2 * Kernel-based Virtual Machine driver for Linux
4 * This module enables machines with Intel VT-x extensions to run virtual
5 * machines without emulation or binary translation.
9 * Copyright (C) 2006 Qumranet, Inc.
10 * Copyright 2010 Red Hat, Inc. and/or its affiliates.
13 * Yaniv Kamay <yaniv@qumranet.com>
14 * Avi Kivity <avi@qumranet.com>
16 * This work is licensed under the terms of the GNU GPL, version 2. See
17 * the COPYING file in the top-level directory.
24 #include "kvm_cache_regs.h"
27 #include <linux/kvm_host.h>
28 #include <linux/types.h>
29 #include <linux/string.h>
31 #include <linux/highmem.h>
32 #include <linux/module.h>
33 #include <linux/swap.h>
34 #include <linux/hugetlb.h>
35 #include <linux/compiler.h>
36 #include <linux/srcu.h>
37 #include <linux/slab.h>
38 #include <linux/uaccess.h>
41 #include <asm/cmpxchg.h>
46 * When setting this variable to true it enables Two-Dimensional-Paging
47 * where the hardware walks 2 page tables:
48 * 1. the guest-virtual to guest-physical
49 * 2. while doing 1. it walks guest-physical to host-physical
50 * If the hardware supports that we don't need to do shadow paging.
52 bool tdp_enabled = false;
56 AUDIT_POST_PAGE_FAULT,
67 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
68 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
72 #define pgprintk(x...) do { } while (0)
73 #define rmap_printk(x...) do { } while (0)
79 module_param(dbg, bool, 0644);
83 #define ASSERT(x) do { } while (0)
87 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
88 __FILE__, __LINE__, #x); \
92 #define PTE_PREFETCH_NUM 8
94 #define PT_FIRST_AVAIL_BITS_SHIFT 10
95 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
97 #define PT64_LEVEL_BITS 9
99 #define PT64_LEVEL_SHIFT(level) \
100 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
102 #define PT64_INDEX(address, level)\
103 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
106 #define PT32_LEVEL_BITS 10
108 #define PT32_LEVEL_SHIFT(level) \
109 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
111 #define PT32_LVL_OFFSET_MASK(level) \
112 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
113 * PT32_LEVEL_BITS))) - 1))
115 #define PT32_INDEX(address, level)\
116 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
119 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
120 #define PT64_DIR_BASE_ADDR_MASK \
121 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
122 #define PT64_LVL_ADDR_MASK(level) \
123 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
124 * PT64_LEVEL_BITS))) - 1))
125 #define PT64_LVL_OFFSET_MASK(level) \
126 (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
127 * PT64_LEVEL_BITS))) - 1))
129 #define PT32_BASE_ADDR_MASK PAGE_MASK
130 #define PT32_DIR_BASE_ADDR_MASK \
131 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
132 #define PT32_LVL_ADDR_MASK(level) \
133 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
134 * PT32_LEVEL_BITS))) - 1))
136 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | shadow_user_mask \
137 | shadow_x_mask | shadow_nx_mask)
139 #define ACC_EXEC_MASK 1
140 #define ACC_WRITE_MASK PT_WRITABLE_MASK
141 #define ACC_USER_MASK PT_USER_MASK
142 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
144 #include <trace/events/kvm.h>
146 #define CREATE_TRACE_POINTS
147 #include "mmutrace.h"
149 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
150 #define SPTE_MMU_WRITEABLE (1ULL << (PT_FIRST_AVAIL_BITS_SHIFT + 1))
152 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
154 /* make pte_list_desc fit well in cache line */
155 #define PTE_LIST_EXT 3
157 struct pte_list_desc {
158 u64 *sptes[PTE_LIST_EXT];
159 struct pte_list_desc *more;
162 struct kvm_shadow_walk_iterator {
170 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
171 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
172 shadow_walk_okay(&(_walker)); \
173 shadow_walk_next(&(_walker)))
175 #define for_each_shadow_entry_lockless(_vcpu, _addr, _walker, spte) \
176 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
177 shadow_walk_okay(&(_walker)) && \
178 ({ spte = mmu_spte_get_lockless(_walker.sptep); 1; }); \
179 __shadow_walk_next(&(_walker), spte))
181 static struct kmem_cache *pte_list_desc_cache;
182 static struct kmem_cache *mmu_page_header_cache;
183 static struct percpu_counter kvm_total_used_mmu_pages;
185 static u64 __read_mostly shadow_nx_mask;
186 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
187 static u64 __read_mostly shadow_user_mask;
188 static u64 __read_mostly shadow_accessed_mask;
189 static u64 __read_mostly shadow_dirty_mask;
190 static u64 __read_mostly shadow_mmio_mask;
192 static void mmu_spte_set(u64 *sptep, u64 spte);
193 static void mmu_free_roots(struct kvm_vcpu *vcpu);
195 void kvm_mmu_set_mmio_spte_mask(u64 mmio_mask)
197 shadow_mmio_mask = mmio_mask;
199 EXPORT_SYMBOL_GPL(kvm_mmu_set_mmio_spte_mask);
202 * the low bit of the generation number is always presumed to be zero.
203 * This disables mmio caching during memslot updates. The concept is
204 * similar to a seqcount but instead of retrying the access we just punt
205 * and ignore the cache.
207 * spte bits 3-11 are used as bits 1-9 of the generation number,
208 * the bits 52-61 are used as bits 10-19 of the generation number.
210 #define MMIO_SPTE_GEN_LOW_SHIFT 2
211 #define MMIO_SPTE_GEN_HIGH_SHIFT 52
213 #define MMIO_GEN_SHIFT 20
214 #define MMIO_GEN_LOW_SHIFT 10
215 #define MMIO_GEN_LOW_MASK ((1 << MMIO_GEN_LOW_SHIFT) - 2)
216 #define MMIO_GEN_MASK ((1 << MMIO_GEN_SHIFT) - 1)
217 #define MMIO_MAX_GEN ((1 << MMIO_GEN_SHIFT) - 1)
219 static u64 generation_mmio_spte_mask(unsigned int gen)
223 WARN_ON(gen > MMIO_MAX_GEN);
225 mask = (gen & MMIO_GEN_LOW_MASK) << MMIO_SPTE_GEN_LOW_SHIFT;
226 mask |= ((u64)gen >> MMIO_GEN_LOW_SHIFT) << MMIO_SPTE_GEN_HIGH_SHIFT;
230 static unsigned int get_mmio_spte_generation(u64 spte)
234 spte &= ~shadow_mmio_mask;
236 gen = (spte >> MMIO_SPTE_GEN_LOW_SHIFT) & MMIO_GEN_LOW_MASK;
237 gen |= (spte >> MMIO_SPTE_GEN_HIGH_SHIFT) << MMIO_GEN_LOW_SHIFT;
241 static unsigned int kvm_current_mmio_generation(struct kvm *kvm)
243 return kvm_memslots(kvm)->generation & MMIO_GEN_MASK;
246 static void mark_mmio_spte(struct kvm *kvm, u64 *sptep, u64 gfn,
249 unsigned int gen = kvm_current_mmio_generation(kvm);
250 u64 mask = generation_mmio_spte_mask(gen);
252 access &= ACC_WRITE_MASK | ACC_USER_MASK;
253 mask |= shadow_mmio_mask | access | gfn << PAGE_SHIFT;
255 trace_mark_mmio_spte(sptep, gfn, access, gen);
256 mmu_spte_set(sptep, mask);
259 static bool is_mmio_spte(u64 spte)
261 return (spte & shadow_mmio_mask) == shadow_mmio_mask;
264 static gfn_t get_mmio_spte_gfn(u64 spte)
266 u64 mask = generation_mmio_spte_mask(MMIO_MAX_GEN) | shadow_mmio_mask;
267 return (spte & ~mask) >> PAGE_SHIFT;
270 static unsigned get_mmio_spte_access(u64 spte)
272 u64 mask = generation_mmio_spte_mask(MMIO_MAX_GEN) | shadow_mmio_mask;
273 return (spte & ~mask) & ~PAGE_MASK;
276 static bool set_mmio_spte(struct kvm *kvm, u64 *sptep, gfn_t gfn,
277 pfn_t pfn, unsigned access)
279 if (unlikely(is_noslot_pfn(pfn))) {
280 mark_mmio_spte(kvm, sptep, gfn, access);
287 static bool check_mmio_spte(struct kvm *kvm, u64 spte)
289 unsigned int kvm_gen, spte_gen;
291 kvm_gen = kvm_current_mmio_generation(kvm);
292 spte_gen = get_mmio_spte_generation(spte);
294 trace_check_mmio_spte(spte, kvm_gen, spte_gen);
295 return likely(kvm_gen == spte_gen);
298 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
299 u64 dirty_mask, u64 nx_mask, u64 x_mask)
301 shadow_user_mask = user_mask;
302 shadow_accessed_mask = accessed_mask;
303 shadow_dirty_mask = dirty_mask;
304 shadow_nx_mask = nx_mask;
305 shadow_x_mask = x_mask;
307 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
309 static int is_cpuid_PSE36(void)
314 static int is_nx(struct kvm_vcpu *vcpu)
316 return vcpu->arch.efer & EFER_NX;
319 static int is_shadow_present_pte(u64 pte)
321 return pte & PT_PRESENT_MASK && !is_mmio_spte(pte);
324 static int is_large_pte(u64 pte)
326 return pte & PT_PAGE_SIZE_MASK;
329 static int is_rmap_spte(u64 pte)
331 return is_shadow_present_pte(pte);
334 static int is_last_spte(u64 pte, int level)
336 if (level == PT_PAGE_TABLE_LEVEL)
338 if (is_large_pte(pte))
343 static pfn_t spte_to_pfn(u64 pte)
345 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
348 static gfn_t pse36_gfn_delta(u32 gpte)
350 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
352 return (gpte & PT32_DIR_PSE36_MASK) << shift;
356 static void __set_spte(u64 *sptep, u64 spte)
361 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
366 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
368 return xchg(sptep, spte);
371 static u64 __get_spte_lockless(u64 *sptep)
373 return ACCESS_ONCE(*sptep);
376 static bool __check_direct_spte_mmio_pf(u64 spte)
378 /* It is valid if the spte is zapped. */
390 static void count_spte_clear(u64 *sptep, u64 spte)
392 struct kvm_mmu_page *sp = page_header(__pa(sptep));
394 if (is_shadow_present_pte(spte))
397 /* Ensure the spte is completely set before we increase the count */
399 sp->clear_spte_count++;
402 static void __set_spte(u64 *sptep, u64 spte)
404 union split_spte *ssptep, sspte;
406 ssptep = (union split_spte *)sptep;
407 sspte = (union split_spte)spte;
409 ssptep->spte_high = sspte.spte_high;
412 * If we map the spte from nonpresent to present, We should store
413 * the high bits firstly, then set present bit, so cpu can not
414 * fetch this spte while we are setting the spte.
418 ssptep->spte_low = sspte.spte_low;
421 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
423 union split_spte *ssptep, sspte;
425 ssptep = (union split_spte *)sptep;
426 sspte = (union split_spte)spte;
428 ssptep->spte_low = sspte.spte_low;
431 * If we map the spte from present to nonpresent, we should clear
432 * present bit firstly to avoid vcpu fetch the old high bits.
436 ssptep->spte_high = sspte.spte_high;
437 count_spte_clear(sptep, spte);
440 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
442 union split_spte *ssptep, sspte, orig;
444 ssptep = (union split_spte *)sptep;
445 sspte = (union split_spte)spte;
447 /* xchg acts as a barrier before the setting of the high bits */
448 orig.spte_low = xchg(&ssptep->spte_low, sspte.spte_low);
449 orig.spte_high = ssptep->spte_high;
450 ssptep->spte_high = sspte.spte_high;
451 count_spte_clear(sptep, spte);
457 * The idea using the light way get the spte on x86_32 guest is from
458 * gup_get_pte(arch/x86/mm/gup.c).
460 * An spte tlb flush may be pending, because kvm_set_pte_rmapp
461 * coalesces them and we are running out of the MMU lock. Therefore
462 * we need to protect against in-progress updates of the spte.
464 * Reading the spte while an update is in progress may get the old value
465 * for the high part of the spte. The race is fine for a present->non-present
466 * change (because the high part of the spte is ignored for non-present spte),
467 * but for a present->present change we must reread the spte.
469 * All such changes are done in two steps (present->non-present and
470 * non-present->present), hence it is enough to count the number of
471 * present->non-present updates: if it changed while reading the spte,
472 * we might have hit the race. This is done using clear_spte_count.
474 static u64 __get_spte_lockless(u64 *sptep)
476 struct kvm_mmu_page *sp = page_header(__pa(sptep));
477 union split_spte spte, *orig = (union split_spte *)sptep;
481 count = sp->clear_spte_count;
484 spte.spte_low = orig->spte_low;
487 spte.spte_high = orig->spte_high;
490 if (unlikely(spte.spte_low != orig->spte_low ||
491 count != sp->clear_spte_count))
497 static bool __check_direct_spte_mmio_pf(u64 spte)
499 union split_spte sspte = (union split_spte)spte;
500 u32 high_mmio_mask = shadow_mmio_mask >> 32;
502 /* It is valid if the spte is zapped. */
506 /* It is valid if the spte is being zapped. */
507 if (sspte.spte_low == 0ull &&
508 (sspte.spte_high & high_mmio_mask) == high_mmio_mask)
515 static bool spte_is_locklessly_modifiable(u64 spte)
517 return (spte & (SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE)) ==
518 (SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE);
521 static bool spte_has_volatile_bits(u64 spte)
524 * Always atomicly update spte if it can be updated
525 * out of mmu-lock, it can ensure dirty bit is not lost,
526 * also, it can help us to get a stable is_writable_pte()
527 * to ensure tlb flush is not missed.
529 if (spte_is_locklessly_modifiable(spte))
532 if (!shadow_accessed_mask)
535 if (!is_shadow_present_pte(spte))
538 if ((spte & shadow_accessed_mask) &&
539 (!is_writable_pte(spte) || (spte & shadow_dirty_mask)))
545 static bool spte_is_bit_cleared(u64 old_spte, u64 new_spte, u64 bit_mask)
547 return (old_spte & bit_mask) && !(new_spte & bit_mask);
550 /* Rules for using mmu_spte_set:
551 * Set the sptep from nonpresent to present.
552 * Note: the sptep being assigned *must* be either not present
553 * or in a state where the hardware will not attempt to update
556 static void mmu_spte_set(u64 *sptep, u64 new_spte)
558 WARN_ON(is_shadow_present_pte(*sptep));
559 __set_spte(sptep, new_spte);
562 /* Rules for using mmu_spte_update:
563 * Update the state bits, it means the mapped pfn is not changged.
565 * Whenever we overwrite a writable spte with a read-only one we
566 * should flush remote TLBs. Otherwise rmap_write_protect
567 * will find a read-only spte, even though the writable spte
568 * might be cached on a CPU's TLB, the return value indicates this
571 static bool mmu_spte_update(u64 *sptep, u64 new_spte)
573 u64 old_spte = *sptep;
576 WARN_ON(!is_rmap_spte(new_spte));
578 if (!is_shadow_present_pte(old_spte)) {
579 mmu_spte_set(sptep, new_spte);
583 if (!spte_has_volatile_bits(old_spte))
584 __update_clear_spte_fast(sptep, new_spte);
586 old_spte = __update_clear_spte_slow(sptep, new_spte);
589 * For the spte updated out of mmu-lock is safe, since
590 * we always atomicly update it, see the comments in
591 * spte_has_volatile_bits().
593 if (spte_is_locklessly_modifiable(old_spte) &&
594 !is_writable_pte(new_spte))
597 if (!shadow_accessed_mask)
600 if (spte_is_bit_cleared(old_spte, new_spte, shadow_accessed_mask))
601 kvm_set_pfn_accessed(spte_to_pfn(old_spte));
602 if (spte_is_bit_cleared(old_spte, new_spte, shadow_dirty_mask))
603 kvm_set_pfn_dirty(spte_to_pfn(old_spte));
609 * Rules for using mmu_spte_clear_track_bits:
610 * It sets the sptep from present to nonpresent, and track the
611 * state bits, it is used to clear the last level sptep.
613 static int mmu_spte_clear_track_bits(u64 *sptep)
616 u64 old_spte = *sptep;
618 if (!spte_has_volatile_bits(old_spte))
619 __update_clear_spte_fast(sptep, 0ull);
621 old_spte = __update_clear_spte_slow(sptep, 0ull);
623 if (!is_rmap_spte(old_spte))
626 pfn = spte_to_pfn(old_spte);
629 * KVM does not hold the refcount of the page used by
630 * kvm mmu, before reclaiming the page, we should
631 * unmap it from mmu first.
633 WARN_ON(!kvm_is_mmio_pfn(pfn) && !page_count(pfn_to_page(pfn)));
635 if (!shadow_accessed_mask || old_spte & shadow_accessed_mask)
636 kvm_set_pfn_accessed(pfn);
637 if (!shadow_dirty_mask || (old_spte & shadow_dirty_mask))
638 kvm_set_pfn_dirty(pfn);
643 * Rules for using mmu_spte_clear_no_track:
644 * Directly clear spte without caring the state bits of sptep,
645 * it is used to set the upper level spte.
647 static void mmu_spte_clear_no_track(u64 *sptep)
649 __update_clear_spte_fast(sptep, 0ull);
652 static u64 mmu_spte_get_lockless(u64 *sptep)
654 return __get_spte_lockless(sptep);
657 static void walk_shadow_page_lockless_begin(struct kvm_vcpu *vcpu)
660 * Prevent page table teardown by making any free-er wait during
661 * kvm_flush_remote_tlbs() IPI to all active vcpus.
664 vcpu->mode = READING_SHADOW_PAGE_TABLES;
666 * Make sure a following spte read is not reordered ahead of the write
672 static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
675 * Make sure the write to vcpu->mode is not reordered in front of
676 * reads to sptes. If it does, kvm_commit_zap_page() can see us
677 * OUTSIDE_GUEST_MODE and proceed to free the shadow page table.
680 vcpu->mode = OUTSIDE_GUEST_MODE;
684 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
685 struct kmem_cache *base_cache, int min)
689 if (cache->nobjs >= min)
691 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
692 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
695 cache->objects[cache->nobjs++] = obj;
700 static int mmu_memory_cache_free_objects(struct kvm_mmu_memory_cache *cache)
705 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc,
706 struct kmem_cache *cache)
709 kmem_cache_free(cache, mc->objects[--mc->nobjs]);
712 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
717 if (cache->nobjs >= min)
719 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
720 page = (void *)__get_free_page(GFP_KERNEL);
723 cache->objects[cache->nobjs++] = page;
728 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
731 free_page((unsigned long)mc->objects[--mc->nobjs]);
734 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
738 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
739 pte_list_desc_cache, 8 + PTE_PREFETCH_NUM);
742 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
745 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
746 mmu_page_header_cache, 4);
751 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
753 mmu_free_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
754 pte_list_desc_cache);
755 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
756 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache,
757 mmu_page_header_cache);
760 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
765 p = mc->objects[--mc->nobjs];
769 static struct pte_list_desc *mmu_alloc_pte_list_desc(struct kvm_vcpu *vcpu)
771 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_list_desc_cache);
774 static void mmu_free_pte_list_desc(struct pte_list_desc *pte_list_desc)
776 kmem_cache_free(pte_list_desc_cache, pte_list_desc);
779 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
781 if (!sp->role.direct)
782 return sp->gfns[index];
784 return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
787 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
790 BUG_ON(gfn != kvm_mmu_page_get_gfn(sp, index));
792 sp->gfns[index] = gfn;
796 * Return the pointer to the large page information for a given gfn,
797 * handling slots that are not large page aligned.
799 static struct kvm_lpage_info *lpage_info_slot(gfn_t gfn,
800 struct kvm_memory_slot *slot,
805 idx = gfn_to_index(gfn, slot->base_gfn, level);
806 return &slot->arch.lpage_info[level - 2][idx];
809 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
811 struct kvm_memory_slot *slot;
812 struct kvm_lpage_info *linfo;
815 slot = gfn_to_memslot(kvm, gfn);
816 for (i = PT_DIRECTORY_LEVEL;
817 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
818 linfo = lpage_info_slot(gfn, slot, i);
819 linfo->write_count += 1;
821 kvm->arch.indirect_shadow_pages++;
824 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
826 struct kvm_memory_slot *slot;
827 struct kvm_lpage_info *linfo;
830 slot = gfn_to_memslot(kvm, gfn);
831 for (i = PT_DIRECTORY_LEVEL;
832 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
833 linfo = lpage_info_slot(gfn, slot, i);
834 linfo->write_count -= 1;
835 WARN_ON(linfo->write_count < 0);
837 kvm->arch.indirect_shadow_pages--;
840 static int has_wrprotected_page(struct kvm *kvm,
844 struct kvm_memory_slot *slot;
845 struct kvm_lpage_info *linfo;
847 slot = gfn_to_memslot(kvm, gfn);
849 linfo = lpage_info_slot(gfn, slot, level);
850 return linfo->write_count;
856 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
858 unsigned long page_size;
861 page_size = kvm_host_page_size(kvm, gfn);
863 for (i = PT_PAGE_TABLE_LEVEL;
864 i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
865 if (page_size >= KVM_HPAGE_SIZE(i))
874 static struct kvm_memory_slot *
875 gfn_to_memslot_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t gfn,
878 struct kvm_memory_slot *slot;
880 slot = gfn_to_memslot(vcpu->kvm, gfn);
881 if (!slot || slot->flags & KVM_MEMSLOT_INVALID ||
882 (no_dirty_log && slot->dirty_bitmap))
888 static bool mapping_level_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t large_gfn)
890 return !gfn_to_memslot_dirty_bitmap(vcpu, large_gfn, true);
893 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
895 int host_level, level, max_level;
897 host_level = host_mapping_level(vcpu->kvm, large_gfn);
899 if (host_level == PT_PAGE_TABLE_LEVEL)
902 max_level = min(kvm_x86_ops->get_lpage_level(), host_level);
904 for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
905 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
912 * Pte mapping structures:
914 * If pte_list bit zero is zero, then pte_list point to the spte.
916 * If pte_list bit zero is one, (then pte_list & ~1) points to a struct
917 * pte_list_desc containing more mappings.
919 * Returns the number of pte entries before the spte was added or zero if
920 * the spte was not added.
923 static int pte_list_add(struct kvm_vcpu *vcpu, u64 *spte,
924 unsigned long *pte_list)
926 struct pte_list_desc *desc;
930 rmap_printk("pte_list_add: %p %llx 0->1\n", spte, *spte);
931 *pte_list = (unsigned long)spte;
932 } else if (!(*pte_list & 1)) {
933 rmap_printk("pte_list_add: %p %llx 1->many\n", spte, *spte);
934 desc = mmu_alloc_pte_list_desc(vcpu);
935 desc->sptes[0] = (u64 *)*pte_list;
936 desc->sptes[1] = spte;
937 *pte_list = (unsigned long)desc | 1;
940 rmap_printk("pte_list_add: %p %llx many->many\n", spte, *spte);
941 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
942 while (desc->sptes[PTE_LIST_EXT-1] && desc->more) {
944 count += PTE_LIST_EXT;
946 if (desc->sptes[PTE_LIST_EXT-1]) {
947 desc->more = mmu_alloc_pte_list_desc(vcpu);
950 for (i = 0; desc->sptes[i]; ++i)
952 desc->sptes[i] = spte;
958 pte_list_desc_remove_entry(unsigned long *pte_list, struct pte_list_desc *desc,
959 int i, struct pte_list_desc *prev_desc)
963 for (j = PTE_LIST_EXT - 1; !desc->sptes[j] && j > i; --j)
965 desc->sptes[i] = desc->sptes[j];
966 desc->sptes[j] = NULL;
969 if (!prev_desc && !desc->more)
970 *pte_list = (unsigned long)desc->sptes[0];
973 prev_desc->more = desc->more;
975 *pte_list = (unsigned long)desc->more | 1;
976 mmu_free_pte_list_desc(desc);
979 static void pte_list_remove(u64 *spte, unsigned long *pte_list)
981 struct pte_list_desc *desc;
982 struct pte_list_desc *prev_desc;
986 printk(KERN_ERR "pte_list_remove: %p 0->BUG\n", spte);
988 } else if (!(*pte_list & 1)) {
989 rmap_printk("pte_list_remove: %p 1->0\n", spte);
990 if ((u64 *)*pte_list != spte) {
991 printk(KERN_ERR "pte_list_remove: %p 1->BUG\n", spte);
996 rmap_printk("pte_list_remove: %p many->many\n", spte);
997 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
1000 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i)
1001 if (desc->sptes[i] == spte) {
1002 pte_list_desc_remove_entry(pte_list,
1010 pr_err("pte_list_remove: %p many->many\n", spte);
1015 typedef void (*pte_list_walk_fn) (u64 *spte);
1016 static void pte_list_walk(unsigned long *pte_list, pte_list_walk_fn fn)
1018 struct pte_list_desc *desc;
1024 if (!(*pte_list & 1))
1025 return fn((u64 *)*pte_list);
1027 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
1029 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i)
1035 static unsigned long *__gfn_to_rmap(gfn_t gfn, int level,
1036 struct kvm_memory_slot *slot)
1040 idx = gfn_to_index(gfn, slot->base_gfn, level);
1041 return &slot->arch.rmap[level - PT_PAGE_TABLE_LEVEL][idx];
1045 * Take gfn and return the reverse mapping to it.
1047 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
1049 struct kvm_memory_slot *slot;
1051 slot = gfn_to_memslot(kvm, gfn);
1052 return __gfn_to_rmap(gfn, level, slot);
1055 static bool rmap_can_add(struct kvm_vcpu *vcpu)
1057 struct kvm_mmu_memory_cache *cache;
1059 cache = &vcpu->arch.mmu_pte_list_desc_cache;
1060 return mmu_memory_cache_free_objects(cache);
1063 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
1065 struct kvm_mmu_page *sp;
1066 unsigned long *rmapp;
1068 sp = page_header(__pa(spte));
1069 kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
1070 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
1071 return pte_list_add(vcpu, spte, rmapp);
1074 static void rmap_remove(struct kvm *kvm, u64 *spte)
1076 struct kvm_mmu_page *sp;
1078 unsigned long *rmapp;
1080 sp = page_header(__pa(spte));
1081 gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
1082 rmapp = gfn_to_rmap(kvm, gfn, sp->role.level);
1083 pte_list_remove(spte, rmapp);
1087 * Used by the following functions to iterate through the sptes linked by a
1088 * rmap. All fields are private and not assumed to be used outside.
1090 struct rmap_iterator {
1091 /* private fields */
1092 struct pte_list_desc *desc; /* holds the sptep if not NULL */
1093 int pos; /* index of the sptep */
1097 * Iteration must be started by this function. This should also be used after
1098 * removing/dropping sptes from the rmap link because in such cases the
1099 * information in the itererator may not be valid.
1101 * Returns sptep if found, NULL otherwise.
1103 static u64 *rmap_get_first(unsigned long rmap, struct rmap_iterator *iter)
1113 iter->desc = (struct pte_list_desc *)(rmap & ~1ul);
1115 return iter->desc->sptes[iter->pos];
1119 * Must be used with a valid iterator: e.g. after rmap_get_first().
1121 * Returns sptep if found, NULL otherwise.
1123 static u64 *rmap_get_next(struct rmap_iterator *iter)
1126 if (iter->pos < PTE_LIST_EXT - 1) {
1130 sptep = iter->desc->sptes[iter->pos];
1135 iter->desc = iter->desc->more;
1139 /* desc->sptes[0] cannot be NULL */
1140 return iter->desc->sptes[iter->pos];
1147 static void drop_spte(struct kvm *kvm, u64 *sptep)
1149 if (mmu_spte_clear_track_bits(sptep))
1150 rmap_remove(kvm, sptep);
1154 static bool __drop_large_spte(struct kvm *kvm, u64 *sptep)
1156 if (is_large_pte(*sptep)) {
1157 WARN_ON(page_header(__pa(sptep))->role.level ==
1158 PT_PAGE_TABLE_LEVEL);
1159 drop_spte(kvm, sptep);
1167 static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1169 if (__drop_large_spte(vcpu->kvm, sptep))
1170 kvm_flush_remote_tlbs(vcpu->kvm);
1174 * Write-protect on the specified @sptep, @pt_protect indicates whether
1175 * spte write-protection is caused by protecting shadow page table.
1177 * Note: write protection is difference between drity logging and spte
1179 * - for dirty logging, the spte can be set to writable at anytime if
1180 * its dirty bitmap is properly set.
1181 * - for spte protection, the spte can be writable only after unsync-ing
1184 * Return true if tlb need be flushed.
1186 static bool spte_write_protect(struct kvm *kvm, u64 *sptep, bool pt_protect)
1190 if (!is_writable_pte(spte) &&
1191 !(pt_protect && spte_is_locklessly_modifiable(spte)))
1194 rmap_printk("rmap_write_protect: spte %p %llx\n", sptep, *sptep);
1197 spte &= ~SPTE_MMU_WRITEABLE;
1198 spte = spte & ~PT_WRITABLE_MASK;
1200 return mmu_spte_update(sptep, spte);
1203 static bool __rmap_write_protect(struct kvm *kvm, unsigned long *rmapp,
1207 struct rmap_iterator iter;
1210 for (sptep = rmap_get_first(*rmapp, &iter); sptep;) {
1211 BUG_ON(!(*sptep & PT_PRESENT_MASK));
1213 flush |= spte_write_protect(kvm, sptep, pt_protect);
1214 sptep = rmap_get_next(&iter);
1221 * kvm_mmu_write_protect_pt_masked - write protect selected PT level pages
1222 * @kvm: kvm instance
1223 * @slot: slot to protect
1224 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1225 * @mask: indicates which pages we should protect
1227 * Used when we do not need to care about huge page mappings: e.g. during dirty
1228 * logging we do not have any such mappings.
1230 void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1231 struct kvm_memory_slot *slot,
1232 gfn_t gfn_offset, unsigned long mask)
1234 unsigned long *rmapp;
1237 rmapp = __gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1238 PT_PAGE_TABLE_LEVEL, slot);
1239 __rmap_write_protect(kvm, rmapp, false);
1241 /* clear the first set bit */
1246 static bool rmap_write_protect(struct kvm *kvm, u64 gfn)
1248 struct kvm_memory_slot *slot;
1249 unsigned long *rmapp;
1251 bool write_protected = false;
1253 slot = gfn_to_memslot(kvm, gfn);
1255 for (i = PT_PAGE_TABLE_LEVEL;
1256 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
1257 rmapp = __gfn_to_rmap(gfn, i, slot);
1258 write_protected |= __rmap_write_protect(kvm, rmapp, true);
1261 return write_protected;
1264 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
1265 struct kvm_memory_slot *slot, unsigned long data)
1268 struct rmap_iterator iter;
1269 int need_tlb_flush = 0;
1271 while ((sptep = rmap_get_first(*rmapp, &iter))) {
1272 BUG_ON(!(*sptep & PT_PRESENT_MASK));
1273 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", sptep, *sptep);
1275 drop_spte(kvm, sptep);
1279 return need_tlb_flush;
1282 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
1283 struct kvm_memory_slot *slot, unsigned long data)
1286 struct rmap_iterator iter;
1289 pte_t *ptep = (pte_t *)data;
1292 WARN_ON(pte_huge(*ptep));
1293 new_pfn = pte_pfn(*ptep);
1295 for (sptep = rmap_get_first(*rmapp, &iter); sptep;) {
1296 BUG_ON(!is_shadow_present_pte(*sptep));
1297 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", sptep, *sptep);
1301 if (pte_write(*ptep)) {
1302 drop_spte(kvm, sptep);
1303 sptep = rmap_get_first(*rmapp, &iter);
1305 new_spte = *sptep & ~PT64_BASE_ADDR_MASK;
1306 new_spte |= (u64)new_pfn << PAGE_SHIFT;
1308 new_spte &= ~PT_WRITABLE_MASK;
1309 new_spte &= ~SPTE_HOST_WRITEABLE;
1310 new_spte &= ~shadow_accessed_mask;
1312 mmu_spte_clear_track_bits(sptep);
1313 mmu_spte_set(sptep, new_spte);
1314 sptep = rmap_get_next(&iter);
1319 kvm_flush_remote_tlbs(kvm);
1324 static int kvm_handle_hva_range(struct kvm *kvm,
1325 unsigned long start,
1328 int (*handler)(struct kvm *kvm,
1329 unsigned long *rmapp,
1330 struct kvm_memory_slot *slot,
1331 unsigned long data))
1335 struct kvm_memslots *slots;
1336 struct kvm_memory_slot *memslot;
1338 slots = kvm_memslots(kvm);
1340 kvm_for_each_memslot(memslot, slots) {
1341 unsigned long hva_start, hva_end;
1342 gfn_t gfn_start, gfn_end;
1344 hva_start = max(start, memslot->userspace_addr);
1345 hva_end = min(end, memslot->userspace_addr +
1346 (memslot->npages << PAGE_SHIFT));
1347 if (hva_start >= hva_end)
1350 * {gfn(page) | page intersects with [hva_start, hva_end)} =
1351 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
1353 gfn_start = hva_to_gfn_memslot(hva_start, memslot);
1354 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
1356 for (j = PT_PAGE_TABLE_LEVEL;
1357 j < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++j) {
1358 unsigned long idx, idx_end;
1359 unsigned long *rmapp;
1362 * {idx(page_j) | page_j intersects with
1363 * [hva_start, hva_end)} = {idx, idx+1, ..., idx_end}.
1365 idx = gfn_to_index(gfn_start, memslot->base_gfn, j);
1366 idx_end = gfn_to_index(gfn_end - 1, memslot->base_gfn, j);
1368 rmapp = __gfn_to_rmap(gfn_start, j, memslot);
1370 for (; idx <= idx_end; ++idx)
1371 ret |= handler(kvm, rmapp++, memslot, data);
1378 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
1380 int (*handler)(struct kvm *kvm, unsigned long *rmapp,
1381 struct kvm_memory_slot *slot,
1382 unsigned long data))
1384 return kvm_handle_hva_range(kvm, hva, hva + 1, data, handler);
1387 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1389 return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
1392 int kvm_unmap_hva_range(struct kvm *kvm, unsigned long start, unsigned long end)
1394 return kvm_handle_hva_range(kvm, start, end, 0, kvm_unmap_rmapp);
1397 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1399 kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
1402 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1403 struct kvm_memory_slot *slot, unsigned long data)
1406 struct rmap_iterator uninitialized_var(iter);
1410 * In case of absence of EPT Access and Dirty Bits supports,
1411 * emulate the accessed bit for EPT, by checking if this page has
1412 * an EPT mapping, and clearing it if it does. On the next access,
1413 * a new EPT mapping will be established.
1414 * This has some overhead, but not as much as the cost of swapping
1415 * out actively used pages or breaking up actively used hugepages.
1417 if (!shadow_accessed_mask) {
1418 young = kvm_unmap_rmapp(kvm, rmapp, slot, data);
1422 for (sptep = rmap_get_first(*rmapp, &iter); sptep;
1423 sptep = rmap_get_next(&iter)) {
1424 BUG_ON(!is_shadow_present_pte(*sptep));
1426 if (*sptep & shadow_accessed_mask) {
1428 clear_bit((ffs(shadow_accessed_mask) - 1),
1429 (unsigned long *)sptep);
1433 /* @data has hva passed to kvm_age_hva(). */
1434 trace_kvm_age_page(data, slot, young);
1438 static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1439 struct kvm_memory_slot *slot, unsigned long data)
1442 struct rmap_iterator iter;
1446 * If there's no access bit in the secondary pte set by the
1447 * hardware it's up to gup-fast/gup to set the access bit in
1448 * the primary pte or in the page structure.
1450 if (!shadow_accessed_mask)
1453 for (sptep = rmap_get_first(*rmapp, &iter); sptep;
1454 sptep = rmap_get_next(&iter)) {
1455 BUG_ON(!is_shadow_present_pte(*sptep));
1457 if (*sptep & shadow_accessed_mask) {
1466 #define RMAP_RECYCLE_THRESHOLD 1000
1468 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
1470 unsigned long *rmapp;
1471 struct kvm_mmu_page *sp;
1473 sp = page_header(__pa(spte));
1475 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
1477 kvm_unmap_rmapp(vcpu->kvm, rmapp, NULL, 0);
1478 kvm_flush_remote_tlbs(vcpu->kvm);
1481 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
1483 return kvm_handle_hva(kvm, hva, hva, kvm_age_rmapp);
1486 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1488 return kvm_handle_hva(kvm, hva, 0, kvm_test_age_rmapp);
1492 static int is_empty_shadow_page(u64 *spt)
1497 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
1498 if (is_shadow_present_pte(*pos)) {
1499 printk(KERN_ERR "%s: %p %llx\n", __func__,
1508 * This value is the sum of all of the kvm instances's
1509 * kvm->arch.n_used_mmu_pages values. We need a global,
1510 * aggregate version in order to make the slab shrinker
1513 static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, int nr)
1515 kvm->arch.n_used_mmu_pages += nr;
1516 percpu_counter_add(&kvm_total_used_mmu_pages, nr);
1519 static void kvm_mmu_free_page(struct kvm_mmu_page *sp)
1521 ASSERT(is_empty_shadow_page(sp->spt));
1522 hlist_del(&sp->hash_link);
1523 list_del(&sp->link);
1524 free_page((unsigned long)sp->spt);
1525 if (!sp->role.direct)
1526 free_page((unsigned long)sp->gfns);
1527 kmem_cache_free(mmu_page_header_cache, sp);
1530 static unsigned kvm_page_table_hashfn(gfn_t gfn)
1532 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
1535 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
1536 struct kvm_mmu_page *sp, u64 *parent_pte)
1541 pte_list_add(vcpu, parent_pte, &sp->parent_ptes);
1544 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
1547 pte_list_remove(parent_pte, &sp->parent_ptes);
1550 static void drop_parent_pte(struct kvm_mmu_page *sp,
1553 mmu_page_remove_parent_pte(sp, parent_pte);
1554 mmu_spte_clear_no_track(parent_pte);
1557 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
1558 u64 *parent_pte, int direct)
1560 struct kvm_mmu_page *sp;
1562 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache);
1563 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache);
1565 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache);
1566 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
1569 * The active_mmu_pages list is the FIFO list, do not move the
1570 * page until it is zapped. kvm_zap_obsolete_pages depends on
1571 * this feature. See the comments in kvm_zap_obsolete_pages().
1573 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
1574 sp->parent_ptes = 0;
1575 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1576 kvm_mod_used_mmu_pages(vcpu->kvm, +1);
1580 static void mark_unsync(u64 *spte);
1581 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1583 pte_list_walk(&sp->parent_ptes, mark_unsync);
1586 static void mark_unsync(u64 *spte)
1588 struct kvm_mmu_page *sp;
1591 sp = page_header(__pa(spte));
1592 index = spte - sp->spt;
1593 if (__test_and_set_bit(index, sp->unsync_child_bitmap))
1595 if (sp->unsync_children++)
1597 kvm_mmu_mark_parents_unsync(sp);
1600 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1601 struct kvm_mmu_page *sp)
1606 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1610 static void nonpaging_update_pte(struct kvm_vcpu *vcpu,
1611 struct kvm_mmu_page *sp, u64 *spte,
1617 #define KVM_PAGE_ARRAY_NR 16
1619 struct kvm_mmu_pages {
1620 struct mmu_page_and_offset {
1621 struct kvm_mmu_page *sp;
1623 } page[KVM_PAGE_ARRAY_NR];
1627 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1633 for (i=0; i < pvec->nr; i++)
1634 if (pvec->page[i].sp == sp)
1637 pvec->page[pvec->nr].sp = sp;
1638 pvec->page[pvec->nr].idx = idx;
1640 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1643 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1644 struct kvm_mmu_pages *pvec)
1646 int i, ret, nr_unsync_leaf = 0;
1648 for_each_set_bit(i, sp->unsync_child_bitmap, 512) {
1649 struct kvm_mmu_page *child;
1650 u64 ent = sp->spt[i];
1652 if (!is_shadow_present_pte(ent) || is_large_pte(ent))
1653 goto clear_child_bitmap;
1655 child = page_header(ent & PT64_BASE_ADDR_MASK);
1657 if (child->unsync_children) {
1658 if (mmu_pages_add(pvec, child, i))
1661 ret = __mmu_unsync_walk(child, pvec);
1663 goto clear_child_bitmap;
1665 nr_unsync_leaf += ret;
1668 } else if (child->unsync) {
1670 if (mmu_pages_add(pvec, child, i))
1673 goto clear_child_bitmap;
1678 __clear_bit(i, sp->unsync_child_bitmap);
1679 sp->unsync_children--;
1680 WARN_ON((int)sp->unsync_children < 0);
1684 return nr_unsync_leaf;
1687 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1688 struct kvm_mmu_pages *pvec)
1690 if (!sp->unsync_children)
1693 mmu_pages_add(pvec, sp, 0);
1694 return __mmu_unsync_walk(sp, pvec);
1697 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1699 WARN_ON(!sp->unsync);
1700 trace_kvm_mmu_sync_page(sp);
1702 --kvm->stat.mmu_unsync;
1705 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1706 struct list_head *invalid_list);
1707 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1708 struct list_head *invalid_list);
1711 * NOTE: we should pay more attention on the zapped-obsolete page
1712 * (is_obsolete_sp(sp) && sp->role.invalid) when you do hash list walk
1713 * since it has been deleted from active_mmu_pages but still can be found
1716 * for_each_gfn_indirect_valid_sp has skipped that kind of page and
1717 * kvm_mmu_get_page(), the only user of for_each_gfn_sp(), has skipped
1718 * all the obsolete pages.
1720 #define for_each_gfn_sp(_kvm, _sp, _gfn) \
1721 hlist_for_each_entry(_sp, \
1722 &(_kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(_gfn)], hash_link) \
1723 if ((_sp)->gfn != (_gfn)) {} else
1725 #define for_each_gfn_indirect_valid_sp(_kvm, _sp, _gfn) \
1726 for_each_gfn_sp(_kvm, _sp, _gfn) \
1727 if ((_sp)->role.direct || (_sp)->role.invalid) {} else
1729 /* @sp->gfn should be write-protected at the call site */
1730 static int __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1731 struct list_head *invalid_list, bool clear_unsync)
1733 if (sp->role.cr4_pae != !!is_pae(vcpu)) {
1734 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1739 kvm_unlink_unsync_page(vcpu->kvm, sp);
1741 if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1742 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1746 kvm_mmu_flush_tlb(vcpu);
1750 static int kvm_sync_page_transient(struct kvm_vcpu *vcpu,
1751 struct kvm_mmu_page *sp)
1753 LIST_HEAD(invalid_list);
1756 ret = __kvm_sync_page(vcpu, sp, &invalid_list, false);
1758 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1763 #ifdef CONFIG_KVM_MMU_AUDIT
1764 #include "mmu_audit.c"
1766 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, int point) { }
1767 static void mmu_audit_disable(void) { }
1770 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1771 struct list_head *invalid_list)
1773 return __kvm_sync_page(vcpu, sp, invalid_list, true);
1776 /* @gfn should be write-protected at the call site */
1777 static void kvm_sync_pages(struct kvm_vcpu *vcpu, gfn_t gfn)
1779 struct kvm_mmu_page *s;
1780 LIST_HEAD(invalid_list);
1783 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn) {
1787 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1788 kvm_unlink_unsync_page(vcpu->kvm, s);
1789 if ((s->role.cr4_pae != !!is_pae(vcpu)) ||
1790 (vcpu->arch.mmu.sync_page(vcpu, s))) {
1791 kvm_mmu_prepare_zap_page(vcpu->kvm, s, &invalid_list);
1797 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1799 kvm_mmu_flush_tlb(vcpu);
1802 struct mmu_page_path {
1803 struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1804 unsigned int idx[PT64_ROOT_LEVEL-1];
1807 #define for_each_sp(pvec, sp, parents, i) \
1808 for (i = mmu_pages_next(&pvec, &parents, -1), \
1809 sp = pvec.page[i].sp; \
1810 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1811 i = mmu_pages_next(&pvec, &parents, i))
1813 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1814 struct mmu_page_path *parents,
1819 for (n = i+1; n < pvec->nr; n++) {
1820 struct kvm_mmu_page *sp = pvec->page[n].sp;
1822 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1823 parents->idx[0] = pvec->page[n].idx;
1827 parents->parent[sp->role.level-2] = sp;
1828 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1834 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1836 struct kvm_mmu_page *sp;
1837 unsigned int level = 0;
1840 unsigned int idx = parents->idx[level];
1842 sp = parents->parent[level];
1846 --sp->unsync_children;
1847 WARN_ON((int)sp->unsync_children < 0);
1848 __clear_bit(idx, sp->unsync_child_bitmap);
1850 } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1853 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1854 struct mmu_page_path *parents,
1855 struct kvm_mmu_pages *pvec)
1857 parents->parent[parent->role.level-1] = NULL;
1861 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1862 struct kvm_mmu_page *parent)
1865 struct kvm_mmu_page *sp;
1866 struct mmu_page_path parents;
1867 struct kvm_mmu_pages pages;
1868 LIST_HEAD(invalid_list);
1870 kvm_mmu_pages_init(parent, &parents, &pages);
1871 while (mmu_unsync_walk(parent, &pages)) {
1872 bool protected = false;
1874 for_each_sp(pages, sp, parents, i)
1875 protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1878 kvm_flush_remote_tlbs(vcpu->kvm);
1880 for_each_sp(pages, sp, parents, i) {
1881 kvm_sync_page(vcpu, sp, &invalid_list);
1882 mmu_pages_clear_parents(&parents);
1884 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1885 cond_resched_lock(&vcpu->kvm->mmu_lock);
1886 kvm_mmu_pages_init(parent, &parents, &pages);
1890 static void init_shadow_page_table(struct kvm_mmu_page *sp)
1894 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1898 static void __clear_sp_write_flooding_count(struct kvm_mmu_page *sp)
1900 sp->write_flooding_count = 0;
1903 static void clear_sp_write_flooding_count(u64 *spte)
1905 struct kvm_mmu_page *sp = page_header(__pa(spte));
1907 __clear_sp_write_flooding_count(sp);
1910 static bool is_obsolete_sp(struct kvm *kvm, struct kvm_mmu_page *sp)
1912 return unlikely(sp->mmu_valid_gen != kvm->arch.mmu_valid_gen);
1915 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1923 union kvm_mmu_page_role role;
1925 struct kvm_mmu_page *sp;
1926 bool need_sync = false;
1928 role = vcpu->arch.mmu.base_role;
1930 role.direct = direct;
1933 role.access = access;
1934 if (!vcpu->arch.mmu.direct_map
1935 && vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1936 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1937 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1938 role.quadrant = quadrant;
1940 for_each_gfn_sp(vcpu->kvm, sp, gfn) {
1941 if (is_obsolete_sp(vcpu->kvm, sp))
1944 if (!need_sync && sp->unsync)
1947 if (sp->role.word != role.word)
1950 if (sp->unsync && kvm_sync_page_transient(vcpu, sp))
1953 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1954 if (sp->unsync_children) {
1955 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
1956 kvm_mmu_mark_parents_unsync(sp);
1957 } else if (sp->unsync)
1958 kvm_mmu_mark_parents_unsync(sp);
1960 __clear_sp_write_flooding_count(sp);
1961 trace_kvm_mmu_get_page(sp, false);
1964 ++vcpu->kvm->stat.mmu_cache_miss;
1965 sp = kvm_mmu_alloc_page(vcpu, parent_pte, direct);
1970 hlist_add_head(&sp->hash_link,
1971 &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
1973 if (rmap_write_protect(vcpu->kvm, gfn))
1974 kvm_flush_remote_tlbs(vcpu->kvm);
1975 if (level > PT_PAGE_TABLE_LEVEL && need_sync)
1976 kvm_sync_pages(vcpu, gfn);
1978 account_shadowed(vcpu->kvm, gfn);
1980 sp->mmu_valid_gen = vcpu->kvm->arch.mmu_valid_gen;
1981 init_shadow_page_table(sp);
1982 trace_kvm_mmu_get_page(sp, true);
1986 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1987 struct kvm_vcpu *vcpu, u64 addr)
1989 iterator->addr = addr;
1990 iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1991 iterator->level = vcpu->arch.mmu.shadow_root_level;
1993 if (iterator->level == PT64_ROOT_LEVEL &&
1994 vcpu->arch.mmu.root_level < PT64_ROOT_LEVEL &&
1995 !vcpu->arch.mmu.direct_map)
1998 if (iterator->level == PT32E_ROOT_LEVEL) {
1999 iterator->shadow_addr
2000 = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
2001 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
2003 if (!iterator->shadow_addr)
2004 iterator->level = 0;
2008 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
2010 if (iterator->level < PT_PAGE_TABLE_LEVEL)
2013 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
2014 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
2018 static void __shadow_walk_next(struct kvm_shadow_walk_iterator *iterator,
2021 if (is_last_spte(spte, iterator->level)) {
2022 iterator->level = 0;
2026 iterator->shadow_addr = spte & PT64_BASE_ADDR_MASK;
2030 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
2032 return __shadow_walk_next(iterator, *iterator->sptep);
2035 static void link_shadow_page(u64 *sptep, struct kvm_mmu_page *sp, bool accessed)
2039 BUILD_BUG_ON(VMX_EPT_READABLE_MASK != PT_PRESENT_MASK ||
2040 VMX_EPT_WRITABLE_MASK != PT_WRITABLE_MASK);
2042 spte = __pa(sp->spt) | PT_PRESENT_MASK | PT_WRITABLE_MASK |
2043 shadow_user_mask | shadow_x_mask;
2046 spte |= shadow_accessed_mask;
2048 mmu_spte_set(sptep, spte);
2051 static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2052 unsigned direct_access)
2054 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
2055 struct kvm_mmu_page *child;
2058 * For the direct sp, if the guest pte's dirty bit
2059 * changed form clean to dirty, it will corrupt the
2060 * sp's access: allow writable in the read-only sp,
2061 * so we should update the spte at this point to get
2062 * a new sp with the correct access.
2064 child = page_header(*sptep & PT64_BASE_ADDR_MASK);
2065 if (child->role.access == direct_access)
2068 drop_parent_pte(child, sptep);
2069 kvm_flush_remote_tlbs(vcpu->kvm);
2073 static bool mmu_page_zap_pte(struct kvm *kvm, struct kvm_mmu_page *sp,
2077 struct kvm_mmu_page *child;
2080 if (is_shadow_present_pte(pte)) {
2081 if (is_last_spte(pte, sp->role.level)) {
2082 drop_spte(kvm, spte);
2083 if (is_large_pte(pte))
2086 child = page_header(pte & PT64_BASE_ADDR_MASK);
2087 drop_parent_pte(child, spte);
2092 if (is_mmio_spte(pte))
2093 mmu_spte_clear_no_track(spte);
2098 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
2099 struct kvm_mmu_page *sp)
2103 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2104 mmu_page_zap_pte(kvm, sp, sp->spt + i);
2107 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
2109 mmu_page_remove_parent_pte(sp, parent_pte);
2112 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
2115 struct rmap_iterator iter;
2117 while ((sptep = rmap_get_first(sp->parent_ptes, &iter)))
2118 drop_parent_pte(sp, sptep);
2121 static int mmu_zap_unsync_children(struct kvm *kvm,
2122 struct kvm_mmu_page *parent,
2123 struct list_head *invalid_list)
2126 struct mmu_page_path parents;
2127 struct kvm_mmu_pages pages;
2129 if (parent->role.level == PT_PAGE_TABLE_LEVEL)
2132 kvm_mmu_pages_init(parent, &parents, &pages);
2133 while (mmu_unsync_walk(parent, &pages)) {
2134 struct kvm_mmu_page *sp;
2136 for_each_sp(pages, sp, parents, i) {
2137 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
2138 mmu_pages_clear_parents(&parents);
2141 kvm_mmu_pages_init(parent, &parents, &pages);
2147 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
2148 struct list_head *invalid_list)
2152 trace_kvm_mmu_prepare_zap_page(sp);
2153 ++kvm->stat.mmu_shadow_zapped;
2154 ret = mmu_zap_unsync_children(kvm, sp, invalid_list);
2155 kvm_mmu_page_unlink_children(kvm, sp);
2156 kvm_mmu_unlink_parents(kvm, sp);
2158 if (!sp->role.invalid && !sp->role.direct)
2159 unaccount_shadowed(kvm, sp->gfn);
2162 kvm_unlink_unsync_page(kvm, sp);
2163 if (!sp->root_count) {
2166 list_move(&sp->link, invalid_list);
2167 kvm_mod_used_mmu_pages(kvm, -1);
2169 list_move(&sp->link, &kvm->arch.active_mmu_pages);
2172 * The obsolete pages can not be used on any vcpus.
2173 * See the comments in kvm_mmu_invalidate_zap_all_pages().
2175 if (!sp->role.invalid && !is_obsolete_sp(kvm, sp))
2176 kvm_reload_remote_mmus(kvm);
2179 sp->role.invalid = 1;
2183 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
2184 struct list_head *invalid_list)
2186 struct kvm_mmu_page *sp, *nsp;
2188 if (list_empty(invalid_list))
2192 * wmb: make sure everyone sees our modifications to the page tables
2193 * rmb: make sure we see changes to vcpu->mode
2198 * Wait for all vcpus to exit guest mode and/or lockless shadow
2201 kvm_flush_remote_tlbs(kvm);
2203 list_for_each_entry_safe(sp, nsp, invalid_list, link) {
2204 WARN_ON(!sp->role.invalid || sp->root_count);
2205 kvm_mmu_free_page(sp);
2209 static bool prepare_zap_oldest_mmu_page(struct kvm *kvm,
2210 struct list_head *invalid_list)
2212 struct kvm_mmu_page *sp;
2214 if (list_empty(&kvm->arch.active_mmu_pages))
2217 sp = list_entry(kvm->arch.active_mmu_pages.prev,
2218 struct kvm_mmu_page, link);
2219 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
2225 * Changing the number of mmu pages allocated to the vm
2226 * Note: if goal_nr_mmu_pages is too small, you will get dead lock
2228 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int goal_nr_mmu_pages)
2230 LIST_HEAD(invalid_list);
2232 spin_lock(&kvm->mmu_lock);
2234 if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) {
2235 /* Need to free some mmu pages to achieve the goal. */
2236 while (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages)
2237 if (!prepare_zap_oldest_mmu_page(kvm, &invalid_list))
2240 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2241 goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages;
2244 kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages;
2246 spin_unlock(&kvm->mmu_lock);
2249 int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
2251 struct kvm_mmu_page *sp;
2252 LIST_HEAD(invalid_list);
2255 pgprintk("%s: looking for gfn %llx\n", __func__, gfn);
2257 spin_lock(&kvm->mmu_lock);
2258 for_each_gfn_indirect_valid_sp(kvm, sp, gfn) {
2259 pgprintk("%s: gfn %llx role %x\n", __func__, gfn,
2262 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
2264 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2265 spin_unlock(&kvm->mmu_lock);
2269 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page);
2272 * The function is based on mtrr_type_lookup() in
2273 * arch/x86/kernel/cpu/mtrr/generic.c
2275 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
2280 u8 prev_match, curr_match;
2281 int num_var_ranges = KVM_NR_VAR_MTRR;
2283 if (!mtrr_state->enabled)
2286 /* Make end inclusive end, instead of exclusive */
2289 /* Look in fixed ranges. Just return the type as per start */
2290 if (mtrr_state->have_fixed && (start < 0x100000)) {
2293 if (start < 0x80000) {
2295 idx += (start >> 16);
2296 return mtrr_state->fixed_ranges[idx];
2297 } else if (start < 0xC0000) {
2299 idx += ((start - 0x80000) >> 14);
2300 return mtrr_state->fixed_ranges[idx];
2301 } else if (start < 0x1000000) {
2303 idx += ((start - 0xC0000) >> 12);
2304 return mtrr_state->fixed_ranges[idx];
2309 * Look in variable ranges
2310 * Look of multiple ranges matching this address and pick type
2311 * as per MTRR precedence
2313 if (!(mtrr_state->enabled & 2))
2314 return mtrr_state->def_type;
2317 for (i = 0; i < num_var_ranges; ++i) {
2318 unsigned short start_state, end_state;
2320 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
2323 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
2324 (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
2325 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
2326 (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
2328 start_state = ((start & mask) == (base & mask));
2329 end_state = ((end & mask) == (base & mask));
2330 if (start_state != end_state)
2333 if ((start & mask) != (base & mask))
2336 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
2337 if (prev_match == 0xFF) {
2338 prev_match = curr_match;
2342 if (prev_match == MTRR_TYPE_UNCACHABLE ||
2343 curr_match == MTRR_TYPE_UNCACHABLE)
2344 return MTRR_TYPE_UNCACHABLE;
2346 if ((prev_match == MTRR_TYPE_WRBACK &&
2347 curr_match == MTRR_TYPE_WRTHROUGH) ||
2348 (prev_match == MTRR_TYPE_WRTHROUGH &&
2349 curr_match == MTRR_TYPE_WRBACK)) {
2350 prev_match = MTRR_TYPE_WRTHROUGH;
2351 curr_match = MTRR_TYPE_WRTHROUGH;
2354 if (prev_match != curr_match)
2355 return MTRR_TYPE_UNCACHABLE;
2358 if (prev_match != 0xFF)
2361 return mtrr_state->def_type;
2364 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
2368 mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
2369 (gfn << PAGE_SHIFT) + PAGE_SIZE);
2370 if (mtrr == 0xfe || mtrr == 0xff)
2371 mtrr = MTRR_TYPE_WRBACK;
2374 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
2376 static void __kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
2378 trace_kvm_mmu_unsync_page(sp);
2379 ++vcpu->kvm->stat.mmu_unsync;
2382 kvm_mmu_mark_parents_unsync(sp);
2385 static void kvm_unsync_pages(struct kvm_vcpu *vcpu, gfn_t gfn)
2387 struct kvm_mmu_page *s;
2389 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn) {
2392 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
2393 __kvm_unsync_page(vcpu, s);
2397 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
2400 struct kvm_mmu_page *s;
2401 bool need_unsync = false;
2403 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn) {
2407 if (s->role.level != PT_PAGE_TABLE_LEVEL)
2414 kvm_unsync_pages(vcpu, gfn);
2418 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2419 unsigned pte_access, int level,
2420 gfn_t gfn, pfn_t pfn, bool speculative,
2421 bool can_unsync, bool host_writable)
2426 if (set_mmio_spte(vcpu->kvm, sptep, gfn, pfn, pte_access))
2429 spte = PT_PRESENT_MASK;
2431 spte |= shadow_accessed_mask;
2433 if (pte_access & ACC_EXEC_MASK)
2434 spte |= shadow_x_mask;
2436 spte |= shadow_nx_mask;
2438 if (pte_access & ACC_USER_MASK)
2439 spte |= shadow_user_mask;
2441 if (level > PT_PAGE_TABLE_LEVEL)
2442 spte |= PT_PAGE_SIZE_MASK;
2444 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
2445 kvm_is_mmio_pfn(pfn));
2448 spte |= SPTE_HOST_WRITEABLE;
2450 pte_access &= ~ACC_WRITE_MASK;
2452 spte |= (u64)pfn << PAGE_SHIFT;
2454 if (pte_access & ACC_WRITE_MASK) {
2457 * Other vcpu creates new sp in the window between
2458 * mapping_level() and acquiring mmu-lock. We can
2459 * allow guest to retry the access, the mapping can
2460 * be fixed if guest refault.
2462 if (level > PT_PAGE_TABLE_LEVEL &&
2463 has_wrprotected_page(vcpu->kvm, gfn, level))
2466 spte |= PT_WRITABLE_MASK | SPTE_MMU_WRITEABLE;
2469 * Optimization: for pte sync, if spte was writable the hash
2470 * lookup is unnecessary (and expensive). Write protection
2471 * is responsibility of mmu_get_page / kvm_sync_page.
2472 * Same reasoning can be applied to dirty page accounting.
2474 if (!can_unsync && is_writable_pte(*sptep))
2477 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
2478 pgprintk("%s: found shadow page for %llx, marking ro\n",
2481 pte_access &= ~ACC_WRITE_MASK;
2482 spte &= ~(PT_WRITABLE_MASK | SPTE_MMU_WRITEABLE);
2486 if (pte_access & ACC_WRITE_MASK)
2487 mark_page_dirty(vcpu->kvm, gfn);
2490 if (mmu_spte_update(sptep, spte))
2491 kvm_flush_remote_tlbs(vcpu->kvm);
2496 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2497 unsigned pte_access, int write_fault, int *emulate,
2498 int level, gfn_t gfn, pfn_t pfn, bool speculative,
2501 int was_rmapped = 0;
2504 pgprintk("%s: spte %llx write_fault %d gfn %llx\n", __func__,
2505 *sptep, write_fault, gfn);
2507 if (is_rmap_spte(*sptep)) {
2509 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
2510 * the parent of the now unreachable PTE.
2512 if (level > PT_PAGE_TABLE_LEVEL &&
2513 !is_large_pte(*sptep)) {
2514 struct kvm_mmu_page *child;
2517 child = page_header(pte & PT64_BASE_ADDR_MASK);
2518 drop_parent_pte(child, sptep);
2519 kvm_flush_remote_tlbs(vcpu->kvm);
2520 } else if (pfn != spte_to_pfn(*sptep)) {
2521 pgprintk("hfn old %llx new %llx\n",
2522 spte_to_pfn(*sptep), pfn);
2523 drop_spte(vcpu->kvm, sptep);
2524 kvm_flush_remote_tlbs(vcpu->kvm);
2529 if (set_spte(vcpu, sptep, pte_access, level, gfn, pfn, speculative,
2530 true, host_writable)) {
2533 kvm_mmu_flush_tlb(vcpu);
2536 if (unlikely(is_mmio_spte(*sptep) && emulate))
2539 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
2540 pgprintk("instantiating %s PTE (%s) at %llx (%llx) addr %p\n",
2541 is_large_pte(*sptep)? "2MB" : "4kB",
2542 *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
2544 if (!was_rmapped && is_large_pte(*sptep))
2545 ++vcpu->kvm->stat.lpages;
2547 if (is_shadow_present_pte(*sptep)) {
2549 rmap_count = rmap_add(vcpu, sptep, gfn);
2550 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
2551 rmap_recycle(vcpu, sptep, gfn);
2555 kvm_release_pfn_clean(pfn);
2558 static pfn_t pte_prefetch_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn,
2561 struct kvm_memory_slot *slot;
2563 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, no_dirty_log);
2565 return KVM_PFN_ERR_FAULT;
2567 return gfn_to_pfn_memslot_atomic(slot, gfn);
2570 static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu,
2571 struct kvm_mmu_page *sp,
2572 u64 *start, u64 *end)
2574 struct page *pages[PTE_PREFETCH_NUM];
2575 unsigned access = sp->role.access;
2579 gfn = kvm_mmu_page_get_gfn(sp, start - sp->spt);
2580 if (!gfn_to_memslot_dirty_bitmap(vcpu, gfn, access & ACC_WRITE_MASK))
2583 ret = gfn_to_page_many_atomic(vcpu->kvm, gfn, pages, end - start);
2587 for (i = 0; i < ret; i++, gfn++, start++)
2588 mmu_set_spte(vcpu, start, access, 0, NULL,
2589 sp->role.level, gfn, page_to_pfn(pages[i]),
2595 static void __direct_pte_prefetch(struct kvm_vcpu *vcpu,
2596 struct kvm_mmu_page *sp, u64 *sptep)
2598 u64 *spte, *start = NULL;
2601 WARN_ON(!sp->role.direct);
2603 i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
2606 for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
2607 if (is_shadow_present_pte(*spte) || spte == sptep) {
2610 if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0)
2618 static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
2620 struct kvm_mmu_page *sp;
2623 * Since it's no accessed bit on EPT, it's no way to
2624 * distinguish between actually accessed translations
2625 * and prefetched, so disable pte prefetch if EPT is
2628 if (!shadow_accessed_mask)
2631 sp = page_header(__pa(sptep));
2632 if (sp->role.level > PT_PAGE_TABLE_LEVEL)
2635 __direct_pte_prefetch(vcpu, sp, sptep);
2638 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
2639 int map_writable, int level, gfn_t gfn, pfn_t pfn,
2642 struct kvm_shadow_walk_iterator iterator;
2643 struct kvm_mmu_page *sp;
2647 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2650 for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
2651 if (iterator.level == level) {
2652 mmu_set_spte(vcpu, iterator.sptep, ACC_ALL,
2653 write, &emulate, level, gfn, pfn,
2654 prefault, map_writable);
2655 direct_pte_prefetch(vcpu, iterator.sptep);
2656 ++vcpu->stat.pf_fixed;
2660 drop_large_spte(vcpu, iterator.sptep);
2661 if (!is_shadow_present_pte(*iterator.sptep)) {
2662 u64 base_addr = iterator.addr;
2664 base_addr &= PT64_LVL_ADDR_MASK(iterator.level);
2665 pseudo_gfn = base_addr >> PAGE_SHIFT;
2666 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
2668 1, ACC_ALL, iterator.sptep);
2670 link_shadow_page(iterator.sptep, sp, true);
2676 static void kvm_send_hwpoison_signal(unsigned long address, struct task_struct *tsk)
2680 info.si_signo = SIGBUS;
2682 info.si_code = BUS_MCEERR_AR;
2683 info.si_addr = (void __user *)address;
2684 info.si_addr_lsb = PAGE_SHIFT;
2686 send_sig_info(SIGBUS, &info, tsk);
2689 static int kvm_handle_bad_page(struct kvm_vcpu *vcpu, gfn_t gfn, pfn_t pfn)
2692 * Do not cache the mmio info caused by writing the readonly gfn
2693 * into the spte otherwise read access on readonly gfn also can
2694 * caused mmio page fault and treat it as mmio access.
2695 * Return 1 to tell kvm to emulate it.
2697 if (pfn == KVM_PFN_ERR_RO_FAULT)
2700 if (pfn == KVM_PFN_ERR_HWPOISON) {
2701 kvm_send_hwpoison_signal(gfn_to_hva(vcpu->kvm, gfn), current);
2708 static void transparent_hugepage_adjust(struct kvm_vcpu *vcpu,
2709 gfn_t *gfnp, pfn_t *pfnp, int *levelp)
2713 int level = *levelp;
2716 * Check if it's a transparent hugepage. If this would be an
2717 * hugetlbfs page, level wouldn't be set to
2718 * PT_PAGE_TABLE_LEVEL and there would be no adjustment done
2721 if (!is_error_noslot_pfn(pfn) && !kvm_is_mmio_pfn(pfn) &&
2722 level == PT_PAGE_TABLE_LEVEL &&
2723 PageTransCompound(pfn_to_page(pfn)) &&
2724 !has_wrprotected_page(vcpu->kvm, gfn, PT_DIRECTORY_LEVEL)) {
2727 * mmu_notifier_retry was successful and we hold the
2728 * mmu_lock here, so the pmd can't become splitting
2729 * from under us, and in turn
2730 * __split_huge_page_refcount() can't run from under
2731 * us and we can safely transfer the refcount from
2732 * PG_tail to PG_head as we switch the pfn to tail to
2735 *levelp = level = PT_DIRECTORY_LEVEL;
2736 mask = KVM_PAGES_PER_HPAGE(level) - 1;
2737 VM_BUG_ON((gfn & mask) != (pfn & mask));
2741 kvm_release_pfn_clean(pfn);
2749 static bool handle_abnormal_pfn(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn,
2750 pfn_t pfn, unsigned access, int *ret_val)
2754 /* The pfn is invalid, report the error! */
2755 if (unlikely(is_error_pfn(pfn))) {
2756 *ret_val = kvm_handle_bad_page(vcpu, gfn, pfn);
2760 if (unlikely(is_noslot_pfn(pfn)))
2761 vcpu_cache_mmio_info(vcpu, gva, gfn, access);
2768 static bool page_fault_can_be_fast(u32 error_code)
2771 * Do not fix the mmio spte with invalid generation number which
2772 * need to be updated by slow page fault path.
2774 if (unlikely(error_code & PFERR_RSVD_MASK))
2778 * #PF can be fast only if the shadow page table is present and it
2779 * is caused by write-protect, that means we just need change the
2780 * W bit of the spte which can be done out of mmu-lock.
2782 if (!(error_code & PFERR_PRESENT_MASK) ||
2783 !(error_code & PFERR_WRITE_MASK))
2790 fast_pf_fix_direct_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
2791 u64 *sptep, u64 spte)
2795 WARN_ON(!sp->role.direct);
2798 * The gfn of direct spte is stable since it is calculated
2801 gfn = kvm_mmu_page_get_gfn(sp, sptep - sp->spt);
2803 if (cmpxchg64(sptep, spte, spte | PT_WRITABLE_MASK) == spte)
2804 mark_page_dirty(vcpu->kvm, gfn);
2811 * - true: let the vcpu to access on the same address again.
2812 * - false: let the real page fault path to fix it.
2814 static bool fast_page_fault(struct kvm_vcpu *vcpu, gva_t gva, int level,
2817 struct kvm_shadow_walk_iterator iterator;
2818 struct kvm_mmu_page *sp;
2822 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2825 if (!page_fault_can_be_fast(error_code))
2828 walk_shadow_page_lockless_begin(vcpu);
2829 for_each_shadow_entry_lockless(vcpu, gva, iterator, spte)
2830 if (!is_shadow_present_pte(spte) || iterator.level < level)
2834 * If the mapping has been changed, let the vcpu fault on the
2835 * same address again.
2837 if (!is_rmap_spte(spte)) {
2842 sp = page_header(__pa(iterator.sptep));
2843 if (!is_last_spte(spte, sp->role.level))
2847 * Check if it is a spurious fault caused by TLB lazily flushed.
2849 * Need not check the access of upper level table entries since
2850 * they are always ACC_ALL.
2852 if (is_writable_pte(spte)) {
2858 * Currently, to simplify the code, only the spte write-protected
2859 * by dirty-log can be fast fixed.
2861 if (!spte_is_locklessly_modifiable(spte))
2865 * Do not fix write-permission on the large spte since we only dirty
2866 * the first page into the dirty-bitmap in fast_pf_fix_direct_spte()
2867 * that means other pages are missed if its slot is dirty-logged.
2869 * Instead, we let the slow page fault path create a normal spte to
2872 * See the comments in kvm_arch_commit_memory_region().
2874 if (sp->role.level > PT_PAGE_TABLE_LEVEL)
2878 * Currently, fast page fault only works for direct mapping since
2879 * the gfn is not stable for indirect shadow page.
2880 * See Documentation/virtual/kvm/locking.txt to get more detail.
2882 ret = fast_pf_fix_direct_spte(vcpu, sp, iterator.sptep, spte);
2884 trace_fast_page_fault(vcpu, gva, error_code, iterator.sptep,
2886 walk_shadow_page_lockless_end(vcpu);
2891 static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
2892 gva_t gva, pfn_t *pfn, bool write, bool *writable);
2893 static void make_mmu_pages_available(struct kvm_vcpu *vcpu);
2895 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, u32 error_code,
2896 gfn_t gfn, bool prefault)
2902 unsigned long mmu_seq;
2903 bool map_writable, write = error_code & PFERR_WRITE_MASK;
2905 force_pt_level = mapping_level_dirty_bitmap(vcpu, gfn);
2906 if (likely(!force_pt_level)) {
2907 level = mapping_level(vcpu, gfn);
2909 * This path builds a PAE pagetable - so we can map
2910 * 2mb pages at maximum. Therefore check if the level
2911 * is larger than that.
2913 if (level > PT_DIRECTORY_LEVEL)
2914 level = PT_DIRECTORY_LEVEL;
2916 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2918 level = PT_PAGE_TABLE_LEVEL;
2920 if (fast_page_fault(vcpu, v, level, error_code))
2923 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2926 if (try_async_pf(vcpu, prefault, gfn, v, &pfn, write, &map_writable))
2929 if (handle_abnormal_pfn(vcpu, v, gfn, pfn, ACC_ALL, &r))
2932 spin_lock(&vcpu->kvm->mmu_lock);
2933 if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
2935 make_mmu_pages_available(vcpu);
2936 if (likely(!force_pt_level))
2937 transparent_hugepage_adjust(vcpu, &gfn, &pfn, &level);
2938 r = __direct_map(vcpu, v, write, map_writable, level, gfn, pfn,
2940 spin_unlock(&vcpu->kvm->mmu_lock);
2946 spin_unlock(&vcpu->kvm->mmu_lock);
2947 kvm_release_pfn_clean(pfn);
2952 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2955 struct kvm_mmu_page *sp;
2956 LIST_HEAD(invalid_list);
2958 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2961 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL &&
2962 (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL ||
2963 vcpu->arch.mmu.direct_map)) {
2964 hpa_t root = vcpu->arch.mmu.root_hpa;
2966 spin_lock(&vcpu->kvm->mmu_lock);
2967 sp = page_header(root);
2969 if (!sp->root_count && sp->role.invalid) {
2970 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
2971 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2973 spin_unlock(&vcpu->kvm->mmu_lock);
2974 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2978 spin_lock(&vcpu->kvm->mmu_lock);
2979 for (i = 0; i < 4; ++i) {
2980 hpa_t root = vcpu->arch.mmu.pae_root[i];
2983 root &= PT64_BASE_ADDR_MASK;
2984 sp = page_header(root);
2986 if (!sp->root_count && sp->role.invalid)
2987 kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2990 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2992 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2993 spin_unlock(&vcpu->kvm->mmu_lock);
2994 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2997 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
3001 if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
3002 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
3009 static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
3011 struct kvm_mmu_page *sp;
3014 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3015 spin_lock(&vcpu->kvm->mmu_lock);
3016 make_mmu_pages_available(vcpu);
3017 sp = kvm_mmu_get_page(vcpu, 0, 0, PT64_ROOT_LEVEL,
3020 spin_unlock(&vcpu->kvm->mmu_lock);
3021 vcpu->arch.mmu.root_hpa = __pa(sp->spt);
3022 } else if (vcpu->arch.mmu.shadow_root_level == PT32E_ROOT_LEVEL) {
3023 for (i = 0; i < 4; ++i) {
3024 hpa_t root = vcpu->arch.mmu.pae_root[i];
3026 ASSERT(!VALID_PAGE(root));
3027 spin_lock(&vcpu->kvm->mmu_lock);
3028 make_mmu_pages_available(vcpu);
3029 sp = kvm_mmu_get_page(vcpu, i << (30 - PAGE_SHIFT),
3031 PT32_ROOT_LEVEL, 1, ACC_ALL,
3033 root = __pa(sp->spt);
3035 spin_unlock(&vcpu->kvm->mmu_lock);
3036 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
3038 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
3045 static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
3047 struct kvm_mmu_page *sp;
3052 root_gfn = vcpu->arch.mmu.get_cr3(vcpu) >> PAGE_SHIFT;
3054 if (mmu_check_root(vcpu, root_gfn))
3058 * Do we shadow a long mode page table? If so we need to
3059 * write-protect the guests page table root.
3061 if (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL) {
3062 hpa_t root = vcpu->arch.mmu.root_hpa;
3064 ASSERT(!VALID_PAGE(root));
3066 spin_lock(&vcpu->kvm->mmu_lock);
3067 make_mmu_pages_available(vcpu);
3068 sp = kvm_mmu_get_page(vcpu, root_gfn, 0, PT64_ROOT_LEVEL,
3070 root = __pa(sp->spt);
3072 spin_unlock(&vcpu->kvm->mmu_lock);
3073 vcpu->arch.mmu.root_hpa = root;
3078 * We shadow a 32 bit page table. This may be a legacy 2-level
3079 * or a PAE 3-level page table. In either case we need to be aware that
3080 * the shadow page table may be a PAE or a long mode page table.
3082 pm_mask = PT_PRESENT_MASK;
3083 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL)
3084 pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK;
3086 for (i = 0; i < 4; ++i) {
3087 hpa_t root = vcpu->arch.mmu.pae_root[i];
3089 ASSERT(!VALID_PAGE(root));
3090 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
3091 pdptr = vcpu->arch.mmu.get_pdptr(vcpu, i);
3092 if (!is_present_gpte(pdptr)) {
3093 vcpu->arch.mmu.pae_root[i] = 0;
3096 root_gfn = pdptr >> PAGE_SHIFT;
3097 if (mmu_check_root(vcpu, root_gfn))
3100 spin_lock(&vcpu->kvm->mmu_lock);
3101 make_mmu_pages_available(vcpu);
3102 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
3105 root = __pa(sp->spt);
3107 spin_unlock(&vcpu->kvm->mmu_lock);
3109 vcpu->arch.mmu.pae_root[i] = root | pm_mask;
3111 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
3114 * If we shadow a 32 bit page table with a long mode page
3115 * table we enter this path.
3117 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3118 if (vcpu->arch.mmu.lm_root == NULL) {
3120 * The additional page necessary for this is only
3121 * allocated on demand.
3126 lm_root = (void*)get_zeroed_page(GFP_KERNEL);
3127 if (lm_root == NULL)
3130 lm_root[0] = __pa(vcpu->arch.mmu.pae_root) | pm_mask;
3132 vcpu->arch.mmu.lm_root = lm_root;
3135 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.lm_root);
3141 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
3143 if (vcpu->arch.mmu.direct_map)
3144 return mmu_alloc_direct_roots(vcpu);
3146 return mmu_alloc_shadow_roots(vcpu);
3149 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
3152 struct kvm_mmu_page *sp;
3154 if (vcpu->arch.mmu.direct_map)
3157 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3160 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
3161 kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
3162 if (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL) {
3163 hpa_t root = vcpu->arch.mmu.root_hpa;
3164 sp = page_header(root);
3165 mmu_sync_children(vcpu, sp);
3166 kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
3169 for (i = 0; i < 4; ++i) {
3170 hpa_t root = vcpu->arch.mmu.pae_root[i];
3172 if (root && VALID_PAGE(root)) {
3173 root &= PT64_BASE_ADDR_MASK;
3174 sp = page_header(root);
3175 mmu_sync_children(vcpu, sp);
3178 kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
3181 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
3183 spin_lock(&vcpu->kvm->mmu_lock);
3184 mmu_sync_roots(vcpu);
3185 spin_unlock(&vcpu->kvm->mmu_lock);
3187 EXPORT_SYMBOL_GPL(kvm_mmu_sync_roots);
3189 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
3190 u32 access, struct x86_exception *exception)
3193 exception->error_code = 0;
3197 static gpa_t nonpaging_gva_to_gpa_nested(struct kvm_vcpu *vcpu, gva_t vaddr,
3199 struct x86_exception *exception)
3202 exception->error_code = 0;
3203 return vcpu->arch.nested_mmu.translate_gpa(vcpu, vaddr, access, exception);
3206 static bool quickly_check_mmio_pf(struct kvm_vcpu *vcpu, u64 addr, bool direct)
3209 return vcpu_match_mmio_gpa(vcpu, addr);
3211 return vcpu_match_mmio_gva(vcpu, addr);
3216 * On direct hosts, the last spte is only allows two states
3217 * for mmio page fault:
3218 * - It is the mmio spte
3219 * - It is zapped or it is being zapped.
3221 * This function completely checks the spte when the last spte
3222 * is not the mmio spte.
3224 static bool check_direct_spte_mmio_pf(u64 spte)
3226 return __check_direct_spte_mmio_pf(spte);
3229 static u64 walk_shadow_page_get_mmio_spte(struct kvm_vcpu *vcpu, u64 addr)
3231 struct kvm_shadow_walk_iterator iterator;
3234 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3237 walk_shadow_page_lockless_begin(vcpu);
3238 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte)
3239 if (!is_shadow_present_pte(spte))
3241 walk_shadow_page_lockless_end(vcpu);
3246 int handle_mmio_page_fault_common(struct kvm_vcpu *vcpu, u64 addr, bool direct)
3250 if (quickly_check_mmio_pf(vcpu, addr, direct))
3251 return RET_MMIO_PF_EMULATE;
3253 spte = walk_shadow_page_get_mmio_spte(vcpu, addr);
3255 if (is_mmio_spte(spte)) {
3256 gfn_t gfn = get_mmio_spte_gfn(spte);
3257 unsigned access = get_mmio_spte_access(spte);
3259 if (!check_mmio_spte(vcpu->kvm, spte))
3260 return RET_MMIO_PF_INVALID;
3265 trace_handle_mmio_page_fault(addr, gfn, access);
3266 vcpu_cache_mmio_info(vcpu, addr, gfn, access);
3267 return RET_MMIO_PF_EMULATE;
3271 * It's ok if the gva is remapped by other cpus on shadow guest,
3272 * it's a BUG if the gfn is not a mmio page.
3274 if (direct && !check_direct_spte_mmio_pf(spte))
3275 return RET_MMIO_PF_BUG;
3278 * If the page table is zapped by other cpus, let CPU fault again on
3281 return RET_MMIO_PF_RETRY;
3283 EXPORT_SYMBOL_GPL(handle_mmio_page_fault_common);
3285 static int handle_mmio_page_fault(struct kvm_vcpu *vcpu, u64 addr,
3286 u32 error_code, bool direct)
3290 ret = handle_mmio_page_fault_common(vcpu, addr, direct);
3291 WARN_ON(ret == RET_MMIO_PF_BUG);
3295 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
3296 u32 error_code, bool prefault)
3301 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
3303 if (unlikely(error_code & PFERR_RSVD_MASK)) {
3304 r = handle_mmio_page_fault(vcpu, gva, error_code, true);
3306 if (likely(r != RET_MMIO_PF_INVALID))
3310 r = mmu_topup_memory_caches(vcpu);
3315 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
3317 gfn = gva >> PAGE_SHIFT;
3319 return nonpaging_map(vcpu, gva & PAGE_MASK,
3320 error_code, gfn, prefault);
3323 static int kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn)
3325 struct kvm_arch_async_pf arch;
3327 arch.token = (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id;
3329 arch.direct_map = vcpu->arch.mmu.direct_map;
3330 arch.cr3 = vcpu->arch.mmu.get_cr3(vcpu);
3332 return kvm_setup_async_pf(vcpu, gva, gfn_to_hva(vcpu->kvm, gfn), &arch);
3335 static bool can_do_async_pf(struct kvm_vcpu *vcpu)
3337 if (unlikely(!irqchip_in_kernel(vcpu->kvm) ||
3338 kvm_event_needs_reinjection(vcpu)))
3341 return kvm_x86_ops->interrupt_allowed(vcpu);
3344 static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
3345 gva_t gva, pfn_t *pfn, bool write, bool *writable)
3349 *pfn = gfn_to_pfn_async(vcpu->kvm, gfn, &async, write, writable);
3352 return false; /* *pfn has correct page already */
3354 if (!prefault && can_do_async_pf(vcpu)) {
3355 trace_kvm_try_async_get_page(gva, gfn);
3356 if (kvm_find_async_pf_gfn(vcpu, gfn)) {
3357 trace_kvm_async_pf_doublefault(gva, gfn);
3358 kvm_make_request(KVM_REQ_APF_HALT, vcpu);
3360 } else if (kvm_arch_setup_async_pf(vcpu, gva, gfn))
3364 *pfn = gfn_to_pfn_prot(vcpu->kvm, gfn, write, writable);
3369 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa, u32 error_code,
3376 gfn_t gfn = gpa >> PAGE_SHIFT;
3377 unsigned long mmu_seq;
3378 int write = error_code & PFERR_WRITE_MASK;
3382 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
3384 if (unlikely(error_code & PFERR_RSVD_MASK)) {
3385 r = handle_mmio_page_fault(vcpu, gpa, error_code, true);
3387 if (likely(r != RET_MMIO_PF_INVALID))
3391 r = mmu_topup_memory_caches(vcpu);
3395 force_pt_level = mapping_level_dirty_bitmap(vcpu, gfn);
3396 if (likely(!force_pt_level)) {
3397 level = mapping_level(vcpu, gfn);
3398 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
3400 level = PT_PAGE_TABLE_LEVEL;
3402 if (fast_page_fault(vcpu, gpa, level, error_code))
3405 mmu_seq = vcpu->kvm->mmu_notifier_seq;
3408 if (try_async_pf(vcpu, prefault, gfn, gpa, &pfn, write, &map_writable))
3411 if (handle_abnormal_pfn(vcpu, 0, gfn, pfn, ACC_ALL, &r))
3414 spin_lock(&vcpu->kvm->mmu_lock);
3415 if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
3417 make_mmu_pages_available(vcpu);
3418 if (likely(!force_pt_level))
3419 transparent_hugepage_adjust(vcpu, &gfn, &pfn, &level);
3420 r = __direct_map(vcpu, gpa, write, map_writable,
3421 level, gfn, pfn, prefault);
3422 spin_unlock(&vcpu->kvm->mmu_lock);
3427 spin_unlock(&vcpu->kvm->mmu_lock);
3428 kvm_release_pfn_clean(pfn);
3432 static void nonpaging_init_context(struct kvm_vcpu *vcpu,
3433 struct kvm_mmu *context)
3435 context->page_fault = nonpaging_page_fault;
3436 context->gva_to_gpa = nonpaging_gva_to_gpa;
3437 context->sync_page = nonpaging_sync_page;
3438 context->invlpg = nonpaging_invlpg;
3439 context->update_pte = nonpaging_update_pte;
3440 context->root_level = 0;
3441 context->shadow_root_level = PT32E_ROOT_LEVEL;
3442 context->root_hpa = INVALID_PAGE;
3443 context->direct_map = true;
3444 context->nx = false;
3447 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3449 ++vcpu->stat.tlb_flush;
3450 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
3452 EXPORT_SYMBOL_GPL(kvm_mmu_flush_tlb);
3454 void kvm_mmu_new_cr3(struct kvm_vcpu *vcpu)
3456 mmu_free_roots(vcpu);
3459 static unsigned long get_cr3(struct kvm_vcpu *vcpu)
3461 return kvm_read_cr3(vcpu);
3464 static void inject_page_fault(struct kvm_vcpu *vcpu,
3465 struct x86_exception *fault)
3467 vcpu->arch.mmu.inject_page_fault(vcpu, fault);
3470 static bool sync_mmio_spte(struct kvm *kvm, u64 *sptep, gfn_t gfn,
3471 unsigned access, int *nr_present)
3473 if (unlikely(is_mmio_spte(*sptep))) {
3474 if (gfn != get_mmio_spte_gfn(*sptep)) {
3475 mmu_spte_clear_no_track(sptep);
3480 mark_mmio_spte(kvm, sptep, gfn, access);
3487 static inline bool is_last_gpte(struct kvm_mmu *mmu, unsigned level, unsigned gpte)
3492 index |= (gpte & PT_PAGE_SIZE_MASK) >> (PT_PAGE_SIZE_SHIFT - 2);
3493 return mmu->last_pte_bitmap & (1 << index);
3496 #define PTTYPE_EPT 18 /* arbitrary */
3497 #define PTTYPE PTTYPE_EPT
3498 #include "paging_tmpl.h"
3502 #include "paging_tmpl.h"
3506 #include "paging_tmpl.h"
3509 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
3510 struct kvm_mmu *context)
3512 int maxphyaddr = cpuid_maxphyaddr(vcpu);
3513 u64 exb_bit_rsvd = 0;
3514 u64 gbpages_bit_rsvd = 0;
3515 u64 nonleaf_bit8_rsvd = 0;
3517 context->bad_mt_xwr = 0;
3520 exb_bit_rsvd = rsvd_bits(63, 63);
3521 if (!guest_cpuid_has_gbpages(vcpu))
3522 gbpages_bit_rsvd = rsvd_bits(7, 7);
3525 * Non-leaf PML4Es and PDPEs reserve bit 8 (which would be the G bit for
3526 * leaf entries) on AMD CPUs only.
3528 if (guest_cpuid_is_amd(vcpu))
3529 nonleaf_bit8_rsvd = rsvd_bits(8, 8);
3531 switch (context->root_level) {
3532 case PT32_ROOT_LEVEL:
3533 /* no rsvd bits for 2 level 4K page table entries */
3534 context->rsvd_bits_mask[0][1] = 0;
3535 context->rsvd_bits_mask[0][0] = 0;
3536 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3538 if (!is_pse(vcpu)) {
3539 context->rsvd_bits_mask[1][1] = 0;
3543 if (is_cpuid_PSE36())
3544 /* 36bits PSE 4MB page */
3545 context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
3547 /* 32 bits PSE 4MB page */
3548 context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
3550 case PT32E_ROOT_LEVEL:
3551 context->rsvd_bits_mask[0][2] =
3552 rsvd_bits(maxphyaddr, 63) |
3553 rsvd_bits(5, 8) | rsvd_bits(1, 2); /* PDPTE */
3554 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
3555 rsvd_bits(maxphyaddr, 62); /* PDE */
3556 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
3557 rsvd_bits(maxphyaddr, 62); /* PTE */
3558 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
3559 rsvd_bits(maxphyaddr, 62) |
3560 rsvd_bits(13, 20); /* large page */
3561 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3563 case PT64_ROOT_LEVEL:
3564 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
3565 nonleaf_bit8_rsvd | rsvd_bits(7, 7) | rsvd_bits(maxphyaddr, 51);
3566 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
3567 nonleaf_bit8_rsvd | gbpages_bit_rsvd | rsvd_bits(maxphyaddr, 51);
3568 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
3569 rsvd_bits(maxphyaddr, 51);
3570 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
3571 rsvd_bits(maxphyaddr, 51);
3572 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
3573 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
3574 gbpages_bit_rsvd | rsvd_bits(maxphyaddr, 51) |
3576 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
3577 rsvd_bits(maxphyaddr, 51) |
3578 rsvd_bits(13, 20); /* large page */
3579 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3584 static void reset_rsvds_bits_mask_ept(struct kvm_vcpu *vcpu,
3585 struct kvm_mmu *context, bool execonly)
3587 int maxphyaddr = cpuid_maxphyaddr(vcpu);
3590 context->rsvd_bits_mask[0][3] =
3591 rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 7);
3592 context->rsvd_bits_mask[0][2] =
3593 rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 6);
3594 context->rsvd_bits_mask[0][1] =
3595 rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 6);
3596 context->rsvd_bits_mask[0][0] = rsvd_bits(maxphyaddr, 51);
3599 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
3600 context->rsvd_bits_mask[1][2] =
3601 rsvd_bits(maxphyaddr, 51) | rsvd_bits(12, 29);
3602 context->rsvd_bits_mask[1][1] =
3603 rsvd_bits(maxphyaddr, 51) | rsvd_bits(12, 20);
3604 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3606 for (pte = 0; pte < 64; pte++) {
3607 int rwx_bits = pte & 7;
3609 if (mt == 0x2 || mt == 0x3 || mt == 0x7 ||
3610 rwx_bits == 0x2 || rwx_bits == 0x6 ||
3611 (rwx_bits == 0x4 && !execonly))
3612 context->bad_mt_xwr |= (1ull << pte);
3616 void update_permission_bitmask(struct kvm_vcpu *vcpu,
3617 struct kvm_mmu *mmu, bool ept)
3619 unsigned bit, byte, pfec;
3621 bool fault, x, w, u, wf, uf, ff, smapf, cr4_smap, cr4_smep, smap = 0;
3623 cr4_smep = kvm_read_cr4_bits(vcpu, X86_CR4_SMEP);
3624 cr4_smap = kvm_read_cr4_bits(vcpu, X86_CR4_SMAP);
3625 for (byte = 0; byte < ARRAY_SIZE(mmu->permissions); ++byte) {
3628 wf = pfec & PFERR_WRITE_MASK;
3629 uf = pfec & PFERR_USER_MASK;
3630 ff = pfec & PFERR_FETCH_MASK;
3632 * PFERR_RSVD_MASK bit is set in PFEC if the access is not
3633 * subject to SMAP restrictions, and cleared otherwise. The
3634 * bit is only meaningful if the SMAP bit is set in CR4.
3636 smapf = !(pfec & PFERR_RSVD_MASK);
3637 for (bit = 0; bit < 8; ++bit) {
3638 x = bit & ACC_EXEC_MASK;
3639 w = bit & ACC_WRITE_MASK;
3640 u = bit & ACC_USER_MASK;
3643 /* Not really needed: !nx will cause pte.nx to fault */
3645 /* Allow supervisor writes if !cr0.wp */
3646 w |= !is_write_protection(vcpu) && !uf;
3647 /* Disallow supervisor fetches of user code if cr4.smep */
3648 x &= !(cr4_smep && u && !uf);
3651 * SMAP:kernel-mode data accesses from user-mode
3652 * mappings should fault. A fault is considered
3653 * as a SMAP violation if all of the following
3654 * conditions are ture:
3655 * - X86_CR4_SMAP is set in CR4
3656 * - An user page is accessed
3657 * - Page fault in kernel mode
3658 * - if CPL = 3 or X86_EFLAGS_AC is clear
3660 * Here, we cover the first three conditions.
3661 * The fourth is computed dynamically in
3662 * permission_fault() and is in smapf.
3664 * Also, SMAP does not affect instruction
3665 * fetches, add the !ff check here to make it
3668 smap = cr4_smap && u && !uf && !ff;
3670 /* Not really needed: no U/S accesses on ept */
3673 fault = (ff && !x) || (uf && !u) || (wf && !w) ||
3675 map |= fault << bit;
3677 mmu->permissions[byte] = map;
3681 static void update_last_pte_bitmap(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
3684 unsigned level, root_level = mmu->root_level;
3685 const unsigned ps_set_index = 1 << 2; /* bit 2 of index: ps */
3687 if (root_level == PT32E_ROOT_LEVEL)
3689 /* PT_PAGE_TABLE_LEVEL always terminates */
3690 map = 1 | (1 << ps_set_index);
3691 for (level = PT_DIRECTORY_LEVEL; level <= root_level; ++level) {
3692 if (level <= PT_PDPE_LEVEL
3693 && (mmu->root_level >= PT32E_ROOT_LEVEL || is_pse(vcpu)))
3694 map |= 1 << (ps_set_index | (level - 1));
3696 mmu->last_pte_bitmap = map;
3699 static void paging64_init_context_common(struct kvm_vcpu *vcpu,
3700 struct kvm_mmu *context,
3703 context->nx = is_nx(vcpu);
3704 context->root_level = level;
3706 reset_rsvds_bits_mask(vcpu, context);
3707 update_permission_bitmask(vcpu, context, false);
3708 update_last_pte_bitmap(vcpu, context);
3710 ASSERT(is_pae(vcpu));
3711 context->page_fault = paging64_page_fault;
3712 context->gva_to_gpa = paging64_gva_to_gpa;
3713 context->sync_page = paging64_sync_page;
3714 context->invlpg = paging64_invlpg;
3715 context->update_pte = paging64_update_pte;
3716 context->shadow_root_level = level;
3717 context->root_hpa = INVALID_PAGE;
3718 context->direct_map = false;
3721 static void paging64_init_context(struct kvm_vcpu *vcpu,
3722 struct kvm_mmu *context)
3724 paging64_init_context_common(vcpu, context, PT64_ROOT_LEVEL);
3727 static void paging32_init_context(struct kvm_vcpu *vcpu,
3728 struct kvm_mmu *context)
3730 context->nx = false;
3731 context->root_level = PT32_ROOT_LEVEL;
3733 reset_rsvds_bits_mask(vcpu, context);
3734 update_permission_bitmask(vcpu, context, false);
3735 update_last_pte_bitmap(vcpu, context);
3737 context->page_fault = paging32_page_fault;
3738 context->gva_to_gpa = paging32_gva_to_gpa;
3739 context->sync_page = paging32_sync_page;
3740 context->invlpg = paging32_invlpg;
3741 context->update_pte = paging32_update_pte;
3742 context->shadow_root_level = PT32E_ROOT_LEVEL;
3743 context->root_hpa = INVALID_PAGE;
3744 context->direct_map = false;
3747 static void paging32E_init_context(struct kvm_vcpu *vcpu,
3748 struct kvm_mmu *context)
3750 paging64_init_context_common(vcpu, context, PT32E_ROOT_LEVEL);
3753 static void init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
3755 struct kvm_mmu *context = vcpu->arch.walk_mmu;
3757 context->base_role.word = 0;
3758 context->page_fault = tdp_page_fault;
3759 context->sync_page = nonpaging_sync_page;
3760 context->invlpg = nonpaging_invlpg;
3761 context->update_pte = nonpaging_update_pte;
3762 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
3763 context->root_hpa = INVALID_PAGE;
3764 context->direct_map = true;
3765 context->set_cr3 = kvm_x86_ops->set_tdp_cr3;
3766 context->get_cr3 = get_cr3;
3767 context->get_pdptr = kvm_pdptr_read;
3768 context->inject_page_fault = kvm_inject_page_fault;
3770 if (!is_paging(vcpu)) {
3771 context->nx = false;
3772 context->gva_to_gpa = nonpaging_gva_to_gpa;
3773 context->root_level = 0;
3774 } else if (is_long_mode(vcpu)) {
3775 context->nx = is_nx(vcpu);
3776 context->root_level = PT64_ROOT_LEVEL;
3777 reset_rsvds_bits_mask(vcpu, context);
3778 context->gva_to_gpa = paging64_gva_to_gpa;
3779 } else if (is_pae(vcpu)) {
3780 context->nx = is_nx(vcpu);
3781 context->root_level = PT32E_ROOT_LEVEL;
3782 reset_rsvds_bits_mask(vcpu, context);
3783 context->gva_to_gpa = paging64_gva_to_gpa;
3785 context->nx = false;
3786 context->root_level = PT32_ROOT_LEVEL;
3787 reset_rsvds_bits_mask(vcpu, context);
3788 context->gva_to_gpa = paging32_gva_to_gpa;
3791 update_permission_bitmask(vcpu, context, false);
3792 update_last_pte_bitmap(vcpu, context);
3795 void kvm_init_shadow_mmu(struct kvm_vcpu *vcpu, struct kvm_mmu *context)
3797 bool smep = kvm_read_cr4_bits(vcpu, X86_CR4_SMEP);
3799 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3801 if (!is_paging(vcpu))
3802 nonpaging_init_context(vcpu, context);
3803 else if (is_long_mode(vcpu))
3804 paging64_init_context(vcpu, context);
3805 else if (is_pae(vcpu))
3806 paging32E_init_context(vcpu, context);
3808 paging32_init_context(vcpu, context);
3810 vcpu->arch.mmu.base_role.nxe = is_nx(vcpu);
3811 vcpu->arch.mmu.base_role.cr4_pae = !!is_pae(vcpu);
3812 vcpu->arch.mmu.base_role.cr0_wp = is_write_protection(vcpu);
3813 vcpu->arch.mmu.base_role.smep_andnot_wp
3814 = smep && !is_write_protection(vcpu);
3816 EXPORT_SYMBOL_GPL(kvm_init_shadow_mmu);
3818 void kvm_init_shadow_ept_mmu(struct kvm_vcpu *vcpu, struct kvm_mmu *context,
3822 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3824 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
3827 context->page_fault = ept_page_fault;
3828 context->gva_to_gpa = ept_gva_to_gpa;
3829 context->sync_page = ept_sync_page;
3830 context->invlpg = ept_invlpg;
3831 context->update_pte = ept_update_pte;
3832 context->root_level = context->shadow_root_level;
3833 context->root_hpa = INVALID_PAGE;
3834 context->direct_map = false;
3836 update_permission_bitmask(vcpu, context, true);
3837 reset_rsvds_bits_mask_ept(vcpu, context, execonly);
3839 EXPORT_SYMBOL_GPL(kvm_init_shadow_ept_mmu);
3841 static void init_kvm_softmmu(struct kvm_vcpu *vcpu)
3843 kvm_init_shadow_mmu(vcpu, vcpu->arch.walk_mmu);
3844 vcpu->arch.walk_mmu->set_cr3 = kvm_x86_ops->set_cr3;
3845 vcpu->arch.walk_mmu->get_cr3 = get_cr3;
3846 vcpu->arch.walk_mmu->get_pdptr = kvm_pdptr_read;
3847 vcpu->arch.walk_mmu->inject_page_fault = kvm_inject_page_fault;
3850 static void init_kvm_nested_mmu(struct kvm_vcpu *vcpu)
3852 struct kvm_mmu *g_context = &vcpu->arch.nested_mmu;
3854 g_context->get_cr3 = get_cr3;
3855 g_context->get_pdptr = kvm_pdptr_read;
3856 g_context->inject_page_fault = kvm_inject_page_fault;
3859 * Note that arch.mmu.gva_to_gpa translates l2_gva to l1_gpa. The
3860 * translation of l2_gpa to l1_gpa addresses is done using the
3861 * arch.nested_mmu.gva_to_gpa function. Basically the gva_to_gpa
3862 * functions between mmu and nested_mmu are swapped.
3864 if (!is_paging(vcpu)) {
3865 g_context->nx = false;
3866 g_context->root_level = 0;
3867 g_context->gva_to_gpa = nonpaging_gva_to_gpa_nested;
3868 } else if (is_long_mode(vcpu)) {
3869 g_context->nx = is_nx(vcpu);
3870 g_context->root_level = PT64_ROOT_LEVEL;
3871 reset_rsvds_bits_mask(vcpu, g_context);
3872 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
3873 } else if (is_pae(vcpu)) {
3874 g_context->nx = is_nx(vcpu);
3875 g_context->root_level = PT32E_ROOT_LEVEL;
3876 reset_rsvds_bits_mask(vcpu, g_context);
3877 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
3879 g_context->nx = false;
3880 g_context->root_level = PT32_ROOT_LEVEL;
3881 reset_rsvds_bits_mask(vcpu, g_context);
3882 g_context->gva_to_gpa = paging32_gva_to_gpa_nested;
3885 update_permission_bitmask(vcpu, g_context, false);
3886 update_last_pte_bitmap(vcpu, g_context);
3889 static void init_kvm_mmu(struct kvm_vcpu *vcpu)
3891 if (mmu_is_nested(vcpu))
3892 return init_kvm_nested_mmu(vcpu);
3893 else if (tdp_enabled)
3894 return init_kvm_tdp_mmu(vcpu);
3896 return init_kvm_softmmu(vcpu);
3899 void kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
3903 kvm_mmu_unload(vcpu);
3906 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
3908 int kvm_mmu_load(struct kvm_vcpu *vcpu)
3912 r = mmu_topup_memory_caches(vcpu);
3915 r = mmu_alloc_roots(vcpu);
3916 kvm_mmu_sync_roots(vcpu);
3919 /* set_cr3() should ensure TLB has been flushed */
3920 vcpu->arch.mmu.set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
3924 EXPORT_SYMBOL_GPL(kvm_mmu_load);
3926 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
3928 mmu_free_roots(vcpu);
3929 WARN_ON(VALID_PAGE(vcpu->arch.mmu.root_hpa));
3931 EXPORT_SYMBOL_GPL(kvm_mmu_unload);
3933 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
3934 struct kvm_mmu_page *sp, u64 *spte,
3937 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
3938 ++vcpu->kvm->stat.mmu_pde_zapped;
3942 ++vcpu->kvm->stat.mmu_pte_updated;
3943 vcpu->arch.mmu.update_pte(vcpu, sp, spte, new);
3946 static bool need_remote_flush(u64 old, u64 new)
3948 if (!is_shadow_present_pte(old))
3950 if (!is_shadow_present_pte(new))
3952 if ((old ^ new) & PT64_BASE_ADDR_MASK)
3954 old ^= shadow_nx_mask;
3955 new ^= shadow_nx_mask;
3956 return (old & ~new & PT64_PERM_MASK) != 0;
3959 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, bool zap_page,
3960 bool remote_flush, bool local_flush)
3966 kvm_flush_remote_tlbs(vcpu->kvm);
3967 else if (local_flush)
3968 kvm_mmu_flush_tlb(vcpu);
3971 static u64 mmu_pte_write_fetch_gpte(struct kvm_vcpu *vcpu, gpa_t *gpa,
3972 const u8 *new, int *bytes)
3978 * Assume that the pte write on a page table of the same type
3979 * as the current vcpu paging mode since we update the sptes only
3980 * when they have the same mode.
3982 if (is_pae(vcpu) && *bytes == 4) {
3983 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
3986 r = kvm_read_guest(vcpu->kvm, *gpa, &gentry, 8);
3989 new = (const u8 *)&gentry;
3994 gentry = *(const u32 *)new;
3997 gentry = *(const u64 *)new;
4008 * If we're seeing too many writes to a page, it may no longer be a page table,
4009 * or we may be forking, in which case it is better to unmap the page.
4011 static bool detect_write_flooding(struct kvm_mmu_page *sp)
4014 * Skip write-flooding detected for the sp whose level is 1, because
4015 * it can become unsync, then the guest page is not write-protected.
4017 if (sp->role.level == PT_PAGE_TABLE_LEVEL)
4020 return ++sp->write_flooding_count >= 3;
4024 * Misaligned accesses are too much trouble to fix up; also, they usually
4025 * indicate a page is not used as a page table.
4027 static bool detect_write_misaligned(struct kvm_mmu_page *sp, gpa_t gpa,
4030 unsigned offset, pte_size, misaligned;
4032 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
4033 gpa, bytes, sp->role.word);
4035 offset = offset_in_page(gpa);
4036 pte_size = sp->role.cr4_pae ? 8 : 4;
4039 * Sometimes, the OS only writes the last one bytes to update status
4040 * bits, for example, in linux, andb instruction is used in clear_bit().
4042 if (!(offset & (pte_size - 1)) && bytes == 1)
4045 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
4046 misaligned |= bytes < 4;
4051 static u64 *get_written_sptes(struct kvm_mmu_page *sp, gpa_t gpa, int *nspte)
4053 unsigned page_offset, quadrant;
4057 page_offset = offset_in_page(gpa);
4058 level = sp->role.level;
4060 if (!sp->role.cr4_pae) {
4061 page_offset <<= 1; /* 32->64 */
4063 * A 32-bit pde maps 4MB while the shadow pdes map
4064 * only 2MB. So we need to double the offset again
4065 * and zap two pdes instead of one.
4067 if (level == PT32_ROOT_LEVEL) {
4068 page_offset &= ~7; /* kill rounding error */
4072 quadrant = page_offset >> PAGE_SHIFT;
4073 page_offset &= ~PAGE_MASK;
4074 if (quadrant != sp->role.quadrant)
4078 spte = &sp->spt[page_offset / sizeof(*spte)];
4082 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
4083 const u8 *new, int bytes)
4085 gfn_t gfn = gpa >> PAGE_SHIFT;
4086 union kvm_mmu_page_role mask = { .word = 0 };
4087 struct kvm_mmu_page *sp;
4088 LIST_HEAD(invalid_list);
4089 u64 entry, gentry, *spte;
4091 bool remote_flush, local_flush, zap_page;
4094 * If we don't have indirect shadow pages, it means no page is
4095 * write-protected, so we can exit simply.
4097 if (!ACCESS_ONCE(vcpu->kvm->arch.indirect_shadow_pages))
4100 zap_page = remote_flush = local_flush = false;
4102 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
4104 gentry = mmu_pte_write_fetch_gpte(vcpu, &gpa, new, &bytes);
4107 * No need to care whether allocation memory is successful
4108 * or not since pte prefetch is skiped if it does not have
4109 * enough objects in the cache.
4111 mmu_topup_memory_caches(vcpu);
4113 spin_lock(&vcpu->kvm->mmu_lock);
4114 ++vcpu->kvm->stat.mmu_pte_write;
4115 kvm_mmu_audit(vcpu, AUDIT_PRE_PTE_WRITE);
4117 mask.cr0_wp = mask.cr4_pae = mask.nxe = 1;
4118 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn) {
4119 if (detect_write_misaligned(sp, gpa, bytes) ||
4120 detect_write_flooding(sp)) {
4121 zap_page |= !!kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
4123 ++vcpu->kvm->stat.mmu_flooded;
4127 spte = get_written_sptes(sp, gpa, &npte);
4134 mmu_page_zap_pte(vcpu->kvm, sp, spte);
4136 !((sp->role.word ^ vcpu->arch.mmu.base_role.word)
4137 & mask.word) && rmap_can_add(vcpu))
4138 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
4139 if (need_remote_flush(entry, *spte))
4140 remote_flush = true;
4144 mmu_pte_write_flush_tlb(vcpu, zap_page, remote_flush, local_flush);
4145 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
4146 kvm_mmu_audit(vcpu, AUDIT_POST_PTE_WRITE);
4147 spin_unlock(&vcpu->kvm->mmu_lock);
4150 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
4155 if (vcpu->arch.mmu.direct_map)
4158 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
4160 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
4164 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
4166 static void make_mmu_pages_available(struct kvm_vcpu *vcpu)
4168 LIST_HEAD(invalid_list);
4170 if (likely(kvm_mmu_available_pages(vcpu->kvm) >= KVM_MIN_FREE_MMU_PAGES))
4173 while (kvm_mmu_available_pages(vcpu->kvm) < KVM_REFILL_PAGES) {
4174 if (!prepare_zap_oldest_mmu_page(vcpu->kvm, &invalid_list))
4177 ++vcpu->kvm->stat.mmu_recycled;
4179 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
4182 static bool is_mmio_page_fault(struct kvm_vcpu *vcpu, gva_t addr)
4184 if (vcpu->arch.mmu.direct_map || mmu_is_nested(vcpu))
4185 return vcpu_match_mmio_gpa(vcpu, addr);
4187 return vcpu_match_mmio_gva(vcpu, addr);
4190 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code,
4191 void *insn, int insn_len)
4193 int r, emulation_type = EMULTYPE_RETRY;
4194 enum emulation_result er;
4196 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code, false);
4205 if (is_mmio_page_fault(vcpu, cr2))
4208 er = x86_emulate_instruction(vcpu, cr2, emulation_type, insn, insn_len);
4213 case EMULATE_USER_EXIT:
4214 ++vcpu->stat.mmio_exits;
4224 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
4226 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
4228 vcpu->arch.mmu.invlpg(vcpu, gva);
4229 kvm_mmu_flush_tlb(vcpu);
4230 ++vcpu->stat.invlpg;
4232 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
4234 void kvm_enable_tdp(void)
4238 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
4240 void kvm_disable_tdp(void)
4242 tdp_enabled = false;
4244 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
4246 static void free_mmu_pages(struct kvm_vcpu *vcpu)
4248 free_page((unsigned long)vcpu->arch.mmu.pae_root);
4249 if (vcpu->arch.mmu.lm_root != NULL)
4250 free_page((unsigned long)vcpu->arch.mmu.lm_root);
4253 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
4261 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
4262 * Therefore we need to allocate shadow page tables in the first
4263 * 4GB of memory, which happens to fit the DMA32 zone.
4265 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
4269 vcpu->arch.mmu.pae_root = page_address(page);
4270 for (i = 0; i < 4; ++i)
4271 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
4276 int kvm_mmu_create(struct kvm_vcpu *vcpu)
4280 vcpu->arch.walk_mmu = &vcpu->arch.mmu;
4281 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
4282 vcpu->arch.mmu.translate_gpa = translate_gpa;
4283 vcpu->arch.nested_mmu.translate_gpa = translate_nested_gpa;
4285 return alloc_mmu_pages(vcpu);
4288 void kvm_mmu_setup(struct kvm_vcpu *vcpu)
4291 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
4296 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
4298 struct kvm_memory_slot *memslot;
4302 memslot = id_to_memslot(kvm->memslots, slot);
4303 last_gfn = memslot->base_gfn + memslot->npages - 1;
4305 spin_lock(&kvm->mmu_lock);
4307 for (i = PT_PAGE_TABLE_LEVEL;
4308 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
4309 unsigned long *rmapp;
4310 unsigned long last_index, index;
4312 rmapp = memslot->arch.rmap[i - PT_PAGE_TABLE_LEVEL];
4313 last_index = gfn_to_index(last_gfn, memslot->base_gfn, i);
4315 for (index = 0; index <= last_index; ++index, ++rmapp) {
4317 __rmap_write_protect(kvm, rmapp, false);
4319 if (need_resched() || spin_needbreak(&kvm->mmu_lock))
4320 cond_resched_lock(&kvm->mmu_lock);
4324 spin_unlock(&kvm->mmu_lock);
4327 * kvm_mmu_slot_remove_write_access() and kvm_vm_ioctl_get_dirty_log()
4328 * which do tlb flush out of mmu-lock should be serialized by
4329 * kvm->slots_lock otherwise tlb flush would be missed.
4331 lockdep_assert_held(&kvm->slots_lock);
4334 * We can flush all the TLBs out of the mmu lock without TLB
4335 * corruption since we just change the spte from writable to
4336 * readonly so that we only need to care the case of changing
4337 * spte from present to present (changing the spte from present
4338 * to nonpresent will flush all the TLBs immediately), in other
4339 * words, the only case we care is mmu_spte_update() where we
4340 * haved checked SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE
4341 * instead of PT_WRITABLE_MASK, that means it does not depend
4342 * on PT_WRITABLE_MASK anymore.
4344 kvm_flush_remote_tlbs(kvm);
4347 #define BATCH_ZAP_PAGES 10
4348 static void kvm_zap_obsolete_pages(struct kvm *kvm)
4350 struct kvm_mmu_page *sp, *node;
4354 list_for_each_entry_safe_reverse(sp, node,
4355 &kvm->arch.active_mmu_pages, link) {
4359 * No obsolete page exists before new created page since
4360 * active_mmu_pages is the FIFO list.
4362 if (!is_obsolete_sp(kvm, sp))
4366 * Since we are reversely walking the list and the invalid
4367 * list will be moved to the head, skip the invalid page
4368 * can help us to avoid the infinity list walking.
4370 if (sp->role.invalid)
4374 * Need not flush tlb since we only zap the sp with invalid
4375 * generation number.
4377 if (batch >= BATCH_ZAP_PAGES &&
4378 cond_resched_lock(&kvm->mmu_lock)) {
4383 ret = kvm_mmu_prepare_zap_page(kvm, sp,
4384 &kvm->arch.zapped_obsolete_pages);
4392 * Should flush tlb before free page tables since lockless-walking
4393 * may use the pages.
4395 kvm_mmu_commit_zap_page(kvm, &kvm->arch.zapped_obsolete_pages);
4399 * Fast invalidate all shadow pages and use lock-break technique
4400 * to zap obsolete pages.
4402 * It's required when memslot is being deleted or VM is being
4403 * destroyed, in these cases, we should ensure that KVM MMU does
4404 * not use any resource of the being-deleted slot or all slots
4405 * after calling the function.
4407 void kvm_mmu_invalidate_zap_all_pages(struct kvm *kvm)
4409 spin_lock(&kvm->mmu_lock);
4410 trace_kvm_mmu_invalidate_zap_all_pages(kvm);
4411 kvm->arch.mmu_valid_gen++;
4414 * Notify all vcpus to reload its shadow page table
4415 * and flush TLB. Then all vcpus will switch to new
4416 * shadow page table with the new mmu_valid_gen.
4418 * Note: we should do this under the protection of
4419 * mmu-lock, otherwise, vcpu would purge shadow page
4420 * but miss tlb flush.
4422 kvm_reload_remote_mmus(kvm);
4424 kvm_zap_obsolete_pages(kvm);
4425 spin_unlock(&kvm->mmu_lock);
4428 static bool kvm_has_zapped_obsolete_pages(struct kvm *kvm)
4430 return unlikely(!list_empty_careful(&kvm->arch.zapped_obsolete_pages));
4433 void kvm_mmu_invalidate_mmio_sptes(struct kvm *kvm)
4436 * The very rare case: if the generation-number is round,
4437 * zap all shadow pages.
4439 if (unlikely(kvm_current_mmio_generation(kvm) == 0)) {
4440 printk_ratelimited(KERN_INFO "kvm: zapping shadow pages for mmio generation wraparound\n");
4441 kvm_mmu_invalidate_zap_all_pages(kvm);
4445 static unsigned long
4446 mmu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
4449 int nr_to_scan = sc->nr_to_scan;
4450 unsigned long freed = 0;
4452 spin_lock(&kvm_lock);
4454 list_for_each_entry(kvm, &vm_list, vm_list) {
4456 LIST_HEAD(invalid_list);
4459 * Never scan more than sc->nr_to_scan VM instances.
4460 * Will not hit this condition practically since we do not try
4461 * to shrink more than one VM and it is very unlikely to see
4462 * !n_used_mmu_pages so many times.
4467 * n_used_mmu_pages is accessed without holding kvm->mmu_lock
4468 * here. We may skip a VM instance errorneosly, but we do not
4469 * want to shrink a VM that only started to populate its MMU
4472 if (!kvm->arch.n_used_mmu_pages &&
4473 !kvm_has_zapped_obsolete_pages(kvm))
4476 idx = srcu_read_lock(&kvm->srcu);
4477 spin_lock(&kvm->mmu_lock);
4479 if (kvm_has_zapped_obsolete_pages(kvm)) {
4480 kvm_mmu_commit_zap_page(kvm,
4481 &kvm->arch.zapped_obsolete_pages);
4485 if (prepare_zap_oldest_mmu_page(kvm, &invalid_list))
4487 kvm_mmu_commit_zap_page(kvm, &invalid_list);
4490 spin_unlock(&kvm->mmu_lock);
4491 srcu_read_unlock(&kvm->srcu, idx);
4494 * unfair on small ones
4495 * per-vm shrinkers cry out
4496 * sadness comes quickly
4498 list_move_tail(&kvm->vm_list, &vm_list);
4502 spin_unlock(&kvm_lock);
4506 static unsigned long
4507 mmu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
4509 return percpu_counter_read_positive(&kvm_total_used_mmu_pages);
4512 static struct shrinker mmu_shrinker = {
4513 .count_objects = mmu_shrink_count,
4514 .scan_objects = mmu_shrink_scan,
4515 .seeks = DEFAULT_SEEKS * 10,
4518 static void mmu_destroy_caches(void)
4520 if (pte_list_desc_cache)
4521 kmem_cache_destroy(pte_list_desc_cache);
4522 if (mmu_page_header_cache)
4523 kmem_cache_destroy(mmu_page_header_cache);
4526 int kvm_mmu_module_init(void)
4528 pte_list_desc_cache = kmem_cache_create("pte_list_desc",
4529 sizeof(struct pte_list_desc),
4531 if (!pte_list_desc_cache)
4534 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
4535 sizeof(struct kvm_mmu_page),
4537 if (!mmu_page_header_cache)
4540 if (percpu_counter_init(&kvm_total_used_mmu_pages, 0))
4543 register_shrinker(&mmu_shrinker);
4548 mmu_destroy_caches();
4553 * Caculate mmu pages needed for kvm.
4555 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
4557 unsigned int nr_mmu_pages;
4558 unsigned int nr_pages = 0;
4559 struct kvm_memslots *slots;
4560 struct kvm_memory_slot *memslot;
4562 slots = kvm_memslots(kvm);
4564 kvm_for_each_memslot(memslot, slots)
4565 nr_pages += memslot->npages;
4567 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
4568 nr_mmu_pages = max(nr_mmu_pages,
4569 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
4571 return nr_mmu_pages;
4574 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
4576 struct kvm_shadow_walk_iterator iterator;
4580 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
4583 walk_shadow_page_lockless_begin(vcpu);
4584 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte) {
4585 sptes[iterator.level-1] = spte;
4587 if (!is_shadow_present_pte(spte))
4590 walk_shadow_page_lockless_end(vcpu);
4594 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
4596 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
4600 kvm_mmu_unload(vcpu);
4601 free_mmu_pages(vcpu);
4602 mmu_free_memory_caches(vcpu);
4605 void kvm_mmu_module_exit(void)
4607 mmu_destroy_caches();
4608 percpu_counter_destroy(&kvm_total_used_mmu_pages);
4609 unregister_shrinker(&mmu_shrinker);
4610 mmu_audit_disable();