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"
26 #include <linux/kvm_host.h>
27 #include <linux/types.h>
28 #include <linux/string.h>
30 #include <linux/highmem.h>
31 #include <linux/module.h>
32 #include <linux/swap.h>
33 #include <linux/hugetlb.h>
34 #include <linux/compiler.h>
35 #include <linux/srcu.h>
36 #include <linux/slab.h>
37 #include <linux/uaccess.h>
40 #include <asm/cmpxchg.h>
45 * When setting this variable to true it enables Two-Dimensional-Paging
46 * where the hardware walks 2 page tables:
47 * 1. the guest-virtual to guest-physical
48 * 2. while doing 1. it walks guest-physical to host-physical
49 * If the hardware supports that we don't need to do shadow paging.
51 bool tdp_enabled = false;
55 AUDIT_POST_PAGE_FAULT,
66 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
67 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
71 #define pgprintk(x...) do { } while (0)
72 #define rmap_printk(x...) do { } while (0)
78 module_param(dbg, bool, 0644);
82 #define ASSERT(x) do { } while (0)
86 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
87 __FILE__, __LINE__, #x); \
91 #define PTE_PREFETCH_NUM 8
93 #define PT_FIRST_AVAIL_BITS_SHIFT 9
94 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
96 #define PT64_LEVEL_BITS 9
98 #define PT64_LEVEL_SHIFT(level) \
99 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
101 #define PT64_INDEX(address, level)\
102 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
105 #define PT32_LEVEL_BITS 10
107 #define PT32_LEVEL_SHIFT(level) \
108 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
110 #define PT32_LVL_OFFSET_MASK(level) \
111 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
112 * PT32_LEVEL_BITS))) - 1))
114 #define PT32_INDEX(address, level)\
115 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
118 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
119 #define PT64_DIR_BASE_ADDR_MASK \
120 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
121 #define PT64_LVL_ADDR_MASK(level) \
122 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
123 * PT64_LEVEL_BITS))) - 1))
124 #define PT64_LVL_OFFSET_MASK(level) \
125 (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
126 * PT64_LEVEL_BITS))) - 1))
128 #define PT32_BASE_ADDR_MASK PAGE_MASK
129 #define PT32_DIR_BASE_ADDR_MASK \
130 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
131 #define PT32_LVL_ADDR_MASK(level) \
132 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
133 * PT32_LEVEL_BITS))) - 1))
135 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
138 #define PTE_LIST_EXT 4
140 #define ACC_EXEC_MASK 1
141 #define ACC_WRITE_MASK PT_WRITABLE_MASK
142 #define ACC_USER_MASK PT_USER_MASK
143 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
145 #include <trace/events/kvm.h>
147 #define CREATE_TRACE_POINTS
148 #include "mmutrace.h"
150 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
152 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
154 struct pte_list_desc {
155 u64 *sptes[PTE_LIST_EXT];
156 struct pte_list_desc *more;
159 struct kvm_shadow_walk_iterator {
167 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
168 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
169 shadow_walk_okay(&(_walker)); \
170 shadow_walk_next(&(_walker)))
172 #define for_each_shadow_entry_lockless(_vcpu, _addr, _walker, spte) \
173 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
174 shadow_walk_okay(&(_walker)) && \
175 ({ spte = mmu_spte_get_lockless(_walker.sptep); 1; }); \
176 __shadow_walk_next(&(_walker), spte))
178 static struct kmem_cache *pte_list_desc_cache;
179 static struct kmem_cache *mmu_page_header_cache;
180 static struct percpu_counter kvm_total_used_mmu_pages;
182 static u64 __read_mostly shadow_nx_mask;
183 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
184 static u64 __read_mostly shadow_user_mask;
185 static u64 __read_mostly shadow_accessed_mask;
186 static u64 __read_mostly shadow_dirty_mask;
187 static u64 __read_mostly shadow_mmio_mask;
189 static void mmu_spte_set(u64 *sptep, u64 spte);
191 void kvm_mmu_set_mmio_spte_mask(u64 mmio_mask)
193 shadow_mmio_mask = mmio_mask;
195 EXPORT_SYMBOL_GPL(kvm_mmu_set_mmio_spte_mask);
197 static void mark_mmio_spte(u64 *sptep, u64 gfn, unsigned access)
199 access &= ACC_WRITE_MASK | ACC_USER_MASK;
201 trace_mark_mmio_spte(sptep, gfn, access);
202 mmu_spte_set(sptep, shadow_mmio_mask | access | gfn << PAGE_SHIFT);
205 static bool is_mmio_spte(u64 spte)
207 return (spte & shadow_mmio_mask) == shadow_mmio_mask;
210 static gfn_t get_mmio_spte_gfn(u64 spte)
212 return (spte & ~shadow_mmio_mask) >> PAGE_SHIFT;
215 static unsigned get_mmio_spte_access(u64 spte)
217 return (spte & ~shadow_mmio_mask) & ~PAGE_MASK;
220 static bool set_mmio_spte(u64 *sptep, gfn_t gfn, pfn_t pfn, unsigned access)
222 if (unlikely(is_noslot_pfn(pfn))) {
223 mark_mmio_spte(sptep, gfn, access);
230 static inline u64 rsvd_bits(int s, int e)
232 return ((1ULL << (e - s + 1)) - 1) << s;
235 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
236 u64 dirty_mask, u64 nx_mask, u64 x_mask)
238 shadow_user_mask = user_mask;
239 shadow_accessed_mask = accessed_mask;
240 shadow_dirty_mask = dirty_mask;
241 shadow_nx_mask = nx_mask;
242 shadow_x_mask = x_mask;
244 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
246 static int is_cpuid_PSE36(void)
251 static int is_nx(struct kvm_vcpu *vcpu)
253 return vcpu->arch.efer & EFER_NX;
256 static int is_shadow_present_pte(u64 pte)
258 return pte & PT_PRESENT_MASK && !is_mmio_spte(pte);
261 static int is_large_pte(u64 pte)
263 return pte & PT_PAGE_SIZE_MASK;
266 static int is_dirty_gpte(unsigned long pte)
268 return pte & PT_DIRTY_MASK;
271 static int is_rmap_spte(u64 pte)
273 return is_shadow_present_pte(pte);
276 static int is_last_spte(u64 pte, int level)
278 if (level == PT_PAGE_TABLE_LEVEL)
280 if (is_large_pte(pte))
285 static pfn_t spte_to_pfn(u64 pte)
287 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
290 static gfn_t pse36_gfn_delta(u32 gpte)
292 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
294 return (gpte & PT32_DIR_PSE36_MASK) << shift;
298 static void __set_spte(u64 *sptep, u64 spte)
303 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
308 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
310 return xchg(sptep, spte);
313 static u64 __get_spte_lockless(u64 *sptep)
315 return ACCESS_ONCE(*sptep);
318 static bool __check_direct_spte_mmio_pf(u64 spte)
320 /* It is valid if the spte is zapped. */
332 static void count_spte_clear(u64 *sptep, u64 spte)
334 struct kvm_mmu_page *sp = page_header(__pa(sptep));
336 if (is_shadow_present_pte(spte))
339 /* Ensure the spte is completely set before we increase the count */
341 sp->clear_spte_count++;
344 static void __set_spte(u64 *sptep, u64 spte)
346 union split_spte *ssptep, sspte;
348 ssptep = (union split_spte *)sptep;
349 sspte = (union split_spte)spte;
351 ssptep->spte_high = sspte.spte_high;
354 * If we map the spte from nonpresent to present, We should store
355 * the high bits firstly, then set present bit, so cpu can not
356 * fetch this spte while we are setting the spte.
360 ssptep->spte_low = sspte.spte_low;
363 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
365 union split_spte *ssptep, sspte;
367 ssptep = (union split_spte *)sptep;
368 sspte = (union split_spte)spte;
370 ssptep->spte_low = sspte.spte_low;
373 * If we map the spte from present to nonpresent, we should clear
374 * present bit firstly to avoid vcpu fetch the old high bits.
378 ssptep->spte_high = sspte.spte_high;
379 count_spte_clear(sptep, spte);
382 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
384 union split_spte *ssptep, sspte, orig;
386 ssptep = (union split_spte *)sptep;
387 sspte = (union split_spte)spte;
389 /* xchg acts as a barrier before the setting of the high bits */
390 orig.spte_low = xchg(&ssptep->spte_low, sspte.spte_low);
391 orig.spte_high = ssptep->spte_high;
392 ssptep->spte_high = sspte.spte_high;
393 count_spte_clear(sptep, spte);
399 * The idea using the light way get the spte on x86_32 guest is from
400 * gup_get_pte(arch/x86/mm/gup.c).
401 * The difference is we can not catch the spte tlb flush if we leave
402 * guest mode, so we emulate it by increase clear_spte_count when spte
405 static u64 __get_spte_lockless(u64 *sptep)
407 struct kvm_mmu_page *sp = page_header(__pa(sptep));
408 union split_spte spte, *orig = (union split_spte *)sptep;
412 count = sp->clear_spte_count;
415 spte.spte_low = orig->spte_low;
418 spte.spte_high = orig->spte_high;
421 if (unlikely(spte.spte_low != orig->spte_low ||
422 count != sp->clear_spte_count))
428 static bool __check_direct_spte_mmio_pf(u64 spte)
430 union split_spte sspte = (union split_spte)spte;
431 u32 high_mmio_mask = shadow_mmio_mask >> 32;
433 /* It is valid if the spte is zapped. */
437 /* It is valid if the spte is being zapped. */
438 if (sspte.spte_low == 0ull &&
439 (sspte.spte_high & high_mmio_mask) == high_mmio_mask)
446 static bool spte_has_volatile_bits(u64 spte)
448 if (!shadow_accessed_mask)
451 if (!is_shadow_present_pte(spte))
454 if ((spte & shadow_accessed_mask) &&
455 (!is_writable_pte(spte) || (spte & shadow_dirty_mask)))
461 static bool spte_is_bit_cleared(u64 old_spte, u64 new_spte, u64 bit_mask)
463 return (old_spte & bit_mask) && !(new_spte & bit_mask);
466 /* Rules for using mmu_spte_set:
467 * Set the sptep from nonpresent to present.
468 * Note: the sptep being assigned *must* be either not present
469 * or in a state where the hardware will not attempt to update
472 static void mmu_spte_set(u64 *sptep, u64 new_spte)
474 WARN_ON(is_shadow_present_pte(*sptep));
475 __set_spte(sptep, new_spte);
478 /* Rules for using mmu_spte_update:
479 * Update the state bits, it means the mapped pfn is not changged.
481 static void mmu_spte_update(u64 *sptep, u64 new_spte)
483 u64 mask, old_spte = *sptep;
485 WARN_ON(!is_rmap_spte(new_spte));
487 if (!is_shadow_present_pte(old_spte))
488 return mmu_spte_set(sptep, new_spte);
490 new_spte |= old_spte & shadow_dirty_mask;
492 mask = shadow_accessed_mask;
493 if (is_writable_pte(old_spte))
494 mask |= shadow_dirty_mask;
496 if (!spte_has_volatile_bits(old_spte) || (new_spte & mask) == mask)
497 __update_clear_spte_fast(sptep, new_spte);
499 old_spte = __update_clear_spte_slow(sptep, new_spte);
501 if (!shadow_accessed_mask)
504 if (spte_is_bit_cleared(old_spte, new_spte, shadow_accessed_mask))
505 kvm_set_pfn_accessed(spte_to_pfn(old_spte));
506 if (spte_is_bit_cleared(old_spte, new_spte, shadow_dirty_mask))
507 kvm_set_pfn_dirty(spte_to_pfn(old_spte));
511 * Rules for using mmu_spte_clear_track_bits:
512 * It sets the sptep from present to nonpresent, and track the
513 * state bits, it is used to clear the last level sptep.
515 static int mmu_spte_clear_track_bits(u64 *sptep)
518 u64 old_spte = *sptep;
520 if (!spte_has_volatile_bits(old_spte))
521 __update_clear_spte_fast(sptep, 0ull);
523 old_spte = __update_clear_spte_slow(sptep, 0ull);
525 if (!is_rmap_spte(old_spte))
528 pfn = spte_to_pfn(old_spte);
529 if (!shadow_accessed_mask || old_spte & shadow_accessed_mask)
530 kvm_set_pfn_accessed(pfn);
531 if (!shadow_dirty_mask || (old_spte & shadow_dirty_mask))
532 kvm_set_pfn_dirty(pfn);
537 * Rules for using mmu_spte_clear_no_track:
538 * Directly clear spte without caring the state bits of sptep,
539 * it is used to set the upper level spte.
541 static void mmu_spte_clear_no_track(u64 *sptep)
543 __update_clear_spte_fast(sptep, 0ull);
546 static u64 mmu_spte_get_lockless(u64 *sptep)
548 return __get_spte_lockless(sptep);
551 static void walk_shadow_page_lockless_begin(struct kvm_vcpu *vcpu)
554 atomic_inc(&vcpu->kvm->arch.reader_counter);
556 /* Increase the counter before walking shadow page table */
557 smp_mb__after_atomic_inc();
560 static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
562 /* Decrease the counter after walking shadow page table finished */
563 smp_mb__before_atomic_dec();
564 atomic_dec(&vcpu->kvm->arch.reader_counter);
568 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
569 struct kmem_cache *base_cache, int min)
573 if (cache->nobjs >= min)
575 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
576 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
579 cache->objects[cache->nobjs++] = obj;
584 static int mmu_memory_cache_free_objects(struct kvm_mmu_memory_cache *cache)
589 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc,
590 struct kmem_cache *cache)
593 kmem_cache_free(cache, mc->objects[--mc->nobjs]);
596 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
601 if (cache->nobjs >= min)
603 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
604 page = (void *)__get_free_page(GFP_KERNEL);
607 cache->objects[cache->nobjs++] = page;
612 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
615 free_page((unsigned long)mc->objects[--mc->nobjs]);
618 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
622 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
623 pte_list_desc_cache, 8 + PTE_PREFETCH_NUM);
626 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
629 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
630 mmu_page_header_cache, 4);
635 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
637 mmu_free_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
638 pte_list_desc_cache);
639 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
640 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache,
641 mmu_page_header_cache);
644 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
650 p = mc->objects[--mc->nobjs];
654 static struct pte_list_desc *mmu_alloc_pte_list_desc(struct kvm_vcpu *vcpu)
656 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_list_desc_cache,
657 sizeof(struct pte_list_desc));
660 static void mmu_free_pte_list_desc(struct pte_list_desc *pte_list_desc)
662 kmem_cache_free(pte_list_desc_cache, pte_list_desc);
665 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
667 if (!sp->role.direct)
668 return sp->gfns[index];
670 return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
673 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
676 BUG_ON(gfn != kvm_mmu_page_get_gfn(sp, index));
678 sp->gfns[index] = gfn;
682 * Return the pointer to the large page information for a given gfn,
683 * handling slots that are not large page aligned.
685 static struct kvm_lpage_info *lpage_info_slot(gfn_t gfn,
686 struct kvm_memory_slot *slot,
691 idx = (gfn >> KVM_HPAGE_GFN_SHIFT(level)) -
692 (slot->base_gfn >> KVM_HPAGE_GFN_SHIFT(level));
693 return &slot->lpage_info[level - 2][idx];
696 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
698 struct kvm_memory_slot *slot;
699 struct kvm_lpage_info *linfo;
702 slot = gfn_to_memslot(kvm, gfn);
703 for (i = PT_DIRECTORY_LEVEL;
704 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
705 linfo = lpage_info_slot(gfn, slot, i);
706 linfo->write_count += 1;
708 kvm->arch.indirect_shadow_pages++;
711 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
713 struct kvm_memory_slot *slot;
714 struct kvm_lpage_info *linfo;
717 slot = gfn_to_memslot(kvm, gfn);
718 for (i = PT_DIRECTORY_LEVEL;
719 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
720 linfo = lpage_info_slot(gfn, slot, i);
721 linfo->write_count -= 1;
722 WARN_ON(linfo->write_count < 0);
724 kvm->arch.indirect_shadow_pages--;
727 static int has_wrprotected_page(struct kvm *kvm,
731 struct kvm_memory_slot *slot;
732 struct kvm_lpage_info *linfo;
734 slot = gfn_to_memslot(kvm, gfn);
736 linfo = lpage_info_slot(gfn, slot, level);
737 return linfo->write_count;
743 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
745 unsigned long page_size;
748 page_size = kvm_host_page_size(kvm, gfn);
750 for (i = PT_PAGE_TABLE_LEVEL;
751 i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
752 if (page_size >= KVM_HPAGE_SIZE(i))
761 static struct kvm_memory_slot *
762 gfn_to_memslot_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t gfn,
765 struct kvm_memory_slot *slot;
767 slot = gfn_to_memslot(vcpu->kvm, gfn);
768 if (!slot || slot->flags & KVM_MEMSLOT_INVALID ||
769 (no_dirty_log && slot->dirty_bitmap))
775 static bool mapping_level_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t large_gfn)
777 return !gfn_to_memslot_dirty_bitmap(vcpu, large_gfn, true);
780 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
782 int host_level, level, max_level;
784 host_level = host_mapping_level(vcpu->kvm, large_gfn);
786 if (host_level == PT_PAGE_TABLE_LEVEL)
789 max_level = kvm_x86_ops->get_lpage_level() < host_level ?
790 kvm_x86_ops->get_lpage_level() : host_level;
792 for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
793 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
800 * Pte mapping structures:
802 * If pte_list bit zero is zero, then pte_list point to the spte.
804 * If pte_list bit zero is one, (then pte_list & ~1) points to a struct
805 * pte_list_desc containing more mappings.
807 * Returns the number of pte entries before the spte was added or zero if
808 * the spte was not added.
811 static int pte_list_add(struct kvm_vcpu *vcpu, u64 *spte,
812 unsigned long *pte_list)
814 struct pte_list_desc *desc;
818 rmap_printk("pte_list_add: %p %llx 0->1\n", spte, *spte);
819 *pte_list = (unsigned long)spte;
820 } else if (!(*pte_list & 1)) {
821 rmap_printk("pte_list_add: %p %llx 1->many\n", spte, *spte);
822 desc = mmu_alloc_pte_list_desc(vcpu);
823 desc->sptes[0] = (u64 *)*pte_list;
824 desc->sptes[1] = spte;
825 *pte_list = (unsigned long)desc | 1;
828 rmap_printk("pte_list_add: %p %llx many->many\n", spte, *spte);
829 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
830 while (desc->sptes[PTE_LIST_EXT-1] && desc->more) {
832 count += PTE_LIST_EXT;
834 if (desc->sptes[PTE_LIST_EXT-1]) {
835 desc->more = mmu_alloc_pte_list_desc(vcpu);
838 for (i = 0; desc->sptes[i]; ++i)
840 desc->sptes[i] = spte;
845 static u64 *pte_list_next(unsigned long *pte_list, u64 *spte)
847 struct pte_list_desc *desc;
853 else if (!(*pte_list & 1)) {
855 return (u64 *)*pte_list;
858 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
861 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i) {
862 if (prev_spte == spte)
863 return desc->sptes[i];
864 prev_spte = desc->sptes[i];
872 pte_list_desc_remove_entry(unsigned long *pte_list, struct pte_list_desc *desc,
873 int i, struct pte_list_desc *prev_desc)
877 for (j = PTE_LIST_EXT - 1; !desc->sptes[j] && j > i; --j)
879 desc->sptes[i] = desc->sptes[j];
880 desc->sptes[j] = NULL;
883 if (!prev_desc && !desc->more)
884 *pte_list = (unsigned long)desc->sptes[0];
887 prev_desc->more = desc->more;
889 *pte_list = (unsigned long)desc->more | 1;
890 mmu_free_pte_list_desc(desc);
893 static void pte_list_remove(u64 *spte, unsigned long *pte_list)
895 struct pte_list_desc *desc;
896 struct pte_list_desc *prev_desc;
900 printk(KERN_ERR "pte_list_remove: %p 0->BUG\n", spte);
902 } else if (!(*pte_list & 1)) {
903 rmap_printk("pte_list_remove: %p 1->0\n", spte);
904 if ((u64 *)*pte_list != spte) {
905 printk(KERN_ERR "pte_list_remove: %p 1->BUG\n", spte);
910 rmap_printk("pte_list_remove: %p many->many\n", spte);
911 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
914 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i)
915 if (desc->sptes[i] == spte) {
916 pte_list_desc_remove_entry(pte_list,
924 pr_err("pte_list_remove: %p many->many\n", spte);
929 typedef void (*pte_list_walk_fn) (u64 *spte);
930 static void pte_list_walk(unsigned long *pte_list, pte_list_walk_fn fn)
932 struct pte_list_desc *desc;
938 if (!(*pte_list & 1))
939 return fn((u64 *)*pte_list);
941 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
943 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i)
949 static unsigned long *__gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level,
950 struct kvm_memory_slot *slot)
952 struct kvm_lpage_info *linfo;
954 if (likely(level == PT_PAGE_TABLE_LEVEL))
955 return &slot->rmap[gfn - slot->base_gfn];
957 linfo = lpage_info_slot(gfn, slot, level);
958 return &linfo->rmap_pde;
962 * Take gfn and return the reverse mapping to it.
964 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
966 struct kvm_memory_slot *slot;
968 slot = gfn_to_memslot(kvm, gfn);
969 return __gfn_to_rmap(kvm, gfn, level, slot);
972 static bool rmap_can_add(struct kvm_vcpu *vcpu)
974 struct kvm_mmu_memory_cache *cache;
976 cache = &vcpu->arch.mmu_pte_list_desc_cache;
977 return mmu_memory_cache_free_objects(cache);
980 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
982 struct kvm_mmu_page *sp;
983 unsigned long *rmapp;
985 sp = page_header(__pa(spte));
986 kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
987 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
988 return pte_list_add(vcpu, spte, rmapp);
991 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
993 return pte_list_next(rmapp, spte);
996 static void rmap_remove(struct kvm *kvm, u64 *spte)
998 struct kvm_mmu_page *sp;
1000 unsigned long *rmapp;
1002 sp = page_header(__pa(spte));
1003 gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
1004 rmapp = gfn_to_rmap(kvm, gfn, sp->role.level);
1005 pte_list_remove(spte, rmapp);
1008 static void drop_spte(struct kvm *kvm, u64 *sptep)
1010 if (mmu_spte_clear_track_bits(sptep))
1011 rmap_remove(kvm, sptep);
1014 int kvm_mmu_rmap_write_protect(struct kvm *kvm, u64 gfn,
1015 struct kvm_memory_slot *slot)
1017 unsigned long *rmapp;
1019 int i, write_protected = 0;
1021 rmapp = __gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL, slot);
1022 spte = rmap_next(kvm, rmapp, NULL);
1024 BUG_ON(!(*spte & PT_PRESENT_MASK));
1025 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
1026 if (is_writable_pte(*spte)) {
1027 mmu_spte_update(spte, *spte & ~PT_WRITABLE_MASK);
1028 write_protected = 1;
1030 spte = rmap_next(kvm, rmapp, spte);
1033 /* check for huge page mappings */
1034 for (i = PT_DIRECTORY_LEVEL;
1035 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
1036 rmapp = __gfn_to_rmap(kvm, gfn, i, slot);
1037 spte = rmap_next(kvm, rmapp, NULL);
1039 BUG_ON(!(*spte & PT_PRESENT_MASK));
1040 BUG_ON(!is_large_pte(*spte));
1041 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
1042 if (is_writable_pte(*spte)) {
1043 drop_spte(kvm, spte);
1046 write_protected = 1;
1048 spte = rmap_next(kvm, rmapp, spte);
1052 return write_protected;
1055 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
1057 struct kvm_memory_slot *slot;
1059 slot = gfn_to_memslot(kvm, gfn);
1060 return kvm_mmu_rmap_write_protect(kvm, gfn, slot);
1063 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
1067 int need_tlb_flush = 0;
1069 while ((spte = rmap_next(kvm, rmapp, NULL))) {
1070 BUG_ON(!(*spte & PT_PRESENT_MASK));
1071 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
1072 drop_spte(kvm, spte);
1075 return need_tlb_flush;
1078 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
1082 u64 *spte, new_spte;
1083 pte_t *ptep = (pte_t *)data;
1086 WARN_ON(pte_huge(*ptep));
1087 new_pfn = pte_pfn(*ptep);
1088 spte = rmap_next(kvm, rmapp, NULL);
1090 BUG_ON(!is_shadow_present_pte(*spte));
1091 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
1093 if (pte_write(*ptep)) {
1094 drop_spte(kvm, spte);
1095 spte = rmap_next(kvm, rmapp, NULL);
1097 new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
1098 new_spte |= (u64)new_pfn << PAGE_SHIFT;
1100 new_spte &= ~PT_WRITABLE_MASK;
1101 new_spte &= ~SPTE_HOST_WRITEABLE;
1102 new_spte &= ~shadow_accessed_mask;
1103 mmu_spte_clear_track_bits(spte);
1104 mmu_spte_set(spte, new_spte);
1105 spte = rmap_next(kvm, rmapp, spte);
1109 kvm_flush_remote_tlbs(kvm);
1114 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
1116 int (*handler)(struct kvm *kvm, unsigned long *rmapp,
1117 unsigned long data))
1122 struct kvm_memslots *slots;
1123 struct kvm_memory_slot *memslot;
1125 slots = kvm_memslots(kvm);
1127 kvm_for_each_memslot(memslot, slots) {
1128 unsigned long start = memslot->userspace_addr;
1131 end = start + (memslot->npages << PAGE_SHIFT);
1132 if (hva >= start && hva < end) {
1133 gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
1134 gfn_t gfn = memslot->base_gfn + gfn_offset;
1136 ret = handler(kvm, &memslot->rmap[gfn_offset], data);
1138 for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
1139 struct kvm_lpage_info *linfo;
1141 linfo = lpage_info_slot(gfn, memslot,
1142 PT_DIRECTORY_LEVEL + j);
1143 ret |= handler(kvm, &linfo->rmap_pde, data);
1145 trace_kvm_age_page(hva, memslot, ret);
1153 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1155 return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
1158 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1160 kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
1163 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1170 * Emulate the accessed bit for EPT, by checking if this page has
1171 * an EPT mapping, and clearing it if it does. On the next access,
1172 * a new EPT mapping will be established.
1173 * This has some overhead, but not as much as the cost of swapping
1174 * out actively used pages or breaking up actively used hugepages.
1176 if (!shadow_accessed_mask)
1177 return kvm_unmap_rmapp(kvm, rmapp, data);
1179 spte = rmap_next(kvm, rmapp, NULL);
1183 BUG_ON(!(_spte & PT_PRESENT_MASK));
1184 _young = _spte & PT_ACCESSED_MASK;
1187 clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
1189 spte = rmap_next(kvm, rmapp, spte);
1194 static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1201 * If there's no access bit in the secondary pte set by the
1202 * hardware it's up to gup-fast/gup to set the access bit in
1203 * the primary pte or in the page structure.
1205 if (!shadow_accessed_mask)
1208 spte = rmap_next(kvm, rmapp, NULL);
1211 BUG_ON(!(_spte & PT_PRESENT_MASK));
1212 young = _spte & PT_ACCESSED_MASK;
1217 spte = rmap_next(kvm, rmapp, spte);
1223 #define RMAP_RECYCLE_THRESHOLD 1000
1225 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
1227 unsigned long *rmapp;
1228 struct kvm_mmu_page *sp;
1230 sp = page_header(__pa(spte));
1232 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
1234 kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
1235 kvm_flush_remote_tlbs(vcpu->kvm);
1238 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
1240 return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
1243 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1245 return kvm_handle_hva(kvm, hva, 0, kvm_test_age_rmapp);
1249 static int is_empty_shadow_page(u64 *spt)
1254 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
1255 if (is_shadow_present_pte(*pos)) {
1256 printk(KERN_ERR "%s: %p %llx\n", __func__,
1265 * This value is the sum of all of the kvm instances's
1266 * kvm->arch.n_used_mmu_pages values. We need a global,
1267 * aggregate version in order to make the slab shrinker
1270 static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, int nr)
1272 kvm->arch.n_used_mmu_pages += nr;
1273 percpu_counter_add(&kvm_total_used_mmu_pages, nr);
1277 * Remove the sp from shadow page cache, after call it,
1278 * we can not find this sp from the cache, and the shadow
1279 * page table is still valid.
1280 * It should be under the protection of mmu lock.
1282 static void kvm_mmu_isolate_page(struct kvm_mmu_page *sp)
1284 ASSERT(is_empty_shadow_page(sp->spt));
1285 hlist_del(&sp->hash_link);
1286 if (!sp->role.direct)
1287 free_page((unsigned long)sp->gfns);
1291 * Free the shadow page table and the sp, we can do it
1292 * out of the protection of mmu lock.
1294 static void kvm_mmu_free_page(struct kvm_mmu_page *sp)
1296 list_del(&sp->link);
1297 free_page((unsigned long)sp->spt);
1298 kmem_cache_free(mmu_page_header_cache, sp);
1301 static unsigned kvm_page_table_hashfn(gfn_t gfn)
1303 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
1306 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
1307 struct kvm_mmu_page *sp, u64 *parent_pte)
1312 pte_list_add(vcpu, parent_pte, &sp->parent_ptes);
1315 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
1318 pte_list_remove(parent_pte, &sp->parent_ptes);
1321 static void drop_parent_pte(struct kvm_mmu_page *sp,
1324 mmu_page_remove_parent_pte(sp, parent_pte);
1325 mmu_spte_clear_no_track(parent_pte);
1328 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
1329 u64 *parent_pte, int direct)
1331 struct kvm_mmu_page *sp;
1332 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache,
1334 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
1336 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache,
1338 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
1339 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
1340 bitmap_zero(sp->slot_bitmap, KVM_MEM_SLOTS_NUM);
1341 sp->parent_ptes = 0;
1342 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1343 kvm_mod_used_mmu_pages(vcpu->kvm, +1);
1347 static void mark_unsync(u64 *spte);
1348 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1350 pte_list_walk(&sp->parent_ptes, mark_unsync);
1353 static void mark_unsync(u64 *spte)
1355 struct kvm_mmu_page *sp;
1358 sp = page_header(__pa(spte));
1359 index = spte - sp->spt;
1360 if (__test_and_set_bit(index, sp->unsync_child_bitmap))
1362 if (sp->unsync_children++)
1364 kvm_mmu_mark_parents_unsync(sp);
1367 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1368 struct kvm_mmu_page *sp)
1373 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1377 static void nonpaging_update_pte(struct kvm_vcpu *vcpu,
1378 struct kvm_mmu_page *sp, u64 *spte,
1384 #define KVM_PAGE_ARRAY_NR 16
1386 struct kvm_mmu_pages {
1387 struct mmu_page_and_offset {
1388 struct kvm_mmu_page *sp;
1390 } page[KVM_PAGE_ARRAY_NR];
1394 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1400 for (i=0; i < pvec->nr; i++)
1401 if (pvec->page[i].sp == sp)
1404 pvec->page[pvec->nr].sp = sp;
1405 pvec->page[pvec->nr].idx = idx;
1407 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1410 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1411 struct kvm_mmu_pages *pvec)
1413 int i, ret, nr_unsync_leaf = 0;
1415 for_each_set_bit(i, sp->unsync_child_bitmap, 512) {
1416 struct kvm_mmu_page *child;
1417 u64 ent = sp->spt[i];
1419 if (!is_shadow_present_pte(ent) || is_large_pte(ent))
1420 goto clear_child_bitmap;
1422 child = page_header(ent & PT64_BASE_ADDR_MASK);
1424 if (child->unsync_children) {
1425 if (mmu_pages_add(pvec, child, i))
1428 ret = __mmu_unsync_walk(child, pvec);
1430 goto clear_child_bitmap;
1432 nr_unsync_leaf += ret;
1435 } else if (child->unsync) {
1437 if (mmu_pages_add(pvec, child, i))
1440 goto clear_child_bitmap;
1445 __clear_bit(i, sp->unsync_child_bitmap);
1446 sp->unsync_children--;
1447 WARN_ON((int)sp->unsync_children < 0);
1451 return nr_unsync_leaf;
1454 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1455 struct kvm_mmu_pages *pvec)
1457 if (!sp->unsync_children)
1460 mmu_pages_add(pvec, sp, 0);
1461 return __mmu_unsync_walk(sp, pvec);
1464 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1466 WARN_ON(!sp->unsync);
1467 trace_kvm_mmu_sync_page(sp);
1469 --kvm->stat.mmu_unsync;
1472 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1473 struct list_head *invalid_list);
1474 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1475 struct list_head *invalid_list);
1477 #define for_each_gfn_sp(kvm, sp, gfn, pos) \
1478 hlist_for_each_entry(sp, pos, \
1479 &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link) \
1480 if ((sp)->gfn != (gfn)) {} else
1482 #define for_each_gfn_indirect_valid_sp(kvm, sp, gfn, pos) \
1483 hlist_for_each_entry(sp, pos, \
1484 &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link) \
1485 if ((sp)->gfn != (gfn) || (sp)->role.direct || \
1486 (sp)->role.invalid) {} else
1488 /* @sp->gfn should be write-protected at the call site */
1489 static int __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1490 struct list_head *invalid_list, bool clear_unsync)
1492 if (sp->role.cr4_pae != !!is_pae(vcpu)) {
1493 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1498 kvm_unlink_unsync_page(vcpu->kvm, sp);
1500 if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1501 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1505 kvm_mmu_flush_tlb(vcpu);
1509 static int kvm_sync_page_transient(struct kvm_vcpu *vcpu,
1510 struct kvm_mmu_page *sp)
1512 LIST_HEAD(invalid_list);
1515 ret = __kvm_sync_page(vcpu, sp, &invalid_list, false);
1517 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1522 #ifdef CONFIG_KVM_MMU_AUDIT
1523 #include "mmu_audit.c"
1525 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, int point) { }
1526 static void mmu_audit_disable(void) { }
1529 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1530 struct list_head *invalid_list)
1532 return __kvm_sync_page(vcpu, sp, invalid_list, true);
1535 /* @gfn should be write-protected at the call site */
1536 static void kvm_sync_pages(struct kvm_vcpu *vcpu, gfn_t gfn)
1538 struct kvm_mmu_page *s;
1539 struct hlist_node *node;
1540 LIST_HEAD(invalid_list);
1543 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1547 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1548 kvm_unlink_unsync_page(vcpu->kvm, s);
1549 if ((s->role.cr4_pae != !!is_pae(vcpu)) ||
1550 (vcpu->arch.mmu.sync_page(vcpu, s))) {
1551 kvm_mmu_prepare_zap_page(vcpu->kvm, s, &invalid_list);
1557 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1559 kvm_mmu_flush_tlb(vcpu);
1562 struct mmu_page_path {
1563 struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1564 unsigned int idx[PT64_ROOT_LEVEL-1];
1567 #define for_each_sp(pvec, sp, parents, i) \
1568 for (i = mmu_pages_next(&pvec, &parents, -1), \
1569 sp = pvec.page[i].sp; \
1570 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1571 i = mmu_pages_next(&pvec, &parents, i))
1573 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1574 struct mmu_page_path *parents,
1579 for (n = i+1; n < pvec->nr; n++) {
1580 struct kvm_mmu_page *sp = pvec->page[n].sp;
1582 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1583 parents->idx[0] = pvec->page[n].idx;
1587 parents->parent[sp->role.level-2] = sp;
1588 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1594 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1596 struct kvm_mmu_page *sp;
1597 unsigned int level = 0;
1600 unsigned int idx = parents->idx[level];
1602 sp = parents->parent[level];
1606 --sp->unsync_children;
1607 WARN_ON((int)sp->unsync_children < 0);
1608 __clear_bit(idx, sp->unsync_child_bitmap);
1610 } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1613 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1614 struct mmu_page_path *parents,
1615 struct kvm_mmu_pages *pvec)
1617 parents->parent[parent->role.level-1] = NULL;
1621 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1622 struct kvm_mmu_page *parent)
1625 struct kvm_mmu_page *sp;
1626 struct mmu_page_path parents;
1627 struct kvm_mmu_pages pages;
1628 LIST_HEAD(invalid_list);
1630 kvm_mmu_pages_init(parent, &parents, &pages);
1631 while (mmu_unsync_walk(parent, &pages)) {
1634 for_each_sp(pages, sp, parents, i)
1635 protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1638 kvm_flush_remote_tlbs(vcpu->kvm);
1640 for_each_sp(pages, sp, parents, i) {
1641 kvm_sync_page(vcpu, sp, &invalid_list);
1642 mmu_pages_clear_parents(&parents);
1644 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1645 cond_resched_lock(&vcpu->kvm->mmu_lock);
1646 kvm_mmu_pages_init(parent, &parents, &pages);
1650 static void init_shadow_page_table(struct kvm_mmu_page *sp)
1654 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1658 static void __clear_sp_write_flooding_count(struct kvm_mmu_page *sp)
1660 sp->write_flooding_count = 0;
1663 static void clear_sp_write_flooding_count(u64 *spte)
1665 struct kvm_mmu_page *sp = page_header(__pa(spte));
1667 __clear_sp_write_flooding_count(sp);
1670 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1678 union kvm_mmu_page_role role;
1680 struct kvm_mmu_page *sp;
1681 struct hlist_node *node;
1682 bool need_sync = false;
1684 role = vcpu->arch.mmu.base_role;
1686 role.direct = direct;
1689 role.access = access;
1690 if (!vcpu->arch.mmu.direct_map
1691 && vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1692 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1693 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1694 role.quadrant = quadrant;
1696 for_each_gfn_sp(vcpu->kvm, sp, gfn, node) {
1697 if (!need_sync && sp->unsync)
1700 if (sp->role.word != role.word)
1703 if (sp->unsync && kvm_sync_page_transient(vcpu, sp))
1706 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1707 if (sp->unsync_children) {
1708 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
1709 kvm_mmu_mark_parents_unsync(sp);
1710 } else if (sp->unsync)
1711 kvm_mmu_mark_parents_unsync(sp);
1713 __clear_sp_write_flooding_count(sp);
1714 trace_kvm_mmu_get_page(sp, false);
1717 ++vcpu->kvm->stat.mmu_cache_miss;
1718 sp = kvm_mmu_alloc_page(vcpu, parent_pte, direct);
1723 hlist_add_head(&sp->hash_link,
1724 &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
1726 if (rmap_write_protect(vcpu->kvm, gfn))
1727 kvm_flush_remote_tlbs(vcpu->kvm);
1728 if (level > PT_PAGE_TABLE_LEVEL && need_sync)
1729 kvm_sync_pages(vcpu, gfn);
1731 account_shadowed(vcpu->kvm, gfn);
1733 init_shadow_page_table(sp);
1734 trace_kvm_mmu_get_page(sp, true);
1738 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1739 struct kvm_vcpu *vcpu, u64 addr)
1741 iterator->addr = addr;
1742 iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1743 iterator->level = vcpu->arch.mmu.shadow_root_level;
1745 if (iterator->level == PT64_ROOT_LEVEL &&
1746 vcpu->arch.mmu.root_level < PT64_ROOT_LEVEL &&
1747 !vcpu->arch.mmu.direct_map)
1750 if (iterator->level == PT32E_ROOT_LEVEL) {
1751 iterator->shadow_addr
1752 = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1753 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1755 if (!iterator->shadow_addr)
1756 iterator->level = 0;
1760 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1762 if (iterator->level < PT_PAGE_TABLE_LEVEL)
1765 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1766 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1770 static void __shadow_walk_next(struct kvm_shadow_walk_iterator *iterator,
1773 if (is_last_spte(spte, iterator->level)) {
1774 iterator->level = 0;
1778 iterator->shadow_addr = spte & PT64_BASE_ADDR_MASK;
1782 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1784 return __shadow_walk_next(iterator, *iterator->sptep);
1787 static void link_shadow_page(u64 *sptep, struct kvm_mmu_page *sp)
1791 spte = __pa(sp->spt)
1792 | PT_PRESENT_MASK | PT_ACCESSED_MASK
1793 | PT_WRITABLE_MASK | PT_USER_MASK;
1794 mmu_spte_set(sptep, spte);
1797 static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1799 if (is_large_pte(*sptep)) {
1800 drop_spte(vcpu->kvm, sptep);
1801 --vcpu->kvm->stat.lpages;
1802 kvm_flush_remote_tlbs(vcpu->kvm);
1806 static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1807 unsigned direct_access)
1809 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
1810 struct kvm_mmu_page *child;
1813 * For the direct sp, if the guest pte's dirty bit
1814 * changed form clean to dirty, it will corrupt the
1815 * sp's access: allow writable in the read-only sp,
1816 * so we should update the spte at this point to get
1817 * a new sp with the correct access.
1819 child = page_header(*sptep & PT64_BASE_ADDR_MASK);
1820 if (child->role.access == direct_access)
1823 drop_parent_pte(child, sptep);
1824 kvm_flush_remote_tlbs(vcpu->kvm);
1828 static bool mmu_page_zap_pte(struct kvm *kvm, struct kvm_mmu_page *sp,
1832 struct kvm_mmu_page *child;
1835 if (is_shadow_present_pte(pte)) {
1836 if (is_last_spte(pte, sp->role.level)) {
1837 drop_spte(kvm, spte);
1838 if (is_large_pte(pte))
1841 child = page_header(pte & PT64_BASE_ADDR_MASK);
1842 drop_parent_pte(child, spte);
1847 if (is_mmio_spte(pte))
1848 mmu_spte_clear_no_track(spte);
1853 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1854 struct kvm_mmu_page *sp)
1858 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1859 mmu_page_zap_pte(kvm, sp, sp->spt + i);
1862 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1864 mmu_page_remove_parent_pte(sp, parent_pte);
1867 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1871 while ((parent_pte = pte_list_next(&sp->parent_ptes, NULL)))
1872 drop_parent_pte(sp, parent_pte);
1875 static int mmu_zap_unsync_children(struct kvm *kvm,
1876 struct kvm_mmu_page *parent,
1877 struct list_head *invalid_list)
1880 struct mmu_page_path parents;
1881 struct kvm_mmu_pages pages;
1883 if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1886 kvm_mmu_pages_init(parent, &parents, &pages);
1887 while (mmu_unsync_walk(parent, &pages)) {
1888 struct kvm_mmu_page *sp;
1890 for_each_sp(pages, sp, parents, i) {
1891 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
1892 mmu_pages_clear_parents(&parents);
1895 kvm_mmu_pages_init(parent, &parents, &pages);
1901 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1902 struct list_head *invalid_list)
1906 trace_kvm_mmu_prepare_zap_page(sp);
1907 ++kvm->stat.mmu_shadow_zapped;
1908 ret = mmu_zap_unsync_children(kvm, sp, invalid_list);
1909 kvm_mmu_page_unlink_children(kvm, sp);
1910 kvm_mmu_unlink_parents(kvm, sp);
1911 if (!sp->role.invalid && !sp->role.direct)
1912 unaccount_shadowed(kvm, sp->gfn);
1914 kvm_unlink_unsync_page(kvm, sp);
1915 if (!sp->root_count) {
1918 list_move(&sp->link, invalid_list);
1919 kvm_mod_used_mmu_pages(kvm, -1);
1921 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1922 kvm_reload_remote_mmus(kvm);
1925 sp->role.invalid = 1;
1929 static void kvm_mmu_isolate_pages(struct list_head *invalid_list)
1931 struct kvm_mmu_page *sp;
1933 list_for_each_entry(sp, invalid_list, link)
1934 kvm_mmu_isolate_page(sp);
1937 static void free_pages_rcu(struct rcu_head *head)
1939 struct kvm_mmu_page *next, *sp;
1941 sp = container_of(head, struct kvm_mmu_page, rcu);
1943 if (!list_empty(&sp->link))
1944 next = list_first_entry(&sp->link,
1945 struct kvm_mmu_page, link);
1948 kvm_mmu_free_page(sp);
1953 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1954 struct list_head *invalid_list)
1956 struct kvm_mmu_page *sp;
1958 if (list_empty(invalid_list))
1961 kvm_flush_remote_tlbs(kvm);
1963 if (atomic_read(&kvm->arch.reader_counter)) {
1964 kvm_mmu_isolate_pages(invalid_list);
1965 sp = list_first_entry(invalid_list, struct kvm_mmu_page, link);
1966 list_del_init(invalid_list);
1968 trace_kvm_mmu_delay_free_pages(sp);
1969 call_rcu(&sp->rcu, free_pages_rcu);
1974 sp = list_first_entry(invalid_list, struct kvm_mmu_page, link);
1975 WARN_ON(!sp->role.invalid || sp->root_count);
1976 kvm_mmu_isolate_page(sp);
1977 kvm_mmu_free_page(sp);
1978 } while (!list_empty(invalid_list));
1983 * Changing the number of mmu pages allocated to the vm
1984 * Note: if goal_nr_mmu_pages is too small, you will get dead lock
1986 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int goal_nr_mmu_pages)
1988 LIST_HEAD(invalid_list);
1990 * If we set the number of mmu pages to be smaller be than the
1991 * number of actived pages , we must to free some mmu pages before we
1995 if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) {
1996 while (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages &&
1997 !list_empty(&kvm->arch.active_mmu_pages)) {
1998 struct kvm_mmu_page *page;
2000 page = container_of(kvm->arch.active_mmu_pages.prev,
2001 struct kvm_mmu_page, link);
2002 kvm_mmu_prepare_zap_page(kvm, page, &invalid_list);
2004 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2005 goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages;
2008 kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages;
2011 int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
2013 struct kvm_mmu_page *sp;
2014 struct hlist_node *node;
2015 LIST_HEAD(invalid_list);
2018 pgprintk("%s: looking for gfn %llx\n", __func__, gfn);
2020 spin_lock(&kvm->mmu_lock);
2021 for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
2022 pgprintk("%s: gfn %llx role %x\n", __func__, gfn,
2025 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
2027 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2028 spin_unlock(&kvm->mmu_lock);
2032 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page);
2034 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
2036 int slot = memslot_id(kvm, gfn);
2037 struct kvm_mmu_page *sp = page_header(__pa(pte));
2039 __set_bit(slot, sp->slot_bitmap);
2043 * The function is based on mtrr_type_lookup() in
2044 * arch/x86/kernel/cpu/mtrr/generic.c
2046 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
2051 u8 prev_match, curr_match;
2052 int num_var_ranges = KVM_NR_VAR_MTRR;
2054 if (!mtrr_state->enabled)
2057 /* Make end inclusive end, instead of exclusive */
2060 /* Look in fixed ranges. Just return the type as per start */
2061 if (mtrr_state->have_fixed && (start < 0x100000)) {
2064 if (start < 0x80000) {
2066 idx += (start >> 16);
2067 return mtrr_state->fixed_ranges[idx];
2068 } else if (start < 0xC0000) {
2070 idx += ((start - 0x80000) >> 14);
2071 return mtrr_state->fixed_ranges[idx];
2072 } else if (start < 0x1000000) {
2074 idx += ((start - 0xC0000) >> 12);
2075 return mtrr_state->fixed_ranges[idx];
2080 * Look in variable ranges
2081 * Look of multiple ranges matching this address and pick type
2082 * as per MTRR precedence
2084 if (!(mtrr_state->enabled & 2))
2085 return mtrr_state->def_type;
2088 for (i = 0; i < num_var_ranges; ++i) {
2089 unsigned short start_state, end_state;
2091 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
2094 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
2095 (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
2096 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
2097 (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
2099 start_state = ((start & mask) == (base & mask));
2100 end_state = ((end & mask) == (base & mask));
2101 if (start_state != end_state)
2104 if ((start & mask) != (base & mask))
2107 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
2108 if (prev_match == 0xFF) {
2109 prev_match = curr_match;
2113 if (prev_match == MTRR_TYPE_UNCACHABLE ||
2114 curr_match == MTRR_TYPE_UNCACHABLE)
2115 return MTRR_TYPE_UNCACHABLE;
2117 if ((prev_match == MTRR_TYPE_WRBACK &&
2118 curr_match == MTRR_TYPE_WRTHROUGH) ||
2119 (prev_match == MTRR_TYPE_WRTHROUGH &&
2120 curr_match == MTRR_TYPE_WRBACK)) {
2121 prev_match = MTRR_TYPE_WRTHROUGH;
2122 curr_match = MTRR_TYPE_WRTHROUGH;
2125 if (prev_match != curr_match)
2126 return MTRR_TYPE_UNCACHABLE;
2129 if (prev_match != 0xFF)
2132 return mtrr_state->def_type;
2135 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
2139 mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
2140 (gfn << PAGE_SHIFT) + PAGE_SIZE);
2141 if (mtrr == 0xfe || mtrr == 0xff)
2142 mtrr = MTRR_TYPE_WRBACK;
2145 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
2147 static void __kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
2149 trace_kvm_mmu_unsync_page(sp);
2150 ++vcpu->kvm->stat.mmu_unsync;
2153 kvm_mmu_mark_parents_unsync(sp);
2156 static void kvm_unsync_pages(struct kvm_vcpu *vcpu, gfn_t gfn)
2158 struct kvm_mmu_page *s;
2159 struct hlist_node *node;
2161 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
2164 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
2165 __kvm_unsync_page(vcpu, s);
2169 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
2172 struct kvm_mmu_page *s;
2173 struct hlist_node *node;
2174 bool need_unsync = false;
2176 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
2180 if (s->role.level != PT_PAGE_TABLE_LEVEL)
2183 if (!need_unsync && !s->unsync) {
2188 kvm_unsync_pages(vcpu, gfn);
2192 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2193 unsigned pte_access, int user_fault,
2194 int write_fault, int level,
2195 gfn_t gfn, pfn_t pfn, bool speculative,
2196 bool can_unsync, bool host_writable)
2198 u64 spte, entry = *sptep;
2201 if (set_mmio_spte(sptep, gfn, pfn, pte_access))
2204 spte = PT_PRESENT_MASK;
2206 spte |= shadow_accessed_mask;
2208 if (pte_access & ACC_EXEC_MASK)
2209 spte |= shadow_x_mask;
2211 spte |= shadow_nx_mask;
2212 if (pte_access & ACC_USER_MASK)
2213 spte |= shadow_user_mask;
2214 if (level > PT_PAGE_TABLE_LEVEL)
2215 spte |= PT_PAGE_SIZE_MASK;
2217 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
2218 kvm_is_mmio_pfn(pfn));
2221 spte |= SPTE_HOST_WRITEABLE;
2223 pte_access &= ~ACC_WRITE_MASK;
2225 spte |= (u64)pfn << PAGE_SHIFT;
2227 if ((pte_access & ACC_WRITE_MASK)
2228 || (!vcpu->arch.mmu.direct_map && write_fault
2229 && !is_write_protection(vcpu) && !user_fault)) {
2231 if (level > PT_PAGE_TABLE_LEVEL &&
2232 has_wrprotected_page(vcpu->kvm, gfn, level)) {
2234 drop_spte(vcpu->kvm, sptep);
2238 spte |= PT_WRITABLE_MASK;
2240 if (!vcpu->arch.mmu.direct_map
2241 && !(pte_access & ACC_WRITE_MASK)) {
2242 spte &= ~PT_USER_MASK;
2244 * If we converted a user page to a kernel page,
2245 * so that the kernel can write to it when cr0.wp=0,
2246 * then we should prevent the kernel from executing it
2247 * if SMEP is enabled.
2249 if (kvm_read_cr4_bits(vcpu, X86_CR4_SMEP))
2250 spte |= PT64_NX_MASK;
2254 * Optimization: for pte sync, if spte was writable the hash
2255 * lookup is unnecessary (and expensive). Write protection
2256 * is responsibility of mmu_get_page / kvm_sync_page.
2257 * Same reasoning can be applied to dirty page accounting.
2259 if (!can_unsync && is_writable_pte(*sptep))
2262 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
2263 pgprintk("%s: found shadow page for %llx, marking ro\n",
2266 pte_access &= ~ACC_WRITE_MASK;
2267 if (is_writable_pte(spte))
2268 spte &= ~PT_WRITABLE_MASK;
2272 if (pte_access & ACC_WRITE_MASK)
2273 mark_page_dirty(vcpu->kvm, gfn);
2276 mmu_spte_update(sptep, spte);
2278 * If we overwrite a writable spte with a read-only one we
2279 * should flush remote TLBs. Otherwise rmap_write_protect
2280 * will find a read-only spte, even though the writable spte
2281 * might be cached on a CPU's TLB.
2283 if (is_writable_pte(entry) && !is_writable_pte(*sptep))
2284 kvm_flush_remote_tlbs(vcpu->kvm);
2289 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2290 unsigned pt_access, unsigned pte_access,
2291 int user_fault, int write_fault,
2292 int *emulate, int level, gfn_t gfn,
2293 pfn_t pfn, bool speculative,
2296 int was_rmapped = 0;
2299 pgprintk("%s: spte %llx access %x write_fault %d"
2300 " user_fault %d gfn %llx\n",
2301 __func__, *sptep, pt_access,
2302 write_fault, user_fault, gfn);
2304 if (is_rmap_spte(*sptep)) {
2306 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
2307 * the parent of the now unreachable PTE.
2309 if (level > PT_PAGE_TABLE_LEVEL &&
2310 !is_large_pte(*sptep)) {
2311 struct kvm_mmu_page *child;
2314 child = page_header(pte & PT64_BASE_ADDR_MASK);
2315 drop_parent_pte(child, sptep);
2316 kvm_flush_remote_tlbs(vcpu->kvm);
2317 } else if (pfn != spte_to_pfn(*sptep)) {
2318 pgprintk("hfn old %llx new %llx\n",
2319 spte_to_pfn(*sptep), pfn);
2320 drop_spte(vcpu->kvm, sptep);
2321 kvm_flush_remote_tlbs(vcpu->kvm);
2326 if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
2327 level, gfn, pfn, speculative, true,
2331 kvm_mmu_flush_tlb(vcpu);
2334 if (unlikely(is_mmio_spte(*sptep) && emulate))
2337 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
2338 pgprintk("instantiating %s PTE (%s) at %llx (%llx) addr %p\n",
2339 is_large_pte(*sptep)? "2MB" : "4kB",
2340 *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
2342 if (!was_rmapped && is_large_pte(*sptep))
2343 ++vcpu->kvm->stat.lpages;
2345 if (is_shadow_present_pte(*sptep)) {
2346 page_header_update_slot(vcpu->kvm, sptep, gfn);
2348 rmap_count = rmap_add(vcpu, sptep, gfn);
2349 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
2350 rmap_recycle(vcpu, sptep, gfn);
2353 kvm_release_pfn_clean(pfn);
2356 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
2360 static pfn_t pte_prefetch_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn,
2363 struct kvm_memory_slot *slot;
2366 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, no_dirty_log);
2368 get_page(fault_page);
2369 return page_to_pfn(fault_page);
2372 hva = gfn_to_hva_memslot(slot, gfn);
2374 return hva_to_pfn_atomic(vcpu->kvm, hva);
2377 static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu,
2378 struct kvm_mmu_page *sp,
2379 u64 *start, u64 *end)
2381 struct page *pages[PTE_PREFETCH_NUM];
2382 unsigned access = sp->role.access;
2386 gfn = kvm_mmu_page_get_gfn(sp, start - sp->spt);
2387 if (!gfn_to_memslot_dirty_bitmap(vcpu, gfn, access & ACC_WRITE_MASK))
2390 ret = gfn_to_page_many_atomic(vcpu->kvm, gfn, pages, end - start);
2394 for (i = 0; i < ret; i++, gfn++, start++)
2395 mmu_set_spte(vcpu, start, ACC_ALL,
2397 sp->role.level, gfn,
2398 page_to_pfn(pages[i]), true, true);
2403 static void __direct_pte_prefetch(struct kvm_vcpu *vcpu,
2404 struct kvm_mmu_page *sp, u64 *sptep)
2406 u64 *spte, *start = NULL;
2409 WARN_ON(!sp->role.direct);
2411 i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
2414 for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
2415 if (is_shadow_present_pte(*spte) || spte == sptep) {
2418 if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0)
2426 static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
2428 struct kvm_mmu_page *sp;
2431 * Since it's no accessed bit on EPT, it's no way to
2432 * distinguish between actually accessed translations
2433 * and prefetched, so disable pte prefetch if EPT is
2436 if (!shadow_accessed_mask)
2439 sp = page_header(__pa(sptep));
2440 if (sp->role.level > PT_PAGE_TABLE_LEVEL)
2443 __direct_pte_prefetch(vcpu, sp, sptep);
2446 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
2447 int map_writable, int level, gfn_t gfn, pfn_t pfn,
2450 struct kvm_shadow_walk_iterator iterator;
2451 struct kvm_mmu_page *sp;
2455 for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
2456 if (iterator.level == level) {
2457 unsigned pte_access = ACC_ALL;
2459 mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, pte_access,
2461 level, gfn, pfn, prefault, map_writable);
2462 direct_pte_prefetch(vcpu, iterator.sptep);
2463 ++vcpu->stat.pf_fixed;
2467 if (!is_shadow_present_pte(*iterator.sptep)) {
2468 u64 base_addr = iterator.addr;
2470 base_addr &= PT64_LVL_ADDR_MASK(iterator.level);
2471 pseudo_gfn = base_addr >> PAGE_SHIFT;
2472 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
2474 1, ACC_ALL, iterator.sptep);
2476 pgprintk("nonpaging_map: ENOMEM\n");
2477 kvm_release_pfn_clean(pfn);
2481 mmu_spte_set(iterator.sptep,
2483 | PT_PRESENT_MASK | PT_WRITABLE_MASK
2484 | shadow_user_mask | shadow_x_mask
2485 | shadow_accessed_mask);
2491 static void kvm_send_hwpoison_signal(unsigned long address, struct task_struct *tsk)
2495 info.si_signo = SIGBUS;
2497 info.si_code = BUS_MCEERR_AR;
2498 info.si_addr = (void __user *)address;
2499 info.si_addr_lsb = PAGE_SHIFT;
2501 send_sig_info(SIGBUS, &info, tsk);
2504 static int kvm_handle_bad_page(struct kvm_vcpu *vcpu, gfn_t gfn, pfn_t pfn)
2506 kvm_release_pfn_clean(pfn);
2507 if (is_hwpoison_pfn(pfn)) {
2508 kvm_send_hwpoison_signal(gfn_to_hva(vcpu->kvm, gfn), current);
2515 static void transparent_hugepage_adjust(struct kvm_vcpu *vcpu,
2516 gfn_t *gfnp, pfn_t *pfnp, int *levelp)
2520 int level = *levelp;
2523 * Check if it's a transparent hugepage. If this would be an
2524 * hugetlbfs page, level wouldn't be set to
2525 * PT_PAGE_TABLE_LEVEL and there would be no adjustment done
2528 if (!is_error_pfn(pfn) && !kvm_is_mmio_pfn(pfn) &&
2529 level == PT_PAGE_TABLE_LEVEL &&
2530 PageTransCompound(pfn_to_page(pfn)) &&
2531 !has_wrprotected_page(vcpu->kvm, gfn, PT_DIRECTORY_LEVEL)) {
2534 * mmu_notifier_retry was successful and we hold the
2535 * mmu_lock here, so the pmd can't become splitting
2536 * from under us, and in turn
2537 * __split_huge_page_refcount() can't run from under
2538 * us and we can safely transfer the refcount from
2539 * PG_tail to PG_head as we switch the pfn to tail to
2542 *levelp = level = PT_DIRECTORY_LEVEL;
2543 mask = KVM_PAGES_PER_HPAGE(level) - 1;
2544 VM_BUG_ON((gfn & mask) != (pfn & mask));
2548 kvm_release_pfn_clean(pfn);
2550 if (!get_page_unless_zero(pfn_to_page(pfn)))
2557 static bool mmu_invalid_pfn(pfn_t pfn)
2559 return unlikely(is_invalid_pfn(pfn));
2562 static bool handle_abnormal_pfn(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn,
2563 pfn_t pfn, unsigned access, int *ret_val)
2567 /* The pfn is invalid, report the error! */
2568 if (unlikely(is_invalid_pfn(pfn))) {
2569 *ret_val = kvm_handle_bad_page(vcpu, gfn, pfn);
2573 if (unlikely(is_noslot_pfn(pfn)))
2574 vcpu_cache_mmio_info(vcpu, gva, gfn, access);
2581 static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
2582 gva_t gva, pfn_t *pfn, bool write, bool *writable);
2584 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn,
2591 unsigned long mmu_seq;
2594 force_pt_level = mapping_level_dirty_bitmap(vcpu, gfn);
2595 if (likely(!force_pt_level)) {
2596 level = mapping_level(vcpu, gfn);
2598 * This path builds a PAE pagetable - so we can map
2599 * 2mb pages at maximum. Therefore check if the level
2600 * is larger than that.
2602 if (level > PT_DIRECTORY_LEVEL)
2603 level = PT_DIRECTORY_LEVEL;
2605 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2607 level = PT_PAGE_TABLE_LEVEL;
2609 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2612 if (try_async_pf(vcpu, prefault, gfn, v, &pfn, write, &map_writable))
2615 if (handle_abnormal_pfn(vcpu, v, gfn, pfn, ACC_ALL, &r))
2618 spin_lock(&vcpu->kvm->mmu_lock);
2619 if (mmu_notifier_retry(vcpu, mmu_seq))
2621 kvm_mmu_free_some_pages(vcpu);
2622 if (likely(!force_pt_level))
2623 transparent_hugepage_adjust(vcpu, &gfn, &pfn, &level);
2624 r = __direct_map(vcpu, v, write, map_writable, level, gfn, pfn,
2626 spin_unlock(&vcpu->kvm->mmu_lock);
2632 spin_unlock(&vcpu->kvm->mmu_lock);
2633 kvm_release_pfn_clean(pfn);
2638 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2641 struct kvm_mmu_page *sp;
2642 LIST_HEAD(invalid_list);
2644 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2646 spin_lock(&vcpu->kvm->mmu_lock);
2647 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL &&
2648 (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL ||
2649 vcpu->arch.mmu.direct_map)) {
2650 hpa_t root = vcpu->arch.mmu.root_hpa;
2652 sp = page_header(root);
2654 if (!sp->root_count && sp->role.invalid) {
2655 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
2656 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2658 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2659 spin_unlock(&vcpu->kvm->mmu_lock);
2662 for (i = 0; i < 4; ++i) {
2663 hpa_t root = vcpu->arch.mmu.pae_root[i];
2666 root &= PT64_BASE_ADDR_MASK;
2667 sp = page_header(root);
2669 if (!sp->root_count && sp->role.invalid)
2670 kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2673 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2675 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2676 spin_unlock(&vcpu->kvm->mmu_lock);
2677 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2680 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2684 if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2685 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
2692 static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
2694 struct kvm_mmu_page *sp;
2697 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2698 spin_lock(&vcpu->kvm->mmu_lock);
2699 kvm_mmu_free_some_pages(vcpu);
2700 sp = kvm_mmu_get_page(vcpu, 0, 0, PT64_ROOT_LEVEL,
2703 spin_unlock(&vcpu->kvm->mmu_lock);
2704 vcpu->arch.mmu.root_hpa = __pa(sp->spt);
2705 } else if (vcpu->arch.mmu.shadow_root_level == PT32E_ROOT_LEVEL) {
2706 for (i = 0; i < 4; ++i) {
2707 hpa_t root = vcpu->arch.mmu.pae_root[i];
2709 ASSERT(!VALID_PAGE(root));
2710 spin_lock(&vcpu->kvm->mmu_lock);
2711 kvm_mmu_free_some_pages(vcpu);
2712 sp = kvm_mmu_get_page(vcpu, i << (30 - PAGE_SHIFT),
2714 PT32_ROOT_LEVEL, 1, ACC_ALL,
2716 root = __pa(sp->spt);
2718 spin_unlock(&vcpu->kvm->mmu_lock);
2719 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2721 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2728 static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
2730 struct kvm_mmu_page *sp;
2735 root_gfn = vcpu->arch.mmu.get_cr3(vcpu) >> PAGE_SHIFT;
2737 if (mmu_check_root(vcpu, root_gfn))
2741 * Do we shadow a long mode page table? If so we need to
2742 * write-protect the guests page table root.
2744 if (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL) {
2745 hpa_t root = vcpu->arch.mmu.root_hpa;
2747 ASSERT(!VALID_PAGE(root));
2749 spin_lock(&vcpu->kvm->mmu_lock);
2750 kvm_mmu_free_some_pages(vcpu);
2751 sp = kvm_mmu_get_page(vcpu, root_gfn, 0, PT64_ROOT_LEVEL,
2753 root = __pa(sp->spt);
2755 spin_unlock(&vcpu->kvm->mmu_lock);
2756 vcpu->arch.mmu.root_hpa = root;
2761 * We shadow a 32 bit page table. This may be a legacy 2-level
2762 * or a PAE 3-level page table. In either case we need to be aware that
2763 * the shadow page table may be a PAE or a long mode page table.
2765 pm_mask = PT_PRESENT_MASK;
2766 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL)
2767 pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK;
2769 for (i = 0; i < 4; ++i) {
2770 hpa_t root = vcpu->arch.mmu.pae_root[i];
2772 ASSERT(!VALID_PAGE(root));
2773 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2774 pdptr = vcpu->arch.mmu.get_pdptr(vcpu, i);
2775 if (!is_present_gpte(pdptr)) {
2776 vcpu->arch.mmu.pae_root[i] = 0;
2779 root_gfn = pdptr >> PAGE_SHIFT;
2780 if (mmu_check_root(vcpu, root_gfn))
2783 spin_lock(&vcpu->kvm->mmu_lock);
2784 kvm_mmu_free_some_pages(vcpu);
2785 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2788 root = __pa(sp->spt);
2790 spin_unlock(&vcpu->kvm->mmu_lock);
2792 vcpu->arch.mmu.pae_root[i] = root | pm_mask;
2794 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2797 * If we shadow a 32 bit page table with a long mode page
2798 * table we enter this path.
2800 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2801 if (vcpu->arch.mmu.lm_root == NULL) {
2803 * The additional page necessary for this is only
2804 * allocated on demand.
2809 lm_root = (void*)get_zeroed_page(GFP_KERNEL);
2810 if (lm_root == NULL)
2813 lm_root[0] = __pa(vcpu->arch.mmu.pae_root) | pm_mask;
2815 vcpu->arch.mmu.lm_root = lm_root;
2818 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.lm_root);
2824 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2826 if (vcpu->arch.mmu.direct_map)
2827 return mmu_alloc_direct_roots(vcpu);
2829 return mmu_alloc_shadow_roots(vcpu);
2832 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2835 struct kvm_mmu_page *sp;
2837 if (vcpu->arch.mmu.direct_map)
2840 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2843 vcpu_clear_mmio_info(vcpu, ~0ul);
2844 kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
2845 if (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL) {
2846 hpa_t root = vcpu->arch.mmu.root_hpa;
2847 sp = page_header(root);
2848 mmu_sync_children(vcpu, sp);
2849 kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
2852 for (i = 0; i < 4; ++i) {
2853 hpa_t root = vcpu->arch.mmu.pae_root[i];
2855 if (root && VALID_PAGE(root)) {
2856 root &= PT64_BASE_ADDR_MASK;
2857 sp = page_header(root);
2858 mmu_sync_children(vcpu, sp);
2861 kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
2864 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2866 spin_lock(&vcpu->kvm->mmu_lock);
2867 mmu_sync_roots(vcpu);
2868 spin_unlock(&vcpu->kvm->mmu_lock);
2871 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2872 u32 access, struct x86_exception *exception)
2875 exception->error_code = 0;
2879 static gpa_t nonpaging_gva_to_gpa_nested(struct kvm_vcpu *vcpu, gva_t vaddr,
2881 struct x86_exception *exception)
2884 exception->error_code = 0;
2885 return vcpu->arch.nested_mmu.translate_gpa(vcpu, vaddr, access);
2888 static bool quickly_check_mmio_pf(struct kvm_vcpu *vcpu, u64 addr, bool direct)
2891 return vcpu_match_mmio_gpa(vcpu, addr);
2893 return vcpu_match_mmio_gva(vcpu, addr);
2898 * On direct hosts, the last spte is only allows two states
2899 * for mmio page fault:
2900 * - It is the mmio spte
2901 * - It is zapped or it is being zapped.
2903 * This function completely checks the spte when the last spte
2904 * is not the mmio spte.
2906 static bool check_direct_spte_mmio_pf(u64 spte)
2908 return __check_direct_spte_mmio_pf(spte);
2911 static u64 walk_shadow_page_get_mmio_spte(struct kvm_vcpu *vcpu, u64 addr)
2913 struct kvm_shadow_walk_iterator iterator;
2916 walk_shadow_page_lockless_begin(vcpu);
2917 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte)
2918 if (!is_shadow_present_pte(spte))
2920 walk_shadow_page_lockless_end(vcpu);
2926 * If it is a real mmio page fault, return 1 and emulat the instruction
2927 * directly, return 0 to let CPU fault again on the address, -1 is
2928 * returned if bug is detected.
2930 int handle_mmio_page_fault_common(struct kvm_vcpu *vcpu, u64 addr, bool direct)
2934 if (quickly_check_mmio_pf(vcpu, addr, direct))
2937 spte = walk_shadow_page_get_mmio_spte(vcpu, addr);
2939 if (is_mmio_spte(spte)) {
2940 gfn_t gfn = get_mmio_spte_gfn(spte);
2941 unsigned access = get_mmio_spte_access(spte);
2946 trace_handle_mmio_page_fault(addr, gfn, access);
2947 vcpu_cache_mmio_info(vcpu, addr, gfn, access);
2952 * It's ok if the gva is remapped by other cpus on shadow guest,
2953 * it's a BUG if the gfn is not a mmio page.
2955 if (direct && !check_direct_spte_mmio_pf(spte))
2959 * If the page table is zapped by other cpus, let CPU fault again on
2964 EXPORT_SYMBOL_GPL(handle_mmio_page_fault_common);
2966 static int handle_mmio_page_fault(struct kvm_vcpu *vcpu, u64 addr,
2967 u32 error_code, bool direct)
2971 ret = handle_mmio_page_fault_common(vcpu, addr, direct);
2976 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2977 u32 error_code, bool prefault)
2982 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2984 if (unlikely(error_code & PFERR_RSVD_MASK))
2985 return handle_mmio_page_fault(vcpu, gva, error_code, true);
2987 r = mmu_topup_memory_caches(vcpu);
2992 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2994 gfn = gva >> PAGE_SHIFT;
2996 return nonpaging_map(vcpu, gva & PAGE_MASK,
2997 error_code & PFERR_WRITE_MASK, gfn, prefault);
3000 static int kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn)
3002 struct kvm_arch_async_pf arch;
3004 arch.token = (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id;
3006 arch.direct_map = vcpu->arch.mmu.direct_map;
3007 arch.cr3 = vcpu->arch.mmu.get_cr3(vcpu);
3009 return kvm_setup_async_pf(vcpu, gva, gfn, &arch);
3012 static bool can_do_async_pf(struct kvm_vcpu *vcpu)
3014 if (unlikely(!irqchip_in_kernel(vcpu->kvm) ||
3015 kvm_event_needs_reinjection(vcpu)))
3018 return kvm_x86_ops->interrupt_allowed(vcpu);
3021 static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
3022 gva_t gva, pfn_t *pfn, bool write, bool *writable)
3026 *pfn = gfn_to_pfn_async(vcpu->kvm, gfn, &async, write, writable);
3029 return false; /* *pfn has correct page already */
3031 put_page(pfn_to_page(*pfn));
3033 if (!prefault && can_do_async_pf(vcpu)) {
3034 trace_kvm_try_async_get_page(gva, gfn);
3035 if (kvm_find_async_pf_gfn(vcpu, gfn)) {
3036 trace_kvm_async_pf_doublefault(gva, gfn);
3037 kvm_make_request(KVM_REQ_APF_HALT, vcpu);
3039 } else if (kvm_arch_setup_async_pf(vcpu, gva, gfn))
3043 *pfn = gfn_to_pfn_prot(vcpu->kvm, gfn, write, writable);
3048 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa, u32 error_code,
3055 gfn_t gfn = gpa >> PAGE_SHIFT;
3056 unsigned long mmu_seq;
3057 int write = error_code & PFERR_WRITE_MASK;
3061 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
3063 if (unlikely(error_code & PFERR_RSVD_MASK))
3064 return handle_mmio_page_fault(vcpu, gpa, error_code, true);
3066 r = mmu_topup_memory_caches(vcpu);
3070 force_pt_level = mapping_level_dirty_bitmap(vcpu, gfn);
3071 if (likely(!force_pt_level)) {
3072 level = mapping_level(vcpu, gfn);
3073 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
3075 level = PT_PAGE_TABLE_LEVEL;
3077 mmu_seq = vcpu->kvm->mmu_notifier_seq;
3080 if (try_async_pf(vcpu, prefault, gfn, gpa, &pfn, write, &map_writable))
3083 if (handle_abnormal_pfn(vcpu, 0, gfn, pfn, ACC_ALL, &r))
3086 spin_lock(&vcpu->kvm->mmu_lock);
3087 if (mmu_notifier_retry(vcpu, mmu_seq))
3089 kvm_mmu_free_some_pages(vcpu);
3090 if (likely(!force_pt_level))
3091 transparent_hugepage_adjust(vcpu, &gfn, &pfn, &level);
3092 r = __direct_map(vcpu, gpa, write, map_writable,
3093 level, gfn, pfn, prefault);
3094 spin_unlock(&vcpu->kvm->mmu_lock);
3099 spin_unlock(&vcpu->kvm->mmu_lock);
3100 kvm_release_pfn_clean(pfn);
3104 static void nonpaging_free(struct kvm_vcpu *vcpu)
3106 mmu_free_roots(vcpu);
3109 static int nonpaging_init_context(struct kvm_vcpu *vcpu,
3110 struct kvm_mmu *context)
3112 context->new_cr3 = nonpaging_new_cr3;
3113 context->page_fault = nonpaging_page_fault;
3114 context->gva_to_gpa = nonpaging_gva_to_gpa;
3115 context->free = nonpaging_free;
3116 context->sync_page = nonpaging_sync_page;
3117 context->invlpg = nonpaging_invlpg;
3118 context->update_pte = nonpaging_update_pte;
3119 context->root_level = 0;
3120 context->shadow_root_level = PT32E_ROOT_LEVEL;
3121 context->root_hpa = INVALID_PAGE;
3122 context->direct_map = true;
3123 context->nx = false;
3127 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3129 ++vcpu->stat.tlb_flush;
3130 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
3133 static void paging_new_cr3(struct kvm_vcpu *vcpu)
3135 pgprintk("%s: cr3 %lx\n", __func__, kvm_read_cr3(vcpu));
3136 mmu_free_roots(vcpu);
3139 static unsigned long get_cr3(struct kvm_vcpu *vcpu)
3141 return kvm_read_cr3(vcpu);
3144 static void inject_page_fault(struct kvm_vcpu *vcpu,
3145 struct x86_exception *fault)
3147 vcpu->arch.mmu.inject_page_fault(vcpu, fault);
3150 static void paging_free(struct kvm_vcpu *vcpu)
3152 nonpaging_free(vcpu);
3155 static bool is_rsvd_bits_set(struct kvm_mmu *mmu, u64 gpte, int level)
3159 bit7 = (gpte >> 7) & 1;
3160 return (gpte & mmu->rsvd_bits_mask[bit7][level-1]) != 0;
3163 static bool sync_mmio_spte(u64 *sptep, gfn_t gfn, unsigned access,
3166 if (unlikely(is_mmio_spte(*sptep))) {
3167 if (gfn != get_mmio_spte_gfn(*sptep)) {
3168 mmu_spte_clear_no_track(sptep);
3173 mark_mmio_spte(sptep, gfn, access);
3181 #include "paging_tmpl.h"
3185 #include "paging_tmpl.h"
3188 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
3189 struct kvm_mmu *context,
3192 int maxphyaddr = cpuid_maxphyaddr(vcpu);
3193 u64 exb_bit_rsvd = 0;
3196 exb_bit_rsvd = rsvd_bits(63, 63);
3198 case PT32_ROOT_LEVEL:
3199 /* no rsvd bits for 2 level 4K page table entries */
3200 context->rsvd_bits_mask[0][1] = 0;
3201 context->rsvd_bits_mask[0][0] = 0;
3202 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3204 if (!is_pse(vcpu)) {
3205 context->rsvd_bits_mask[1][1] = 0;
3209 if (is_cpuid_PSE36())
3210 /* 36bits PSE 4MB page */
3211 context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
3213 /* 32 bits PSE 4MB page */
3214 context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
3216 case PT32E_ROOT_LEVEL:
3217 context->rsvd_bits_mask[0][2] =
3218 rsvd_bits(maxphyaddr, 63) |
3219 rsvd_bits(7, 8) | rsvd_bits(1, 2); /* PDPTE */
3220 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
3221 rsvd_bits(maxphyaddr, 62); /* PDE */
3222 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
3223 rsvd_bits(maxphyaddr, 62); /* PTE */
3224 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
3225 rsvd_bits(maxphyaddr, 62) |
3226 rsvd_bits(13, 20); /* large page */
3227 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3229 case PT64_ROOT_LEVEL:
3230 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
3231 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
3232 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
3233 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
3234 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
3235 rsvd_bits(maxphyaddr, 51);
3236 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
3237 rsvd_bits(maxphyaddr, 51);
3238 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
3239 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
3240 rsvd_bits(maxphyaddr, 51) |
3242 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
3243 rsvd_bits(maxphyaddr, 51) |
3244 rsvd_bits(13, 20); /* large page */
3245 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3250 static int paging64_init_context_common(struct kvm_vcpu *vcpu,
3251 struct kvm_mmu *context,
3254 context->nx = is_nx(vcpu);
3256 reset_rsvds_bits_mask(vcpu, context, level);
3258 ASSERT(is_pae(vcpu));
3259 context->new_cr3 = paging_new_cr3;
3260 context->page_fault = paging64_page_fault;
3261 context->gva_to_gpa = paging64_gva_to_gpa;
3262 context->sync_page = paging64_sync_page;
3263 context->invlpg = paging64_invlpg;
3264 context->update_pte = paging64_update_pte;
3265 context->free = paging_free;
3266 context->root_level = level;
3267 context->shadow_root_level = level;
3268 context->root_hpa = INVALID_PAGE;
3269 context->direct_map = false;
3273 static int paging64_init_context(struct kvm_vcpu *vcpu,
3274 struct kvm_mmu *context)
3276 return paging64_init_context_common(vcpu, context, PT64_ROOT_LEVEL);
3279 static int paging32_init_context(struct kvm_vcpu *vcpu,
3280 struct kvm_mmu *context)
3282 context->nx = false;
3284 reset_rsvds_bits_mask(vcpu, context, PT32_ROOT_LEVEL);
3286 context->new_cr3 = paging_new_cr3;
3287 context->page_fault = paging32_page_fault;
3288 context->gva_to_gpa = paging32_gva_to_gpa;
3289 context->free = paging_free;
3290 context->sync_page = paging32_sync_page;
3291 context->invlpg = paging32_invlpg;
3292 context->update_pte = paging32_update_pte;
3293 context->root_level = PT32_ROOT_LEVEL;
3294 context->shadow_root_level = PT32E_ROOT_LEVEL;
3295 context->root_hpa = INVALID_PAGE;
3296 context->direct_map = false;
3300 static int paging32E_init_context(struct kvm_vcpu *vcpu,
3301 struct kvm_mmu *context)
3303 return paging64_init_context_common(vcpu, context, PT32E_ROOT_LEVEL);
3306 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
3308 struct kvm_mmu *context = vcpu->arch.walk_mmu;
3310 context->base_role.word = 0;
3311 context->new_cr3 = nonpaging_new_cr3;
3312 context->page_fault = tdp_page_fault;
3313 context->free = nonpaging_free;
3314 context->sync_page = nonpaging_sync_page;
3315 context->invlpg = nonpaging_invlpg;
3316 context->update_pte = nonpaging_update_pte;
3317 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
3318 context->root_hpa = INVALID_PAGE;
3319 context->direct_map = true;
3320 context->set_cr3 = kvm_x86_ops->set_tdp_cr3;
3321 context->get_cr3 = get_cr3;
3322 context->get_pdptr = kvm_pdptr_read;
3323 context->inject_page_fault = kvm_inject_page_fault;
3324 context->nx = is_nx(vcpu);
3326 if (!is_paging(vcpu)) {
3327 context->nx = false;
3328 context->gva_to_gpa = nonpaging_gva_to_gpa;
3329 context->root_level = 0;
3330 } else if (is_long_mode(vcpu)) {
3331 context->nx = is_nx(vcpu);
3332 reset_rsvds_bits_mask(vcpu, context, PT64_ROOT_LEVEL);
3333 context->gva_to_gpa = paging64_gva_to_gpa;
3334 context->root_level = PT64_ROOT_LEVEL;
3335 } else if (is_pae(vcpu)) {
3336 context->nx = is_nx(vcpu);
3337 reset_rsvds_bits_mask(vcpu, context, PT32E_ROOT_LEVEL);
3338 context->gva_to_gpa = paging64_gva_to_gpa;
3339 context->root_level = PT32E_ROOT_LEVEL;
3341 context->nx = false;
3342 reset_rsvds_bits_mask(vcpu, context, PT32_ROOT_LEVEL);
3343 context->gva_to_gpa = paging32_gva_to_gpa;
3344 context->root_level = PT32_ROOT_LEVEL;
3350 int kvm_init_shadow_mmu(struct kvm_vcpu *vcpu, struct kvm_mmu *context)
3353 bool smep = kvm_read_cr4_bits(vcpu, X86_CR4_SMEP);
3355 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3357 if (!is_paging(vcpu))
3358 r = nonpaging_init_context(vcpu, context);
3359 else if (is_long_mode(vcpu))
3360 r = paging64_init_context(vcpu, context);
3361 else if (is_pae(vcpu))
3362 r = paging32E_init_context(vcpu, context);
3364 r = paging32_init_context(vcpu, context);
3366 vcpu->arch.mmu.base_role.cr4_pae = !!is_pae(vcpu);
3367 vcpu->arch.mmu.base_role.cr0_wp = is_write_protection(vcpu);
3368 vcpu->arch.mmu.base_role.smep_andnot_wp
3369 = smep && !is_write_protection(vcpu);
3373 EXPORT_SYMBOL_GPL(kvm_init_shadow_mmu);
3375 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
3377 int r = kvm_init_shadow_mmu(vcpu, vcpu->arch.walk_mmu);
3379 vcpu->arch.walk_mmu->set_cr3 = kvm_x86_ops->set_cr3;
3380 vcpu->arch.walk_mmu->get_cr3 = get_cr3;
3381 vcpu->arch.walk_mmu->get_pdptr = kvm_pdptr_read;
3382 vcpu->arch.walk_mmu->inject_page_fault = kvm_inject_page_fault;
3387 static int init_kvm_nested_mmu(struct kvm_vcpu *vcpu)
3389 struct kvm_mmu *g_context = &vcpu->arch.nested_mmu;
3391 g_context->get_cr3 = get_cr3;
3392 g_context->get_pdptr = kvm_pdptr_read;
3393 g_context->inject_page_fault = kvm_inject_page_fault;
3396 * Note that arch.mmu.gva_to_gpa translates l2_gva to l1_gpa. The
3397 * translation of l2_gpa to l1_gpa addresses is done using the
3398 * arch.nested_mmu.gva_to_gpa function. Basically the gva_to_gpa
3399 * functions between mmu and nested_mmu are swapped.
3401 if (!is_paging(vcpu)) {
3402 g_context->nx = false;
3403 g_context->root_level = 0;
3404 g_context->gva_to_gpa = nonpaging_gva_to_gpa_nested;
3405 } else if (is_long_mode(vcpu)) {
3406 g_context->nx = is_nx(vcpu);
3407 reset_rsvds_bits_mask(vcpu, g_context, PT64_ROOT_LEVEL);
3408 g_context->root_level = PT64_ROOT_LEVEL;
3409 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
3410 } else if (is_pae(vcpu)) {
3411 g_context->nx = is_nx(vcpu);
3412 reset_rsvds_bits_mask(vcpu, g_context, PT32E_ROOT_LEVEL);
3413 g_context->root_level = PT32E_ROOT_LEVEL;
3414 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
3416 g_context->nx = false;
3417 reset_rsvds_bits_mask(vcpu, g_context, PT32_ROOT_LEVEL);
3418 g_context->root_level = PT32_ROOT_LEVEL;
3419 g_context->gva_to_gpa = paging32_gva_to_gpa_nested;
3425 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
3427 if (mmu_is_nested(vcpu))
3428 return init_kvm_nested_mmu(vcpu);
3429 else if (tdp_enabled)
3430 return init_kvm_tdp_mmu(vcpu);
3432 return init_kvm_softmmu(vcpu);
3435 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
3438 if (VALID_PAGE(vcpu->arch.mmu.root_hpa))
3439 /* mmu.free() should set root_hpa = INVALID_PAGE */
3440 vcpu->arch.mmu.free(vcpu);
3443 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
3445 destroy_kvm_mmu(vcpu);
3446 return init_kvm_mmu(vcpu);
3448 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
3450 int kvm_mmu_load(struct kvm_vcpu *vcpu)
3454 r = mmu_topup_memory_caches(vcpu);
3457 r = mmu_alloc_roots(vcpu);
3458 spin_lock(&vcpu->kvm->mmu_lock);
3459 mmu_sync_roots(vcpu);
3460 spin_unlock(&vcpu->kvm->mmu_lock);
3463 /* set_cr3() should ensure TLB has been flushed */
3464 vcpu->arch.mmu.set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
3468 EXPORT_SYMBOL_GPL(kvm_mmu_load);
3470 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
3472 mmu_free_roots(vcpu);
3474 EXPORT_SYMBOL_GPL(kvm_mmu_unload);
3476 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
3477 struct kvm_mmu_page *sp, u64 *spte,
3480 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
3481 ++vcpu->kvm->stat.mmu_pde_zapped;
3485 ++vcpu->kvm->stat.mmu_pte_updated;
3486 vcpu->arch.mmu.update_pte(vcpu, sp, spte, new);
3489 static bool need_remote_flush(u64 old, u64 new)
3491 if (!is_shadow_present_pte(old))
3493 if (!is_shadow_present_pte(new))
3495 if ((old ^ new) & PT64_BASE_ADDR_MASK)
3497 old ^= PT64_NX_MASK;
3498 new ^= PT64_NX_MASK;
3499 return (old & ~new & PT64_PERM_MASK) != 0;
3502 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, bool zap_page,
3503 bool remote_flush, bool local_flush)
3509 kvm_flush_remote_tlbs(vcpu->kvm);
3510 else if (local_flush)
3511 kvm_mmu_flush_tlb(vcpu);
3514 static u64 mmu_pte_write_fetch_gpte(struct kvm_vcpu *vcpu, gpa_t *gpa,
3515 const u8 *new, int *bytes)
3521 * Assume that the pte write on a page table of the same type
3522 * as the current vcpu paging mode since we update the sptes only
3523 * when they have the same mode.
3525 if (is_pae(vcpu) && *bytes == 4) {
3526 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
3529 r = kvm_read_guest(vcpu->kvm, *gpa, &gentry, min(*bytes, 8));
3532 new = (const u8 *)&gentry;
3537 gentry = *(const u32 *)new;
3540 gentry = *(const u64 *)new;
3551 * If we're seeing too many writes to a page, it may no longer be a page table,
3552 * or we may be forking, in which case it is better to unmap the page.
3554 static bool detect_write_flooding(struct kvm_mmu_page *sp, u64 *spte)
3557 * Skip write-flooding detected for the sp whose level is 1, because
3558 * it can become unsync, then the guest page is not write-protected.
3560 if (sp->role.level == 1)
3563 return ++sp->write_flooding_count >= 3;
3567 * Misaligned accesses are too much trouble to fix up; also, they usually
3568 * indicate a page is not used as a page table.
3570 static bool detect_write_misaligned(struct kvm_mmu_page *sp, gpa_t gpa,
3573 unsigned offset, pte_size, misaligned;
3575 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
3576 gpa, bytes, sp->role.word);
3578 offset = offset_in_page(gpa);
3579 pte_size = sp->role.cr4_pae ? 8 : 4;
3582 * Sometimes, the OS only writes the last one bytes to update status
3583 * bits, for example, in linux, andb instruction is used in clear_bit().
3585 if (!(offset & (pte_size - 1)) && bytes == 1)
3588 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
3589 misaligned |= bytes < 4;
3594 static u64 *get_written_sptes(struct kvm_mmu_page *sp, gpa_t gpa, int *nspte)
3596 unsigned page_offset, quadrant;
3600 page_offset = offset_in_page(gpa);
3601 level = sp->role.level;
3603 if (!sp->role.cr4_pae) {
3604 page_offset <<= 1; /* 32->64 */
3606 * A 32-bit pde maps 4MB while the shadow pdes map
3607 * only 2MB. So we need to double the offset again
3608 * and zap two pdes instead of one.
3610 if (level == PT32_ROOT_LEVEL) {
3611 page_offset &= ~7; /* kill rounding error */
3615 quadrant = page_offset >> PAGE_SHIFT;
3616 page_offset &= ~PAGE_MASK;
3617 if (quadrant != sp->role.quadrant)
3621 spte = &sp->spt[page_offset / sizeof(*spte)];
3625 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
3626 const u8 *new, int bytes)
3628 gfn_t gfn = gpa >> PAGE_SHIFT;
3629 union kvm_mmu_page_role mask = { .word = 0 };
3630 struct kvm_mmu_page *sp;
3631 struct hlist_node *node;
3632 LIST_HEAD(invalid_list);
3633 u64 entry, gentry, *spte;
3635 bool remote_flush, local_flush, zap_page;
3638 * If we don't have indirect shadow pages, it means no page is
3639 * write-protected, so we can exit simply.
3641 if (!ACCESS_ONCE(vcpu->kvm->arch.indirect_shadow_pages))
3644 zap_page = remote_flush = local_flush = false;
3646 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
3648 gentry = mmu_pte_write_fetch_gpte(vcpu, &gpa, new, &bytes);
3651 * No need to care whether allocation memory is successful
3652 * or not since pte prefetch is skiped if it does not have
3653 * enough objects in the cache.
3655 mmu_topup_memory_caches(vcpu);
3657 spin_lock(&vcpu->kvm->mmu_lock);
3658 ++vcpu->kvm->stat.mmu_pte_write;
3659 kvm_mmu_audit(vcpu, AUDIT_PRE_PTE_WRITE);
3661 mask.cr0_wp = mask.cr4_pae = mask.nxe = 1;
3662 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn, node) {
3663 spte = get_written_sptes(sp, gpa, &npte);
3665 if (detect_write_misaligned(sp, gpa, bytes) ||
3666 detect_write_flooding(sp, spte)) {
3667 zap_page |= !!kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
3669 ++vcpu->kvm->stat.mmu_flooded;
3673 spte = get_written_sptes(sp, gpa, &npte);
3680 mmu_page_zap_pte(vcpu->kvm, sp, spte);
3682 !((sp->role.word ^ vcpu->arch.mmu.base_role.word)
3683 & mask.word) && rmap_can_add(vcpu))
3684 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
3685 if (!remote_flush && need_remote_flush(entry, *spte))
3686 remote_flush = true;
3690 mmu_pte_write_flush_tlb(vcpu, zap_page, remote_flush, local_flush);
3691 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
3692 kvm_mmu_audit(vcpu, AUDIT_POST_PTE_WRITE);
3693 spin_unlock(&vcpu->kvm->mmu_lock);
3696 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
3701 if (vcpu->arch.mmu.direct_map)
3704 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
3706 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
3710 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
3712 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
3714 LIST_HEAD(invalid_list);
3716 while (kvm_mmu_available_pages(vcpu->kvm) < KVM_REFILL_PAGES &&
3717 !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
3718 struct kvm_mmu_page *sp;
3720 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
3721 struct kvm_mmu_page, link);
3722 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
3723 ++vcpu->kvm->stat.mmu_recycled;
3725 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
3728 static bool is_mmio_page_fault(struct kvm_vcpu *vcpu, gva_t addr)
3730 if (vcpu->arch.mmu.direct_map || mmu_is_nested(vcpu))
3731 return vcpu_match_mmio_gpa(vcpu, addr);
3733 return vcpu_match_mmio_gva(vcpu, addr);
3736 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code,
3737 void *insn, int insn_len)
3739 int r, emulation_type = EMULTYPE_RETRY;
3740 enum emulation_result er;
3742 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code, false);
3751 if (is_mmio_page_fault(vcpu, cr2))
3754 er = x86_emulate_instruction(vcpu, cr2, emulation_type, insn, insn_len);
3759 case EMULATE_DO_MMIO:
3760 ++vcpu->stat.mmio_exits;
3770 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
3772 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
3774 vcpu->arch.mmu.invlpg(vcpu, gva);
3775 kvm_mmu_flush_tlb(vcpu);
3776 ++vcpu->stat.invlpg;
3778 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
3780 void kvm_enable_tdp(void)
3784 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
3786 void kvm_disable_tdp(void)
3788 tdp_enabled = false;
3790 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
3792 static void free_mmu_pages(struct kvm_vcpu *vcpu)
3794 free_page((unsigned long)vcpu->arch.mmu.pae_root);
3795 if (vcpu->arch.mmu.lm_root != NULL)
3796 free_page((unsigned long)vcpu->arch.mmu.lm_root);
3799 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
3807 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
3808 * Therefore we need to allocate shadow page tables in the first
3809 * 4GB of memory, which happens to fit the DMA32 zone.
3811 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
3815 vcpu->arch.mmu.pae_root = page_address(page);
3816 for (i = 0; i < 4; ++i)
3817 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
3822 int kvm_mmu_create(struct kvm_vcpu *vcpu)
3826 vcpu->arch.walk_mmu = &vcpu->arch.mmu;
3827 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
3828 vcpu->arch.mmu.translate_gpa = translate_gpa;
3829 vcpu->arch.nested_mmu.translate_gpa = translate_nested_gpa;
3831 return alloc_mmu_pages(vcpu);
3834 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
3837 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3839 return init_kvm_mmu(vcpu);
3842 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
3844 struct kvm_mmu_page *sp;
3846 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
3850 if (!test_bit(slot, sp->slot_bitmap))
3854 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3855 if (!is_shadow_present_pte(pt[i]) ||
3856 !is_last_spte(pt[i], sp->role.level))
3859 if (is_large_pte(pt[i])) {
3860 drop_spte(kvm, &pt[i]);
3866 if (is_writable_pte(pt[i]))
3867 mmu_spte_update(&pt[i],
3868 pt[i] & ~PT_WRITABLE_MASK);
3871 kvm_flush_remote_tlbs(kvm);
3874 void kvm_mmu_zap_all(struct kvm *kvm)
3876 struct kvm_mmu_page *sp, *node;
3877 LIST_HEAD(invalid_list);
3879 spin_lock(&kvm->mmu_lock);
3881 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
3882 if (kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list))
3885 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3886 spin_unlock(&kvm->mmu_lock);
3889 static void kvm_mmu_remove_some_alloc_mmu_pages(struct kvm *kvm,
3890 struct list_head *invalid_list)
3892 struct kvm_mmu_page *page;
3894 page = container_of(kvm->arch.active_mmu_pages.prev,
3895 struct kvm_mmu_page, link);
3896 kvm_mmu_prepare_zap_page(kvm, page, invalid_list);
3899 static int mmu_shrink(struct shrinker *shrink, struct shrink_control *sc)
3902 struct kvm *kvm_freed = NULL;
3903 int nr_to_scan = sc->nr_to_scan;
3905 if (nr_to_scan == 0)
3908 raw_spin_lock(&kvm_lock);
3910 list_for_each_entry(kvm, &vm_list, vm_list) {
3912 LIST_HEAD(invalid_list);
3914 idx = srcu_read_lock(&kvm->srcu);
3915 spin_lock(&kvm->mmu_lock);
3916 if (!kvm_freed && nr_to_scan > 0 &&
3917 kvm->arch.n_used_mmu_pages > 0) {
3918 kvm_mmu_remove_some_alloc_mmu_pages(kvm,
3924 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3925 spin_unlock(&kvm->mmu_lock);
3926 srcu_read_unlock(&kvm->srcu, idx);
3929 list_move_tail(&kvm_freed->vm_list, &vm_list);
3931 raw_spin_unlock(&kvm_lock);
3934 return percpu_counter_read_positive(&kvm_total_used_mmu_pages);
3937 static struct shrinker mmu_shrinker = {
3938 .shrink = mmu_shrink,
3939 .seeks = DEFAULT_SEEKS * 10,
3942 static void mmu_destroy_caches(void)
3944 if (pte_list_desc_cache)
3945 kmem_cache_destroy(pte_list_desc_cache);
3946 if (mmu_page_header_cache)
3947 kmem_cache_destroy(mmu_page_header_cache);
3950 int kvm_mmu_module_init(void)
3952 pte_list_desc_cache = kmem_cache_create("pte_list_desc",
3953 sizeof(struct pte_list_desc),
3955 if (!pte_list_desc_cache)
3958 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
3959 sizeof(struct kvm_mmu_page),
3961 if (!mmu_page_header_cache)
3964 if (percpu_counter_init(&kvm_total_used_mmu_pages, 0))
3967 register_shrinker(&mmu_shrinker);
3972 mmu_destroy_caches();
3977 * Caculate mmu pages needed for kvm.
3979 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3981 unsigned int nr_mmu_pages;
3982 unsigned int nr_pages = 0;
3983 struct kvm_memslots *slots;
3984 struct kvm_memory_slot *memslot;
3986 slots = kvm_memslots(kvm);
3988 kvm_for_each_memslot(memslot, slots)
3989 nr_pages += memslot->npages;
3991 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3992 nr_mmu_pages = max(nr_mmu_pages,
3993 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3995 return nr_mmu_pages;
3998 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
4000 struct kvm_shadow_walk_iterator iterator;
4004 walk_shadow_page_lockless_begin(vcpu);
4005 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte) {
4006 sptes[iterator.level-1] = spte;
4008 if (!is_shadow_present_pte(spte))
4011 walk_shadow_page_lockless_end(vcpu);
4015 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
4017 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
4021 destroy_kvm_mmu(vcpu);
4022 free_mmu_pages(vcpu);
4023 mmu_free_memory_caches(vcpu);
4026 void kvm_mmu_module_exit(void)
4028 mmu_destroy_caches();
4029 percpu_counter_destroy(&kvm_total_used_mmu_pages);
4030 unregister_shrinker(&mmu_shrinker);
4031 mmu_audit_disable();