_vcpu_gfn_to_hva(vcpu, gfn), current);
2938 return RET_PF_RETRY;
2944 static bool handle_abnormal_pfn(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn,
2945 kvm_pfn_t pfn, unsigned int access,
2948 /* The pfn is invalid, report the error! */
2949 if (unlikely(is_error_pfn(pfn))) {
2950 *ret_val = kvm_handle_bad_page(vcpu, gfn, pfn);
2954 if (unlikely(is_noslot_pfn(pfn)))
2955 vcpu_cache_mmio_info(vcpu, gva, gfn,
2956 access & shadow_mmio_access_mask);
2961 static bool page_fault_can_be_fast(u32 error_code)
2964 * Do not fix the mmio spte with invalid generation number which
2965 * need to be updated by slow page fault path.
2967 if (unlikely(error_code & PFERR_RSVD_MASK))
2970 /* See if the page fault is due to an NX violation */
2971 if (unlikely(((error_code & (PFERR_FETCH_MASK | PFERR_PRESENT_MASK))
2972 == (PFERR_FETCH_MASK | PFERR_PRESENT_MASK))))
2976 * #PF can be fast if:
2977 * 1. The shadow page table entry is not present, which could mean that
2978 * the fault is potentially caused by access tracking (if enabled).
2979 * 2. The shadow page table entry is present and the fault
2980 * is caused by write-protect, that means we just need change the W
2981 * bit of the spte which can be done out of mmu-lock.
2983 * However, if access tracking is disabled we know that a non-present
2984 * page must be a genuine page fault where we have to create a new SPTE.
2985 * So, if access tracking is disabled, we return true only for write
2986 * accesses to a present page.
2989 return shadow_acc_track_mask != 0 ||
2990 ((error_code & (PFERR_WRITE_MASK | PFERR_PRESENT_MASK))
2991 == (PFERR_WRITE_MASK | PFERR_PRESENT_MASK));
2995 * Returns true if the SPTE was fixed successfully. Otherwise,
2996 * someone else modified the SPTE from its original value.
2999 fast_pf_fix_direct_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
3000 u64 *sptep, u64 old_spte, u64 new_spte)
3004 WARN_ON(!sp->role.direct);
3007 * Theoretically we could also set dirty bit (and flush TLB) here in
3008 * order to eliminate unnecessary PML logging. See comments in
3009 * set_spte. But fast_page_fault is very unlikely to happen with PML
3010 * enabled, so we do not do this. This might result in the same GPA
3011 * to be logged in PML buffer again when the write really happens, and
3012 * eventually to be called by mark_page_dirty twice. But it's also no
3013 * harm. This also avoids the TLB flush needed after setting dirty bit
3014 * so non-PML cases won't be impacted.
3016 * Compare with set_spte where instead shadow_dirty_mask is set.
3018 if (cmpxchg64(sptep, old_spte, new_spte) != old_spte)
3021 if (is_writable_pte(new_spte) && !is_writable_pte(old_spte)) {
3023 * The gfn of direct spte is stable since it is
3024 * calculated by sp->gfn.
3026 gfn = kvm_mmu_page_get_gfn(sp, sptep - sp->spt);
3027 kvm_vcpu_mark_page_dirty(vcpu, gfn);
3033 static bool is_access_allowed(u32 fault_err_code, u64 spte)
3035 if (fault_err_code & PFERR_FETCH_MASK)
3036 return is_executable_pte(spte);
3038 if (fault_err_code & PFERR_WRITE_MASK)
3039 return is_writable_pte(spte);
3041 /* Fault was on Read access */
3042 return spte & PT_PRESENT_MASK;
3046 * Returns one of RET_PF_INVALID, RET_PF_FIXED or RET_PF_SPURIOUS.
3048 static int fast_page_fault(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
3051 struct kvm_shadow_walk_iterator iterator;
3052 struct kvm_mmu_page *sp;
3053 int ret = RET_PF_INVALID;
3055 uint retry_count = 0;
3057 if (!page_fault_can_be_fast(error_code))
3060 walk_shadow_page_lockless_begin(vcpu);
3065 for_each_shadow_entry_lockless(vcpu, cr2_or_gpa, iterator, spte)
3066 if (!is_shadow_present_pte(spte))
3069 sp = sptep_to_sp(iterator.sptep);
3070 if (!is_last_spte(spte, sp->role.level))
3074 * Check whether the memory access that caused the fault would
3075 * still cause it if it were to be performed right now. If not,
3076 * then this is a spurious fault caused by TLB lazily flushed,
3077 * or some other CPU has already fixed the PTE after the
3078 * current CPU took the fault.
3080 * Need not check the access of upper level table entries since
3081 * they are always ACC_ALL.
3083 if (is_access_allowed(error_code, spte)) {
3084 ret = RET_PF_SPURIOUS;
3090 if (is_access_track_spte(spte))
3091 new_spte = restore_acc_track_spte(new_spte);
3094 * Currently, to simplify the code, write-protection can
3095 * be removed in the fast path only if the SPTE was
3096 * write-protected for dirty-logging or access tracking.
3098 if ((error_code & PFERR_WRITE_MASK) &&
3099 spte_can_locklessly_be_made_writable(spte)) {
3100 new_spte |= PT_WRITABLE_MASK;
3103 * Do not fix write-permission on the large spte. Since
3104 * we only dirty the first page into the dirty-bitmap in
3105 * fast_pf_fix_direct_spte(), other pages are missed
3106 * if its slot has dirty logging enabled.
3108 * Instead, we let the slow page fault path create a
3109 * normal spte to fix the access.
3111 * See the comments in kvm_arch_commit_memory_region().
3113 if (sp->role.level > PG_LEVEL_4K)
3117 /* Verify that the fault can be handled in the fast path */
3118 if (new_spte == spte ||
3119 !is_access_allowed(error_code, new_spte))
3123 * Currently, fast page fault only works for direct mapping
3124 * since the gfn is not stable for indirect shadow page. See
3125 * Documentation/virt/kvm/locking.rst to get more detail.
3127 if (fast_pf_fix_direct_spte(vcpu, sp, iterator.sptep, spte,
3133 if (++retry_count > 4) {
3134 printk_once(KERN_WARNING
3135 "kvm: Fast #PF retrying more than 4 times.\n");
3141 trace_fast_page_fault(vcpu, cr2_or_gpa, error_code, iterator.sptep,
3143 walk_shadow_page_lockless_end(vcpu);
3148 static void mmu_free_root_page(struct kvm *kvm, hpa_t *root_hpa,
3149 struct list_head *invalid_list)
3151 struct kvm_mmu_page *sp;
3153 if (!VALID_PAGE(*root_hpa))
3156 sp = to_shadow_page(*root_hpa & PT64_BASE_ADDR_MASK);
3158 if (kvm_mmu_put_root(kvm, sp)) {
3159 if (is_tdp_mmu_page(sp))
3160 kvm_tdp_mmu_free_root(kvm, sp);
3161 else if (sp->role.invalid)
3162 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
3165 *root_hpa = INVALID_PAGE;
3168 /* roots_to_free must be some combination of the KVM_MMU_ROOT_* flags */
3169 void kvm_mmu_free_roots(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
3170 ulong roots_to_free)
3172 struct kvm *kvm = vcpu->kvm;
3174 LIST_HEAD(invalid_list);
3175 bool free_active_root = roots_to_free & KVM_MMU_ROOT_CURRENT;
3177 BUILD_BUG_ON(KVM_MMU_NUM_PREV_ROOTS >= BITS_PER_LONG);
3179 /* Before acquiring the MMU lock, see if we need to do any real work. */
3180 if (!(free_active_root && VALID_PAGE(mmu->root_hpa))) {
3181 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3182 if ((roots_to_free & KVM_MMU_ROOT_PREVIOUS(i)) &&
3183 VALID_PAGE(mmu->prev_roots[i].hpa))
3186 if (i == KVM_MMU_NUM_PREV_ROOTS)
3190 write_lock(&kvm->mmu_lock);
3192 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3193 if (roots_to_free & KVM_MMU_ROOT_PREVIOUS(i))
3194 mmu_free_root_page(kvm, &mmu->prev_roots[i].hpa,
3197 if (free_active_root) {
3198 if (mmu->shadow_root_level >= PT64_ROOT_4LEVEL &&
3199 (mmu->root_level >= PT64_ROOT_4LEVEL || mmu->direct_map)) {
3200 mmu_free_root_page(kvm, &mmu->root_hpa, &invalid_list);
3202 for (i = 0; i < 4; ++i)
3203 if (mmu->pae_root[i] != 0)
3204 mmu_free_root_page(kvm,
3207 mmu->root_hpa = INVALID_PAGE;
3212 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3213 write_unlock(&kvm->mmu_lock);
3215 EXPORT_SYMBOL_GPL(kvm_mmu_free_roots);
3217 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
3221 if (!kvm_vcpu_is_visible_gfn(vcpu, root_gfn)) {
3222 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
3229 static hpa_t mmu_alloc_root(struct kvm_vcpu *vcpu, gfn_t gfn, gva_t gva,
3230 u8 level, bool direct)
3232 struct kvm_mmu_page *sp;
3234 write_lock(&vcpu->kvm->mmu_lock);
3236 if (make_mmu_pages_available(vcpu)) {
3237 write_unlock(&vcpu->kvm->mmu_lock);
3238 return INVALID_PAGE;
3240 sp = kvm_mmu_get_page(vcpu, gfn, gva, level, direct, ACC_ALL);
3243 write_unlock(&vcpu->kvm->mmu_lock);
3244 return __pa(sp->spt);
3247 static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
3249 u8 shadow_root_level = vcpu->arch.mmu->shadow_root_level;
3253 if (is_tdp_mmu_enabled(vcpu->kvm)) {
3254 root = kvm_tdp_mmu_get_vcpu_root_hpa(vcpu);
3256 if (!VALID_PAGE(root))
3258 vcpu->arch.mmu->root_hpa = root;
3259 } else if (shadow_root_level >= PT64_ROOT_4LEVEL) {
3260 root = mmu_alloc_root(vcpu, 0, 0, shadow_root_level,
3263 if (!VALID_PAGE(root))
3265 vcpu->arch.mmu->root_hpa = root;
3266 } else if (shadow_root_level == PT32E_ROOT_LEVEL) {
3267 for (i = 0; i < 4; ++i) {
3268 MMU_WARN_ON(VALID_PAGE(vcpu->arch.mmu->pae_root[i]));
3270 root = mmu_alloc_root(vcpu, i << (30 - PAGE_SHIFT),
3271 i << 30, PT32_ROOT_LEVEL, true);
3272 if (!VALID_PAGE(root))
3274 vcpu->arch.mmu->pae_root[i] = root | PT_PRESENT_MASK;
3276 vcpu->arch.mmu->root_hpa = __pa(vcpu->arch.mmu->pae_root);
3280 /* root_pgd is ignored for direct MMUs. */
3281 vcpu->arch.mmu->root_pgd = 0;
3286 static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
3289 gfn_t root_gfn, root_pgd;
3293 root_pgd = vcpu->arch.mmu->get_guest_pgd(vcpu);
3294 root_gfn = root_pgd >> PAGE_SHIFT;
3296 if (mmu_check_root(vcpu, root_gfn))
3300 * Do we shadow a long mode page table? If so we need to
3301 * write-protect the guests page table root.
3303 if (vcpu->arch.mmu->root_level >= PT64_ROOT_4LEVEL) {
3304 MMU_WARN_ON(VALID_PAGE(vcpu->arch.mmu->root_hpa));
3306 root = mmu_alloc_root(vcpu, root_gfn, 0,
3307 vcpu->arch.mmu->shadow_root_level, false);
3308 if (!VALID_PAGE(root))
3310 vcpu->arch.mmu->root_hpa = root;
3315 * We shadow a 32 bit page table. This may be a legacy 2-level
3316 * or a PAE 3-level page table. In either case we need to be aware that
3317 * the shadow page table may be a PAE or a long mode page table.
3319 pm_mask = PT_PRESENT_MASK;
3320 if (vcpu->arch.mmu->shadow_root_level == PT64_ROOT_4LEVEL)
3321 pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK;
3323 for (i = 0; i < 4; ++i) {
3324 MMU_WARN_ON(VALID_PAGE(vcpu->arch.mmu->pae_root[i]));
3325 if (vcpu->arch.mmu->root_level == PT32E_ROOT_LEVEL) {
3326 pdptr = vcpu->arch.mmu->get_pdptr(vcpu, i);
3327 if (!(pdptr & PT_PRESENT_MASK)) {
3328 vcpu->arch.mmu->pae_root[i] = 0;
3331 root_gfn = pdptr >> PAGE_SHIFT;
3332 if (mmu_check_root(vcpu, root_gfn))
3336 root = mmu_alloc_root(vcpu, root_gfn, i << 30,
3337 PT32_ROOT_LEVEL, false);
3338 if (!VALID_PAGE(root))
3340 vcpu->arch.mmu->pae_root[i] = root | pm_mask;
3342 vcpu->arch.mmu->root_hpa = __pa(vcpu->arch.mmu->pae_root);
3345 * If we shadow a 32 bit page table with a long mode page
3346 * table we enter this path.
3348 if (vcpu->arch.mmu->shadow_root_level == PT64_ROOT_4LEVEL) {
3349 if (vcpu->arch.mmu->lm_root == NULL) {
3351 * The additional page necessary for this is only
3352 * allocated on demand.
3357 lm_root = (void*)get_zeroed_page(GFP_KERNEL_ACCOUNT);
3358 if (lm_root == NULL)
3361 lm_root[0] = __pa(vcpu->arch.mmu->pae_root) | pm_mask;
3363 vcpu->arch.mmu->lm_root = lm_root;
3366 vcpu->arch.mmu->root_hpa = __pa(vcpu->arch.mmu->lm_root);
3370 vcpu->arch.mmu->root_pgd = root_pgd;
3375 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
3377 if (vcpu->arch.mmu->direct_map)
3378 return mmu_alloc_direct_roots(vcpu);
3380 return mmu_alloc_shadow_roots(vcpu);
3383 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
3386 struct kvm_mmu_page *sp;
3388 if (vcpu->arch.mmu->direct_map)
3391 if (!VALID_PAGE(vcpu->arch.mmu->root_hpa))
3394 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
3396 if (vcpu->arch.mmu->root_level >= PT64_ROOT_4LEVEL) {
3397 hpa_t root = vcpu->arch.mmu->root_hpa;
3398 sp = to_shadow_page(root);
3401 * Even if another CPU was marking the SP as unsync-ed
3402 * simultaneously, any guest page table changes are not
3403 * guaranteed to be visible anyway until this VCPU issues a TLB
3404 * flush strictly after those changes are made. We only need to
3405 * ensure that the other CPU sets these flags before any actual
3406 * changes to the page tables are made. The comments in
3407 * mmu_need_write_protect() describe what could go wrong if this
3408 * requirement isn't satisfied.
3410 if (!smp_load_acquire(&sp->unsync) &&
3411 !smp_load_acquire(&sp->unsync_children))
3414 write_lock(&vcpu->kvm->mmu_lock);
3415 kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
3417 mmu_sync_children(vcpu, sp);
3419 kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
3420 write_unlock(&vcpu->kvm->mmu_lock);
3424 write_lock(&vcpu->kvm->mmu_lock);
3425 kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
3427 for (i = 0; i < 4; ++i) {
3428 hpa_t root = vcpu->arch.mmu->pae_root[i];
3430 if (root && VALID_PAGE(root)) {
3431 root &= PT64_BASE_ADDR_MASK;
3432 sp = to_shadow_page(root);
3433 mmu_sync_children(vcpu, sp);
3437 kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
3438 write_unlock(&vcpu->kvm->mmu_lock);
3440 EXPORT_SYMBOL_GPL(kvm_mmu_sync_roots);
3442 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gpa_t vaddr,
3443 u32 access, struct x86_exception *exception)
3446 exception->error_code = 0;
3450 static gpa_t nonpaging_gva_to_gpa_nested(struct kvm_vcpu *vcpu, gpa_t vaddr,
3452 struct x86_exception *exception)
3455 exception->error_code = 0;
3456 return vcpu->arch.nested_mmu.translate_gpa(vcpu, vaddr, access, exception);
3460 __is_rsvd_bits_set(struct rsvd_bits_validate *rsvd_check, u64 pte, int level)
3462 int bit7 = (pte >> 7) & 1;
3464 return pte & rsvd_check->rsvd_bits_mask[bit7][level-1];
3467 static bool __is_bad_mt_xwr(struct rsvd_bits_validate *rsvd_check, u64 pte)
3469 return rsvd_check->bad_mt_xwr & BIT_ULL(pte & 0x3f);
3472 static bool mmio_info_in_cache(struct kvm_vcpu *vcpu, u64 addr, bool direct)
3475 * A nested guest cannot use the MMIO cache if it is using nested
3476 * page tables, because cr2 is a nGPA while the cache stores GPAs.
3478 if (mmu_is_nested(vcpu))
3482 return vcpu_match_mmio_gpa(vcpu, addr);
3484 return vcpu_match_mmio_gva(vcpu, addr);
3488 * Return the level of the lowest level SPTE added to sptes.
3489 * That SPTE may be non-present.
3491 static int get_walk(struct kvm_vcpu *vcpu, u64 addr, u64 *sptes, int *root_level)
3493 struct kvm_shadow_walk_iterator iterator;
3497 walk_shadow_page_lockless_begin(vcpu);
3499 for (shadow_walk_init(&iterator, vcpu, addr),
3500 *root_level = iterator.level;
3501 shadow_walk_okay(&iterator);
3502 __shadow_walk_next(&iterator, spte)) {
3503 leaf = iterator.level;
3504 spte = mmu_spte_get_lockless(iterator.sptep);
3508 if (!is_shadow_present_pte(spte))
3512 walk_shadow_page_lockless_end(vcpu);
3517 /* return true if reserved bit(s) are detected on a valid, non-MMIO SPTE. */
3518 static bool get_mmio_spte(struct kvm_vcpu *vcpu, u64 addr, u64 *sptep)
3520 u64 sptes[PT64_ROOT_MAX_LEVEL + 1];
3521 struct rsvd_bits_validate *rsvd_check;
3522 int root, leaf, level;
3523 bool reserved = false;
3525 if (!VALID_PAGE(vcpu->arch.mmu->root_hpa)) {
3530 if (is_tdp_mmu_root(vcpu->kvm, vcpu->arch.mmu->root_hpa))
3531 leaf = kvm_tdp_mmu_get_walk(vcpu, addr, sptes, &root);
3533 leaf = get_walk(vcpu, addr, sptes, &root);
3535 if (unlikely(leaf < 0)) {
3540 *sptep = sptes[leaf];
3543 * Skip reserved bits checks on the terminal leaf if it's not a valid
3544 * SPTE. Note, this also (intentionally) skips MMIO SPTEs, which, by
3545 * design, always have reserved bits set. The purpose of the checks is
3546 * to detect reserved bits on non-MMIO SPTEs. i.e. buggy SPTEs.
3548 if (!is_shadow_present_pte(sptes[leaf]))
3551 rsvd_check = &vcpu->arch.mmu->shadow_zero_check;
3553 for (level = root; level >= leaf; level--)
3555 * Use a bitwise-OR instead of a logical-OR to aggregate the
3556 * reserved bit and EPT's invalid memtype/XWR checks to avoid
3557 * adding a Jcc in the loop.
3559 reserved |= __is_bad_mt_xwr(rsvd_check, sptes[level]) |
3560 __is_rsvd_bits_set(rsvd_check, sptes[level], level);
3563 pr_err("%s: detect reserved bits on spte, addr 0x%llx, dump hierarchy:\n",
3565 for (level = root; level >= leaf; level--)
3566 pr_err("------ spte 0x%llx level %d.\n",
3567 sptes[level], level);
3573 static int handle_mmio_page_fault(struct kvm_vcpu *vcpu, u64 addr, bool direct)
3578 if (mmio_info_in_cache(vcpu, addr, direct))
3579 return RET_PF_EMULATE;
3581 reserved = get_mmio_spte(vcpu, addr, &spte);
3582 if (WARN_ON(reserved))
3585 if (is_mmio_spte(spte)) {
3586 gfn_t gfn = get_mmio_spte_gfn(spte);
3587 unsigned int access = get_mmio_spte_access(spte);
3589 if (!check_mmio_spte(vcpu, spte))
3590 return RET_PF_INVALID;
3595 trace_handle_mmio_page_fault(addr, gfn, access);
3596 vcpu_cache_mmio_info(vcpu, addr, gfn, access);
3597 return RET_PF_EMULATE;
3601 * If the page table is zapped by other cpus, let CPU fault again on
3604 return RET_PF_RETRY;
3607 static bool page_fault_handle_page_track(struct kvm_vcpu *vcpu,
3608 u32 error_code, gfn_t gfn)
3610 if (unlikely(error_code & PFERR_RSVD_MASK))
3613 if (!(error_code & PFERR_PRESENT_MASK) ||
3614 !(error_code & PFERR_WRITE_MASK))
3618 * guest is writing the page which is write tracked which can
3619 * not be fixed by page fault handler.
3621 if (kvm_page_track_is_active(vcpu, gfn, KVM_PAGE_TRACK_WRITE))
3627 static void shadow_page_table_clear_flood(struct kvm_vcpu *vcpu, gva_t addr)
3629 struct kvm_shadow_walk_iterator iterator;
3632 walk_shadow_page_lockless_begin(vcpu);
3633 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte) {
3634 clear_sp_write_flooding_count(iterator.sptep);
3635 if (!is_shadow_present_pte(spte))
3638 walk_shadow_page_lockless_end(vcpu);
3641 static bool kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
3644 struct kvm_arch_async_pf arch;
3646 arch.token = (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id;
3648 arch.direct_map = vcpu->arch.mmu->direct_map;
3649 arch.cr3 = vcpu->arch.mmu->get_guest_pgd(vcpu);
3651 return kvm_setup_async_pf(vcpu, cr2_or_gpa,
3652 kvm_vcpu_gfn_to_hva(vcpu, gfn), &arch);
3655 static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
3656 gpa_t cr2_or_gpa, kvm_pfn_t *pfn, bool write,
3659 struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
3662 /* Don't expose private memslots to L2. */
3663 if (is_guest_mode(vcpu) && !kvm_is_visible_memslot(slot)) {
3664 *pfn = KVM_PFN_NOSLOT;
3670 *pfn = __gfn_to_pfn_memslot(slot, gfn, false, &async, write, writable);
3672 return false; /* *pfn has correct page already */
3674 if (!prefault && kvm_can_do_async_pf(vcpu)) {
3675 trace_kvm_try_async_get_page(cr2_or_gpa, gfn);
3676 if (kvm_find_async_pf_gfn(vcpu, gfn)) {
3677 trace_kvm_async_pf_doublefault(cr2_or_gpa, gfn);
3678 kvm_make_request(KVM_REQ_APF_HALT, vcpu);
3680 } else if (kvm_arch_setup_async_pf(vcpu, cr2_or_gpa, gfn))
3684 *pfn = __gfn_to_pfn_memslot(slot, gfn, false, NULL, write, writable);
3688 static int direct_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
3689 bool prefault, int max_level, bool is_tdp)
3691 bool write = error_code & PFERR_WRITE_MASK;
3694 gfn_t gfn = gpa >> PAGE_SHIFT;
3695 unsigned long mmu_seq;
3699 if (page_fault_handle_page_track(vcpu, error_code, gfn))
3700 return RET_PF_EMULATE;
3702 if (!is_tdp_mmu_root(vcpu->kvm, vcpu->arch.mmu->root_hpa)) {
3703 r = fast_page_fault(vcpu, gpa, error_code);
3704 if (r != RET_PF_INVALID)
3708 r = mmu_topup_memory_caches(vcpu, false);
3712 mmu_seq = vcpu->kvm->mmu_notifier_seq;
3715 if (try_async_pf(vcpu, prefault, gfn, gpa, &pfn, write, &map_writable))
3716 return RET_PF_RETRY;
3718 if (handle_abnormal_pfn(vcpu, is_tdp ? 0 : gpa, gfn, pfn, ACC_ALL, &r))
3723 if (is_tdp_mmu_root(vcpu->kvm, vcpu->arch.mmu->root_hpa))
3724 read_lock(&vcpu->kvm->mmu_lock);
3726 write_lock(&vcpu->kvm->mmu_lock);
3728 if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
3730 r = make_mmu_pages_available(vcpu);
3734 if (is_tdp_mmu_root(vcpu->kvm, vcpu->arch.mmu->root_hpa))
3735 r = kvm_tdp_mmu_map(vcpu, gpa, error_code, map_writable, max_level,
3738 r = __direct_map(vcpu, gpa, error_code, map_writable, max_level, pfn,
3742 if (is_tdp_mmu_root(vcpu->kvm, vcpu->arch.mmu->root_hpa))
3743 read_unlock(&vcpu->kvm->mmu_lock);
3745 write_unlock(&vcpu->kvm->mmu_lock);
3746 kvm_release_pfn_clean(pfn);
3750 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa,
3751 u32 error_code, bool prefault)
3753 pgprintk("%s: gva %lx error %x\n", __func__, gpa, error_code);
3755 /* This path builds a PAE pagetable, we can map 2mb pages at maximum. */
3756 return direct_page_fault(vcpu, gpa & PAGE_MASK, error_code, prefault,
3757 PG_LEVEL_2M, false);
3760 int kvm_handle_page_fault(struct kvm_vcpu *vcpu, u64 error_code,
3761 u64 fault_address, char *insn, int insn_len)
3764 u32 flags = vcpu->arch.apf.host_apf_flags;
3766 #ifndef CONFIG_X86_64
3767 /* A 64-bit CR2 should be impossible on 32-bit KVM. */
3768 if (WARN_ON_ONCE(fault_address >> 32))
3772 vcpu->arch.l1tf_flush_l1d = true;
3774 trace_kvm_page_fault(fault_address, error_code);
3776 if (kvm_event_needs_reinjection(vcpu))
3777 kvm_mmu_unprotect_page_virt(vcpu, fault_address);
3778 r = kvm_mmu_page_fault(vcpu, fault_address, error_code, insn,
3780 } else if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
3781 vcpu->arch.apf.host_apf_flags = 0;
3782 local_irq_disable();
3783 kvm_async_pf_task_wait_schedule(fault_address);
3786 WARN_ONCE(1, "Unexpected host async PF flags: %x\n", flags);
3791 EXPORT_SYMBOL_GPL(kvm_handle_page_fault);
3793 int kvm_tdp_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
3798 for (max_level = KVM_MAX_HUGEPAGE_LEVEL;
3799 max_level > PG_LEVEL_4K;
3801 int page_num = KVM_PAGES_PER_HPAGE(max_level);
3802 gfn_t base = (gpa >> PAGE_SHIFT) & ~(page_num - 1);
3804 if (kvm_mtrr_check_gfn_range_consistency(vcpu, base, page_num))
3808 return direct_page_fault(vcpu, gpa, error_code, prefault,
3812 static void nonpaging_init_context(struct kvm_vcpu *vcpu,
3813 struct kvm_mmu *context)
3815 context->page_fault = nonpaging_page_fault;
3816 context->gva_to_gpa = nonpaging_gva_to_gpa;
3817 context->sync_page = nonpaging_sync_page;
3818 context->invlpg = NULL;
3819 context->root_level = 0;
3820 context->shadow_root_level = PT32E_ROOT_LEVEL;
3821 context->direct_map = true;
3822 context->nx = false;
3825 static inline bool is_root_usable(struct kvm_mmu_root_info *root, gpa_t pgd,
3826 union kvm_mmu_page_role role)
3828 return (role.direct || pgd == root->pgd) &&
3829 VALID_PAGE(root->hpa) && to_shadow_page(root->hpa) &&
3830 role.word == to_shadow_page(root->hpa)->role.word;
3834 * Find out if a previously cached root matching the new pgd/role is available.
3835 * The current root is also inserted into the cache.
3836 * If a matching root was found, it is assigned to kvm_mmu->root_hpa and true is
3838 * Otherwise, the LRU root from the cache is assigned to kvm_mmu->root_hpa and
3839 * false is returned. This root should now be freed by the caller.
3841 static bool cached_root_available(struct kvm_vcpu *vcpu, gpa_t new_pgd,
3842 union kvm_mmu_page_role new_role)
3845 struct kvm_mmu_root_info root;
3846 struct kvm_mmu *mmu = vcpu->arch.mmu;
3848 root.pgd = mmu->root_pgd;
3849 root.hpa = mmu->root_hpa;
3851 if (is_root_usable(&root, new_pgd, new_role))
3854 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
3855 swap(root, mmu->prev_roots[i]);
3857 if (is_root_usable(&root, new_pgd, new_role))
3861 mmu->root_hpa = root.hpa;
3862 mmu->root_pgd = root.pgd;
3864 return i < KVM_MMU_NUM_PREV_ROOTS;
3867 static bool fast_pgd_switch(struct kvm_vcpu *vcpu, gpa_t new_pgd,
3868 union kvm_mmu_page_role new_role)
3870 struct kvm_mmu *mmu = vcpu->arch.mmu;
3873 * For now, limit the fast switch to 64-bit hosts+VMs in order to avoid
3874 * having to deal with PDPTEs. We may add support for 32-bit hosts/VMs
3875 * later if necessary.
3877 if (mmu->shadow_root_level >= PT64_ROOT_4LEVEL &&
3878 mmu->root_level >= PT64_ROOT_4LEVEL)
3879 return cached_root_available(vcpu, new_pgd, new_role);
3884 static void __kvm_mmu_new_pgd(struct kvm_vcpu *vcpu, gpa_t new_pgd,
3885 union kvm_mmu_page_role new_role,
3886 bool skip_tlb_flush, bool skip_mmu_sync)
3888 if (!fast_pgd_switch(vcpu, new_pgd, new_role)) {
3889 kvm_mmu_free_roots(vcpu, vcpu->arch.mmu, KVM_MMU_ROOT_CURRENT);
3894 * It's possible that the cached previous root page is obsolete because
3895 * of a change in the MMU generation number. However, changing the
3896 * generation number is accompanied by KVM_REQ_MMU_RELOAD, which will
3897 * free the root set here and allocate a new one.
3899 kvm_make_request(KVM_REQ_LOAD_MMU_PGD, vcpu);
3901 if (!skip_mmu_sync || force_flush_and_sync_on_reuse)
3902 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
3903 if (!skip_tlb_flush || force_flush_and_sync_on_reuse)
3904 kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
3907 * The last MMIO access's GVA and GPA are cached in the VCPU. When
3908 * switching to a new CR3, that GVA->GPA mapping may no longer be
3909 * valid. So clear any cached MMIO info even when we don't need to sync
3910 * the shadow page tables.
3912 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
3915 * If this is a direct root page, it doesn't have a write flooding
3916 * count. Otherwise, clear the write flooding count.
3918 if (!new_role.direct)
3919 __clear_sp_write_flooding_count(
3920 to_shadow_page(vcpu->arch.mmu->root_hpa));
3923 void kvm_mmu_new_pgd(struct kvm_vcpu *vcpu, gpa_t new_pgd, bool skip_tlb_flush,
3926 __kvm_mmu_new_pgd(vcpu, new_pgd, kvm_mmu_calc_root_page_role(vcpu),
3927 skip_tlb_flush, skip_mmu_sync);
3929 EXPORT_SYMBOL_GPL(kvm_mmu_new_pgd);
3931 static unsigned long get_cr3(struct kvm_vcpu *vcpu)
3933 return kvm_read_cr3(vcpu);
3936 static bool sync_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn,
3937 unsigned int access, int *nr_present)
3939 if (unlikely(is_mmio_spte(*sptep))) {
3940 if (gfn != get_mmio_spte_gfn(*sptep)) {
3941 mmu_spte_clear_no_track(sptep);
3946 mark_mmio_spte(vcpu, sptep, gfn, access);
3953 static inline bool is_last_gpte(struct kvm_mmu *mmu,
3954 unsigned level, unsigned gpte)
3957 * The RHS has bit 7 set iff level < mmu->last_nonleaf_level.
3958 * If it is clear, there are no large pages at this level, so clear
3959 * PT_PAGE_SIZE_MASK in gpte if that is the case.
3961 gpte &= level - mmu->last_nonleaf_level;
3964 * PG_LEVEL_4K always terminates. The RHS has bit 7 set
3965 * iff level <= PG_LEVEL_4K, which for our purpose means
3966 * level == PG_LEVEL_4K; set PT_PAGE_SIZE_MASK in gpte then.
3968 gpte |= level - PG_LEVEL_4K - 1;
3970 return gpte & PT_PAGE_SIZE_MASK;
3973 #define PTTYPE_EPT 18 /* arbitrary */
3974 #define PTTYPE PTTYPE_EPT
3975 #include "paging_tmpl.h"
3979 #include "paging_tmpl.h"
3983 #include "paging_tmpl.h"
3987 __reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
3988 struct rsvd_bits_validate *rsvd_check,
3989 u64 pa_bits_rsvd, int level, bool nx, bool gbpages,
3992 u64 gbpages_bit_rsvd = 0;
3993 u64 nonleaf_bit8_rsvd = 0;
3996 rsvd_check->bad_mt_xwr = 0;
3999 gbpages_bit_rsvd = rsvd_bits(7, 7);
4001 if (level == PT32E_ROOT_LEVEL)
4002 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 62);
4004 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 51);
4006 /* Note, NX doesn't exist in PDPTEs, this is handled below. */
4008 high_bits_rsvd |= rsvd_bits(63, 63);
4011 * Non-leaf PML4Es and PDPEs reserve bit 8 (which would be the G bit for
4012 * leaf entries) on AMD CPUs only.
4015 nonleaf_bit8_rsvd = rsvd_bits(8, 8);
4018 case PT32_ROOT_LEVEL:
4019 /* no rsvd bits for 2 level 4K page table entries */
4020 rsvd_check->rsvd_bits_mask[0][1] = 0;
4021 rsvd_check->rsvd_bits_mask[0][0] = 0;
4022 rsvd_check->rsvd_bits_mask[1][0] =
4023 rsvd_check->rsvd_bits_mask[0][0];
4026 rsvd_check->rsvd_bits_mask[1][1] = 0;
4030 if (is_cpuid_PSE36())
4031 /* 36bits PSE 4MB page */
4032 rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
4034 /* 32 bits PSE 4MB page */
4035 rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
4037 case PT32E_ROOT_LEVEL:
4038 rsvd_check->rsvd_bits_mask[0][2] = rsvd_bits(63, 63) |
4041 rsvd_bits(1, 2); /* PDPTE */
4042 rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd; /* PDE */
4043 rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd; /* PTE */
4044 rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd |
4045 rsvd_bits(13, 20); /* large page */
4046 rsvd_check->rsvd_bits_mask[1][0] =
4047 rsvd_check->rsvd_bits_mask[0][0];
4049 case PT64_ROOT_5LEVEL:
4050 rsvd_check->rsvd_bits_mask[0][4] = high_bits_rsvd |
4053 rsvd_check->rsvd_bits_mask[1][4] =
4054 rsvd_check->rsvd_bits_mask[0][4];
4056 case PT64_ROOT_4LEVEL:
4057 rsvd_check->rsvd_bits_mask[0][3] = high_bits_rsvd |
4060 rsvd_check->rsvd_bits_mask[0][2] = high_bits_rsvd |
4062 rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd;
4063 rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd;
4064 rsvd_check->rsvd_bits_mask[1][3] =
4065 rsvd_check->rsvd_bits_mask[0][3];
4066 rsvd_check->rsvd_bits_mask[1][2] = high_bits_rsvd |
4069 rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd |
4070 rsvd_bits(13, 20); /* large page */
4071 rsvd_check->rsvd_bits_mask[1][0] =
4072 rsvd_check->rsvd_bits_mask[0][0];
4077 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
4078 struct kvm_mmu *context)
4080 __reset_rsvds_bits_mask(vcpu, &context->guest_rsvd_check,
4081 vcpu->arch.reserved_gpa_bits,
4082 context->root_level, context->nx,
4083 guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES),
4085 guest_cpuid_is_amd_or_hygon(vcpu));
4089 __reset_rsvds_bits_mask_ept(struct rsvd_bits_validate *rsvd_check,
4090 u64 pa_bits_rsvd, bool execonly)
4092 u64 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 51);
4095 rsvd_check->rsvd_bits_mask[0][4] = high_bits_rsvd | rsvd_bits(3, 7);
4096 rsvd_check->rsvd_bits_mask[0][3] = high_bits_rsvd | rsvd_bits(3, 7);
4097 rsvd_check->rsvd_bits_mask[0][2] = high_bits_rsvd | rsvd_bits(3, 6);
4098 rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd | rsvd_bits(3, 6);
4099 rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd;
4102 rsvd_check->rsvd_bits_mask[1][4] = rsvd_check->rsvd_bits_mask[0][4];
4103 rsvd_check->rsvd_bits_mask[1][3] = rsvd_check->rsvd_bits_mask[0][3];
4104 rsvd_check->rsvd_bits_mask[1][2] = high_bits_rsvd | rsvd_bits(12, 29);
4105 rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd | rsvd_bits(12, 20);
4106 rsvd_check->rsvd_bits_mask[1][0] = rsvd_check->rsvd_bits_mask[0][0];
4108 bad_mt_xwr = 0xFFull << (2 * 8); /* bits 3..5 must not be 2 */
4109 bad_mt_xwr |= 0xFFull << (3 * 8); /* bits 3..5 must not be 3 */
4110 bad_mt_xwr |= 0xFFull << (7 * 8); /* bits 3..5 must not be 7 */
4111 bad_mt_xwr |= REPEAT_BYTE(1ull << 2); /* bits 0..2 must not be 010 */
4112 bad_mt_xwr |= REPEAT_BYTE(1ull << 6); /* bits 0..2 must not be 110 */
4114 /* bits 0..2 must not be 100 unless VMX capabilities allow it */
4115 bad_mt_xwr |= REPEAT_BYTE(1ull << 4);
4117 rsvd_check->bad_mt_xwr = bad_mt_xwr;
4120 static void reset_rsvds_bits_mask_ept(struct kvm_vcpu *vcpu,
4121 struct kvm_mmu *context, bool execonly)
4123 __reset_rsvds_bits_mask_ept(&context->guest_rsvd_check,
4124 vcpu->arch.reserved_gpa_bits, execonly);
4127 static inline u64 reserved_hpa_bits(void)
4129 return rsvd_bits(shadow_phys_bits, 63);
4133 * the page table on host is the shadow page table for the page
4134 * table in guest or amd nested guest, its mmu features completely
4135 * follow the features in guest.
4138 reset_shadow_zero_bits_mask(struct kvm_vcpu *vcpu, struct kvm_mmu *context)
4140 bool uses_nx = context->nx ||
4141 context->mmu_role.base.smep_andnot_wp;
4142 struct rsvd_bits_validate *shadow_zero_check;
4146 * Passing "true" to the last argument is okay; it adds a check
4147 * on bit 8 of the SPTEs which KVM doesn't use anyway.
4149 shadow_zero_check = &context->shadow_zero_check;
4150 __reset_rsvds_bits_mask(vcpu, shadow_zero_check,
4151 reserved_hpa_bits(),
4152 context->shadow_root_level, uses_nx,
4153 guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES),
4154 is_pse(vcpu), true);
4156 if (!shadow_me_mask)
4159 for (i = context->shadow_root_level; --i >= 0;) {
4160 shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask;
4161 shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask;
4165 EXPORT_SYMBOL_GPL(reset_shadow_zero_bits_mask);
4167 static inline bool boot_cpu_is_amd(void)
4169 WARN_ON_ONCE(!tdp_enabled);
4170 return shadow_x_mask == 0;
4174 * the direct page table on host, use as much mmu features as
4175 * possible, however, kvm currently does not do execution-protection.
4178 reset_tdp_shadow_zero_bits_mask(struct kvm_vcpu *vcpu,
4179 struct kvm_mmu *context)
4181 struct rsvd_bits_validate *shadow_zero_check;
4184 shadow_zero_check = &context->shadow_zero_check;
4186 if (boot_cpu_is_amd())
4187 __reset_rsvds_bits_mask(vcpu, shadow_zero_check,
4188 reserved_hpa_bits(),
4189 context->shadow_root_level, false,
4190 boot_cpu_has(X86_FEATURE_GBPAGES),
4193 __reset_rsvds_bits_mask_ept(shadow_zero_check,
4194 reserved_hpa_bits(), false);
4196 if (!shadow_me_mask)
4199 for (i = context->shadow_root_level; --i >= 0;) {
4200 shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask;
4201 shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask;
4206 * as the comments in reset_shadow_zero_bits_mask() except it
4207 * is the shadow page table for intel nested guest.
4210 reset_ept_shadow_zero_bits_mask(struct kvm_vcpu *vcpu,
4211 struct kvm_mmu *context, bool execonly)
4213 __reset_rsvds_bits_mask_ept(&context->shadow_zero_check,
4214 reserved_hpa_bits(), execonly);
4217 #define BYTE_MASK(access) \
4218 ((1 & (access) ? 2 : 0) | \
4219 (2 & (access) ? 4 : 0) | \
4220 (3 & (access) ? 8 : 0) | \
4221 (4 & (access) ? 16 : 0) | \
4222 (5 & (access) ? 32 : 0) | \
4223 (6 & (access) ? 64 : 0) | \
4224 (7 & (access) ? 128 : 0))
4227 static void update_permission_bitmask(struct kvm_vcpu *vcpu,
4228 struct kvm_mmu *mmu, bool ept)
4232 const u8 x = BYTE_MASK(ACC_EXEC_MASK);
4233 const u8 w = BYTE_MASK(ACC_WRITE_MASK);
4234 const u8 u = BYTE_MASK(ACC_USER_MASK);
4236 bool cr4_smep = kvm_read_cr4_bits(vcpu, X86_CR4_SMEP) != 0;
4237 bool cr4_smap = kvm_read_cr4_bits(vcpu, X86_CR4_SMAP) != 0;
4238 bool cr0_wp = is_write_protection(vcpu);
4240 for (byte = 0; byte < ARRAY_SIZE(mmu->permissions); ++byte) {
4241 unsigned pfec = byte << 1;
4244 * Each "*f" variable has a 1 bit for each UWX value
4245 * that causes a fault with the given PFEC.
4248 /* Faults from writes to non-writable pages */
4249 u8 wf = (pfec & PFERR_WRITE_MASK) ? (u8)~w : 0;
4250 /* Faults from user mode accesses to supervisor pages */
4251 u8 uf = (pfec & PFERR_USER_MASK) ? (u8)~u : 0;
4252 /* Faults from fetches of non-executable pages*/
4253 u8 ff = (pfec & PFERR_FETCH_MASK) ? (u8)~x : 0;
4254 /* Faults from kernel mode fetches of user pages */
4256 /* Faults from kernel mode accesses of user pages */
4260 /* Faults from kernel mode accesses to user pages */
4261 u8 kf = (pfec & PFERR_USER_MASK) ? 0 : u;
4263 /* Not really needed: !nx will cause pte.nx to fault */
4267 /* Allow supervisor writes if !cr0.wp */
4269 wf = (pfec & PFERR_USER_MASK) ? wf : 0;
4271 /* Disallow supervisor fetches of user code if cr4.smep */
4273 smepf = (pfec & PFERR_FETCH_MASK) ? kf : 0;
4276 * SMAP:kernel-mode data accesses from user-mode
4277 * mappings should fault. A fault is considered
4278 * as a SMAP violation if all of the following
4279 * conditions are true:
4280 * - X86_CR4_SMAP is set in CR4
4281 * - A user page is accessed
4282 * - The access is not a fetch
4283 * - Page fault in kernel mode
4284 * - if CPL = 3 or X86_EFLAGS_AC is clear
4286 * Here, we cover the first three conditions.
4287 * The fourth is computed dynamically in permission_fault();
4288 * PFERR_RSVD_MASK bit will be set in PFEC if the access is
4289 * *not* subject to SMAP restrictions.
4292 smapf = (pfec & (PFERR_RSVD_MASK|PFERR_FETCH_MASK)) ? 0 : kf;
4295 mmu->permissions[byte] = ff | uf | wf | smepf | smapf;
4300 * PKU is an additional mechanism by which the paging controls access to
4301 * user-mode addresses based on the value in the PKRU register. Protection
4302 * key violations are reported through a bit in the page fault error code.
4303 * Unlike other bits of the error code, the PK bit is not known at the
4304 * call site of e.g. gva_to_gpa; it must be computed directly in
4305 * permission_fault based on two bits of PKRU, on some machine state (CR4,
4306 * CR0, EFER, CPL), and on other bits of the error code and the page tables.
4308 * In particular the following conditions come from the error code, the
4309 * page tables and the machine state:
4310 * - PK is always zero unless CR4.PKE=1 and EFER.LMA=1
4311 * - PK is always zero if RSVD=1 (reserved bit set) or F=1 (instruction fetch)
4312 * - PK is always zero if U=0 in the page tables
4313 * - PKRU.WD is ignored if CR0.WP=0 and the access is a supervisor access.
4315 * The PKRU bitmask caches the result of these four conditions. The error
4316 * code (minus the P bit) and the page table's U bit form an index into the
4317 * PKRU bitmask. Two bits of the PKRU bitmask are then extracted and ANDed
4318 * with the two bits of the PKRU register corresponding to the protection key.
4319 * For the first three conditions above the bits will be 00, thus masking
4320 * away both AD and WD. For all reads or if the last condition holds, WD
4321 * only will be masked away.
4323 static void update_pkru_bitmask(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
4334 /* PKEY is enabled only if CR4.PKE and EFER.LMA are both set. */
4335 if (!kvm_read_cr4_bits(vcpu, X86_CR4_PKE) || !is_long_mode(vcpu)) {
4340 wp = is_write_protection(vcpu);
4342 for (bit = 0; bit < ARRAY_SIZE(mmu->permissions); ++bit) {
4343 unsigned pfec, pkey_bits;
4344 bool check_pkey, check_write, ff, uf, wf, pte_user;
4347 ff = pfec & PFERR_FETCH_MASK;
4348 uf = pfec & PFERR_USER_MASK;
4349 wf = pfec & PFERR_WRITE_MASK;
4351 /* PFEC.RSVD is replaced by ACC_USER_MASK. */
4352 pte_user = pfec & PFERR_RSVD_MASK;
4355 * Only need to check the access which is not an
4356 * instruction fetch and is to a user page.
4358 check_pkey = (!ff && pte_user);
4360 * write access is controlled by PKRU if it is a
4361 * user access or CR0.WP = 1.
4363 check_write = check_pkey && wf && (uf || wp);
4365 /* PKRU.AD stops both read and write access. */
4366 pkey_bits = !!check_pkey;
4367 /* PKRU.WD stops write access. */
4368 pkey_bits |= (!!check_write) << 1;
4370 mmu->pkru_mask |= (pkey_bits & 3) << pfec;
4374 static void update_last_nonleaf_level(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
4376 unsigned root_level = mmu->root_level;
4378 mmu->last_nonleaf_level = root_level;
4379 if (root_level == PT32_ROOT_LEVEL && is_pse(vcpu))
4380 mmu->last_nonleaf_level++;
4383 static void paging64_init_context_common(struct kvm_vcpu *vcpu,
4384 struct kvm_mmu *context,
4387 context->nx = is_nx(vcpu);
4388 context->root_level = level;
4390 reset_rsvds_bits_mask(vcpu, context);
4391 update_permission_bitmask(vcpu, context, false);
4392 update_pkru_bitmask(vcpu, context, false);
4393 update_last_nonleaf_level(vcpu, context);
4395 MMU_WARN_ON(!is_pae(vcpu));
4396 context->page_fault = paging64_page_fault;
4397 context->gva_to_gpa = paging64_gva_to_gpa;
4398 context->sync_page = paging64_sync_page;
4399 context->invlpg = paging64_invlpg;
4400 context->shadow_root_level = level;
4401 context->direct_map = false;
4404 static void paging64_init_context(struct kvm_vcpu *vcpu,
4405 struct kvm_mmu *context)
4407 int root_level = is_la57_mode(vcpu) ?
4408 PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
4410 paging64_init_context_common(vcpu, context, root_level);
4413 static void paging32_init_context(struct kvm_vcpu *vcpu,
4414 struct kvm_mmu *context)
4416 context->nx = false;
4417 context->root_level = PT32_ROOT_LEVEL;
4419 reset_rsvds_bits_mask(vcpu, context);
4420 update_permission_bitmask(vcpu, context, false);
4421 update_pkru_bitmask(vcpu, context, false);
4422 update_last_nonleaf_level(vcpu, context);
4424 context->page_fault = paging32_page_fault;
4425 context->gva_to_gpa = paging32_gva_to_gpa;
4426 context->sync_page = paging32_sync_page;
4427 context->invlpg = paging32_invlpg;
4428 context->shadow_root_level = PT32E_ROOT_LEVEL;
4429 context->direct_map = false;
4432 static void paging32E_init_context(struct kvm_vcpu *vcpu,
4433 struct kvm_mmu *context)
4435 paging64_init_context_common(vcpu, context, PT32E_ROOT_LEVEL);
4438 static union kvm_mmu_extended_role kvm_calc_mmu_role_ext(struct kvm_vcpu *vcpu)
4440 union kvm_mmu_extended_role ext = {0};
4442 ext.cr0_pg = !!is_paging(vcpu);
4443 ext.cr4_pae = !!is_pae(vcpu);
4444 ext.cr4_smep = !!kvm_read_cr4_bits(vcpu, X86_CR4_SMEP);
4445 ext.cr4_smap = !!kvm_read_cr4_bits(vcpu, X86_CR4_SMAP);
4446 ext.cr4_pse = !!is_pse(vcpu);
4447 ext.cr4_pke = !!kvm_read_cr4_bits(vcpu, X86_CR4_PKE);
4448 ext.maxphyaddr = cpuid_maxphyaddr(vcpu);
4455 static union kvm_mmu_role kvm_calc_mmu_role_common(struct kvm_vcpu *vcpu,
4458 union kvm_mmu_role role = {0};
4460 role.base.access = ACC_ALL;
4461 role.base.nxe = !!is_nx(vcpu);
4462 role.base.cr0_wp = is_write_protection(vcpu);
4463 role.base.smm = is_smm(vcpu);
4464 role.base.guest_mode = is_guest_mode(vcpu);
4469 role.ext = kvm_calc_mmu_role_ext(vcpu);
4474 static inline int kvm_mmu_get_tdp_level(struct kvm_vcpu *vcpu)
4476 /* Use 5-level TDP if and only if it's useful/necessary. */
4477 if (max_tdp_level == 5 && cpuid_maxphyaddr(vcpu) <= 48)
4480 return max_tdp_level;
4483 static union kvm_mmu_role
4484 kvm_calc_tdp_mmu_root_page_role(struct kvm_vcpu *vcpu, bool base_only)
4486 union kvm_mmu_role role = kvm_calc_mmu_role_common(vcpu, base_only);
4488 role.base.ad_disabled = (shadow_accessed_mask == 0);
4489 role.base.level = kvm_mmu_get_tdp_level(vcpu);
4490 role.base.direct = true;
4491 role.base.gpte_is_8_bytes = true;
4496 static void init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
4498 struct kvm_mmu *context = &vcpu->arch.root_mmu;
4499 union kvm_mmu_role new_role =
4500 kvm_calc_tdp_mmu_root_page_role(vcpu, false);
4502 if (new_role.as_u64 == context->mmu_role.as_u64)
4505 context->mmu_role.as_u64 = new_role.as_u64;
4506 context->page_fault = kvm_tdp_page_fault;
4507 context->sync_page = nonpaging_sync_page;
4508 context->invlpg = NULL;
4509 context->shadow_root_level = kvm_mmu_get_tdp_level(vcpu);
4510 context->direct_map = true;
4511 context->get_guest_pgd = get_cr3;
4512 context->get_pdptr = kvm_pdptr_read;
4513 context->inject_page_fault = kvm_inject_page_fault;
4515 if (!is_paging(vcpu)) {
4516 context->nx = false;
4517 context->gva_to_gpa = nonpaging_gva_to_gpa;
4518 context->root_level = 0;
4519 } else if (is_long_mode(vcpu)) {
4520 context->nx = is_nx(vcpu);
4521 context->root_level = is_la57_mode(vcpu) ?
4522 PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
4523 reset_rsvds_bits_mask(vcpu, context);
4524 context->gva_to_gpa = paging64_gva_to_gpa;
4525 } else if (is_pae(vcpu)) {
4526 context->nx = is_nx(vcpu);
4527 context->root_level = PT32E_ROOT_LEVEL;
4528 reset_rsvds_bits_mask(vcpu, context);
4529 context->gva_to_gpa = paging64_gva_to_gpa;
4531 context->nx = false;
4532 context->root_level = PT32_ROOT_LEVEL;
4533 reset_rsvds_bits_mask(vcpu, context);
4534 context->gva_to_gpa = paging32_gva_to_gpa;
4537 update_permission_bitmask(vcpu, context, false);
4538 update_pkru_bitmask(vcpu, context, false);
4539 update_last_nonleaf_level(vcpu, context);
4540 reset_tdp_shadow_zero_bits_mask(vcpu, context);
4543 static union kvm_mmu_role
4544 kvm_calc_shadow_root_page_role_common(struct kvm_vcpu *vcpu, bool base_only)
4546 union kvm_mmu_role role = kvm_calc_mmu_role_common(vcpu, base_only);
4548 role.base.smep_andnot_wp = role.ext.cr4_smep &&
4549 !is_write_protection(vcpu);
4550 role.base.smap_andnot_wp = role.ext.cr4_smap &&
4551 !is_write_protection(vcpu);
4552 role.base.gpte_is_8_bytes = !!is_pae(vcpu);
4557 static union kvm_mmu_role
4558 kvm_calc_shadow_mmu_root_page_role(struct kvm_vcpu *vcpu, bool base_only)
4560 union kvm_mmu_role role =
4561 kvm_calc_shadow_root_page_role_common(vcpu, base_only);
4563 role.base.direct = !is_paging(vcpu);
4565 if (!is_long_mode(vcpu))
4566 role.base.level = PT32E_ROOT_LEVEL;
4567 else if (is_la57_mode(vcpu))
4568 role.base.level = PT64_ROOT_5LEVEL;
4570 role.base.level = PT64_ROOT_4LEVEL;
4575 static void shadow_mmu_init_context(struct kvm_vcpu *vcpu, struct kvm_mmu *context,
4576 u32 cr0, u32 cr4, u32 efer,
4577 union kvm_mmu_role new_role)
4579 if (!(cr0 & X86_CR0_PG))
4580 nonpaging_init_context(vcpu, context);
4581 else if (efer & EFER_LMA)
4582 paging64_init_context(vcpu, context);
4583 else if (cr4 & X86_CR4_PAE)
4584 paging32E_init_context(vcpu, context);
4586 paging32_init_context(vcpu, context);
4588 context->mmu_role.as_u64 = new_role.as_u64;
4589 reset_shadow_zero_bits_mask(vcpu, context);
4592 static void kvm_init_shadow_mmu(struct kvm_vcpu *vcpu, u32 cr0, u32 cr4, u32 efer)
4594 struct kvm_mmu *context = &vcpu->arch.root_mmu;
4595 union kvm_mmu_role new_role =
4596 kvm_calc_shadow_mmu_root_page_role(vcpu, false);
4598 if (new_role.as_u64 != context->mmu_role.as_u64)
4599 shadow_mmu_init_context(vcpu, context, cr0, cr4, efer, new_role);
4602 static union kvm_mmu_role
4603 kvm_calc_shadow_npt_root_page_role(struct kvm_vcpu *vcpu)
4605 union kvm_mmu_role role =
4606 kvm_calc_shadow_root_page_role_common(vcpu, false);
4608 role.base.direct = false;
4609 role.base.level = kvm_mmu_get_tdp_level(vcpu);
4614 void kvm_init_shadow_npt_mmu(struct kvm_vcpu *vcpu, u32 cr0, u32 cr4, u32 efer,
4617 struct kvm_mmu *context = &vcpu->arch.guest_mmu;
4618 union kvm_mmu_role new_role = kvm_calc_shadow_npt_root_page_role(vcpu);
4620 context->shadow_root_level = new_role.base.level;
4622 __kvm_mmu_new_pgd(vcpu, nested_cr3, new_role.base, false, false);
4624 if (new_role.as_u64 != context->mmu_role.as_u64)
4625 shadow_mmu_init_context(vcpu, context, cr0, cr4, efer, new_role);
4627 EXPORT_SYMBOL_GPL(kvm_init_shadow_npt_mmu);
4629 static union kvm_mmu_role
4630 kvm_calc_shadow_ept_root_page_role(struct kvm_vcpu *vcpu, bool accessed_dirty,
4631 bool execonly, u8 level)
4633 union kvm_mmu_role role = {0};
4635 /* SMM flag is inherited from root_mmu */
4636 role.base.smm = vcpu->arch.root_mmu.mmu_role.base.smm;
4638 role.base.level = level;
4639 role.base.gpte_is_8_bytes = true;
4640 role.base.direct = false;
4641 role.base.ad_disabled = !accessed_dirty;
4642 role.base.guest_mode = true;
4643 role.base.access = ACC_ALL;
4646 * WP=1 and NOT_WP=1 is an impossible combination, use WP and the
4647 * SMAP variation to denote shadow EPT entries.
4649 role.base.cr0_wp = true;
4650 role.base.smap_andnot_wp = true;
4652 role.ext = kvm_calc_mmu_role_ext(vcpu);
4653 role.ext.execonly = execonly;
4658 void kvm_init_shadow_ept_mmu(struct kvm_vcpu *vcpu, bool execonly,
4659 bool accessed_dirty, gpa_t new_eptp)
4661 struct kvm_mmu *context = &vcpu->arch.guest_mmu;
4662 u8 level = vmx_eptp_page_walk_level(new_eptp);
4663 union kvm_mmu_role new_role =
4664 kvm_calc_shadow_ept_root_page_role(vcpu, accessed_dirty,
4667 __kvm_mmu_new_pgd(vcpu, new_eptp, new_role.base, true, true);
4669 if (new_role.as_u64 == context->mmu_role.as_u64)
4672 context->shadow_root_level = level;
4675 context->ept_ad = accessed_dirty;
4676 context->page_fault = ept_page_fault;
4677 context->gva_to_gpa = ept_gva_to_gpa;
4678 context->sync_page = ept_sync_page;
4679 context->invlpg = ept_invlpg;
4680 context->root_level = level;
4681 context->direct_map = false;
4682 context->mmu_role.as_u64 = new_role.as_u64;
4684 update_permission_bitmask(vcpu, context, true);
4685 update_pkru_bitmask(vcpu, context, true);
4686 update_last_nonleaf_level(vcpu, context);
4687 reset_rsvds_bits_mask_ept(vcpu, context, execonly);
4688 reset_ept_shadow_zero_bits_mask(vcpu, context, execonly);
4690 EXPORT_SYMBOL_GPL(kvm_init_shadow_ept_mmu);
4692 static void init_kvm_softmmu(struct kvm_vcpu *vcpu)
4694 struct kvm_mmu *context = &vcpu->arch.root_mmu;
4696 kvm_init_shadow_mmu(vcpu,
4697 kvm_read_cr0_bits(vcpu, X86_CR0_PG),
4698 kvm_read_cr4_bits(vcpu, X86_CR4_PAE),
4701 context->get_guest_pgd = get_cr3;
4702 context->get_pdptr = kvm_pdptr_read;
4703 context->inject_page_fault = kvm_inject_page_fault;
4706 static void init_kvm_nested_mmu(struct kvm_vcpu *vcpu)
4708 union kvm_mmu_role new_role = kvm_calc_mmu_role_common(vcpu, false);
4709 struct kvm_mmu *g_context = &vcpu->arch.nested_mmu;
4711 if (new_role.as_u64 == g_context->mmu_role.as_u64)
4714 g_context->mmu_role.as_u64 = new_role.as_u64;
4715 g_context->get_guest_pgd = get_cr3;
4716 g_context->get_pdptr = kvm_pdptr_read;
4717 g_context->inject_page_fault = kvm_inject_page_fault;
4720 * L2 page tables are never shadowed, so there is no need to sync
4723 g_context->invlpg = NULL;
4726 * Note that arch.mmu->gva_to_gpa translates l2_gpa to l1_gpa using
4727 * L1's nested page tables (e.g. EPT12). The nested translation
4728 * of l2_gva to l1_gpa is done by arch.nested_mmu.gva_to_gpa using
4729 * L2's page tables as the first level of translation and L1's
4730 * nested page tables as the second level of translation. Basically
4731 * the gva_to_gpa functions between mmu and nested_mmu are swapped.
4733 if (!is_paging(vcpu)) {
4734 g_context->nx = false;
4735 g_context->root_level = 0;
4736 g_context->gva_to_gpa = nonpaging_gva_to_gpa_nested;
4737 } else if (is_long_mode(vcpu)) {
4738 g_context->nx = is_nx(vcpu);
4739 g_context->root_level = is_la57_mode(vcpu) ?
4740 PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
4741 reset_rsvds_bits_mask(vcpu, g_context);
4742 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
4743 } else if (is_pae(vcpu)) {
4744 g_context->nx = is_nx(vcpu);
4745 g_context->root_level = PT32E_ROOT_LEVEL;
4746 reset_rsvds_bits_mask(vcpu, g_context);
4747 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
4749 g_context->nx = false;
4750 g_context->root_level = PT32_ROOT_LEVEL;
4751 reset_rsvds_bits_mask(vcpu, g_context);
4752 g_context->gva_to_gpa = paging32_gva_to_gpa_nested;
4755 update_permission_bitmask(vcpu, g_context, false);
4756 update_pkru_bitmask(vcpu, g_context, false);
4757 update_last_nonleaf_level(vcpu, g_context);
4760 void kvm_init_mmu(struct kvm_vcpu *vcpu, bool reset_roots)
4765 vcpu->arch.mmu->root_hpa = INVALID_PAGE;
4767 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
4768 vcpu->arch.mmu->prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
4771 if (mmu_is_nested(vcpu))
4772 init_kvm_nested_mmu(vcpu);
4773 else if (tdp_enabled)
4774 init_kvm_tdp_mmu(vcpu);
4776 init_kvm_softmmu(vcpu);
4778 EXPORT_SYMBOL_GPL(kvm_init_mmu);
4780 static union kvm_mmu_page_role
4781 kvm_mmu_calc_root_page_role(struct kvm_vcpu *vcpu)
4783 union kvm_mmu_role role;
4786 role = kvm_calc_tdp_mmu_root_page_role(vcpu, true);
4788 role = kvm_calc_shadow_mmu_root_page_role(vcpu, true);
4793 void kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
4795 kvm_mmu_unload(vcpu);
4796 kvm_init_mmu(vcpu, true);
4798 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
4800 int kvm_mmu_load(struct kvm_vcpu *vcpu)
4804 r = mmu_topup_memory_caches(vcpu, !vcpu->arch.mmu->direct_map);
4807 r = mmu_alloc_roots(vcpu);
4808 kvm_mmu_sync_roots(vcpu);
4811 kvm_mmu_load_pgd(vcpu);
4812 static_call(kvm_x86_tlb_flush_current)(vcpu);
4816 EXPORT_SYMBOL_GPL(kvm_mmu_load);
4818 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
4820 kvm_mmu_free_roots(vcpu, &vcpu->arch.root_mmu, KVM_MMU_ROOTS_ALL);
4821 WARN_ON(VALID_PAGE(vcpu->arch.root_mmu.root_hpa));
4822 kvm_mmu_free_roots(vcpu, &vcpu->arch.guest_mmu, KVM_MMU_ROOTS_ALL);
4823 WARN_ON(VALID_PAGE(vcpu->arch.guest_mmu.root_hpa));
4825 EXPORT_SYMBOL_GPL(kvm_mmu_unload);
4827 static bool need_remote_flush(u64 old, u64 new)
4829 if (!is_shadow_present_pte(old))
4831 if (!is_shadow_present_pte(new))
4833 if ((old ^ new) & PT64_BASE_ADDR_MASK)
4835 old ^= shadow_nx_mask;
4836 new ^= shadow_nx_mask;
4837 return (old & ~new & PT64_PERM_MASK) != 0;
4840 static u64 mmu_pte_write_fetch_gpte(struct kvm_vcpu *vcpu, gpa_t *gpa,
4847 * Assume that the pte write on a page table of the same type
4848 * as the current vcpu paging mode since we update the sptes only
4849 * when they have the same mode.
4851 if (is_pae(vcpu) && *bytes == 4) {
4852 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
4857 if (*bytes == 4 || *bytes == 8) {
4858 r = kvm_vcpu_read_guest_atomic(vcpu, *gpa, &gentry, *bytes);
4867 * If we're seeing too many writes to a page, it may no longer be a page table,
4868 * or we may be forking, in which case it is better to unmap the page.
4870 static bool detect_write_flooding(struct kvm_mmu_page *sp)
4873 * Skip write-flooding detected for the sp whose level is 1, because
4874 * it can become unsync, then the guest page is not write-protected.
4876 if (sp->role.level == PG_LEVEL_4K)
4879 atomic_inc(&sp->write_flooding_count);
4880 return atomic_read(&sp->write_flooding_count) >= 3;
4884 * Misaligned accesses are too much trouble to fix up; also, they usually
4885 * indicate a page is not used as a page table.
4887 static bool detect_write_misaligned(struct kvm_mmu_page *sp, gpa_t gpa,
4890 unsigned offset, pte_size, misaligned;
4892 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
4893 gpa, bytes, sp->role.word);
4895 offset = offset_in_page(gpa);
4896 pte_size = sp->role.gpte_is_8_bytes ? 8 : 4;
4899 * Sometimes, the OS only writes the last one bytes to update status
4900 * bits, for example, in linux, andb instruction is used in clear_bit().
4902 if (!(offset & (pte_size - 1)) && bytes == 1)
4905 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
4906 misaligned |= bytes < 4;
4911 static u64 *get_written_sptes(struct kvm_mmu_page *sp, gpa_t gpa, int *nspte)
4913 unsigned page_offset, quadrant;
4917 page_offset = offset_in_page(gpa);
4918 level = sp->role.level;
4920 if (!sp->role.gpte_is_8_bytes) {
4921 page_offset <<= 1; /* 32->64 */
4923 * A 32-bit pde maps 4MB while the shadow pdes map
4924 * only 2MB. So we need to double the offset again
4925 * and zap two pdes instead of one.
4927 if (level == PT32_ROOT_LEVEL) {
4928 page_offset &= ~7; /* kill rounding error */
4932 quadrant = page_offset >> PAGE_SHIFT;
4933 page_offset &= ~PAGE_MASK;
4934 if (quadrant != sp->role.quadrant)
4938 spte = &sp->spt[page_offset / sizeof(*spte)];
4942 static void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
4943 const u8 *new, int bytes,
4944 struct kvm_page_track_notifier_node *node)
4946 gfn_t gfn = gpa >> PAGE_SHIFT;
4947 struct kvm_mmu_page *sp;
4948 LIST_HEAD(invalid_list);
4949 u64 entry, gentry, *spte;
4951 bool remote_flush, local_flush;
4954 * If we don't have indirect shadow pages, it means no page is
4955 * write-protected, so we can exit simply.
4957 if (!READ_ONCE(vcpu->kvm->arch.indirect_shadow_pages))
4960 remote_flush = local_flush = false;
4962 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
4965 * No need to care whether allocation memory is successful
4966 * or not since pte prefetch is skiped if it does not have
4967 * enough objects in the cache.
4969 mmu_topup_memory_caches(vcpu, true);
4971 write_lock(&vcpu->kvm->mmu_lock);
4973 gentry = mmu_pte_write_fetch_gpte(vcpu, &gpa, &bytes);
4975 ++vcpu->kvm->stat.mmu_pte_write;
4976 kvm_mmu_audit(vcpu, AUDIT_PRE_PTE_WRITE);
4978 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn) {
4979 if (detect_write_misaligned(sp, gpa, bytes) ||
4980 detect_write_flooding(sp)) {
4981 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
4982 ++vcpu->kvm->stat.mmu_flooded;
4986 spte = get_written_sptes(sp, gpa, &npte);
4993 mmu_page_zap_pte(vcpu->kvm, sp, spte, NULL);
4994 if (gentry && sp->role.level != PG_LEVEL_4K)
4995 ++vcpu->kvm->stat.mmu_pde_zapped;
4996 if (need_remote_flush(entry, *spte))
4997 remote_flush = true;
5001 kvm_mmu_flush_or_zap(vcpu, &invalid_list, remote_flush, local_flush);
5002 kvm_mmu_audit(vcpu, AUDIT_POST_PTE_WRITE);
5003 write_unlock(&vcpu->kvm->mmu_lock);
5006 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
5011 if (vcpu->arch.mmu->direct_map)
5014 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
5016 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
5020 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
5022 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, u64 error_code,
5023 void *insn, int insn_len)
5025 int r, emulation_type = EMULTYPE_PF;
5026 bool direct = vcpu->arch.mmu->direct_map;
5028 if (WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root_hpa)))
5029 return RET_PF_RETRY;
5032 if (unlikely(error_code & PFERR_RSVD_MASK)) {
5033 r = handle_mmio_page_fault(vcpu, cr2_or_gpa, direct);
5034 if (r == RET_PF_EMULATE)
5038 if (r == RET_PF_INVALID) {
5039 r = kvm_mmu_do_page_fault(vcpu, cr2_or_gpa,
5040 lower_32_bits(error_code), false);
5041 if (WARN_ON_ONCE(r == RET_PF_INVALID))
5047 if (r != RET_PF_EMULATE)
5051 * Before emulating the instruction, check if the error code
5052 * was due to a RO violation while translating the guest page.
5053 * This can occur when using nested virtualization with nested
5054 * paging in both guests. If true, we simply unprotect the page
5055 * and resume the guest.
5057 if (vcpu->arch.mmu->direct_map &&
5058 (error_code & PFERR_NESTED_GUEST_PAGE) == PFERR_NESTED_GUEST_PAGE) {
5059 kvm_mmu_unprotect_page(vcpu->kvm, gpa_to_gfn(cr2_or_gpa));
5064 * vcpu->arch.mmu.page_fault returned RET_PF_EMULATE, but we can still
5065 * optimistically try to just unprotect the page and let the processor
5066 * re-execute the instruction that caused the page fault. Do not allow
5067 * retrying MMIO emulation, as it's not only pointless but could also
5068 * cause us to enter an infinite loop because the processor will keep
5069 * faulting on the non-existent MMIO address. Retrying an instruction
5070 * from a nested guest is also pointless and dangerous as we are only
5071 * explicitly shadowing L1's page tables, i.e. unprotecting something
5072 * for L1 isn't going to magically fix whatever issue cause L2 to fail.
5074 if (!mmio_info_in_cache(vcpu, cr2_or_gpa, direct) && !is_guest_mode(vcpu))
5075 emulation_type |= EMULTYPE_ALLOW_RETRY_PF;
5077 return x86_emulate_instruction(vcpu, cr2_or_gpa, emulation_type, insn,
5080 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
5082 void kvm_mmu_invalidate_gva(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
5083 gva_t gva, hpa_t root_hpa)
5087 /* It's actually a GPA for vcpu->arch.guest_mmu. */
5088 if (mmu != &vcpu->arch.guest_mmu) {
5089 /* INVLPG on a non-canonical address is a NOP according to the SDM. */
5090 if (is_noncanonical_address(gva, vcpu))
5093 static_call(kvm_x86_tlb_flush_gva)(vcpu, gva);
5099 if (root_hpa == INVALID_PAGE) {
5100 mmu->invlpg(vcpu, gva, mmu->root_hpa);
5103 * INVLPG is required to invalidate any global mappings for the VA,
5104 * irrespective of PCID. Since it would take us roughly similar amount
5105 * of work to determine whether any of the prev_root mappings of the VA
5106 * is marked global, or to just sync it blindly, so we might as well
5107 * just always sync it.
5109 * Mappings not reachable via the current cr3 or the prev_roots will be
5110 * synced when switching to that cr3, so nothing needs to be done here
5113 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5114 if (VALID_PAGE(mmu->prev_roots[i].hpa))
5115 mmu->invlpg(vcpu, gva, mmu->prev_roots[i].hpa);
5117 mmu->invlpg(vcpu, gva, root_hpa);
5120 EXPORT_SYMBOL_GPL(kvm_mmu_invalidate_gva);
5122 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
5124 kvm_mmu_invalidate_gva(vcpu, vcpu->arch.mmu, gva, INVALID_PAGE);
5125 ++vcpu->stat.invlpg;
5127 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
5130 void kvm_mmu_invpcid_gva(struct kvm_vcpu *vcpu, gva_t gva, unsigned long pcid)
5132 struct kvm_mmu *mmu = vcpu->arch.mmu;
5133 bool tlb_flush = false;
5136 if (pcid == kvm_get_active_pcid(vcpu)) {
5137 mmu->invlpg(vcpu, gva, mmu->root_hpa);
5141 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
5142 if (VALID_PAGE(mmu->prev_roots[i].hpa) &&
5143 pcid == kvm_get_pcid(vcpu, mmu->prev_roots[i].pgd)) {
5144 mmu->invlpg(vcpu, gva, mmu->prev_roots[i].hpa);
5150 static_call(kvm_x86_tlb_flush_gva)(vcpu, gva);
5152 ++vcpu->stat.invlpg;
5155 * Mappings not reachable via the current cr3 or the prev_roots will be
5156 * synced when switching to that cr3, so nothing needs to be done here
5160 EXPORT_SYMBOL_GPL(kvm_mmu_invpcid_gva);
5162 void kvm_configure_mmu(bool enable_tdp, int tdp_max_root_level,
5163 int tdp_huge_page_level)
5165 tdp_enabled = enable_tdp;
5166 max_tdp_level = tdp_max_root_level;
5169 * max_huge_page_level reflects KVM's MMU capabilities irrespective
5170 * of kernel support, e.g. KVM may be capable of using 1GB pages when
5171 * the kernel is not. But, KVM never creates a page size greater than
5172 * what is used by the kernel for any given HVA, i.e. the kernel's
5173 * capabilities are ultimately consulted by kvm_mmu_hugepage_adjust().
5176 max_huge_page_level = tdp_huge_page_level;
5177 else if (boot_cpu_has(X86_FEATURE_GBPAGES))
5178 max_huge_page_level = PG_LEVEL_1G;
5180 max_huge_page_level = PG_LEVEL_2M;
5182 EXPORT_SYMBOL_GPL(kvm_configure_mmu);
5184 /* The return value indicates if tlb flush on all vcpus is needed. */
5185 typedef bool (*slot_level_handler) (struct kvm *kvm, struct kvm_rmap_head *rmap_head);
5187 /* The caller should hold mmu-lock before calling this function. */
5188 static __always_inline bool
5189 slot_handle_level_range(struct kvm *kvm, struct kvm_memory_slot *memslot,
5190 slot_level_handler fn, int start_level, int end_level,
5191 gfn_t start_gfn, gfn_t end_gfn, bool lock_flush_tlb)
5193 struct slot_rmap_walk_iterator iterator;
5196 for_each_slot_rmap_range(memslot, start_level, end_level, start_gfn,
5197 end_gfn, &iterator) {
5199 flush |= fn(kvm, iterator.rmap);
5201 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) {
5202 if (flush && lock_flush_tlb) {
5203 kvm_flush_remote_tlbs_with_address(kvm,
5205 iterator.gfn - start_gfn + 1);
5208 cond_resched_rwlock_write(&kvm->mmu_lock);
5212 if (flush && lock_flush_tlb) {
5213 kvm_flush_remote_tlbs_with_address(kvm, start_gfn,
5214 end_gfn - start_gfn + 1);
5221 static __always_inline bool
5222 slot_handle_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
5223 slot_level_handler fn, int start_level, int end_level,
5224 bool lock_flush_tlb)
5226 return slot_handle_level_range(kvm, memslot, fn, start_level,
5227 end_level, memslot->base_gfn,
5228 memslot->base_gfn + memslot->npages - 1,
5232 static __always_inline bool
5233 slot_handle_all_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
5234 slot_level_handler fn, bool lock_flush_tlb)
5236 return slot_handle_level(kvm, memslot, fn, PG_LEVEL_4K,
5237 KVM_MAX_HUGEPAGE_LEVEL, lock_flush_tlb);
5240 static __always_inline bool
5241 slot_handle_large_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
5242 slot_level_handler fn, bool lock_flush_tlb)
5244 return slot_handle_level(kvm, memslot, fn, PG_LEVEL_4K + 1,
5245 KVM_MAX_HUGEPAGE_LEVEL, lock_flush_tlb);
5248 static __always_inline bool
5249 slot_handle_leaf(struct kvm *kvm, struct kvm_memory_slot *memslot,
5250 slot_level_handler fn, bool lock_flush_tlb)
5252 return slot_handle_level(kvm, memslot, fn, PG_LEVEL_4K,
5253 PG_LEVEL_4K, lock_flush_tlb);
5256 static void free_mmu_pages(struct kvm_mmu *mmu)
5258 free_page((unsigned long)mmu->pae_root);
5259 free_page((unsigned long)mmu->lm_root);
5262 static int __kvm_mmu_create(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
5267 mmu->root_hpa = INVALID_PAGE;
5269 mmu->translate_gpa = translate_gpa;
5270 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5271 mmu->prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
5274 * When using PAE paging, the four PDPTEs are treated as 'root' pages,
5275 * while the PDP table is a per-vCPU construct that's allocated at MMU
5276 * creation. When emulating 32-bit mode, cr3 is only 32 bits even on
5277 * x86_64. Therefore we need to allocate the PDP table in the first
5278 * 4GB of memory, which happens to fit the DMA32 zone. Except for
5279 * SVM's 32-bit NPT support, TDP paging doesn't use PAE paging and can
5280 * skip allocating the PDP table.
5282 if (tdp_enabled && kvm_mmu_get_tdp_level(vcpu) > PT32E_ROOT_LEVEL)
5285 page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_DMA32);
5289 mmu->pae_root = page_address(page);
5290 for (i = 0; i < 4; ++i)
5291 mmu->pae_root[i] = INVALID_PAGE;
5296 int kvm_mmu_create(struct kvm_vcpu *vcpu)
5300 vcpu->arch.mmu_pte_list_desc_cache.kmem_cache = pte_list_desc_cache;
5301 vcpu->arch.mmu_pte_list_desc_cache.gfp_zero = __GFP_ZERO;
5303 vcpu->arch.mmu_page_header_cache.kmem_cache = mmu_page_header_cache;
5304 vcpu->arch.mmu_page_header_cache.gfp_zero = __GFP_ZERO;
5306 vcpu->arch.mmu_shadow_page_cache.gfp_zero = __GFP_ZERO;
5308 vcpu->arch.mmu = &vcpu->arch.root_mmu;
5309 vcpu->arch.walk_mmu = &vcpu->arch.root_mmu;
5311 vcpu->arch.nested_mmu.translate_gpa = translate_nested_gpa;
5313 ret = __kvm_mmu_create(vcpu, &vcpu->arch.guest_mmu);
5317 ret = __kvm_mmu_create(vcpu, &vcpu->arch.root_mmu);
5319 goto fail_allocate_root;
5323 free_mmu_pages(&vcpu->arch.guest_mmu);
5327 #define BATCH_ZAP_PAGES 10
5328 static void kvm_zap_obsolete_pages(struct kvm *kvm)
5330 struct kvm_mmu_page *sp, *node;
5331 int nr_zapped, batch = 0;
5334 list_for_each_entry_safe_reverse(sp, node,
5335 &kvm->arch.active_mmu_pages, link) {
5337 * No obsolete valid page exists before a newly created page
5338 * since active_mmu_pages is a FIFO list.
5340 if (!is_obsolete_sp(kvm, sp))
5344 * Invalid pages should never land back on the list of active
5345 * pages. Skip the bogus page, otherwise we'll get stuck in an
5346 * infinite loop if the page gets put back on the list (again).
5348 if (WARN_ON(sp->role.invalid))
5352 * No need to flush the TLB since we're only zapping shadow
5353 * pages with an obsolete generation number and all vCPUS have
5354 * loaded a new root, i.e. the shadow pages being zapped cannot
5355 * be in active use by the guest.
5357 if (batch >= BATCH_ZAP_PAGES &&
5358 cond_resched_rwlock_write(&kvm->mmu_lock)) {
5363 if (__kvm_mmu_prepare_zap_page(kvm, sp,
5364 &kvm->arch.zapped_obsolete_pages, &nr_zapped)) {
5371 * Trigger a remote TLB flush before freeing the page tables to ensure
5372 * KVM is not in the middle of a lockless shadow page table walk, which
5373 * may reference the pages.
5375 kvm_mmu_commit_zap_page(kvm, &kvm->arch.zapped_obsolete_pages);
5379 * Fast invalidate all shadow pages and use lock-break technique
5380 * to zap obsolete pages.
5382 * It's required when memslot is being deleted or VM is being
5383 * destroyed, in these cases, we should ensure that KVM MMU does
5384 * not use any resource of the being-deleted slot or all slots
5385 * after calling the function.
5387 static void kvm_mmu_zap_all_fast(struct kvm *kvm)
5389 lockdep_assert_held(&kvm->slots_lock);
5391 write_lock(&kvm->mmu_lock);
5392 trace_kvm_mmu_zap_all_fast(kvm);
5395 * Toggle mmu_valid_gen between '0' and '1'. Because slots_lock is
5396 * held for the entire duration of zapping obsolete pages, it's
5397 * impossible for there to be multiple invalid generations associated
5398 * with *valid* shadow pages at any given time, i.e. there is exactly
5399 * one valid generation and (at most) one invalid generation.
5401 kvm->arch.mmu_valid_gen = kvm->arch.mmu_valid_gen ? 0 : 1;
5404 * Notify all vcpus to reload its shadow page table and flush TLB.
5405 * Then all vcpus will switch to new shadow page table with the new
5408 * Note: we need to do this under the protection of mmu_lock,
5409 * otherwise, vcpu would purge shadow page but miss tlb flush.
5411 kvm_reload_remote_mmus(kvm);
5413 kvm_zap_obsolete_pages(kvm);
5415 if (is_tdp_mmu_enabled(kvm))
5416 kvm_tdp_mmu_zap_all(kvm);
5418 write_unlock(&kvm->mmu_lock);
5421 static bool kvm_has_zapped_obsolete_pages(struct kvm *kvm)
5423 return unlikely(!list_empty_careful(&kvm->arch.zapped_obsolete_pages));
5426 static void kvm_mmu_invalidate_zap_pages_in_memslot(struct kvm *kvm,
5427 struct kvm_memory_slot *slot,
5428 struct kvm_page_track_notifier_node *node)
5430 kvm_mmu_zap_all_fast(kvm);
5433 void kvm_mmu_init_vm(struct kvm *kvm)
5435 struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
5437 kvm_mmu_init_tdp_mmu(kvm);
5439 node->track_write = kvm_mmu_pte_write;
5440 node->track_flush_slot = kvm_mmu_invalidate_zap_pages_in_memslot;
5441 kvm_page_track_register_notifier(kvm, node);
5444 void kvm_mmu_uninit_vm(struct kvm *kvm)
5446 struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
5448 kvm_page_track_unregister_notifier(kvm, node);
5450 kvm_mmu_uninit_tdp_mmu(kvm);
5453 void kvm_zap_gfn_range(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end)
5455 struct kvm_memslots *slots;
5456 struct kvm_memory_slot *memslot;
5460 write_lock(&kvm->mmu_lock);
5461 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
5462 slots = __kvm_memslots(kvm, i);
5463 kvm_for_each_memslot(memslot, slots) {
5466 start = max(gfn_start, memslot->base_gfn);
5467 end = min(gfn_end, memslot->base_gfn + memslot->npages);
5471 slot_handle_level_range(kvm, memslot, kvm_zap_rmapp,
5473 KVM_MAX_HUGEPAGE_LEVEL,
5474 start, end - 1, true);
5478 if (is_tdp_mmu_enabled(kvm)) {
5479 flush = kvm_tdp_mmu_zap_gfn_range(kvm, gfn_start, gfn_end);
5481 kvm_flush_remote_tlbs(kvm);
5484 write_unlock(&kvm->mmu_lock);
5487 static bool slot_rmap_write_protect(struct kvm *kvm,
5488 struct kvm_rmap_head *rmap_head)
5490 return __rmap_write_protect(kvm, rmap_head, false);
5493 void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
5494 struct kvm_memory_slot *memslot,
5499 write_lock(&kvm->mmu_lock);
5500 flush = slot_handle_level(kvm, memslot, slot_rmap_write_protect,
5501 start_level, KVM_MAX_HUGEPAGE_LEVEL, false);
5502 if (is_tdp_mmu_enabled(kvm))
5503 flush |= kvm_tdp_mmu_wrprot_slot(kvm, memslot, PG_LEVEL_4K);
5504 write_unlock(&kvm->mmu_lock);
5507 * We can flush all the TLBs out of the mmu lock without TLB
5508 * corruption since we just change the spte from writable to
5509 * readonly so that we only need to care the case of changing
5510 * spte from present to present (changing the spte from present
5511 * to nonpresent will flush all the TLBs immediately), in other
5512 * words, the only case we care is mmu_spte_update() where we
5513 * have checked SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE
5514 * instead of PT_WRITABLE_MASK, that means it does not depend
5515 * on PT_WRITABLE_MASK anymore.
5518 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
5521 static bool kvm_mmu_zap_collapsible_spte(struct kvm *kvm,
5522 struct kvm_rmap_head *rmap_head)
5525 struct rmap_iterator iter;
5526 int need_tlb_flush = 0;
5528 struct kvm_mmu_page *sp;
5531 for_each_rmap_spte(rmap_head, &iter, sptep) {
5532 sp = sptep_to_sp(sptep);
5533 pfn = spte_to_pfn(*sptep);
5536 * We cannot do huge page mapping for indirect shadow pages,
5537 * which are found on the last rmap (level = 1) when not using
5538 * tdp; such shadow pages are synced with the page table in
5539 * the guest, and the guest page table is using 4K page size
5540 * mapping if the indirect sp has level = 1.
5542 if (sp->role.direct && !kvm_is_reserved_pfn(pfn) &&
5543 (kvm_is_zone_device_pfn(pfn) ||
5544 PageCompound(pfn_to_page(pfn)))) {
5545 pte_list_remove(rmap_head, sptep);
5547 if (kvm_available_flush_tlb_with_range())
5548 kvm_flush_remote_tlbs_with_address(kvm, sp->gfn,
5549 KVM_PAGES_PER_HPAGE(sp->role.level));
5557 return need_tlb_flush;
5560 void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm,
5561 const struct kvm_memory_slot *memslot)
5563 /* FIXME: const-ify all uses of struct kvm_memory_slot. */
5564 write_lock(&kvm->mmu_lock);
5565 slot_handle_leaf(kvm, (struct kvm_memory_slot *)memslot,
5566 kvm_mmu_zap_collapsible_spte, true);
5568 if (is_tdp_mmu_enabled(kvm))
5569 kvm_tdp_mmu_zap_collapsible_sptes(kvm, memslot);
5570 write_unlock(&kvm->mmu_lock);
5573 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm,
5574 struct kvm_memory_slot *memslot)
5577 * All current use cases for flushing the TLBs for a specific memslot
5578 * are related to dirty logging, and do the TLB flush out of mmu_lock.
5579 * The interaction between the various operations on memslot must be
5580 * serialized by slots_locks to ensure the TLB flush from one operation
5581 * is observed by any other operation on the same memslot.
5583 lockdep_assert_held(&kvm->slots_lock);
5584 kvm_flush_remote_tlbs_with_address(kvm, memslot->base_gfn,
5588 void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm,
5589 struct kvm_memory_slot *memslot)
5593 write_lock(&kvm->mmu_lock);
5594 flush = slot_handle_leaf(kvm, memslot, __rmap_clear_dirty, false);
5595 if (is_tdp_mmu_enabled(kvm))
5596 flush |= kvm_tdp_mmu_clear_dirty_slot(kvm, memslot);
5597 write_unlock(&kvm->mmu_lock);
5600 * It's also safe to flush TLBs out of mmu lock here as currently this
5601 * function is only used for dirty logging, in which case flushing TLB
5602 * out of mmu lock also guarantees no dirty pages will be lost in
5606 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
5608 EXPORT_SYMBOL_GPL(kvm_mmu_slot_leaf_clear_dirty);
5610 void kvm_mmu_slot_largepage_remove_write_access(struct kvm *kvm,
5611 struct kvm_memory_slot *memslot)
5615 write_lock(&kvm->mmu_lock);
5616 flush = slot_handle_large_level(kvm, memslot, slot_rmap_write_protect,
5618 if (is_tdp_mmu_enabled(kvm))
5619 flush |= kvm_tdp_mmu_wrprot_slot(kvm, memslot, PG_LEVEL_2M);
5620 write_unlock(&kvm->mmu_lock);
5623 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
5625 EXPORT_SYMBOL_GPL(kvm_mmu_slot_largepage_remove_write_access);
5627 void kvm_mmu_slot_set_dirty(struct kvm *kvm,
5628 struct kvm_memory_slot *memslot)
5632 write_lock(&kvm->mmu_lock);
5633 flush = slot_handle_all_level(kvm, memslot, __rmap_set_dirty, false);
5634 if (is_tdp_mmu_enabled(kvm))
5635 flush |= kvm_tdp_mmu_slot_set_dirty(kvm, memslot);
5636 write_unlock(&kvm->mmu_lock);
5639 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
5641 EXPORT_SYMBOL_GPL(kvm_mmu_slot_set_dirty);
5643 void kvm_mmu_zap_all(struct kvm *kvm)
5645 struct kvm_mmu_page *sp, *node;
5646 LIST_HEAD(invalid_list);
5649 write_lock(&kvm->mmu_lock);
5651 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link) {
5652 if (WARN_ON(sp->role.invalid))
5654 if (__kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list, &ign))
5656 if (cond_resched_rwlock_write(&kvm->mmu_lock))
5660 kvm_mmu_commit_zap_page(kvm, &invalid_list);
5662 if (is_tdp_mmu_enabled(kvm))
5663 kvm_tdp_mmu_zap_all(kvm);
5665 write_unlock(&kvm->mmu_lock);
5668 void kvm_mmu_invalidate_mmio_sptes(struct kvm *kvm, u64 gen)
5670 WARN_ON(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS);
5672 gen &= MMIO_SPTE_GEN_MASK;
5675 * Generation numbers are incremented in multiples of the number of
5676 * address spaces in order to provide unique generations across all
5677 * address spaces. Strip what is effectively the address space
5678 * modifier prior to checking for a wrap of the MMIO generation so
5679 * that a wrap in any address space is detected.
5681 gen &= ~((u64)KVM_ADDRESS_SPACE_NUM - 1);
5684 * The very rare case: if the MMIO generation number has wrapped,
5685 * zap all shadow pages.
5687 if (unlikely(gen == 0)) {
5688 kvm_debug_ratelimited("kvm: zapping shadow pages for mmio generation wraparound\n");
5689 kvm_mmu_zap_all_fast(kvm);
5693 static unsigned long
5694 mmu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
5697 int nr_to_scan = sc->nr_to_scan;
5698 unsigned long freed = 0;
5700 mutex_lock(&kvm_lock);
5702 list_for_each_entry(kvm, &vm_list, vm_list) {
5704 LIST_HEAD(invalid_list);
5707 * Never scan more than sc->nr_to_scan VM instances.
5708 * Will not hit this condition practically since we do not try
5709 * to shrink more than one VM and it is very unlikely to see
5710 * !n_used_mmu_pages so many times.
5715 * n_used_mmu_pages is accessed without holding kvm->mmu_lock
5716 * here. We may skip a VM instance errorneosly, but we do not
5717 * want to shrink a VM that only started to populate its MMU
5720 if (!kvm->arch.n_used_mmu_pages &&
5721 !kvm_has_zapped_obsolete_pages(kvm))
5724 idx = srcu_read_lock(&kvm->srcu);
5725 write_lock(&kvm->mmu_lock);
5727 if (kvm_has_zapped_obsolete_pages(kvm)) {
5728 kvm_mmu_commit_zap_page(kvm,
5729 &kvm->arch.zapped_obsolete_pages);
5733 freed = kvm_mmu_zap_oldest_mmu_pages(kvm, sc->nr_to_scan);
5736 write_unlock(&kvm->mmu_lock);
5737 srcu_read_unlock(&kvm->srcu, idx);
5740 * unfair on small ones
5741 * per-vm shrinkers cry out
5742 * sadness comes quickly
5744 list_move_tail(&kvm->vm_list, &vm_list);
5748 mutex_unlock(&kvm_lock);
5752 static unsigned long
5753 mmu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
5755 return percpu_counter_read_positive(&kvm_total_used_mmu_pages);
5758 static struct shrinker mmu_shrinker = {
5759 .count_objects = mmu_shrink_count,
5760 .scan_objects = mmu_shrink_scan,
5761 .seeks = DEFAULT_SEEKS * 10,
5764 static void mmu_destroy_caches(void)
5766 kmem_cache_destroy(pte_list_desc_cache);
5767 kmem_cache_destroy(mmu_page_header_cache);
5770 static void kvm_set_mmio_spte_mask(void)
5775 * Set a reserved PA bit in MMIO SPTEs to generate page faults with
5776 * PFEC.RSVD=1 on MMIO accesses. 64-bit PTEs (PAE, x86-64, and EPT
5777 * paging) support a maximum of 52 bits of PA, i.e. if the CPU supports
5778 * 52-bit physical addresses then there are no reserved PA bits in the
5779 * PTEs and so the reserved PA approach must be disabled.
5781 if (shadow_phys_bits < 52)
5782 mask = BIT_ULL(51) | PT_PRESENT_MASK;
5786 kvm_mmu_set_mmio_spte_mask(mask, ACC_WRITE_MASK | ACC_USER_MASK);
5789 static bool get_nx_auto_mode(void)
5791 /* Return true when CPU has the bug, and mitigations are ON */
5792 return boot_cpu_has_bug(X86_BUG_ITLB_MULTIHIT) && !cpu_mitigations_off();
5795 static void __set_nx_huge_pages(bool val)
5797 nx_huge_pages = itlb_multihit_kvm_mitigation = val;
5800 static int set_nx_huge_pages(const char *val, const struct kernel_param *kp)
5802 bool old_val = nx_huge_pages;
5805 /* In "auto" mode deploy workaround only if CPU has the bug. */
5806 if (sysfs_streq(val, "off"))
5808 else if (sysfs_streq(val, "force"))
5810 else if (sysfs_streq(val, "auto"))
5811 new_val = get_nx_auto_mode();
5812 else if (strtobool(val, &new_val) < 0)
5815 __set_nx_huge_pages(new_val);
5817 if (new_val != old_val) {
5820 mutex_lock(&kvm_lock);
5822 list_for_each_entry(kvm, &vm_list, vm_list) {
5823 mutex_lock(&kvm->slots_lock);
5824 kvm_mmu_zap_all_fast(kvm);
5825 mutex_unlock(&kvm->slots_lock);
5827 wake_up_process(kvm->arch.nx_lpage_recovery_thread);
5829 mutex_unlock(&kvm_lock);
5835 int kvm_mmu_module_init(void)
5839 if (nx_huge_pages == -1)
5840 __set_nx_huge_pages(get_nx_auto_mode());
5843 * MMU roles use union aliasing which is, generally speaking, an
5844 * undefined behavior. However, we supposedly know how compilers behave
5845 * and the current status quo is unlikely to change. Guardians below are
5846 * supposed to let us know if the assumption becomes false.
5848 BUILD_BUG_ON(sizeof(union kvm_mmu_page_role) != sizeof(u32));
5849 BUILD_BUG_ON(sizeof(union kvm_mmu_extended_role) != sizeof(u32));
5850 BUILD_BUG_ON(sizeof(union kvm_mmu_role) != sizeof(u64));
5852 kvm_mmu_reset_all_pte_masks();
5854 kvm_set_mmio_spte_mask();
5856 pte_list_desc_cache = kmem_cache_create("pte_list_desc",
5857 sizeof(struct pte_list_desc),
5858 0, SLAB_ACCOUNT, NULL);
5859 if (!pte_list_desc_cache)
5862 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
5863 sizeof(struct kvm_mmu_page),
5864 0, SLAB_ACCOUNT, NULL);
5865 if (!mmu_page_header_cache)
5868 if (percpu_counter_init(&kvm_total_used_mmu_pages, 0, GFP_KERNEL))
5871 ret = register_shrinker(&mmu_shrinker);
5878 mmu_destroy_caches();
5883 * Calculate mmu pages needed for kvm.
5885 unsigned long kvm_mmu_calculate_default_mmu_pages(struct kvm *kvm)
5887 unsigned long nr_mmu_pages;
5888 unsigned long nr_pages = 0;
5889 struct kvm_memslots *slots;
5890 struct kvm_memory_slot *memslot;
5893 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
5894 slots = __kvm_memslots(kvm, i);
5896 kvm_for_each_memslot(memslot, slots)
5897 nr_pages += memslot->npages;
5900 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
5901 nr_mmu_pages = max(nr_mmu_pages, KVM_MIN_ALLOC_MMU_PAGES);
5903 return nr_mmu_pages;
5906 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
5908 kvm_mmu_unload(vcpu);
5909 free_mmu_pages(&vcpu->arch.root_mmu);
5910 free_mmu_pages(&vcpu->arch.guest_mmu);
5911 mmu_free_memory_caches(vcpu);
5914 void kvm_mmu_module_exit(void)
5916 mmu_destroy_caches();
5917 percpu_counter_destroy(&kvm_total_used_mmu_pages);
5918 unregister_shrinker(&mmu_shrinker);
5919 mmu_audit_disable();
5922 static int set_nx_huge_pages_recovery_ratio(const char *val, const struct kernel_param *kp)
5924 unsigned int old_val;
5927 old_val = nx_huge_pages_recovery_ratio;
5928 err = param_set_uint(val, kp);
5932 if (READ_ONCE(nx_huge_pages) &&
5933 !old_val && nx_huge_pages_recovery_ratio) {
5936 mutex_lock(&kvm_lock);
5938 list_for_each_entry(kvm, &vm_list, vm_list)
5939 wake_up_process(kvm->arch.nx_lpage_recovery_thread);
5941 mutex_unlock(&kvm_lock);
5947 static void kvm_recover_nx_lpages(struct kvm *kvm)
5950 struct kvm_mmu_page *sp;
5952 LIST_HEAD(invalid_list);
5955 rcu_idx = srcu_read_lock(&kvm->srcu);
5956 write_lock(&kvm->mmu_lock);
5958 ratio = READ_ONCE(nx_huge_pages_recovery_ratio);
5959 to_zap = ratio ? DIV_ROUND_UP(kvm->stat.nx_lpage_splits, ratio) : 0;
5960 for ( ; to_zap; --to_zap) {
5961 if (list_empty(&kvm->arch.lpage_disallowed_mmu_pages))
5965 * We use a separate list instead of just using active_mmu_pages
5966 * because the number of lpage_disallowed pages is expected to
5967 * be relatively small compared to the total.
5969 sp = list_first_entry(&kvm->arch.lpage_disallowed_mmu_pages,
5970 struct kvm_mmu_page,
5971 lpage_disallowed_link);
5972 WARN_ON_ONCE(!sp->lpage_disallowed);
5973 if (is_tdp_mmu_page(sp)) {
5974 kvm_tdp_mmu_zap_gfn_range(kvm, sp->gfn,
5975 sp->gfn + KVM_PAGES_PER_HPAGE(sp->role.level));
5977 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
5978 WARN_ON_ONCE(sp->lpage_disallowed);
5981 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) {
5982 kvm_mmu_commit_zap_page(kvm, &invalid_list);
5983 cond_resched_rwlock_write(&kvm->mmu_lock);
5986 kvm_mmu_commit_zap_page(kvm, &invalid_list);
5988 write_unlock(&kvm->mmu_lock);
5989 srcu_read_unlock(&kvm->srcu, rcu_idx);
5992 static long get_nx_lpage_recovery_timeout(u64 start_time)
5994 return READ_ONCE(nx_huge_pages) && READ_ONCE(nx_huge_pages_recovery_ratio)
5995 ? start_time + 60 * HZ - get_jiffies_64()
5996 : MAX_SCHEDULE_TIMEOUT;
5999 static int kvm_nx_lpage_recovery_worker(struct kvm *kvm, uintptr_t data)
6002 long remaining_time;
6005 start_time = get_jiffies_64();
6006 remaining_time = get_nx_lpage_recovery_timeout(start_time);
6008 set_current_state(TASK_INTERRUPTIBLE);
6009 while (!kthread_should_stop() && remaining_time > 0) {
6010 schedule_timeout(remaining_time);
6011 remaining_time = get_nx_lpage_recovery_timeout(start_time);
6012 set_current_state(TASK_INTERRUPTIBLE);
6015 set_current_state(TASK_RUNNING);
6017 if (kthread_should_stop())
6020 kvm_recover_nx_lpages(kvm);
6024 int kvm_mmu_post_init_vm(struct kvm *kvm)
6028 err = kvm_vm_create_worker_thread(kvm, kvm_nx_lpage_recovery_worker, 0,
6029 "kvm-nx-lpage-recovery",
6030 &kvm->arch.nx_lpage_recovery_thread);
6032 kthread_unpark(kvm->arch.nx_lpage_recovery_thread);
6037 void kvm_mmu_pre_destroy_vm(struct kvm *kvm)
6039 if (kvm->arch.nx_lpage_recovery_thread)
6040 kthread_stop(kvm->arch.nx_lpage_recovery_thread);