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 module_param(dbg, bool, 0644);
69 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
70 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
71 #define MMU_WARN_ON(x) WARN_ON(x)
73 #define pgprintk(x...) do { } while (0)
74 #define rmap_printk(x...) do { } while (0)
75 #define MMU_WARN_ON(x) do { } while (0)
78 #define PTE_PREFETCH_NUM 8
80 #define PT_FIRST_AVAIL_BITS_SHIFT 10
81 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
83 #define PT64_LEVEL_BITS 9
85 #define PT64_LEVEL_SHIFT(level) \
86 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
88 #define PT64_INDEX(address, level)\
89 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
92 #define PT32_LEVEL_BITS 10
94 #define PT32_LEVEL_SHIFT(level) \
95 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
97 #define PT32_LVL_OFFSET_MASK(level) \
98 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
99 * PT32_LEVEL_BITS))) - 1))
101 #define PT32_INDEX(address, level)\
102 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
105 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
106 #define PT64_DIR_BASE_ADDR_MASK \
107 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
108 #define PT64_LVL_ADDR_MASK(level) \
109 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
110 * PT64_LEVEL_BITS))) - 1))
111 #define PT64_LVL_OFFSET_MASK(level) \
112 (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
113 * PT64_LEVEL_BITS))) - 1))
115 #define PT32_BASE_ADDR_MASK PAGE_MASK
116 #define PT32_DIR_BASE_ADDR_MASK \
117 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
118 #define PT32_LVL_ADDR_MASK(level) \
119 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
120 * PT32_LEVEL_BITS))) - 1))
122 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | shadow_user_mask \
123 | shadow_x_mask | shadow_nx_mask)
125 #define ACC_EXEC_MASK 1
126 #define ACC_WRITE_MASK PT_WRITABLE_MASK
127 #define ACC_USER_MASK PT_USER_MASK
128 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
130 #include <trace/events/kvm.h>
132 #define CREATE_TRACE_POINTS
133 #include "mmutrace.h"
135 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
136 #define SPTE_MMU_WRITEABLE (1ULL << (PT_FIRST_AVAIL_BITS_SHIFT + 1))
138 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
140 /* make pte_list_desc fit well in cache line */
141 #define PTE_LIST_EXT 3
143 struct pte_list_desc {
144 u64 *sptes[PTE_LIST_EXT];
145 struct pte_list_desc *more;
148 struct kvm_shadow_walk_iterator {
156 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
157 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
158 shadow_walk_okay(&(_walker)); \
159 shadow_walk_next(&(_walker)))
161 #define for_each_shadow_entry_lockless(_vcpu, _addr, _walker, spte) \
162 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
163 shadow_walk_okay(&(_walker)) && \
164 ({ spte = mmu_spte_get_lockless(_walker.sptep); 1; }); \
165 __shadow_walk_next(&(_walker), spte))
167 static struct kmem_cache *pte_list_desc_cache;
168 static struct kmem_cache *mmu_page_header_cache;
169 static struct percpu_counter kvm_total_used_mmu_pages;
171 static u64 __read_mostly shadow_nx_mask;
172 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
173 static u64 __read_mostly shadow_user_mask;
174 static u64 __read_mostly shadow_accessed_mask;
175 static u64 __read_mostly shadow_dirty_mask;
176 static u64 __read_mostly shadow_mmio_mask;
178 static void mmu_spte_set(u64 *sptep, u64 spte);
179 static void mmu_free_roots(struct kvm_vcpu *vcpu);
181 void kvm_mmu_set_mmio_spte_mask(u64 mmio_mask)
183 shadow_mmio_mask = mmio_mask;
185 EXPORT_SYMBOL_GPL(kvm_mmu_set_mmio_spte_mask);
188 * the low bit of the generation number is always presumed to be zero.
189 * This disables mmio caching during memslot updates. The concept is
190 * similar to a seqcount but instead of retrying the access we just punt
191 * and ignore the cache.
193 * spte bits 3-11 are used as bits 1-9 of the generation number,
194 * the bits 52-61 are used as bits 10-19 of the generation number.
196 #define MMIO_SPTE_GEN_LOW_SHIFT 2
197 #define MMIO_SPTE_GEN_HIGH_SHIFT 52
199 #define MMIO_GEN_SHIFT 20
200 #define MMIO_GEN_LOW_SHIFT 10
201 #define MMIO_GEN_LOW_MASK ((1 << MMIO_GEN_LOW_SHIFT) - 2)
202 #define MMIO_GEN_MASK ((1 << MMIO_GEN_SHIFT) - 1)
204 static u64 generation_mmio_spte_mask(unsigned int gen)
208 WARN_ON(gen & ~MMIO_GEN_MASK);
210 mask = (gen & MMIO_GEN_LOW_MASK) << MMIO_SPTE_GEN_LOW_SHIFT;
211 mask |= ((u64)gen >> MMIO_GEN_LOW_SHIFT) << MMIO_SPTE_GEN_HIGH_SHIFT;
215 static unsigned int get_mmio_spte_generation(u64 spte)
219 spte &= ~shadow_mmio_mask;
221 gen = (spte >> MMIO_SPTE_GEN_LOW_SHIFT) & MMIO_GEN_LOW_MASK;
222 gen |= (spte >> MMIO_SPTE_GEN_HIGH_SHIFT) << MMIO_GEN_LOW_SHIFT;
226 static unsigned int kvm_current_mmio_generation(struct kvm_vcpu *vcpu)
228 return kvm_vcpu_memslots(vcpu)->generation & MMIO_GEN_MASK;
231 static void mark_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, u64 gfn,
234 unsigned int gen = kvm_current_mmio_generation(vcpu);
235 u64 mask = generation_mmio_spte_mask(gen);
237 access &= ACC_WRITE_MASK | ACC_USER_MASK;
238 mask |= shadow_mmio_mask | access | gfn << PAGE_SHIFT;
240 trace_mark_mmio_spte(sptep, gfn, access, gen);
241 mmu_spte_set(sptep, mask);
244 static bool is_mmio_spte(u64 spte)
246 return (spte & shadow_mmio_mask) == shadow_mmio_mask;
249 static gfn_t get_mmio_spte_gfn(u64 spte)
251 u64 mask = generation_mmio_spte_mask(MMIO_GEN_MASK) | shadow_mmio_mask;
252 return (spte & ~mask) >> PAGE_SHIFT;
255 static unsigned get_mmio_spte_access(u64 spte)
257 u64 mask = generation_mmio_spte_mask(MMIO_GEN_MASK) | shadow_mmio_mask;
258 return (spte & ~mask) & ~PAGE_MASK;
261 static bool set_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn,
262 pfn_t pfn, unsigned access)
264 if (unlikely(is_noslot_pfn(pfn))) {
265 mark_mmio_spte(vcpu, sptep, gfn, access);
272 static bool check_mmio_spte(struct kvm_vcpu *vcpu, u64 spte)
274 unsigned int kvm_gen, spte_gen;
276 kvm_gen = kvm_current_mmio_generation(vcpu);
277 spte_gen = get_mmio_spte_generation(spte);
279 trace_check_mmio_spte(spte, kvm_gen, spte_gen);
280 return likely(kvm_gen == spte_gen);
283 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
284 u64 dirty_mask, u64 nx_mask, u64 x_mask)
286 shadow_user_mask = user_mask;
287 shadow_accessed_mask = accessed_mask;
288 shadow_dirty_mask = dirty_mask;
289 shadow_nx_mask = nx_mask;
290 shadow_x_mask = x_mask;
292 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
294 static int is_cpuid_PSE36(void)
299 static int is_nx(struct kvm_vcpu *vcpu)
301 return vcpu->arch.efer & EFER_NX;
304 static int is_shadow_present_pte(u64 pte)
306 return pte & PT_PRESENT_MASK && !is_mmio_spte(pte);
309 static int is_large_pte(u64 pte)
311 return pte & PT_PAGE_SIZE_MASK;
314 static int is_rmap_spte(u64 pte)
316 return is_shadow_present_pte(pte);
319 static int is_last_spte(u64 pte, int level)
321 if (level == PT_PAGE_TABLE_LEVEL)
323 if (is_large_pte(pte))
328 static pfn_t spte_to_pfn(u64 pte)
330 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
333 static gfn_t pse36_gfn_delta(u32 gpte)
335 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
337 return (gpte & PT32_DIR_PSE36_MASK) << shift;
341 static void __set_spte(u64 *sptep, u64 spte)
346 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
351 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
353 return xchg(sptep, spte);
356 static u64 __get_spte_lockless(u64 *sptep)
358 return ACCESS_ONCE(*sptep);
361 static bool __check_direct_spte_mmio_pf(u64 spte)
363 /* It is valid if the spte is zapped. */
375 static void count_spte_clear(u64 *sptep, u64 spte)
377 struct kvm_mmu_page *sp = page_header(__pa(sptep));
379 if (is_shadow_present_pte(spte))
382 /* Ensure the spte is completely set before we increase the count */
384 sp->clear_spte_count++;
387 static void __set_spte(u64 *sptep, u64 spte)
389 union split_spte *ssptep, sspte;
391 ssptep = (union split_spte *)sptep;
392 sspte = (union split_spte)spte;
394 ssptep->spte_high = sspte.spte_high;
397 * If we map the spte from nonpresent to present, We should store
398 * the high bits firstly, then set present bit, so cpu can not
399 * fetch this spte while we are setting the spte.
403 ssptep->spte_low = sspte.spte_low;
406 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
408 union split_spte *ssptep, sspte;
410 ssptep = (union split_spte *)sptep;
411 sspte = (union split_spte)spte;
413 ssptep->spte_low = sspte.spte_low;
416 * If we map the spte from present to nonpresent, we should clear
417 * present bit firstly to avoid vcpu fetch the old high bits.
421 ssptep->spte_high = sspte.spte_high;
422 count_spte_clear(sptep, spte);
425 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
427 union split_spte *ssptep, sspte, orig;
429 ssptep = (union split_spte *)sptep;
430 sspte = (union split_spte)spte;
432 /* xchg acts as a barrier before the setting of the high bits */
433 orig.spte_low = xchg(&ssptep->spte_low, sspte.spte_low);
434 orig.spte_high = ssptep->spte_high;
435 ssptep->spte_high = sspte.spte_high;
436 count_spte_clear(sptep, spte);
442 * The idea using the light way get the spte on x86_32 guest is from
443 * gup_get_pte(arch/x86/mm/gup.c).
445 * An spte tlb flush may be pending, because kvm_set_pte_rmapp
446 * coalesces them and we are running out of the MMU lock. Therefore
447 * we need to protect against in-progress updates of the spte.
449 * Reading the spte while an update is in progress may get the old value
450 * for the high part of the spte. The race is fine for a present->non-present
451 * change (because the high part of the spte is ignored for non-present spte),
452 * but for a present->present change we must reread the spte.
454 * All such changes are done in two steps (present->non-present and
455 * non-present->present), hence it is enough to count the number of
456 * present->non-present updates: if it changed while reading the spte,
457 * we might have hit the race. This is done using clear_spte_count.
459 static u64 __get_spte_lockless(u64 *sptep)
461 struct kvm_mmu_page *sp = page_header(__pa(sptep));
462 union split_spte spte, *orig = (union split_spte *)sptep;
466 count = sp->clear_spte_count;
469 spte.spte_low = orig->spte_low;
472 spte.spte_high = orig->spte_high;
475 if (unlikely(spte.spte_low != orig->spte_low ||
476 count != sp->clear_spte_count))
482 static bool __check_direct_spte_mmio_pf(u64 spte)
484 union split_spte sspte = (union split_spte)spte;
485 u32 high_mmio_mask = shadow_mmio_mask >> 32;
487 /* It is valid if the spte is zapped. */
491 /* It is valid if the spte is being zapped. */
492 if (sspte.spte_low == 0ull &&
493 (sspte.spte_high & high_mmio_mask) == high_mmio_mask)
500 static bool spte_is_locklessly_modifiable(u64 spte)
502 return (spte & (SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE)) ==
503 (SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE);
506 static bool spte_has_volatile_bits(u64 spte)
509 * Always atomicly update spte if it can be updated
510 * out of mmu-lock, it can ensure dirty bit is not lost,
511 * also, it can help us to get a stable is_writable_pte()
512 * to ensure tlb flush is not missed.
514 if (spte_is_locklessly_modifiable(spte))
517 if (!shadow_accessed_mask)
520 if (!is_shadow_present_pte(spte))
523 if ((spte & shadow_accessed_mask) &&
524 (!is_writable_pte(spte) || (spte & shadow_dirty_mask)))
530 static bool spte_is_bit_cleared(u64 old_spte, u64 new_spte, u64 bit_mask)
532 return (old_spte & bit_mask) && !(new_spte & bit_mask);
535 static bool spte_is_bit_changed(u64 old_spte, u64 new_spte, u64 bit_mask)
537 return (old_spte & bit_mask) != (new_spte & bit_mask);
540 /* Rules for using mmu_spte_set:
541 * Set the sptep from nonpresent to present.
542 * Note: the sptep being assigned *must* be either not present
543 * or in a state where the hardware will not attempt to update
546 static void mmu_spte_set(u64 *sptep, u64 new_spte)
548 WARN_ON(is_shadow_present_pte(*sptep));
549 __set_spte(sptep, new_spte);
552 /* Rules for using mmu_spte_update:
553 * Update the state bits, it means the mapped pfn is not changged.
555 * Whenever we overwrite a writable spte with a read-only one we
556 * should flush remote TLBs. Otherwise rmap_write_protect
557 * will find a read-only spte, even though the writable spte
558 * might be cached on a CPU's TLB, the return value indicates this
561 static bool mmu_spte_update(u64 *sptep, u64 new_spte)
563 u64 old_spte = *sptep;
566 WARN_ON(!is_rmap_spte(new_spte));
568 if (!is_shadow_present_pte(old_spte)) {
569 mmu_spte_set(sptep, new_spte);
573 if (!spte_has_volatile_bits(old_spte))
574 __update_clear_spte_fast(sptep, new_spte);
576 old_spte = __update_clear_spte_slow(sptep, new_spte);
579 * For the spte updated out of mmu-lock is safe, since
580 * we always atomicly update it, see the comments in
581 * spte_has_volatile_bits().
583 if (spte_is_locklessly_modifiable(old_spte) &&
584 !is_writable_pte(new_spte))
587 if (!shadow_accessed_mask)
591 * Flush TLB when accessed/dirty bits are changed in the page tables,
592 * to guarantee consistency between TLB and page tables.
594 if (spte_is_bit_changed(old_spte, new_spte,
595 shadow_accessed_mask | shadow_dirty_mask))
598 if (spte_is_bit_cleared(old_spte, new_spte, shadow_accessed_mask))
599 kvm_set_pfn_accessed(spte_to_pfn(old_spte));
600 if (spte_is_bit_cleared(old_spte, new_spte, shadow_dirty_mask))
601 kvm_set_pfn_dirty(spte_to_pfn(old_spte));
607 * Rules for using mmu_spte_clear_track_bits:
608 * It sets the sptep from present to nonpresent, and track the
609 * state bits, it is used to clear the last level sptep.
611 static int mmu_spte_clear_track_bits(u64 *sptep)
614 u64 old_spte = *sptep;
616 if (!spte_has_volatile_bits(old_spte))
617 __update_clear_spte_fast(sptep, 0ull);
619 old_spte = __update_clear_spte_slow(sptep, 0ull);
621 if (!is_rmap_spte(old_spte))
624 pfn = spte_to_pfn(old_spte);
627 * KVM does not hold the refcount of the page used by
628 * kvm mmu, before reclaiming the page, we should
629 * unmap it from mmu first.
631 WARN_ON(!kvm_is_reserved_pfn(pfn) && !page_count(pfn_to_page(pfn)));
633 if (!shadow_accessed_mask || old_spte & shadow_accessed_mask)
634 kvm_set_pfn_accessed(pfn);
635 if (!shadow_dirty_mask || (old_spte & shadow_dirty_mask))
636 kvm_set_pfn_dirty(pfn);
641 * Rules for using mmu_spte_clear_no_track:
642 * Directly clear spte without caring the state bits of sptep,
643 * it is used to set the upper level spte.
645 static void mmu_spte_clear_no_track(u64 *sptep)
647 __update_clear_spte_fast(sptep, 0ull);
650 static u64 mmu_spte_get_lockless(u64 *sptep)
652 return __get_spte_lockless(sptep);
655 static void walk_shadow_page_lockless_begin(struct kvm_vcpu *vcpu)
658 * Prevent page table teardown by making any free-er wait during
659 * kvm_flush_remote_tlbs() IPI to all active vcpus.
662 vcpu->mode = READING_SHADOW_PAGE_TABLES;
664 * Make sure a following spte read is not reordered ahead of the write
670 static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
673 * Make sure the write to vcpu->mode is not reordered in front of
674 * reads to sptes. If it does, kvm_commit_zap_page() can see us
675 * OUTSIDE_GUEST_MODE and proceed to free the shadow page table.
678 vcpu->mode = OUTSIDE_GUEST_MODE;
682 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
683 struct kmem_cache *base_cache, int min)
687 if (cache->nobjs >= min)
689 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
690 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
693 cache->objects[cache->nobjs++] = obj;
698 static int mmu_memory_cache_free_objects(struct kvm_mmu_memory_cache *cache)
703 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc,
704 struct kmem_cache *cache)
707 kmem_cache_free(cache, mc->objects[--mc->nobjs]);
710 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
715 if (cache->nobjs >= min)
717 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
718 page = (void *)__get_free_page(GFP_KERNEL);
721 cache->objects[cache->nobjs++] = page;
726 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
729 free_page((unsigned long)mc->objects[--mc->nobjs]);
732 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
736 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
737 pte_list_desc_cache, 8 + PTE_PREFETCH_NUM);
740 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
743 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
744 mmu_page_header_cache, 4);
749 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
751 mmu_free_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
752 pte_list_desc_cache);
753 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
754 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache,
755 mmu_page_header_cache);
758 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
763 p = mc->objects[--mc->nobjs];
767 static struct pte_list_desc *mmu_alloc_pte_list_desc(struct kvm_vcpu *vcpu)
769 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_list_desc_cache);
772 static void mmu_free_pte_list_desc(struct pte_list_desc *pte_list_desc)
774 kmem_cache_free(pte_list_desc_cache, pte_list_desc);
777 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
779 if (!sp->role.direct)
780 return sp->gfns[index];
782 return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
785 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
788 BUG_ON(gfn != kvm_mmu_page_get_gfn(sp, index));
790 sp->gfns[index] = gfn;
794 * Return the pointer to the large page information for a given gfn,
795 * handling slots that are not large page aligned.
797 static struct kvm_lpage_info *lpage_info_slot(gfn_t gfn,
798 struct kvm_memory_slot *slot,
803 idx = gfn_to_index(gfn, slot->base_gfn, level);
804 return &slot->arch.lpage_info[level - 2][idx];
807 static void account_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
809 struct kvm_memslots *slots;
810 struct kvm_memory_slot *slot;
811 struct kvm_lpage_info *linfo;
816 slots = kvm_memslots_for_spte_role(kvm, sp->role);
817 slot = __gfn_to_memslot(slots, gfn);
818 for (i = PT_DIRECTORY_LEVEL; i <= PT_MAX_HUGEPAGE_LEVEL; ++i) {
819 linfo = lpage_info_slot(gfn, slot, i);
820 linfo->write_count += 1;
822 kvm->arch.indirect_shadow_pages++;
825 static void unaccount_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
827 struct kvm_memslots *slots;
828 struct kvm_memory_slot *slot;
829 struct kvm_lpage_info *linfo;
834 slots = kvm_memslots_for_spte_role(kvm, sp->role);
835 slot = __gfn_to_memslot(slots, gfn);
836 for (i = PT_DIRECTORY_LEVEL; i <= PT_MAX_HUGEPAGE_LEVEL; ++i) {
837 linfo = lpage_info_slot(gfn, slot, i);
838 linfo->write_count -= 1;
839 WARN_ON(linfo->write_count < 0);
841 kvm->arch.indirect_shadow_pages--;
844 static int has_wrprotected_page(struct kvm_vcpu *vcpu,
848 struct kvm_memory_slot *slot;
849 struct kvm_lpage_info *linfo;
851 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
853 linfo = lpage_info_slot(gfn, slot, level);
854 return linfo->write_count;
860 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
862 unsigned long page_size;
865 page_size = kvm_host_page_size(kvm, gfn);
867 for (i = PT_PAGE_TABLE_LEVEL; i <= PT_MAX_HUGEPAGE_LEVEL; ++i) {
868 if (page_size >= KVM_HPAGE_SIZE(i))
877 static struct kvm_memory_slot *
878 gfn_to_memslot_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t gfn,
881 struct kvm_memory_slot *slot;
883 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
884 if (!slot || slot->flags & KVM_MEMSLOT_INVALID ||
885 (no_dirty_log && slot->dirty_bitmap))
891 static bool mapping_level_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t large_gfn)
893 return !gfn_to_memslot_dirty_bitmap(vcpu, large_gfn, true);
896 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
898 int host_level, level, max_level;
900 host_level = host_mapping_level(vcpu->kvm, large_gfn);
902 if (host_level == PT_PAGE_TABLE_LEVEL)
905 max_level = min(kvm_x86_ops->get_lpage_level(), host_level);
907 for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
908 if (has_wrprotected_page(vcpu, large_gfn, level))
915 * Pte mapping structures:
917 * If pte_list bit zero is zero, then pte_list point to the spte.
919 * If pte_list bit zero is one, (then pte_list & ~1) points to a struct
920 * pte_list_desc containing more mappings.
922 * Returns the number of pte entries before the spte was added or zero if
923 * the spte was not added.
926 static int pte_list_add(struct kvm_vcpu *vcpu, u64 *spte,
927 unsigned long *pte_list)
929 struct pte_list_desc *desc;
933 rmap_printk("pte_list_add: %p %llx 0->1\n", spte, *spte);
934 *pte_list = (unsigned long)spte;
935 } else if (!(*pte_list & 1)) {
936 rmap_printk("pte_list_add: %p %llx 1->many\n", spte, *spte);
937 desc = mmu_alloc_pte_list_desc(vcpu);
938 desc->sptes[0] = (u64 *)*pte_list;
939 desc->sptes[1] = spte;
940 *pte_list = (unsigned long)desc | 1;
943 rmap_printk("pte_list_add: %p %llx many->many\n", spte, *spte);
944 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
945 while (desc->sptes[PTE_LIST_EXT-1] && desc->more) {
947 count += PTE_LIST_EXT;
949 if (desc->sptes[PTE_LIST_EXT-1]) {
950 desc->more = mmu_alloc_pte_list_desc(vcpu);
953 for (i = 0; desc->sptes[i]; ++i)
955 desc->sptes[i] = spte;
961 pte_list_desc_remove_entry(unsigned long *pte_list, struct pte_list_desc *desc,
962 int i, struct pte_list_desc *prev_desc)
966 for (j = PTE_LIST_EXT - 1; !desc->sptes[j] && j > i; --j)
968 desc->sptes[i] = desc->sptes[j];
969 desc->sptes[j] = NULL;
972 if (!prev_desc && !desc->more)
973 *pte_list = (unsigned long)desc->sptes[0];
976 prev_desc->more = desc->more;
978 *pte_list = (unsigned long)desc->more | 1;
979 mmu_free_pte_list_desc(desc);
982 static void pte_list_remove(u64 *spte, unsigned long *pte_list)
984 struct pte_list_desc *desc;
985 struct pte_list_desc *prev_desc;
989 printk(KERN_ERR "pte_list_remove: %p 0->BUG\n", spte);
991 } else if (!(*pte_list & 1)) {
992 rmap_printk("pte_list_remove: %p 1->0\n", spte);
993 if ((u64 *)*pte_list != spte) {
994 printk(KERN_ERR "pte_list_remove: %p 1->BUG\n", spte);
999 rmap_printk("pte_list_remove: %p many->many\n", spte);
1000 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
1003 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i)
1004 if (desc->sptes[i] == spte) {
1005 pte_list_desc_remove_entry(pte_list,
1013 pr_err("pte_list_remove: %p many->many\n", spte);
1018 typedef void (*pte_list_walk_fn) (u64 *spte);
1019 static void pte_list_walk(unsigned long *pte_list, pte_list_walk_fn fn)
1021 struct pte_list_desc *desc;
1027 if (!(*pte_list & 1))
1028 return fn((u64 *)*pte_list);
1030 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
1032 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i)
1038 static unsigned long *__gfn_to_rmap(gfn_t gfn, int level,
1039 struct kvm_memory_slot *slot)
1043 idx = gfn_to_index(gfn, slot->base_gfn, level);
1044 return &slot->arch.rmap[level - PT_PAGE_TABLE_LEVEL][idx];
1048 * Take gfn and return the reverse mapping to it.
1050 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, struct kvm_mmu_page *sp)
1052 struct kvm_memslots *slots;
1053 struct kvm_memory_slot *slot;
1055 slots = kvm_memslots_for_spte_role(kvm, sp->role);
1056 slot = __gfn_to_memslot(slots, gfn);
1057 return __gfn_to_rmap(gfn, sp->role.level, slot);
1060 static bool rmap_can_add(struct kvm_vcpu *vcpu)
1062 struct kvm_mmu_memory_cache *cache;
1064 cache = &vcpu->arch.mmu_pte_list_desc_cache;
1065 return mmu_memory_cache_free_objects(cache);
1068 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
1070 struct kvm_mmu_page *sp;
1071 unsigned long *rmapp;
1073 sp = page_header(__pa(spte));
1074 kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
1075 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp);
1076 return pte_list_add(vcpu, spte, rmapp);
1079 static void rmap_remove(struct kvm *kvm, u64 *spte)
1081 struct kvm_mmu_page *sp;
1083 unsigned long *rmapp;
1085 sp = page_header(__pa(spte));
1086 gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
1087 rmapp = gfn_to_rmap(kvm, gfn, sp);
1088 pte_list_remove(spte, rmapp);
1092 * Used by the following functions to iterate through the sptes linked by a
1093 * rmap. All fields are private and not assumed to be used outside.
1095 struct rmap_iterator {
1096 /* private fields */
1097 struct pte_list_desc *desc; /* holds the sptep if not NULL */
1098 int pos; /* index of the sptep */
1102 * Iteration must be started by this function. This should also be used after
1103 * removing/dropping sptes from the rmap link because in such cases the
1104 * information in the itererator may not be valid.
1106 * Returns sptep if found, NULL otherwise.
1108 static u64 *rmap_get_first(unsigned long rmap, struct rmap_iterator *iter)
1118 iter->desc = (struct pte_list_desc *)(rmap & ~1ul);
1120 return iter->desc->sptes[iter->pos];
1124 * Must be used with a valid iterator: e.g. after rmap_get_first().
1126 * Returns sptep if found, NULL otherwise.
1128 static u64 *rmap_get_next(struct rmap_iterator *iter)
1131 if (iter->pos < PTE_LIST_EXT - 1) {
1135 sptep = iter->desc->sptes[iter->pos];
1140 iter->desc = iter->desc->more;
1144 /* desc->sptes[0] cannot be NULL */
1145 return iter->desc->sptes[iter->pos];
1152 #define for_each_rmap_spte(_rmap_, _iter_, _spte_) \
1153 for (_spte_ = rmap_get_first(*_rmap_, _iter_); \
1154 _spte_ && ({BUG_ON(!is_shadow_present_pte(*_spte_)); 1;}); \
1155 _spte_ = rmap_get_next(_iter_))
1157 static void drop_spte(struct kvm *kvm, u64 *sptep)
1159 if (mmu_spte_clear_track_bits(sptep))
1160 rmap_remove(kvm, sptep);
1164 static bool __drop_large_spte(struct kvm *kvm, u64 *sptep)
1166 if (is_large_pte(*sptep)) {
1167 WARN_ON(page_header(__pa(sptep))->role.level ==
1168 PT_PAGE_TABLE_LEVEL);
1169 drop_spte(kvm, sptep);
1177 static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1179 if (__drop_large_spte(vcpu->kvm, sptep))
1180 kvm_flush_remote_tlbs(vcpu->kvm);
1184 * Write-protect on the specified @sptep, @pt_protect indicates whether
1185 * spte write-protection is caused by protecting shadow page table.
1187 * Note: write protection is difference between dirty logging and spte
1189 * - for dirty logging, the spte can be set to writable at anytime if
1190 * its dirty bitmap is properly set.
1191 * - for spte protection, the spte can be writable only after unsync-ing
1194 * Return true if tlb need be flushed.
1196 static bool spte_write_protect(struct kvm *kvm, u64 *sptep, bool pt_protect)
1200 if (!is_writable_pte(spte) &&
1201 !(pt_protect && spte_is_locklessly_modifiable(spte)))
1204 rmap_printk("rmap_write_protect: spte %p %llx\n", sptep, *sptep);
1207 spte &= ~SPTE_MMU_WRITEABLE;
1208 spte = spte & ~PT_WRITABLE_MASK;
1210 return mmu_spte_update(sptep, spte);
1213 static bool __rmap_write_protect(struct kvm *kvm, unsigned long *rmapp,
1217 struct rmap_iterator iter;
1220 for_each_rmap_spte(rmapp, &iter, sptep)
1221 flush |= spte_write_protect(kvm, sptep, pt_protect);
1226 static bool spte_clear_dirty(struct kvm *kvm, u64 *sptep)
1230 rmap_printk("rmap_clear_dirty: spte %p %llx\n", sptep, *sptep);
1232 spte &= ~shadow_dirty_mask;
1234 return mmu_spte_update(sptep, spte);
1237 static bool __rmap_clear_dirty(struct kvm *kvm, unsigned long *rmapp)
1240 struct rmap_iterator iter;
1243 for_each_rmap_spte(rmapp, &iter, sptep)
1244 flush |= spte_clear_dirty(kvm, sptep);
1249 static bool spte_set_dirty(struct kvm *kvm, u64 *sptep)
1253 rmap_printk("rmap_set_dirty: spte %p %llx\n", sptep, *sptep);
1255 spte |= shadow_dirty_mask;
1257 return mmu_spte_update(sptep, spte);
1260 static bool __rmap_set_dirty(struct kvm *kvm, unsigned long *rmapp)
1263 struct rmap_iterator iter;
1266 for_each_rmap_spte(rmapp, &iter, sptep)
1267 flush |= spte_set_dirty(kvm, sptep);
1273 * kvm_mmu_write_protect_pt_masked - write protect selected PT level pages
1274 * @kvm: kvm instance
1275 * @slot: slot to protect
1276 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1277 * @mask: indicates which pages we should protect
1279 * Used when we do not need to care about huge page mappings: e.g. during dirty
1280 * logging we do not have any such mappings.
1282 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1283 struct kvm_memory_slot *slot,
1284 gfn_t gfn_offset, unsigned long mask)
1286 unsigned long *rmapp;
1289 rmapp = __gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1290 PT_PAGE_TABLE_LEVEL, slot);
1291 __rmap_write_protect(kvm, rmapp, false);
1293 /* clear the first set bit */
1299 * kvm_mmu_clear_dirty_pt_masked - clear MMU D-bit for PT level pages
1300 * @kvm: kvm instance
1301 * @slot: slot to clear D-bit
1302 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1303 * @mask: indicates which pages we should clear D-bit
1305 * Used for PML to re-log the dirty GPAs after userspace querying dirty_bitmap.
1307 void kvm_mmu_clear_dirty_pt_masked(struct kvm *kvm,
1308 struct kvm_memory_slot *slot,
1309 gfn_t gfn_offset, unsigned long mask)
1311 unsigned long *rmapp;
1314 rmapp = __gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1315 PT_PAGE_TABLE_LEVEL, slot);
1316 __rmap_clear_dirty(kvm, rmapp);
1318 /* clear the first set bit */
1322 EXPORT_SYMBOL_GPL(kvm_mmu_clear_dirty_pt_masked);
1325 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1328 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1329 * enable dirty logging for them.
1331 * Used when we do not need to care about huge page mappings: e.g. during dirty
1332 * logging we do not have any such mappings.
1334 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1335 struct kvm_memory_slot *slot,
1336 gfn_t gfn_offset, unsigned long mask)
1338 if (kvm_x86_ops->enable_log_dirty_pt_masked)
1339 kvm_x86_ops->enable_log_dirty_pt_masked(kvm, slot, gfn_offset,
1342 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1345 static bool rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
1347 struct kvm_memory_slot *slot;
1348 unsigned long *rmapp;
1350 bool write_protected = false;
1352 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1354 for (i = PT_PAGE_TABLE_LEVEL; i <= PT_MAX_HUGEPAGE_LEVEL; ++i) {
1355 rmapp = __gfn_to_rmap(gfn, i, slot);
1356 write_protected |= __rmap_write_protect(vcpu->kvm, rmapp, true);
1359 return write_protected;
1362 static bool kvm_zap_rmapp(struct kvm *kvm, unsigned long *rmapp)
1365 struct rmap_iterator iter;
1368 while ((sptep = rmap_get_first(*rmapp, &iter))) {
1369 BUG_ON(!(*sptep & PT_PRESENT_MASK));
1370 rmap_printk("%s: spte %p %llx.\n", __func__, sptep, *sptep);
1372 drop_spte(kvm, sptep);
1379 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
1380 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1383 return kvm_zap_rmapp(kvm, rmapp);
1386 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
1387 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1391 struct rmap_iterator iter;
1394 pte_t *ptep = (pte_t *)data;
1397 WARN_ON(pte_huge(*ptep));
1398 new_pfn = pte_pfn(*ptep);
1401 for_each_rmap_spte(rmapp, &iter, sptep) {
1402 rmap_printk("kvm_set_pte_rmapp: spte %p %llx gfn %llx (%d)\n",
1403 sptep, *sptep, gfn, level);
1407 if (pte_write(*ptep)) {
1408 drop_spte(kvm, sptep);
1411 new_spte = *sptep & ~PT64_BASE_ADDR_MASK;
1412 new_spte |= (u64)new_pfn << PAGE_SHIFT;
1414 new_spte &= ~PT_WRITABLE_MASK;
1415 new_spte &= ~SPTE_HOST_WRITEABLE;
1416 new_spte &= ~shadow_accessed_mask;
1418 mmu_spte_clear_track_bits(sptep);
1419 mmu_spte_set(sptep, new_spte);
1424 kvm_flush_remote_tlbs(kvm);
1429 struct slot_rmap_walk_iterator {
1431 struct kvm_memory_slot *slot;
1437 /* output fields. */
1439 unsigned long *rmap;
1442 /* private field. */
1443 unsigned long *end_rmap;
1447 rmap_walk_init_level(struct slot_rmap_walk_iterator *iterator, int level)
1449 iterator->level = level;
1450 iterator->gfn = iterator->start_gfn;
1451 iterator->rmap = __gfn_to_rmap(iterator->gfn, level, iterator->slot);
1452 iterator->end_rmap = __gfn_to_rmap(iterator->end_gfn, level,
1457 slot_rmap_walk_init(struct slot_rmap_walk_iterator *iterator,
1458 struct kvm_memory_slot *slot, int start_level,
1459 int end_level, gfn_t start_gfn, gfn_t end_gfn)
1461 iterator->slot = slot;
1462 iterator->start_level = start_level;
1463 iterator->end_level = end_level;
1464 iterator->start_gfn = start_gfn;
1465 iterator->end_gfn = end_gfn;
1467 rmap_walk_init_level(iterator, iterator->start_level);
1470 static bool slot_rmap_walk_okay(struct slot_rmap_walk_iterator *iterator)
1472 return !!iterator->rmap;
1475 static void slot_rmap_walk_next(struct slot_rmap_walk_iterator *iterator)
1477 if (++iterator->rmap <= iterator->end_rmap) {
1478 iterator->gfn += (1UL << KVM_HPAGE_GFN_SHIFT(iterator->level));
1482 if (++iterator->level > iterator->end_level) {
1483 iterator->rmap = NULL;
1487 rmap_walk_init_level(iterator, iterator->level);
1490 #define for_each_slot_rmap_range(_slot_, _start_level_, _end_level_, \
1491 _start_gfn, _end_gfn, _iter_) \
1492 for (slot_rmap_walk_init(_iter_, _slot_, _start_level_, \
1493 _end_level_, _start_gfn, _end_gfn); \
1494 slot_rmap_walk_okay(_iter_); \
1495 slot_rmap_walk_next(_iter_))
1497 static int kvm_handle_hva_range(struct kvm *kvm,
1498 unsigned long start,
1501 int (*handler)(struct kvm *kvm,
1502 unsigned long *rmapp,
1503 struct kvm_memory_slot *slot,
1506 unsigned long data))
1508 struct kvm_memslots *slots;
1509 struct kvm_memory_slot *memslot;
1510 struct slot_rmap_walk_iterator iterator;
1514 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
1515 slots = __kvm_memslots(kvm, i);
1516 kvm_for_each_memslot(memslot, slots) {
1517 unsigned long hva_start, hva_end;
1518 gfn_t gfn_start, gfn_end;
1520 hva_start = max(start, memslot->userspace_addr);
1521 hva_end = min(end, memslot->userspace_addr +
1522 (memslot->npages << PAGE_SHIFT));
1523 if (hva_start >= hva_end)
1526 * {gfn(page) | page intersects with [hva_start, hva_end)} =
1527 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
1529 gfn_start = hva_to_gfn_memslot(hva_start, memslot);
1530 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
1532 for_each_slot_rmap_range(memslot, PT_PAGE_TABLE_LEVEL,
1533 PT_MAX_HUGEPAGE_LEVEL,
1534 gfn_start, gfn_end - 1,
1536 ret |= handler(kvm, iterator.rmap, memslot,
1537 iterator.gfn, iterator.level, data);
1544 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
1546 int (*handler)(struct kvm *kvm, unsigned long *rmapp,
1547 struct kvm_memory_slot *slot,
1548 gfn_t gfn, int level,
1549 unsigned long data))
1551 return kvm_handle_hva_range(kvm, hva, hva + 1, data, handler);
1554 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1556 return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
1559 int kvm_unmap_hva_range(struct kvm *kvm, unsigned long start, unsigned long end)
1561 return kvm_handle_hva_range(kvm, start, end, 0, kvm_unmap_rmapp);
1564 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1566 kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
1569 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1570 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1574 struct rmap_iterator uninitialized_var(iter);
1577 BUG_ON(!shadow_accessed_mask);
1579 for_each_rmap_spte(rmapp, &iter, sptep)
1580 if (*sptep & shadow_accessed_mask) {
1582 clear_bit((ffs(shadow_accessed_mask) - 1),
1583 (unsigned long *)sptep);
1586 trace_kvm_age_page(gfn, level, slot, young);
1590 static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1591 struct kvm_memory_slot *slot, gfn_t gfn,
1592 int level, unsigned long data)
1595 struct rmap_iterator iter;
1599 * If there's no access bit in the secondary pte set by the
1600 * hardware it's up to gup-fast/gup to set the access bit in
1601 * the primary pte or in the page structure.
1603 if (!shadow_accessed_mask)
1606 for_each_rmap_spte(rmapp, &iter, sptep)
1607 if (*sptep & shadow_accessed_mask) {
1615 #define RMAP_RECYCLE_THRESHOLD 1000
1617 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
1619 unsigned long *rmapp;
1620 struct kvm_mmu_page *sp;
1622 sp = page_header(__pa(spte));
1624 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp);
1626 kvm_unmap_rmapp(vcpu->kvm, rmapp, NULL, gfn, sp->role.level, 0);
1627 kvm_flush_remote_tlbs(vcpu->kvm);
1630 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1633 * In case of absence of EPT Access and Dirty Bits supports,
1634 * emulate the accessed bit for EPT, by checking if this page has
1635 * an EPT mapping, and clearing it if it does. On the next access,
1636 * a new EPT mapping will be established.
1637 * This has some overhead, but not as much as the cost of swapping
1638 * out actively used pages or breaking up actively used hugepages.
1640 if (!shadow_accessed_mask) {
1642 * We are holding the kvm->mmu_lock, and we are blowing up
1643 * shadow PTEs. MMU notifier consumers need to be kept at bay.
1644 * This is correct as long as we don't decouple the mmu_lock
1645 * protected regions (like invalidate_range_start|end does).
1647 kvm->mmu_notifier_seq++;
1648 return kvm_handle_hva_range(kvm, start, end, 0,
1652 return kvm_handle_hva_range(kvm, start, end, 0, kvm_age_rmapp);
1655 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1657 return kvm_handle_hva(kvm, hva, 0, kvm_test_age_rmapp);
1661 static int is_empty_shadow_page(u64 *spt)
1666 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
1667 if (is_shadow_present_pte(*pos)) {
1668 printk(KERN_ERR "%s: %p %llx\n", __func__,
1677 * This value is the sum of all of the kvm instances's
1678 * kvm->arch.n_used_mmu_pages values. We need a global,
1679 * aggregate version in order to make the slab shrinker
1682 static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, int nr)
1684 kvm->arch.n_used_mmu_pages += nr;
1685 percpu_counter_add(&kvm_total_used_mmu_pages, nr);
1688 static void kvm_mmu_free_page(struct kvm_mmu_page *sp)
1690 MMU_WARN_ON(!is_empty_shadow_page(sp->spt));
1691 hlist_del(&sp->hash_link);
1692 list_del(&sp->link);
1693 free_page((unsigned long)sp->spt);
1694 if (!sp->role.direct)
1695 free_page((unsigned long)sp->gfns);
1696 kmem_cache_free(mmu_page_header_cache, sp);
1699 static unsigned kvm_page_table_hashfn(gfn_t gfn)
1701 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
1704 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
1705 struct kvm_mmu_page *sp, u64 *parent_pte)
1710 pte_list_add(vcpu, parent_pte, &sp->parent_ptes);
1713 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
1716 pte_list_remove(parent_pte, &sp->parent_ptes);
1719 static void drop_parent_pte(struct kvm_mmu_page *sp,
1722 mmu_page_remove_parent_pte(sp, parent_pte);
1723 mmu_spte_clear_no_track(parent_pte);
1726 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
1727 u64 *parent_pte, int direct)
1729 struct kvm_mmu_page *sp;
1731 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache);
1732 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache);
1734 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache);
1735 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
1738 * The active_mmu_pages list is the FIFO list, do not move the
1739 * page until it is zapped. kvm_zap_obsolete_pages depends on
1740 * this feature. See the comments in kvm_zap_obsolete_pages().
1742 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
1743 sp->parent_ptes = 0;
1744 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1745 kvm_mod_used_mmu_pages(vcpu->kvm, +1);
1749 static void mark_unsync(u64 *spte);
1750 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1752 pte_list_walk(&sp->parent_ptes, mark_unsync);
1755 static void mark_unsync(u64 *spte)
1757 struct kvm_mmu_page *sp;
1760 sp = page_header(__pa(spte));
1761 index = spte - sp->spt;
1762 if (__test_and_set_bit(index, sp->unsync_child_bitmap))
1764 if (sp->unsync_children++)
1766 kvm_mmu_mark_parents_unsync(sp);
1769 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1770 struct kvm_mmu_page *sp)
1775 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1779 static void nonpaging_update_pte(struct kvm_vcpu *vcpu,
1780 struct kvm_mmu_page *sp, u64 *spte,
1786 #define KVM_PAGE_ARRAY_NR 16
1788 struct kvm_mmu_pages {
1789 struct mmu_page_and_offset {
1790 struct kvm_mmu_page *sp;
1792 } page[KVM_PAGE_ARRAY_NR];
1796 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1802 for (i=0; i < pvec->nr; i++)
1803 if (pvec->page[i].sp == sp)
1806 pvec->page[pvec->nr].sp = sp;
1807 pvec->page[pvec->nr].idx = idx;
1809 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1812 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1813 struct kvm_mmu_pages *pvec)
1815 int i, ret, nr_unsync_leaf = 0;
1817 for_each_set_bit(i, sp->unsync_child_bitmap, 512) {
1818 struct kvm_mmu_page *child;
1819 u64 ent = sp->spt[i];
1821 if (!is_shadow_present_pte(ent) || is_large_pte(ent))
1822 goto clear_child_bitmap;
1824 child = page_header(ent & PT64_BASE_ADDR_MASK);
1826 if (child->unsync_children) {
1827 if (mmu_pages_add(pvec, child, i))
1830 ret = __mmu_unsync_walk(child, pvec);
1832 goto clear_child_bitmap;
1834 nr_unsync_leaf += ret;
1837 } else if (child->unsync) {
1839 if (mmu_pages_add(pvec, child, i))
1842 goto clear_child_bitmap;
1847 __clear_bit(i, sp->unsync_child_bitmap);
1848 sp->unsync_children--;
1849 WARN_ON((int)sp->unsync_children < 0);
1853 return nr_unsync_leaf;
1856 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1857 struct kvm_mmu_pages *pvec)
1859 if (!sp->unsync_children)
1862 mmu_pages_add(pvec, sp, 0);
1863 return __mmu_unsync_walk(sp, pvec);
1866 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1868 WARN_ON(!sp->unsync);
1869 trace_kvm_mmu_sync_page(sp);
1871 --kvm->stat.mmu_unsync;
1874 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1875 struct list_head *invalid_list);
1876 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1877 struct list_head *invalid_list);
1880 * NOTE: we should pay more attention on the zapped-obsolete page
1881 * (is_obsolete_sp(sp) && sp->role.invalid) when you do hash list walk
1882 * since it has been deleted from active_mmu_pages but still can be found
1885 * for_each_gfn_indirect_valid_sp has skipped that kind of page and
1886 * kvm_mmu_get_page(), the only user of for_each_gfn_sp(), has skipped
1887 * all the obsolete pages.
1889 #define for_each_gfn_sp(_kvm, _sp, _gfn) \
1890 hlist_for_each_entry(_sp, \
1891 &(_kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(_gfn)], hash_link) \
1892 if ((_sp)->gfn != (_gfn)) {} else
1894 #define for_each_gfn_indirect_valid_sp(_kvm, _sp, _gfn) \
1895 for_each_gfn_sp(_kvm, _sp, _gfn) \
1896 if ((_sp)->role.direct || (_sp)->role.invalid) {} else
1898 /* @sp->gfn should be write-protected at the call site */
1899 static int __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1900 struct list_head *invalid_list, bool clear_unsync)
1902 if (sp->role.cr4_pae != !!is_pae(vcpu)) {
1903 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1908 kvm_unlink_unsync_page(vcpu->kvm, sp);
1910 if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1911 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1915 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
1919 static int kvm_sync_page_transient(struct kvm_vcpu *vcpu,
1920 struct kvm_mmu_page *sp)
1922 LIST_HEAD(invalid_list);
1925 ret = __kvm_sync_page(vcpu, sp, &invalid_list, false);
1927 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1932 #ifdef CONFIG_KVM_MMU_AUDIT
1933 #include "mmu_audit.c"
1935 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, int point) { }
1936 static void mmu_audit_disable(void) { }
1939 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1940 struct list_head *invalid_list)
1942 return __kvm_sync_page(vcpu, sp, invalid_list, true);
1945 /* @gfn should be write-protected at the call site */
1946 static void kvm_sync_pages(struct kvm_vcpu *vcpu, gfn_t gfn)
1948 struct kvm_mmu_page *s;
1949 LIST_HEAD(invalid_list);
1952 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn) {
1956 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1957 kvm_unlink_unsync_page(vcpu->kvm, s);
1958 if ((s->role.cr4_pae != !!is_pae(vcpu)) ||
1959 (vcpu->arch.mmu.sync_page(vcpu, s))) {
1960 kvm_mmu_prepare_zap_page(vcpu->kvm, s, &invalid_list);
1966 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1968 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
1971 struct mmu_page_path {
1972 struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1973 unsigned int idx[PT64_ROOT_LEVEL-1];
1976 #define for_each_sp(pvec, sp, parents, i) \
1977 for (i = mmu_pages_next(&pvec, &parents, -1), \
1978 sp = pvec.page[i].sp; \
1979 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1980 i = mmu_pages_next(&pvec, &parents, i))
1982 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1983 struct mmu_page_path *parents,
1988 for (n = i+1; n < pvec->nr; n++) {
1989 struct kvm_mmu_page *sp = pvec->page[n].sp;
1991 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1992 parents->idx[0] = pvec->page[n].idx;
1996 parents->parent[sp->role.level-2] = sp;
1997 parents->idx[sp->role.level-1] = pvec->page[n].idx;
2003 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
2005 struct kvm_mmu_page *sp;
2006 unsigned int level = 0;
2009 unsigned int idx = parents->idx[level];
2011 sp = parents->parent[level];
2015 --sp->unsync_children;
2016 WARN_ON((int)sp->unsync_children < 0);
2017 __clear_bit(idx, sp->unsync_child_bitmap);
2019 } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
2022 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
2023 struct mmu_page_path *parents,
2024 struct kvm_mmu_pages *pvec)
2026 parents->parent[parent->role.level-1] = NULL;
2030 static void mmu_sync_children(struct kvm_vcpu *vcpu,
2031 struct kvm_mmu_page *parent)
2034 struct kvm_mmu_page *sp;
2035 struct mmu_page_path parents;
2036 struct kvm_mmu_pages pages;
2037 LIST_HEAD(invalid_list);
2039 kvm_mmu_pages_init(parent, &parents, &pages);
2040 while (mmu_unsync_walk(parent, &pages)) {
2041 bool protected = false;
2043 for_each_sp(pages, sp, parents, i)
2044 protected |= rmap_write_protect(vcpu, sp->gfn);
2047 kvm_flush_remote_tlbs(vcpu->kvm);
2049 for_each_sp(pages, sp, parents, i) {
2050 kvm_sync_page(vcpu, sp, &invalid_list);
2051 mmu_pages_clear_parents(&parents);
2053 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2054 cond_resched_lock(&vcpu->kvm->mmu_lock);
2055 kvm_mmu_pages_init(parent, &parents, &pages);
2059 static void init_shadow_page_table(struct kvm_mmu_page *sp)
2063 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2067 static void __clear_sp_write_flooding_count(struct kvm_mmu_page *sp)
2069 sp->write_flooding_count = 0;
2072 static void clear_sp_write_flooding_count(u64 *spte)
2074 struct kvm_mmu_page *sp = page_header(__pa(spte));
2076 __clear_sp_write_flooding_count(sp);
2079 static bool is_obsolete_sp(struct kvm *kvm, struct kvm_mmu_page *sp)
2081 return unlikely(sp->mmu_valid_gen != kvm->arch.mmu_valid_gen);
2084 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
2092 union kvm_mmu_page_role role;
2094 struct kvm_mmu_page *sp;
2095 bool need_sync = false;
2097 role = vcpu->arch.mmu.base_role;
2099 role.direct = direct;
2102 role.access = access;
2103 if (!vcpu->arch.mmu.direct_map
2104 && vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
2105 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
2106 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
2107 role.quadrant = quadrant;
2109 for_each_gfn_sp(vcpu->kvm, sp, gfn) {
2110 if (is_obsolete_sp(vcpu->kvm, sp))
2113 if (!need_sync && sp->unsync)
2116 if (sp->role.word != role.word)
2119 if (sp->unsync && kvm_sync_page_transient(vcpu, sp))
2122 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
2123 if (sp->unsync_children) {
2124 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
2125 kvm_mmu_mark_parents_unsync(sp);
2126 } else if (sp->unsync)
2127 kvm_mmu_mark_parents_unsync(sp);
2129 __clear_sp_write_flooding_count(sp);
2130 trace_kvm_mmu_get_page(sp, false);
2133 ++vcpu->kvm->stat.mmu_cache_miss;
2134 sp = kvm_mmu_alloc_page(vcpu, parent_pte, direct);
2139 hlist_add_head(&sp->hash_link,
2140 &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
2142 if (rmap_write_protect(vcpu, gfn))
2143 kvm_flush_remote_tlbs(vcpu->kvm);
2144 if (level > PT_PAGE_TABLE_LEVEL && need_sync)
2145 kvm_sync_pages(vcpu, gfn);
2147 account_shadowed(vcpu->kvm, sp);
2149 sp->mmu_valid_gen = vcpu->kvm->arch.mmu_valid_gen;
2150 init_shadow_page_table(sp);
2151 trace_kvm_mmu_get_page(sp, true);
2155 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
2156 struct kvm_vcpu *vcpu, u64 addr)
2158 iterator->addr = addr;
2159 iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
2160 iterator->level = vcpu->arch.mmu.shadow_root_level;
2162 if (iterator->level == PT64_ROOT_LEVEL &&
2163 vcpu->arch.mmu.root_level < PT64_ROOT_LEVEL &&
2164 !vcpu->arch.mmu.direct_map)
2167 if (iterator->level == PT32E_ROOT_LEVEL) {
2168 iterator->shadow_addr
2169 = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
2170 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
2172 if (!iterator->shadow_addr)
2173 iterator->level = 0;
2177 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
2179 if (iterator->level < PT_PAGE_TABLE_LEVEL)
2182 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
2183 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
2187 static void __shadow_walk_next(struct kvm_shadow_walk_iterator *iterator,
2190 if (is_last_spte(spte, iterator->level)) {
2191 iterator->level = 0;
2195 iterator->shadow_addr = spte & PT64_BASE_ADDR_MASK;
2199 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
2201 return __shadow_walk_next(iterator, *iterator->sptep);
2204 static void link_shadow_page(u64 *sptep, struct kvm_mmu_page *sp, bool accessed)
2208 BUILD_BUG_ON(VMX_EPT_READABLE_MASK != PT_PRESENT_MASK ||
2209 VMX_EPT_WRITABLE_MASK != PT_WRITABLE_MASK);
2211 spte = __pa(sp->spt) | PT_PRESENT_MASK | PT_WRITABLE_MASK |
2212 shadow_user_mask | shadow_x_mask;
2215 spte |= shadow_accessed_mask;
2217 mmu_spte_set(sptep, spte);
2220 static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2221 unsigned direct_access)
2223 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
2224 struct kvm_mmu_page *child;
2227 * For the direct sp, if the guest pte's dirty bit
2228 * changed form clean to dirty, it will corrupt the
2229 * sp's access: allow writable in the read-only sp,
2230 * so we should update the spte at this point to get
2231 * a new sp with the correct access.
2233 child = page_header(*sptep & PT64_BASE_ADDR_MASK);
2234 if (child->role.access == direct_access)
2237 drop_parent_pte(child, sptep);
2238 kvm_flush_remote_tlbs(vcpu->kvm);
2242 static bool mmu_page_zap_pte(struct kvm *kvm, struct kvm_mmu_page *sp,
2246 struct kvm_mmu_page *child;
2249 if (is_shadow_present_pte(pte)) {
2250 if (is_last_spte(pte, sp->role.level)) {
2251 drop_spte(kvm, spte);
2252 if (is_large_pte(pte))
2255 child = page_header(pte & PT64_BASE_ADDR_MASK);
2256 drop_parent_pte(child, spte);
2261 if (is_mmio_spte(pte))
2262 mmu_spte_clear_no_track(spte);
2267 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
2268 struct kvm_mmu_page *sp)
2272 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2273 mmu_page_zap_pte(kvm, sp, sp->spt + i);
2276 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
2278 mmu_page_remove_parent_pte(sp, parent_pte);
2281 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
2284 struct rmap_iterator iter;
2286 while ((sptep = rmap_get_first(sp->parent_ptes, &iter)))
2287 drop_parent_pte(sp, sptep);
2290 static int mmu_zap_unsync_children(struct kvm *kvm,
2291 struct kvm_mmu_page *parent,
2292 struct list_head *invalid_list)
2295 struct mmu_page_path parents;
2296 struct kvm_mmu_pages pages;
2298 if (parent->role.level == PT_PAGE_TABLE_LEVEL)
2301 kvm_mmu_pages_init(parent, &parents, &pages);
2302 while (mmu_unsync_walk(parent, &pages)) {
2303 struct kvm_mmu_page *sp;
2305 for_each_sp(pages, sp, parents, i) {
2306 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
2307 mmu_pages_clear_parents(&parents);
2310 kvm_mmu_pages_init(parent, &parents, &pages);
2316 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
2317 struct list_head *invalid_list)
2321 trace_kvm_mmu_prepare_zap_page(sp);
2322 ++kvm->stat.mmu_shadow_zapped;
2323 ret = mmu_zap_unsync_children(kvm, sp, invalid_list);
2324 kvm_mmu_page_unlink_children(kvm, sp);
2325 kvm_mmu_unlink_parents(kvm, sp);
2327 if (!sp->role.invalid && !sp->role.direct)
2328 unaccount_shadowed(kvm, sp);
2331 kvm_unlink_unsync_page(kvm, sp);
2332 if (!sp->root_count) {
2335 list_move(&sp->link, invalid_list);
2336 kvm_mod_used_mmu_pages(kvm, -1);
2338 list_move(&sp->link, &kvm->arch.active_mmu_pages);
2341 * The obsolete pages can not be used on any vcpus.
2342 * See the comments in kvm_mmu_invalidate_zap_all_pages().
2344 if (!sp->role.invalid && !is_obsolete_sp(kvm, sp))
2345 kvm_reload_remote_mmus(kvm);
2348 sp->role.invalid = 1;
2352 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
2353 struct list_head *invalid_list)
2355 struct kvm_mmu_page *sp, *nsp;
2357 if (list_empty(invalid_list))
2361 * wmb: make sure everyone sees our modifications to the page tables
2362 * rmb: make sure we see changes to vcpu->mode
2367 * Wait for all vcpus to exit guest mode and/or lockless shadow
2370 kvm_flush_remote_tlbs(kvm);
2372 list_for_each_entry_safe(sp, nsp, invalid_list, link) {
2373 WARN_ON(!sp->role.invalid || sp->root_count);
2374 kvm_mmu_free_page(sp);
2378 static bool prepare_zap_oldest_mmu_page(struct kvm *kvm,
2379 struct list_head *invalid_list)
2381 struct kvm_mmu_page *sp;
2383 if (list_empty(&kvm->arch.active_mmu_pages))
2386 sp = list_entry(kvm->arch.active_mmu_pages.prev,
2387 struct kvm_mmu_page, link);
2388 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
2394 * Changing the number of mmu pages allocated to the vm
2395 * Note: if goal_nr_mmu_pages is too small, you will get dead lock
2397 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int goal_nr_mmu_pages)
2399 LIST_HEAD(invalid_list);
2401 spin_lock(&kvm->mmu_lock);
2403 if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) {
2404 /* Need to free some mmu pages to achieve the goal. */
2405 while (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages)
2406 if (!prepare_zap_oldest_mmu_page(kvm, &invalid_list))
2409 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2410 goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages;
2413 kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages;
2415 spin_unlock(&kvm->mmu_lock);
2418 int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
2420 struct kvm_mmu_page *sp;
2421 LIST_HEAD(invalid_list);
2424 pgprintk("%s: looking for gfn %llx\n", __func__, gfn);
2426 spin_lock(&kvm->mmu_lock);
2427 for_each_gfn_indirect_valid_sp(kvm, sp, gfn) {
2428 pgprintk("%s: gfn %llx role %x\n", __func__, gfn,
2431 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
2433 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2434 spin_unlock(&kvm->mmu_lock);
2438 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page);
2440 static void __kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
2442 trace_kvm_mmu_unsync_page(sp);
2443 ++vcpu->kvm->stat.mmu_unsync;
2446 kvm_mmu_mark_parents_unsync(sp);
2449 static void kvm_unsync_pages(struct kvm_vcpu *vcpu, gfn_t gfn)
2451 struct kvm_mmu_page *s;
2453 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn) {
2456 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
2457 __kvm_unsync_page(vcpu, s);
2461 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
2464 struct kvm_mmu_page *s;
2465 bool need_unsync = false;
2467 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn) {
2471 if (s->role.level != PT_PAGE_TABLE_LEVEL)
2478 kvm_unsync_pages(vcpu, gfn);
2482 static bool kvm_is_mmio_pfn(pfn_t pfn)
2485 return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn));
2490 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2491 unsigned pte_access, int level,
2492 gfn_t gfn, pfn_t pfn, bool speculative,
2493 bool can_unsync, bool host_writable)
2498 if (set_mmio_spte(vcpu, sptep, gfn, pfn, pte_access))
2501 spte = PT_PRESENT_MASK;
2503 spte |= shadow_accessed_mask;
2505 if (pte_access & ACC_EXEC_MASK)
2506 spte |= shadow_x_mask;
2508 spte |= shadow_nx_mask;
2510 if (pte_access & ACC_USER_MASK)
2511 spte |= shadow_user_mask;
2513 if (level > PT_PAGE_TABLE_LEVEL)
2514 spte |= PT_PAGE_SIZE_MASK;
2516 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
2517 kvm_is_mmio_pfn(pfn));
2520 spte |= SPTE_HOST_WRITEABLE;
2522 pte_access &= ~ACC_WRITE_MASK;
2524 spte |= (u64)pfn << PAGE_SHIFT;
2526 if (pte_access & ACC_WRITE_MASK) {
2529 * Other vcpu creates new sp in the window between
2530 * mapping_level() and acquiring mmu-lock. We can
2531 * allow guest to retry the access, the mapping can
2532 * be fixed if guest refault.
2534 if (level > PT_PAGE_TABLE_LEVEL &&
2535 has_wrprotected_page(vcpu, gfn, level))
2538 spte |= PT_WRITABLE_MASK | SPTE_MMU_WRITEABLE;
2541 * Optimization: for pte sync, if spte was writable the hash
2542 * lookup is unnecessary (and expensive). Write protection
2543 * is responsibility of mmu_get_page / kvm_sync_page.
2544 * Same reasoning can be applied to dirty page accounting.
2546 if (!can_unsync && is_writable_pte(*sptep))
2549 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
2550 pgprintk("%s: found shadow page for %llx, marking ro\n",
2553 pte_access &= ~ACC_WRITE_MASK;
2554 spte &= ~(PT_WRITABLE_MASK | SPTE_MMU_WRITEABLE);
2558 if (pte_access & ACC_WRITE_MASK) {
2559 kvm_vcpu_mark_page_dirty(vcpu, gfn);
2560 spte |= shadow_dirty_mask;
2564 if (mmu_spte_update(sptep, spte))
2565 kvm_flush_remote_tlbs(vcpu->kvm);
2570 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2571 unsigned pte_access, int write_fault, int *emulate,
2572 int level, gfn_t gfn, pfn_t pfn, bool speculative,
2575 int was_rmapped = 0;
2578 pgprintk("%s: spte %llx write_fault %d gfn %llx\n", __func__,
2579 *sptep, write_fault, gfn);
2581 if (is_rmap_spte(*sptep)) {
2583 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
2584 * the parent of the now unreachable PTE.
2586 if (level > PT_PAGE_TABLE_LEVEL &&
2587 !is_large_pte(*sptep)) {
2588 struct kvm_mmu_page *child;
2591 child = page_header(pte & PT64_BASE_ADDR_MASK);
2592 drop_parent_pte(child, sptep);
2593 kvm_flush_remote_tlbs(vcpu->kvm);
2594 } else if (pfn != spte_to_pfn(*sptep)) {
2595 pgprintk("hfn old %llx new %llx\n",
2596 spte_to_pfn(*sptep), pfn);
2597 drop_spte(vcpu->kvm, sptep);
2598 kvm_flush_remote_tlbs(vcpu->kvm);
2603 if (set_spte(vcpu, sptep, pte_access, level, gfn, pfn, speculative,
2604 true, host_writable)) {
2607 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
2610 if (unlikely(is_mmio_spte(*sptep) && emulate))
2613 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
2614 pgprintk("instantiating %s PTE (%s) at %llx (%llx) addr %p\n",
2615 is_large_pte(*sptep)? "2MB" : "4kB",
2616 *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
2618 if (!was_rmapped && is_large_pte(*sptep))
2619 ++vcpu->kvm->stat.lpages;
2621 if (is_shadow_present_pte(*sptep)) {
2623 rmap_count = rmap_add(vcpu, sptep, gfn);
2624 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
2625 rmap_recycle(vcpu, sptep, gfn);
2629 kvm_release_pfn_clean(pfn);
2632 static pfn_t pte_prefetch_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn,
2635 struct kvm_memory_slot *slot;
2637 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, no_dirty_log);
2639 return KVM_PFN_ERR_FAULT;
2641 return gfn_to_pfn_memslot_atomic(slot, gfn);
2644 static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu,
2645 struct kvm_mmu_page *sp,
2646 u64 *start, u64 *end)
2648 struct page *pages[PTE_PREFETCH_NUM];
2649 struct kvm_memory_slot *slot;
2650 unsigned access = sp->role.access;
2654 gfn = kvm_mmu_page_get_gfn(sp, start - sp->spt);
2655 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, access & ACC_WRITE_MASK);
2659 ret = gfn_to_page_many_atomic(slot, gfn, pages, end - start);
2663 for (i = 0; i < ret; i++, gfn++, start++)
2664 mmu_set_spte(vcpu, start, access, 0, NULL,
2665 sp->role.level, gfn, page_to_pfn(pages[i]),
2671 static void __direct_pte_prefetch(struct kvm_vcpu *vcpu,
2672 struct kvm_mmu_page *sp, u64 *sptep)
2674 u64 *spte, *start = NULL;
2677 WARN_ON(!sp->role.direct);
2679 i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
2682 for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
2683 if (is_shadow_present_pte(*spte) || spte == sptep) {
2686 if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0)
2694 static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
2696 struct kvm_mmu_page *sp;
2699 * Since it's no accessed bit on EPT, it's no way to
2700 * distinguish between actually accessed translations
2701 * and prefetched, so disable pte prefetch if EPT is
2704 if (!shadow_accessed_mask)
2707 sp = page_header(__pa(sptep));
2708 if (sp->role.level > PT_PAGE_TABLE_LEVEL)
2711 __direct_pte_prefetch(vcpu, sp, sptep);
2714 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
2715 int map_writable, int level, gfn_t gfn, pfn_t pfn,
2718 struct kvm_shadow_walk_iterator iterator;
2719 struct kvm_mmu_page *sp;
2723 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2726 for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
2727 if (iterator.level == level) {
2728 mmu_set_spte(vcpu, iterator.sptep, ACC_ALL,
2729 write, &emulate, level, gfn, pfn,
2730 prefault, map_writable);
2731 direct_pte_prefetch(vcpu, iterator.sptep);
2732 ++vcpu->stat.pf_fixed;
2736 drop_large_spte(vcpu, iterator.sptep);
2737 if (!is_shadow_present_pte(*iterator.sptep)) {
2738 u64 base_addr = iterator.addr;
2740 base_addr &= PT64_LVL_ADDR_MASK(iterator.level);
2741 pseudo_gfn = base_addr >> PAGE_SHIFT;
2742 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
2744 1, ACC_ALL, iterator.sptep);
2746 link_shadow_page(iterator.sptep, sp, true);
2752 static void kvm_send_hwpoison_signal(unsigned long address, struct task_struct *tsk)
2756 info.si_signo = SIGBUS;
2758 info.si_code = BUS_MCEERR_AR;
2759 info.si_addr = (void __user *)address;
2760 info.si_addr_lsb = PAGE_SHIFT;
2762 send_sig_info(SIGBUS, &info, tsk);
2765 static int kvm_handle_bad_page(struct kvm_vcpu *vcpu, gfn_t gfn, pfn_t pfn)
2768 * Do not cache the mmio info caused by writing the readonly gfn
2769 * into the spte otherwise read access on readonly gfn also can
2770 * caused mmio page fault and treat it as mmio access.
2771 * Return 1 to tell kvm to emulate it.
2773 if (pfn == KVM_PFN_ERR_RO_FAULT)
2776 if (pfn == KVM_PFN_ERR_HWPOISON) {
2777 kvm_send_hwpoison_signal(kvm_vcpu_gfn_to_hva(vcpu, gfn), current);
2784 static void transparent_hugepage_adjust(struct kvm_vcpu *vcpu,
2785 gfn_t *gfnp, pfn_t *pfnp, int *levelp)
2789 int level = *levelp;
2792 * Check if it's a transparent hugepage. If this would be an
2793 * hugetlbfs page, level wouldn't be set to
2794 * PT_PAGE_TABLE_LEVEL and there would be no adjustment done
2797 if (!is_error_noslot_pfn(pfn) && !kvm_is_reserved_pfn(pfn) &&
2798 level == PT_PAGE_TABLE_LEVEL &&
2799 PageTransCompound(pfn_to_page(pfn)) &&
2800 !has_wrprotected_page(vcpu, gfn, PT_DIRECTORY_LEVEL)) {
2803 * mmu_notifier_retry was successful and we hold the
2804 * mmu_lock here, so the pmd can't become splitting
2805 * from under us, and in turn
2806 * __split_huge_page_refcount() can't run from under
2807 * us and we can safely transfer the refcount from
2808 * PG_tail to PG_head as we switch the pfn to tail to
2811 *levelp = level = PT_DIRECTORY_LEVEL;
2812 mask = KVM_PAGES_PER_HPAGE(level) - 1;
2813 VM_BUG_ON((gfn & mask) != (pfn & mask));
2817 kvm_release_pfn_clean(pfn);
2825 static bool handle_abnormal_pfn(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn,
2826 pfn_t pfn, unsigned access, int *ret_val)
2830 /* The pfn is invalid, report the error! */
2831 if (unlikely(is_error_pfn(pfn))) {
2832 *ret_val = kvm_handle_bad_page(vcpu, gfn, pfn);
2836 if (unlikely(is_noslot_pfn(pfn)))
2837 vcpu_cache_mmio_info(vcpu, gva, gfn, access);
2844 static bool page_fault_can_be_fast(u32 error_code)
2847 * Do not fix the mmio spte with invalid generation number which
2848 * need to be updated by slow page fault path.
2850 if (unlikely(error_code & PFERR_RSVD_MASK))
2854 * #PF can be fast only if the shadow page table is present and it
2855 * is caused by write-protect, that means we just need change the
2856 * W bit of the spte which can be done out of mmu-lock.
2858 if (!(error_code & PFERR_PRESENT_MASK) ||
2859 !(error_code & PFERR_WRITE_MASK))
2866 fast_pf_fix_direct_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
2867 u64 *sptep, u64 spte)
2871 WARN_ON(!sp->role.direct);
2874 * The gfn of direct spte is stable since it is calculated
2877 gfn = kvm_mmu_page_get_gfn(sp, sptep - sp->spt);
2880 * Theoretically we could also set dirty bit (and flush TLB) here in
2881 * order to eliminate unnecessary PML logging. See comments in
2882 * set_spte. But fast_page_fault is very unlikely to happen with PML
2883 * enabled, so we do not do this. This might result in the same GPA
2884 * to be logged in PML buffer again when the write really happens, and
2885 * eventually to be called by mark_page_dirty twice. But it's also no
2886 * harm. This also avoids the TLB flush needed after setting dirty bit
2887 * so non-PML cases won't be impacted.
2889 * Compare with set_spte where instead shadow_dirty_mask is set.
2891 if (cmpxchg64(sptep, spte, spte | PT_WRITABLE_MASK) == spte)
2892 kvm_vcpu_mark_page_dirty(vcpu, gfn);
2899 * - true: let the vcpu to access on the same address again.
2900 * - false: let the real page fault path to fix it.
2902 static bool fast_page_fault(struct kvm_vcpu *vcpu, gva_t gva, int level,
2905 struct kvm_shadow_walk_iterator iterator;
2906 struct kvm_mmu_page *sp;
2910 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2913 if (!page_fault_can_be_fast(error_code))
2916 walk_shadow_page_lockless_begin(vcpu);
2917 for_each_shadow_entry_lockless(vcpu, gva, iterator, spte)
2918 if (!is_shadow_present_pte(spte) || iterator.level < level)
2922 * If the mapping has been changed, let the vcpu fault on the
2923 * same address again.
2925 if (!is_rmap_spte(spte)) {
2930 sp = page_header(__pa(iterator.sptep));
2931 if (!is_last_spte(spte, sp->role.level))
2935 * Check if it is a spurious fault caused by TLB lazily flushed.
2937 * Need not check the access of upper level table entries since
2938 * they are always ACC_ALL.
2940 if (is_writable_pte(spte)) {
2946 * Currently, to simplify the code, only the spte write-protected
2947 * by dirty-log can be fast fixed.
2949 if (!spte_is_locklessly_modifiable(spte))
2953 * Do not fix write-permission on the large spte since we only dirty
2954 * the first page into the dirty-bitmap in fast_pf_fix_direct_spte()
2955 * that means other pages are missed if its slot is dirty-logged.
2957 * Instead, we let the slow page fault path create a normal spte to
2960 * See the comments in kvm_arch_commit_memory_region().
2962 if (sp->role.level > PT_PAGE_TABLE_LEVEL)
2966 * Currently, fast page fault only works for direct mapping since
2967 * the gfn is not stable for indirect shadow page.
2968 * See Documentation/virtual/kvm/locking.txt to get more detail.
2970 ret = fast_pf_fix_direct_spte(vcpu, sp, iterator.sptep, spte);
2972 trace_fast_page_fault(vcpu, gva, error_code, iterator.sptep,
2974 walk_shadow_page_lockless_end(vcpu);
2979 static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
2980 gva_t gva, pfn_t *pfn, bool write, bool *writable);
2981 static void make_mmu_pages_available(struct kvm_vcpu *vcpu);
2983 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, u32 error_code,
2984 gfn_t gfn, bool prefault)
2990 unsigned long mmu_seq;
2991 bool map_writable, write = error_code & PFERR_WRITE_MASK;
2993 force_pt_level = mapping_level_dirty_bitmap(vcpu, gfn);
2994 if (likely(!force_pt_level)) {
2995 level = mapping_level(vcpu, gfn);
2997 * This path builds a PAE pagetable - so we can map
2998 * 2mb pages at maximum. Therefore check if the level
2999 * is larger than that.
3001 if (level > PT_DIRECTORY_LEVEL)
3002 level = PT_DIRECTORY_LEVEL;
3004 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
3006 level = PT_PAGE_TABLE_LEVEL;
3008 if (fast_page_fault(vcpu, v, level, error_code))
3011 mmu_seq = vcpu->kvm->mmu_notifier_seq;
3014 if (try_async_pf(vcpu, prefault, gfn, v, &pfn, write, &map_writable))
3017 if (handle_abnormal_pfn(vcpu, v, gfn, pfn, ACC_ALL, &r))
3020 spin_lock(&vcpu->kvm->mmu_lock);
3021 if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
3023 make_mmu_pages_available(vcpu);
3024 if (likely(!force_pt_level))
3025 transparent_hugepage_adjust(vcpu, &gfn, &pfn, &level);
3026 r = __direct_map(vcpu, v, write, map_writable, level, gfn, pfn,
3028 spin_unlock(&vcpu->kvm->mmu_lock);
3034 spin_unlock(&vcpu->kvm->mmu_lock);
3035 kvm_release_pfn_clean(pfn);
3040 static void mmu_free_roots(struct kvm_vcpu *vcpu)
3043 struct kvm_mmu_page *sp;
3044 LIST_HEAD(invalid_list);
3046 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3049 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL &&
3050 (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL ||
3051 vcpu->arch.mmu.direct_map)) {
3052 hpa_t root = vcpu->arch.mmu.root_hpa;
3054 spin_lock(&vcpu->kvm->mmu_lock);
3055 sp = page_header(root);
3057 if (!sp->root_count && sp->role.invalid) {
3058 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
3059 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
3061 spin_unlock(&vcpu->kvm->mmu_lock);
3062 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
3066 spin_lock(&vcpu->kvm->mmu_lock);
3067 for (i = 0; i < 4; ++i) {
3068 hpa_t root = vcpu->arch.mmu.pae_root[i];
3071 root &= PT64_BASE_ADDR_MASK;
3072 sp = page_header(root);
3074 if (!sp->root_count && sp->role.invalid)
3075 kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
3078 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
3080 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
3081 spin_unlock(&vcpu->kvm->mmu_lock);
3082 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
3085 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
3089 if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
3090 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
3097 static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
3099 struct kvm_mmu_page *sp;
3102 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3103 spin_lock(&vcpu->kvm->mmu_lock);
3104 make_mmu_pages_available(vcpu);
3105 sp = kvm_mmu_get_page(vcpu, 0, 0, PT64_ROOT_LEVEL,
3108 spin_unlock(&vcpu->kvm->mmu_lock);
3109 vcpu->arch.mmu.root_hpa = __pa(sp->spt);
3110 } else if (vcpu->arch.mmu.shadow_root_level == PT32E_ROOT_LEVEL) {
3111 for (i = 0; i < 4; ++i) {
3112 hpa_t root = vcpu->arch.mmu.pae_root[i];
3114 MMU_WARN_ON(VALID_PAGE(root));
3115 spin_lock(&vcpu->kvm->mmu_lock);
3116 make_mmu_pages_available(vcpu);
3117 sp = kvm_mmu_get_page(vcpu, i << (30 - PAGE_SHIFT),
3119 PT32_ROOT_LEVEL, 1, ACC_ALL,
3121 root = __pa(sp->spt);
3123 spin_unlock(&vcpu->kvm->mmu_lock);
3124 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
3126 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
3133 static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
3135 struct kvm_mmu_page *sp;
3140 root_gfn = vcpu->arch.mmu.get_cr3(vcpu) >> PAGE_SHIFT;
3142 if (mmu_check_root(vcpu, root_gfn))
3146 * Do we shadow a long mode page table? If so we need to
3147 * write-protect the guests page table root.
3149 if (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL) {
3150 hpa_t root = vcpu->arch.mmu.root_hpa;
3152 MMU_WARN_ON(VALID_PAGE(root));
3154 spin_lock(&vcpu->kvm->mmu_lock);
3155 make_mmu_pages_available(vcpu);
3156 sp = kvm_mmu_get_page(vcpu, root_gfn, 0, PT64_ROOT_LEVEL,
3158 root = __pa(sp->spt);
3160 spin_unlock(&vcpu->kvm->mmu_lock);
3161 vcpu->arch.mmu.root_hpa = root;
3166 * We shadow a 32 bit page table. This may be a legacy 2-level
3167 * or a PAE 3-level page table. In either case we need to be aware that
3168 * the shadow page table may be a PAE or a long mode page table.
3170 pm_mask = PT_PRESENT_MASK;
3171 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL)
3172 pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK;
3174 for (i = 0; i < 4; ++i) {
3175 hpa_t root = vcpu->arch.mmu.pae_root[i];
3177 MMU_WARN_ON(VALID_PAGE(root));
3178 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
3179 pdptr = vcpu->arch.mmu.get_pdptr(vcpu, i);
3180 if (!is_present_gpte(pdptr)) {
3181 vcpu->arch.mmu.pae_root[i] = 0;
3184 root_gfn = pdptr >> PAGE_SHIFT;
3185 if (mmu_check_root(vcpu, root_gfn))
3188 spin_lock(&vcpu->kvm->mmu_lock);
3189 make_mmu_pages_available(vcpu);
3190 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
3193 root = __pa(sp->spt);
3195 spin_unlock(&vcpu->kvm->mmu_lock);
3197 vcpu->arch.mmu.pae_root[i] = root | pm_mask;
3199 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
3202 * If we shadow a 32 bit page table with a long mode page
3203 * table we enter this path.
3205 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3206 if (vcpu->arch.mmu.lm_root == NULL) {
3208 * The additional page necessary for this is only
3209 * allocated on demand.
3214 lm_root = (void*)get_zeroed_page(GFP_KERNEL);
3215 if (lm_root == NULL)
3218 lm_root[0] = __pa(vcpu->arch.mmu.pae_root) | pm_mask;
3220 vcpu->arch.mmu.lm_root = lm_root;
3223 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.lm_root);
3229 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
3231 if (vcpu->arch.mmu.direct_map)
3232 return mmu_alloc_direct_roots(vcpu);
3234 return mmu_alloc_shadow_roots(vcpu);
3237 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
3240 struct kvm_mmu_page *sp;
3242 if (vcpu->arch.mmu.direct_map)
3245 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3248 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
3249 kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
3250 if (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL) {
3251 hpa_t root = vcpu->arch.mmu.root_hpa;
3252 sp = page_header(root);
3253 mmu_sync_children(vcpu, sp);
3254 kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
3257 for (i = 0; i < 4; ++i) {
3258 hpa_t root = vcpu->arch.mmu.pae_root[i];
3260 if (root && VALID_PAGE(root)) {
3261 root &= PT64_BASE_ADDR_MASK;
3262 sp = page_header(root);
3263 mmu_sync_children(vcpu, sp);
3266 kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
3269 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
3271 spin_lock(&vcpu->kvm->mmu_lock);
3272 mmu_sync_roots(vcpu);
3273 spin_unlock(&vcpu->kvm->mmu_lock);
3275 EXPORT_SYMBOL_GPL(kvm_mmu_sync_roots);
3277 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
3278 u32 access, struct x86_exception *exception)
3281 exception->error_code = 0;
3285 static gpa_t nonpaging_gva_to_gpa_nested(struct kvm_vcpu *vcpu, gva_t vaddr,
3287 struct x86_exception *exception)
3290 exception->error_code = 0;
3291 return vcpu->arch.nested_mmu.translate_gpa(vcpu, vaddr, access, exception);
3294 static bool quickly_check_mmio_pf(struct kvm_vcpu *vcpu, u64 addr, bool direct)
3297 return vcpu_match_mmio_gpa(vcpu, addr);
3299 return vcpu_match_mmio_gva(vcpu, addr);
3304 * On direct hosts, the last spte is only allows two states
3305 * for mmio page fault:
3306 * - It is the mmio spte
3307 * - It is zapped or it is being zapped.
3309 * This function completely checks the spte when the last spte
3310 * is not the mmio spte.
3312 static bool check_direct_spte_mmio_pf(u64 spte)
3314 return __check_direct_spte_mmio_pf(spte);
3317 static u64 walk_shadow_page_get_mmio_spte(struct kvm_vcpu *vcpu, u64 addr)
3319 struct kvm_shadow_walk_iterator iterator;
3322 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3325 walk_shadow_page_lockless_begin(vcpu);
3326 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte)
3327 if (!is_shadow_present_pte(spte))
3329 walk_shadow_page_lockless_end(vcpu);
3334 int handle_mmio_page_fault_common(struct kvm_vcpu *vcpu, u64 addr, bool direct)
3338 if (quickly_check_mmio_pf(vcpu, addr, direct))
3339 return RET_MMIO_PF_EMULATE;
3341 spte = walk_shadow_page_get_mmio_spte(vcpu, addr);
3343 if (is_mmio_spte(spte)) {
3344 gfn_t gfn = get_mmio_spte_gfn(spte);
3345 unsigned access = get_mmio_spte_access(spte);
3347 if (!check_mmio_spte(vcpu, spte))
3348 return RET_MMIO_PF_INVALID;
3353 trace_handle_mmio_page_fault(addr, gfn, access);
3354 vcpu_cache_mmio_info(vcpu, addr, gfn, access);
3355 return RET_MMIO_PF_EMULATE;
3359 * It's ok if the gva is remapped by other cpus on shadow guest,
3360 * it's a BUG if the gfn is not a mmio page.
3362 if (direct && !check_direct_spte_mmio_pf(spte))
3363 return RET_MMIO_PF_BUG;
3366 * If the page table is zapped by other cpus, let CPU fault again on
3369 return RET_MMIO_PF_RETRY;
3371 EXPORT_SYMBOL_GPL(handle_mmio_page_fault_common);
3373 static int handle_mmio_page_fault(struct kvm_vcpu *vcpu, u64 addr,
3374 u32 error_code, bool direct)
3378 ret = handle_mmio_page_fault_common(vcpu, addr, direct);
3379 WARN_ON(ret == RET_MMIO_PF_BUG);
3383 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
3384 u32 error_code, bool prefault)
3389 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
3391 if (unlikely(error_code & PFERR_RSVD_MASK)) {
3392 r = handle_mmio_page_fault(vcpu, gva, error_code, true);
3394 if (likely(r != RET_MMIO_PF_INVALID))
3398 r = mmu_topup_memory_caches(vcpu);
3402 MMU_WARN_ON(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3404 gfn = gva >> PAGE_SHIFT;
3406 return nonpaging_map(vcpu, gva & PAGE_MASK,
3407 error_code, gfn, prefault);
3410 static int kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn)
3412 struct kvm_arch_async_pf arch;
3414 arch.token = (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id;
3416 arch.direct_map = vcpu->arch.mmu.direct_map;
3417 arch.cr3 = vcpu->arch.mmu.get_cr3(vcpu);
3419 return kvm_setup_async_pf(vcpu, gva, kvm_vcpu_gfn_to_hva(vcpu, gfn), &arch);
3422 static bool can_do_async_pf(struct kvm_vcpu *vcpu)
3424 if (unlikely(!irqchip_in_kernel(vcpu->kvm) ||
3425 kvm_event_needs_reinjection(vcpu)))
3428 return kvm_x86_ops->interrupt_allowed(vcpu);
3431 static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
3432 gva_t gva, pfn_t *pfn, bool write, bool *writable)
3434 struct kvm_memory_slot *slot;
3437 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
3439 *pfn = __gfn_to_pfn_memslot(slot, gfn, false, &async, write, writable);
3441 return false; /* *pfn has correct page already */
3443 if (!prefault && can_do_async_pf(vcpu)) {
3444 trace_kvm_try_async_get_page(gva, gfn);
3445 if (kvm_find_async_pf_gfn(vcpu, gfn)) {
3446 trace_kvm_async_pf_doublefault(gva, gfn);
3447 kvm_make_request(KVM_REQ_APF_HALT, vcpu);
3449 } else if (kvm_arch_setup_async_pf(vcpu, gva, gfn))
3453 *pfn = __gfn_to_pfn_memslot(slot, gfn, false, NULL, write, writable);
3458 check_hugepage_cache_consistency(struct kvm_vcpu *vcpu, gfn_t gfn, int level)
3460 int page_num = KVM_PAGES_PER_HPAGE(level);
3462 gfn &= ~(page_num - 1);
3464 return kvm_mtrr_check_gfn_range_consistency(vcpu, gfn, page_num);
3467 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa, u32 error_code,
3474 gfn_t gfn = gpa >> PAGE_SHIFT;
3475 unsigned long mmu_seq;
3476 int write = error_code & PFERR_WRITE_MASK;
3479 MMU_WARN_ON(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3481 if (unlikely(error_code & PFERR_RSVD_MASK)) {
3482 r = handle_mmio_page_fault(vcpu, gpa, error_code, true);
3484 if (likely(r != RET_MMIO_PF_INVALID))
3488 r = mmu_topup_memory_caches(vcpu);
3492 if (mapping_level_dirty_bitmap(vcpu, gfn) ||
3493 !check_hugepage_cache_consistency(vcpu, gfn, PT_DIRECTORY_LEVEL))
3498 if (likely(!force_pt_level)) {
3499 level = mapping_level(vcpu, gfn);
3500 if (level > PT_DIRECTORY_LEVEL &&
3501 !check_hugepage_cache_consistency(vcpu, gfn, level))
3502 level = PT_DIRECTORY_LEVEL;
3503 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
3505 level = PT_PAGE_TABLE_LEVEL;
3507 if (fast_page_fault(vcpu, gpa, level, error_code))
3510 mmu_seq = vcpu->kvm->mmu_notifier_seq;
3513 if (try_async_pf(vcpu, prefault, gfn, gpa, &pfn, write, &map_writable))
3516 if (handle_abnormal_pfn(vcpu, 0, gfn, pfn, ACC_ALL, &r))
3519 spin_lock(&vcpu->kvm->mmu_lock);
3520 if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
3522 make_mmu_pages_available(vcpu);
3523 if (likely(!force_pt_level))
3524 transparent_hugepage_adjust(vcpu, &gfn, &pfn, &level);
3525 r = __direct_map(vcpu, gpa, write, map_writable,
3526 level, gfn, pfn, prefault);
3527 spin_unlock(&vcpu->kvm->mmu_lock);
3532 spin_unlock(&vcpu->kvm->mmu_lock);
3533 kvm_release_pfn_clean(pfn);
3537 static void nonpaging_init_context(struct kvm_vcpu *vcpu,
3538 struct kvm_mmu *context)
3540 context->page_fault = nonpaging_page_fault;
3541 context->gva_to_gpa = nonpaging_gva_to_gpa;
3542 context->sync_page = nonpaging_sync_page;
3543 context->invlpg = nonpaging_invlpg;
3544 context->update_pte = nonpaging_update_pte;
3545 context->root_level = 0;
3546 context->shadow_root_level = PT32E_ROOT_LEVEL;
3547 context->root_hpa = INVALID_PAGE;
3548 context->direct_map = true;
3549 context->nx = false;
3552 void kvm_mmu_new_cr3(struct kvm_vcpu *vcpu)
3554 mmu_free_roots(vcpu);
3557 static unsigned long get_cr3(struct kvm_vcpu *vcpu)
3559 return kvm_read_cr3(vcpu);
3562 static void inject_page_fault(struct kvm_vcpu *vcpu,
3563 struct x86_exception *fault)
3565 vcpu->arch.mmu.inject_page_fault(vcpu, fault);
3568 static bool sync_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn,
3569 unsigned access, int *nr_present)
3571 if (unlikely(is_mmio_spte(*sptep))) {
3572 if (gfn != get_mmio_spte_gfn(*sptep)) {
3573 mmu_spte_clear_no_track(sptep);
3578 mark_mmio_spte(vcpu, sptep, gfn, access);
3585 static inline bool is_last_gpte(struct kvm_mmu *mmu, unsigned level, unsigned gpte)
3590 index |= (gpte & PT_PAGE_SIZE_MASK) >> (PT_PAGE_SIZE_SHIFT - 2);
3591 return mmu->last_pte_bitmap & (1 << index);
3594 #define PTTYPE_EPT 18 /* arbitrary */
3595 #define PTTYPE PTTYPE_EPT
3596 #include "paging_tmpl.h"
3600 #include "paging_tmpl.h"
3604 #include "paging_tmpl.h"
3607 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
3608 struct kvm_mmu *context)
3610 int maxphyaddr = cpuid_maxphyaddr(vcpu);
3611 u64 exb_bit_rsvd = 0;
3612 u64 gbpages_bit_rsvd = 0;
3613 u64 nonleaf_bit8_rsvd = 0;
3615 context->bad_mt_xwr = 0;
3618 exb_bit_rsvd = rsvd_bits(63, 63);
3619 if (!guest_cpuid_has_gbpages(vcpu))
3620 gbpages_bit_rsvd = rsvd_bits(7, 7);
3623 * Non-leaf PML4Es and PDPEs reserve bit 8 (which would be the G bit for
3624 * leaf entries) on AMD CPUs only.
3626 if (guest_cpuid_is_amd(vcpu))
3627 nonleaf_bit8_rsvd = rsvd_bits(8, 8);
3629 switch (context->root_level) {
3630 case PT32_ROOT_LEVEL:
3631 /* no rsvd bits for 2 level 4K page table entries */
3632 context->rsvd_bits_mask[0][1] = 0;
3633 context->rsvd_bits_mask[0][0] = 0;
3634 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3636 if (!is_pse(vcpu)) {
3637 context->rsvd_bits_mask[1][1] = 0;
3641 if (is_cpuid_PSE36())
3642 /* 36bits PSE 4MB page */
3643 context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
3645 /* 32 bits PSE 4MB page */
3646 context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
3648 case PT32E_ROOT_LEVEL:
3649 context->rsvd_bits_mask[0][2] =
3650 rsvd_bits(maxphyaddr, 63) |
3651 rsvd_bits(5, 8) | rsvd_bits(1, 2); /* PDPTE */
3652 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
3653 rsvd_bits(maxphyaddr, 62); /* PDE */
3654 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
3655 rsvd_bits(maxphyaddr, 62); /* PTE */
3656 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
3657 rsvd_bits(maxphyaddr, 62) |
3658 rsvd_bits(13, 20); /* large page */
3659 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3661 case PT64_ROOT_LEVEL:
3662 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
3663 nonleaf_bit8_rsvd | rsvd_bits(7, 7) | rsvd_bits(maxphyaddr, 51);
3664 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
3665 nonleaf_bit8_rsvd | gbpages_bit_rsvd | rsvd_bits(maxphyaddr, 51);
3666 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
3667 rsvd_bits(maxphyaddr, 51);
3668 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
3669 rsvd_bits(maxphyaddr, 51);
3670 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
3671 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
3672 gbpages_bit_rsvd | rsvd_bits(maxphyaddr, 51) |
3674 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
3675 rsvd_bits(maxphyaddr, 51) |
3676 rsvd_bits(13, 20); /* large page */
3677 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3682 static void reset_rsvds_bits_mask_ept(struct kvm_vcpu *vcpu,
3683 struct kvm_mmu *context, bool execonly)
3685 int maxphyaddr = cpuid_maxphyaddr(vcpu);
3688 context->rsvd_bits_mask[0][3] =
3689 rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 7);
3690 context->rsvd_bits_mask[0][2] =
3691 rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 6);
3692 context->rsvd_bits_mask[0][1] =
3693 rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 6);
3694 context->rsvd_bits_mask[0][0] = rsvd_bits(maxphyaddr, 51);
3697 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
3698 context->rsvd_bits_mask[1][2] =
3699 rsvd_bits(maxphyaddr, 51) | rsvd_bits(12, 29);
3700 context->rsvd_bits_mask[1][1] =
3701 rsvd_bits(maxphyaddr, 51) | rsvd_bits(12, 20);
3702 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3704 for (pte = 0; pte < 64; pte++) {
3705 int rwx_bits = pte & 7;
3707 if (mt == 0x2 || mt == 0x3 || mt == 0x7 ||
3708 rwx_bits == 0x2 || rwx_bits == 0x6 ||
3709 (rwx_bits == 0x4 && !execonly))
3710 context->bad_mt_xwr |= (1ull << pte);
3714 static void update_permission_bitmask(struct kvm_vcpu *vcpu,
3715 struct kvm_mmu *mmu, bool ept)
3717 unsigned bit, byte, pfec;
3719 bool fault, x, w, u, wf, uf, ff, smapf, cr4_smap, cr4_smep, smap = 0;
3721 cr4_smep = kvm_read_cr4_bits(vcpu, X86_CR4_SMEP);
3722 cr4_smap = kvm_read_cr4_bits(vcpu, X86_CR4_SMAP);
3723 for (byte = 0; byte < ARRAY_SIZE(mmu->permissions); ++byte) {
3726 wf = pfec & PFERR_WRITE_MASK;
3727 uf = pfec & PFERR_USER_MASK;
3728 ff = pfec & PFERR_FETCH_MASK;
3730 * PFERR_RSVD_MASK bit is set in PFEC if the access is not
3731 * subject to SMAP restrictions, and cleared otherwise. The
3732 * bit is only meaningful if the SMAP bit is set in CR4.
3734 smapf = !(pfec & PFERR_RSVD_MASK);
3735 for (bit = 0; bit < 8; ++bit) {
3736 x = bit & ACC_EXEC_MASK;
3737 w = bit & ACC_WRITE_MASK;
3738 u = bit & ACC_USER_MASK;
3741 /* Not really needed: !nx will cause pte.nx to fault */
3743 /* Allow supervisor writes if !cr0.wp */
3744 w |= !is_write_protection(vcpu) && !uf;
3745 /* Disallow supervisor fetches of user code if cr4.smep */
3746 x &= !(cr4_smep && u && !uf);
3749 * SMAP:kernel-mode data accesses from user-mode
3750 * mappings should fault. A fault is considered
3751 * as a SMAP violation if all of the following
3752 * conditions are ture:
3753 * - X86_CR4_SMAP is set in CR4
3754 * - An user page is accessed
3755 * - Page fault in kernel mode
3756 * - if CPL = 3 or X86_EFLAGS_AC is clear
3758 * Here, we cover the first three conditions.
3759 * The fourth is computed dynamically in
3760 * permission_fault() and is in smapf.
3762 * Also, SMAP does not affect instruction
3763 * fetches, add the !ff check here to make it
3766 smap = cr4_smap && u && !uf && !ff;
3768 /* Not really needed: no U/S accesses on ept */
3771 fault = (ff && !x) || (uf && !u) || (wf && !w) ||
3773 map |= fault << bit;
3775 mmu->permissions[byte] = map;
3779 static void update_last_pte_bitmap(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
3782 unsigned level, root_level = mmu->root_level;
3783 const unsigned ps_set_index = 1 << 2; /* bit 2 of index: ps */
3785 if (root_level == PT32E_ROOT_LEVEL)
3787 /* PT_PAGE_TABLE_LEVEL always terminates */
3788 map = 1 | (1 << ps_set_index);
3789 for (level = PT_DIRECTORY_LEVEL; level <= root_level; ++level) {
3790 if (level <= PT_PDPE_LEVEL
3791 && (mmu->root_level >= PT32E_ROOT_LEVEL || is_pse(vcpu)))
3792 map |= 1 << (ps_set_index | (level - 1));
3794 mmu->last_pte_bitmap = map;
3797 static void paging64_init_context_common(struct kvm_vcpu *vcpu,
3798 struct kvm_mmu *context,
3801 context->nx = is_nx(vcpu);
3802 context->root_level = level;
3804 reset_rsvds_bits_mask(vcpu, context);
3805 update_permission_bitmask(vcpu, context, false);
3806 update_last_pte_bitmap(vcpu, context);
3808 MMU_WARN_ON(!is_pae(vcpu));
3809 context->page_fault = paging64_page_fault;
3810 context->gva_to_gpa = paging64_gva_to_gpa;
3811 context->sync_page = paging64_sync_page;
3812 context->invlpg = paging64_invlpg;
3813 context->update_pte = paging64_update_pte;
3814 context->shadow_root_level = level;
3815 context->root_hpa = INVALID_PAGE;
3816 context->direct_map = false;
3819 static void paging64_init_context(struct kvm_vcpu *vcpu,
3820 struct kvm_mmu *context)
3822 paging64_init_context_common(vcpu, context, PT64_ROOT_LEVEL);
3825 static void paging32_init_context(struct kvm_vcpu *vcpu,
3826 struct kvm_mmu *context)
3828 context->nx = false;
3829 context->root_level = PT32_ROOT_LEVEL;
3831 reset_rsvds_bits_mask(vcpu, context);
3832 update_permission_bitmask(vcpu, context, false);
3833 update_last_pte_bitmap(vcpu, context);
3835 context->page_fault = paging32_page_fault;
3836 context->gva_to_gpa = paging32_gva_to_gpa;
3837 context->sync_page = paging32_sync_page;
3838 context->invlpg = paging32_invlpg;
3839 context->update_pte = paging32_update_pte;
3840 context->shadow_root_level = PT32E_ROOT_LEVEL;
3841 context->root_hpa = INVALID_PAGE;
3842 context->direct_map = false;
3845 static void paging32E_init_context(struct kvm_vcpu *vcpu,
3846 struct kvm_mmu *context)
3848 paging64_init_context_common(vcpu, context, PT32E_ROOT_LEVEL);
3851 static void init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
3853 struct kvm_mmu *context = &vcpu->arch.mmu;
3855 context->base_role.word = 0;
3856 context->base_role.smm = is_smm(vcpu);
3857 context->page_fault = tdp_page_fault;
3858 context->sync_page = nonpaging_sync_page;
3859 context->invlpg = nonpaging_invlpg;
3860 context->update_pte = nonpaging_update_pte;
3861 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
3862 context->root_hpa = INVALID_PAGE;
3863 context->direct_map = true;
3864 context->set_cr3 = kvm_x86_ops->set_tdp_cr3;
3865 context->get_cr3 = get_cr3;
3866 context->get_pdptr = kvm_pdptr_read;
3867 context->inject_page_fault = kvm_inject_page_fault;
3869 if (!is_paging(vcpu)) {
3870 context->nx = false;
3871 context->gva_to_gpa = nonpaging_gva_to_gpa;
3872 context->root_level = 0;
3873 } else if (is_long_mode(vcpu)) {
3874 context->nx = is_nx(vcpu);
3875 context->root_level = PT64_ROOT_LEVEL;
3876 reset_rsvds_bits_mask(vcpu, context);
3877 context->gva_to_gpa = paging64_gva_to_gpa;
3878 } else if (is_pae(vcpu)) {
3879 context->nx = is_nx(vcpu);
3880 context->root_level = PT32E_ROOT_LEVEL;
3881 reset_rsvds_bits_mask(vcpu, context);
3882 context->gva_to_gpa = paging64_gva_to_gpa;
3884 context->nx = false;
3885 context->root_level = PT32_ROOT_LEVEL;
3886 reset_rsvds_bits_mask(vcpu, context);
3887 context->gva_to_gpa = paging32_gva_to_gpa;
3890 update_permission_bitmask(vcpu, context, false);
3891 update_last_pte_bitmap(vcpu, context);
3894 void kvm_init_shadow_mmu(struct kvm_vcpu *vcpu)
3896 bool smep = kvm_read_cr4_bits(vcpu, X86_CR4_SMEP);
3897 bool smap = kvm_read_cr4_bits(vcpu, X86_CR4_SMAP);
3898 struct kvm_mmu *context = &vcpu->arch.mmu;
3900 MMU_WARN_ON(VALID_PAGE(context->root_hpa));
3902 if (!is_paging(vcpu))
3903 nonpaging_init_context(vcpu, context);
3904 else if (is_long_mode(vcpu))
3905 paging64_init_context(vcpu, context);
3906 else if (is_pae(vcpu))
3907 paging32E_init_context(vcpu, context);
3909 paging32_init_context(vcpu, context);
3911 context->base_role.nxe = is_nx(vcpu);
3912 context->base_role.cr4_pae = !!is_pae(vcpu);
3913 context->base_role.cr0_wp = is_write_protection(vcpu);
3914 context->base_role.smep_andnot_wp
3915 = smep && !is_write_protection(vcpu);
3916 context->base_role.smap_andnot_wp
3917 = smap && !is_write_protection(vcpu);
3918 context->base_role.smm = is_smm(vcpu);
3920 EXPORT_SYMBOL_GPL(kvm_init_shadow_mmu);
3922 void kvm_init_shadow_ept_mmu(struct kvm_vcpu *vcpu, bool execonly)
3924 struct kvm_mmu *context = &vcpu->arch.mmu;
3926 MMU_WARN_ON(VALID_PAGE(context->root_hpa));
3928 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
3931 context->page_fault = ept_page_fault;
3932 context->gva_to_gpa = ept_gva_to_gpa;
3933 context->sync_page = ept_sync_page;
3934 context->invlpg = ept_invlpg;
3935 context->update_pte = ept_update_pte;
3936 context->root_level = context->shadow_root_level;
3937 context->root_hpa = INVALID_PAGE;
3938 context->direct_map = false;
3940 update_permission_bitmask(vcpu, context, true);
3941 reset_rsvds_bits_mask_ept(vcpu, context, execonly);
3943 EXPORT_SYMBOL_GPL(kvm_init_shadow_ept_mmu);
3945 static void init_kvm_softmmu(struct kvm_vcpu *vcpu)
3947 struct kvm_mmu *context = &vcpu->arch.mmu;
3949 kvm_init_shadow_mmu(vcpu);
3950 context->set_cr3 = kvm_x86_ops->set_cr3;
3951 context->get_cr3 = get_cr3;
3952 context->get_pdptr = kvm_pdptr_read;
3953 context->inject_page_fault = kvm_inject_page_fault;
3956 static void init_kvm_nested_mmu(struct kvm_vcpu *vcpu)
3958 struct kvm_mmu *g_context = &vcpu->arch.nested_mmu;
3960 g_context->get_cr3 = get_cr3;
3961 g_context->get_pdptr = kvm_pdptr_read;
3962 g_context->inject_page_fault = kvm_inject_page_fault;
3965 * Note that arch.mmu.gva_to_gpa translates l2_gva to l1_gpa. The
3966 * translation of l2_gpa to l1_gpa addresses is done using the
3967 * arch.nested_mmu.gva_to_gpa function. Basically the gva_to_gpa
3968 * functions between mmu and nested_mmu are swapped.
3970 if (!is_paging(vcpu)) {
3971 g_context->nx = false;
3972 g_context->root_level = 0;
3973 g_context->gva_to_gpa = nonpaging_gva_to_gpa_nested;
3974 } else if (is_long_mode(vcpu)) {
3975 g_context->nx = is_nx(vcpu);
3976 g_context->root_level = PT64_ROOT_LEVEL;
3977 reset_rsvds_bits_mask(vcpu, g_context);
3978 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
3979 } else if (is_pae(vcpu)) {
3980 g_context->nx = is_nx(vcpu);
3981 g_context->root_level = PT32E_ROOT_LEVEL;
3982 reset_rsvds_bits_mask(vcpu, g_context);
3983 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
3985 g_context->nx = false;
3986 g_context->root_level = PT32_ROOT_LEVEL;
3987 reset_rsvds_bits_mask(vcpu, g_context);
3988 g_context->gva_to_gpa = paging32_gva_to_gpa_nested;
3991 update_permission_bitmask(vcpu, g_context, false);
3992 update_last_pte_bitmap(vcpu, g_context);
3995 static void init_kvm_mmu(struct kvm_vcpu *vcpu)
3997 if (mmu_is_nested(vcpu))
3998 init_kvm_nested_mmu(vcpu);
3999 else if (tdp_enabled)
4000 init_kvm_tdp_mmu(vcpu);
4002 init_kvm_softmmu(vcpu);
4005 void kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
4007 kvm_mmu_unload(vcpu);
4010 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
4012 int kvm_mmu_load(struct kvm_vcpu *vcpu)
4016 r = mmu_topup_memory_caches(vcpu);
4019 r = mmu_alloc_roots(vcpu);
4020 kvm_mmu_sync_roots(vcpu);
4023 /* set_cr3() should ensure TLB has been flushed */
4024 vcpu->arch.mmu.set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
4028 EXPORT_SYMBOL_GPL(kvm_mmu_load);
4030 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
4032 mmu_free_roots(vcpu);
4033 WARN_ON(VALID_PAGE(vcpu->arch.mmu.root_hpa));
4035 EXPORT_SYMBOL_GPL(kvm_mmu_unload);
4037 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
4038 struct kvm_mmu_page *sp, u64 *spte,
4041 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
4042 ++vcpu->kvm->stat.mmu_pde_zapped;
4046 ++vcpu->kvm->stat.mmu_pte_updated;
4047 vcpu->arch.mmu.update_pte(vcpu, sp, spte, new);
4050 static bool need_remote_flush(u64 old, u64 new)
4052 if (!is_shadow_present_pte(old))
4054 if (!is_shadow_present_pte(new))
4056 if ((old ^ new) & PT64_BASE_ADDR_MASK)
4058 old ^= shadow_nx_mask;
4059 new ^= shadow_nx_mask;
4060 return (old & ~new & PT64_PERM_MASK) != 0;
4063 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, bool zap_page,
4064 bool remote_flush, bool local_flush)
4070 kvm_flush_remote_tlbs(vcpu->kvm);
4071 else if (local_flush)
4072 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
4075 static u64 mmu_pte_write_fetch_gpte(struct kvm_vcpu *vcpu, gpa_t *gpa,
4076 const u8 *new, int *bytes)
4082 * Assume that the pte write on a page table of the same type
4083 * as the current vcpu paging mode since we update the sptes only
4084 * when they have the same mode.
4086 if (is_pae(vcpu) && *bytes == 4) {
4087 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
4090 r = kvm_vcpu_read_guest(vcpu, *gpa, &gentry, 8);
4093 new = (const u8 *)&gentry;
4098 gentry = *(const u32 *)new;
4101 gentry = *(const u64 *)new;
4112 * If we're seeing too many writes to a page, it may no longer be a page table,
4113 * or we may be forking, in which case it is better to unmap the page.
4115 static bool detect_write_flooding(struct kvm_mmu_page *sp)
4118 * Skip write-flooding detected for the sp whose level is 1, because
4119 * it can become unsync, then the guest page is not write-protected.
4121 if (sp->role.level == PT_PAGE_TABLE_LEVEL)
4124 return ++sp->write_flooding_count >= 3;
4128 * Misaligned accesses are too much trouble to fix up; also, they usually
4129 * indicate a page is not used as a page table.
4131 static bool detect_write_misaligned(struct kvm_mmu_page *sp, gpa_t gpa,
4134 unsigned offset, pte_size, misaligned;
4136 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
4137 gpa, bytes, sp->role.word);
4139 offset = offset_in_page(gpa);
4140 pte_size = sp->role.cr4_pae ? 8 : 4;
4143 * Sometimes, the OS only writes the last one bytes to update status
4144 * bits, for example, in linux, andb instruction is used in clear_bit().
4146 if (!(offset & (pte_size - 1)) && bytes == 1)
4149 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
4150 misaligned |= bytes < 4;
4155 static u64 *get_written_sptes(struct kvm_mmu_page *sp, gpa_t gpa, int *nspte)
4157 unsigned page_offset, quadrant;
4161 page_offset = offset_in_page(gpa);
4162 level = sp->role.level;
4164 if (!sp->role.cr4_pae) {
4165 page_offset <<= 1; /* 32->64 */
4167 * A 32-bit pde maps 4MB while the shadow pdes map
4168 * only 2MB. So we need to double the offset again
4169 * and zap two pdes instead of one.
4171 if (level == PT32_ROOT_LEVEL) {
4172 page_offset &= ~7; /* kill rounding error */
4176 quadrant = page_offset >> PAGE_SHIFT;
4177 page_offset &= ~PAGE_MASK;
4178 if (quadrant != sp->role.quadrant)
4182 spte = &sp->spt[page_offset / sizeof(*spte)];
4186 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
4187 const u8 *new, int bytes)
4189 gfn_t gfn = gpa >> PAGE_SHIFT;
4190 struct kvm_mmu_page *sp;
4191 LIST_HEAD(invalid_list);
4192 u64 entry, gentry, *spte;
4194 bool remote_flush, local_flush, zap_page;
4195 union kvm_mmu_page_role mask = { };
4200 mask.smep_andnot_wp = 1;
4201 mask.smap_andnot_wp = 1;
4205 * If we don't have indirect shadow pages, it means no page is
4206 * write-protected, so we can exit simply.
4208 if (!ACCESS_ONCE(vcpu->kvm->arch.indirect_shadow_pages))
4211 zap_page = remote_flush = local_flush = false;
4213 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
4215 gentry = mmu_pte_write_fetch_gpte(vcpu, &gpa, new, &bytes);
4218 * No need to care whether allocation memory is successful
4219 * or not since pte prefetch is skiped if it does not have
4220 * enough objects in the cache.
4222 mmu_topup_memory_caches(vcpu);
4224 spin_lock(&vcpu->kvm->mmu_lock);
4225 ++vcpu->kvm->stat.mmu_pte_write;
4226 kvm_mmu_audit(vcpu, AUDIT_PRE_PTE_WRITE);
4228 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn) {
4229 if (detect_write_misaligned(sp, gpa, bytes) ||
4230 detect_write_flooding(sp)) {
4231 zap_page |= !!kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
4233 ++vcpu->kvm->stat.mmu_flooded;
4237 spte = get_written_sptes(sp, gpa, &npte);
4244 mmu_page_zap_pte(vcpu->kvm, sp, spte);
4246 !((sp->role.word ^ vcpu->arch.mmu.base_role.word)
4247 & mask.word) && rmap_can_add(vcpu))
4248 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
4249 if (need_remote_flush(entry, *spte))
4250 remote_flush = true;
4254 mmu_pte_write_flush_tlb(vcpu, zap_page, remote_flush, local_flush);
4255 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
4256 kvm_mmu_audit(vcpu, AUDIT_POST_PTE_WRITE);
4257 spin_unlock(&vcpu->kvm->mmu_lock);
4260 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
4265 if (vcpu->arch.mmu.direct_map)
4268 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
4270 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
4274 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
4276 static void make_mmu_pages_available(struct kvm_vcpu *vcpu)
4278 LIST_HEAD(invalid_list);
4280 if (likely(kvm_mmu_available_pages(vcpu->kvm) >= KVM_MIN_FREE_MMU_PAGES))
4283 while (kvm_mmu_available_pages(vcpu->kvm) < KVM_REFILL_PAGES) {
4284 if (!prepare_zap_oldest_mmu_page(vcpu->kvm, &invalid_list))
4287 ++vcpu->kvm->stat.mmu_recycled;
4289 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
4292 static bool is_mmio_page_fault(struct kvm_vcpu *vcpu, gva_t addr)
4294 if (vcpu->arch.mmu.direct_map || mmu_is_nested(vcpu))
4295 return vcpu_match_mmio_gpa(vcpu, addr);
4297 return vcpu_match_mmio_gva(vcpu, addr);
4300 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code,
4301 void *insn, int insn_len)
4303 int r, emulation_type = EMULTYPE_RETRY;
4304 enum emulation_result er;
4306 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code, false);
4315 if (is_mmio_page_fault(vcpu, cr2))
4318 er = x86_emulate_instruction(vcpu, cr2, emulation_type, insn, insn_len);
4323 case EMULATE_USER_EXIT:
4324 ++vcpu->stat.mmio_exits;
4334 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
4336 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
4338 vcpu->arch.mmu.invlpg(vcpu, gva);
4339 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
4340 ++vcpu->stat.invlpg;
4342 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
4344 void kvm_enable_tdp(void)
4348 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
4350 void kvm_disable_tdp(void)
4352 tdp_enabled = false;
4354 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
4356 static void free_mmu_pages(struct kvm_vcpu *vcpu)
4358 free_page((unsigned long)vcpu->arch.mmu.pae_root);
4359 if (vcpu->arch.mmu.lm_root != NULL)
4360 free_page((unsigned long)vcpu->arch.mmu.lm_root);
4363 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
4369 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
4370 * Therefore we need to allocate shadow page tables in the first
4371 * 4GB of memory, which happens to fit the DMA32 zone.
4373 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
4377 vcpu->arch.mmu.pae_root = page_address(page);
4378 for (i = 0; i < 4; ++i)
4379 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
4384 int kvm_mmu_create(struct kvm_vcpu *vcpu)
4386 vcpu->arch.walk_mmu = &vcpu->arch.mmu;
4387 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
4388 vcpu->arch.mmu.translate_gpa = translate_gpa;
4389 vcpu->arch.nested_mmu.translate_gpa = translate_nested_gpa;
4391 return alloc_mmu_pages(vcpu);
4394 void kvm_mmu_setup(struct kvm_vcpu *vcpu)
4396 MMU_WARN_ON(VALID_PAGE(vcpu->arch.mmu.root_hpa));
4401 /* The return value indicates if tlb flush on all vcpus is needed. */
4402 typedef bool (*slot_level_handler) (struct kvm *kvm, unsigned long *rmap);
4404 /* The caller should hold mmu-lock before calling this function. */
4406 slot_handle_level_range(struct kvm *kvm, struct kvm_memory_slot *memslot,
4407 slot_level_handler fn, int start_level, int end_level,
4408 gfn_t start_gfn, gfn_t end_gfn, bool lock_flush_tlb)
4410 struct slot_rmap_walk_iterator iterator;
4413 for_each_slot_rmap_range(memslot, start_level, end_level, start_gfn,
4414 end_gfn, &iterator) {
4416 flush |= fn(kvm, iterator.rmap);
4418 if (need_resched() || spin_needbreak(&kvm->mmu_lock)) {
4419 if (flush && lock_flush_tlb) {
4420 kvm_flush_remote_tlbs(kvm);
4423 cond_resched_lock(&kvm->mmu_lock);
4427 if (flush && lock_flush_tlb) {
4428 kvm_flush_remote_tlbs(kvm);
4436 slot_handle_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
4437 slot_level_handler fn, int start_level, int end_level,
4438 bool lock_flush_tlb)
4440 return slot_handle_level_range(kvm, memslot, fn, start_level,
4441 end_level, memslot->base_gfn,
4442 memslot->base_gfn + memslot->npages - 1,
4447 slot_handle_all_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
4448 slot_level_handler fn, bool lock_flush_tlb)
4450 return slot_handle_level(kvm, memslot, fn, PT_PAGE_TABLE_LEVEL,
4451 PT_MAX_HUGEPAGE_LEVEL, lock_flush_tlb);
4455 slot_handle_large_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
4456 slot_level_handler fn, bool lock_flush_tlb)
4458 return slot_handle_level(kvm, memslot, fn, PT_PAGE_TABLE_LEVEL + 1,
4459 PT_MAX_HUGEPAGE_LEVEL, lock_flush_tlb);
4463 slot_handle_leaf(struct kvm *kvm, struct kvm_memory_slot *memslot,
4464 slot_level_handler fn, bool lock_flush_tlb)
4466 return slot_handle_level(kvm, memslot, fn, PT_PAGE_TABLE_LEVEL,
4467 PT_PAGE_TABLE_LEVEL, lock_flush_tlb);
4470 void kvm_zap_gfn_range(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end)
4472 struct kvm_memslots *slots;
4473 struct kvm_memory_slot *memslot;
4476 spin_lock(&kvm->mmu_lock);
4477 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
4478 slots = __kvm_memslots(kvm, i);
4479 kvm_for_each_memslot(memslot, slots) {
4482 start = max(gfn_start, memslot->base_gfn);
4483 end = min(gfn_end, memslot->base_gfn + memslot->npages);
4487 slot_handle_level_range(kvm, memslot, kvm_zap_rmapp,
4488 PT_PAGE_TABLE_LEVEL, PT_MAX_HUGEPAGE_LEVEL,
4489 start, end - 1, true);
4493 spin_unlock(&kvm->mmu_lock);
4496 static bool slot_rmap_write_protect(struct kvm *kvm, unsigned long *rmapp)
4498 return __rmap_write_protect(kvm, rmapp, false);
4501 void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
4502 struct kvm_memory_slot *memslot)
4506 spin_lock(&kvm->mmu_lock);
4507 flush = slot_handle_all_level(kvm, memslot, slot_rmap_write_protect,
4509 spin_unlock(&kvm->mmu_lock);
4512 * kvm_mmu_slot_remove_write_access() and kvm_vm_ioctl_get_dirty_log()
4513 * which do tlb flush out of mmu-lock should be serialized by
4514 * kvm->slots_lock otherwise tlb flush would be missed.
4516 lockdep_assert_held(&kvm->slots_lock);
4519 * We can flush all the TLBs out of the mmu lock without TLB
4520 * corruption since we just change the spte from writable to
4521 * readonly so that we only need to care the case of changing
4522 * spte from present to present (changing the spte from present
4523 * to nonpresent will flush all the TLBs immediately), in other
4524 * words, the only case we care is mmu_spte_update() where we
4525 * haved checked SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE
4526 * instead of PT_WRITABLE_MASK, that means it does not depend
4527 * on PT_WRITABLE_MASK anymore.
4530 kvm_flush_remote_tlbs(kvm);
4533 static bool kvm_mmu_zap_collapsible_spte(struct kvm *kvm,
4534 unsigned long *rmapp)
4537 struct rmap_iterator iter;
4538 int need_tlb_flush = 0;
4540 struct kvm_mmu_page *sp;
4543 for_each_rmap_spte(rmapp, &iter, sptep) {
4544 sp = page_header(__pa(sptep));
4545 pfn = spte_to_pfn(*sptep);
4548 * We cannot do huge page mapping for indirect shadow pages,
4549 * which are found on the last rmap (level = 1) when not using
4550 * tdp; such shadow pages are synced with the page table in
4551 * the guest, and the guest page table is using 4K page size
4552 * mapping if the indirect sp has level = 1.
4554 if (sp->role.direct &&
4555 !kvm_is_reserved_pfn(pfn) &&
4556 PageTransCompound(pfn_to_page(pfn))) {
4557 drop_spte(kvm, sptep);
4563 return need_tlb_flush;
4566 void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm,
4567 const struct kvm_memory_slot *memslot)
4569 /* FIXME: const-ify all uses of struct kvm_memory_slot. */
4570 spin_lock(&kvm->mmu_lock);
4571 slot_handle_leaf(kvm, (struct kvm_memory_slot *)memslot,
4572 kvm_mmu_zap_collapsible_spte, true);
4573 spin_unlock(&kvm->mmu_lock);
4576 void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm,
4577 struct kvm_memory_slot *memslot)
4581 spin_lock(&kvm->mmu_lock);
4582 flush = slot_handle_leaf(kvm, memslot, __rmap_clear_dirty, false);
4583 spin_unlock(&kvm->mmu_lock);
4585 lockdep_assert_held(&kvm->slots_lock);
4588 * It's also safe to flush TLBs out of mmu lock here as currently this
4589 * function is only used for dirty logging, in which case flushing TLB
4590 * out of mmu lock also guarantees no dirty pages will be lost in
4594 kvm_flush_remote_tlbs(kvm);
4596 EXPORT_SYMBOL_GPL(kvm_mmu_slot_leaf_clear_dirty);
4598 void kvm_mmu_slot_largepage_remove_write_access(struct kvm *kvm,
4599 struct kvm_memory_slot *memslot)
4603 spin_lock(&kvm->mmu_lock);
4604 flush = slot_handle_large_level(kvm, memslot, slot_rmap_write_protect,
4606 spin_unlock(&kvm->mmu_lock);
4608 /* see kvm_mmu_slot_remove_write_access */
4609 lockdep_assert_held(&kvm->slots_lock);
4612 kvm_flush_remote_tlbs(kvm);
4614 EXPORT_SYMBOL_GPL(kvm_mmu_slot_largepage_remove_write_access);
4616 void kvm_mmu_slot_set_dirty(struct kvm *kvm,
4617 struct kvm_memory_slot *memslot)
4621 spin_lock(&kvm->mmu_lock);
4622 flush = slot_handle_all_level(kvm, memslot, __rmap_set_dirty, false);
4623 spin_unlock(&kvm->mmu_lock);
4625 lockdep_assert_held(&kvm->slots_lock);
4627 /* see kvm_mmu_slot_leaf_clear_dirty */
4629 kvm_flush_remote_tlbs(kvm);
4631 EXPORT_SYMBOL_GPL(kvm_mmu_slot_set_dirty);
4633 #define BATCH_ZAP_PAGES 10
4634 static void kvm_zap_obsolete_pages(struct kvm *kvm)
4636 struct kvm_mmu_page *sp, *node;
4640 list_for_each_entry_safe_reverse(sp, node,
4641 &kvm->arch.active_mmu_pages, link) {
4645 * No obsolete page exists before new created page since
4646 * active_mmu_pages is the FIFO list.
4648 if (!is_obsolete_sp(kvm, sp))
4652 * Since we are reversely walking the list and the invalid
4653 * list will be moved to the head, skip the invalid page
4654 * can help us to avoid the infinity list walking.
4656 if (sp->role.invalid)
4660 * Need not flush tlb since we only zap the sp with invalid
4661 * generation number.
4663 if (batch >= BATCH_ZAP_PAGES &&
4664 cond_resched_lock(&kvm->mmu_lock)) {
4669 ret = kvm_mmu_prepare_zap_page(kvm, sp,
4670 &kvm->arch.zapped_obsolete_pages);
4678 * Should flush tlb before free page tables since lockless-walking
4679 * may use the pages.
4681 kvm_mmu_commit_zap_page(kvm, &kvm->arch.zapped_obsolete_pages);
4685 * Fast invalidate all shadow pages and use lock-break technique
4686 * to zap obsolete pages.
4688 * It's required when memslot is being deleted or VM is being
4689 * destroyed, in these cases, we should ensure that KVM MMU does
4690 * not use any resource of the being-deleted slot or all slots
4691 * after calling the function.
4693 void kvm_mmu_invalidate_zap_all_pages(struct kvm *kvm)
4695 spin_lock(&kvm->mmu_lock);
4696 trace_kvm_mmu_invalidate_zap_all_pages(kvm);
4697 kvm->arch.mmu_valid_gen++;
4700 * Notify all vcpus to reload its shadow page table
4701 * and flush TLB. Then all vcpus will switch to new
4702 * shadow page table with the new mmu_valid_gen.
4704 * Note: we should do this under the protection of
4705 * mmu-lock, otherwise, vcpu would purge shadow page
4706 * but miss tlb flush.
4708 kvm_reload_remote_mmus(kvm);
4710 kvm_zap_obsolete_pages(kvm);
4711 spin_unlock(&kvm->mmu_lock);
4714 static bool kvm_has_zapped_obsolete_pages(struct kvm *kvm)
4716 return unlikely(!list_empty_careful(&kvm->arch.zapped_obsolete_pages));
4719 void kvm_mmu_invalidate_mmio_sptes(struct kvm *kvm, struct kvm_memslots *slots)
4722 * The very rare case: if the generation-number is round,
4723 * zap all shadow pages.
4725 if (unlikely((slots->generation & MMIO_GEN_MASK) == 0)) {
4726 printk_ratelimited(KERN_DEBUG "kvm: zapping shadow pages for mmio generation wraparound\n");
4727 kvm_mmu_invalidate_zap_all_pages(kvm);
4731 static unsigned long
4732 mmu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
4735 int nr_to_scan = sc->nr_to_scan;
4736 unsigned long freed = 0;
4738 spin_lock(&kvm_lock);
4740 list_for_each_entry(kvm, &vm_list, vm_list) {
4742 LIST_HEAD(invalid_list);
4745 * Never scan more than sc->nr_to_scan VM instances.
4746 * Will not hit this condition practically since we do not try
4747 * to shrink more than one VM and it is very unlikely to see
4748 * !n_used_mmu_pages so many times.
4753 * n_used_mmu_pages is accessed without holding kvm->mmu_lock
4754 * here. We may skip a VM instance errorneosly, but we do not
4755 * want to shrink a VM that only started to populate its MMU
4758 if (!kvm->arch.n_used_mmu_pages &&
4759 !kvm_has_zapped_obsolete_pages(kvm))
4762 idx = srcu_read_lock(&kvm->srcu);
4763 spin_lock(&kvm->mmu_lock);
4765 if (kvm_has_zapped_obsolete_pages(kvm)) {
4766 kvm_mmu_commit_zap_page(kvm,
4767 &kvm->arch.zapped_obsolete_pages);
4771 if (prepare_zap_oldest_mmu_page(kvm, &invalid_list))
4773 kvm_mmu_commit_zap_page(kvm, &invalid_list);
4776 spin_unlock(&kvm->mmu_lock);
4777 srcu_read_unlock(&kvm->srcu, idx);
4780 * unfair on small ones
4781 * per-vm shrinkers cry out
4782 * sadness comes quickly
4784 list_move_tail(&kvm->vm_list, &vm_list);
4788 spin_unlock(&kvm_lock);
4792 static unsigned long
4793 mmu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
4795 return percpu_counter_read_positive(&kvm_total_used_mmu_pages);
4798 static struct shrinker mmu_shrinker = {
4799 .count_objects = mmu_shrink_count,
4800 .scan_objects = mmu_shrink_scan,
4801 .seeks = DEFAULT_SEEKS * 10,
4804 static void mmu_destroy_caches(void)
4806 if (pte_list_desc_cache)
4807 kmem_cache_destroy(pte_list_desc_cache);
4808 if (mmu_page_header_cache)
4809 kmem_cache_destroy(mmu_page_header_cache);
4812 int kvm_mmu_module_init(void)
4814 pte_list_desc_cache = kmem_cache_create("pte_list_desc",
4815 sizeof(struct pte_list_desc),
4817 if (!pte_list_desc_cache)
4820 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
4821 sizeof(struct kvm_mmu_page),
4823 if (!mmu_page_header_cache)
4826 if (percpu_counter_init(&kvm_total_used_mmu_pages, 0, GFP_KERNEL))
4829 register_shrinker(&mmu_shrinker);
4834 mmu_destroy_caches();
4839 * Caculate mmu pages needed for kvm.
4841 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
4843 unsigned int nr_mmu_pages;
4844 unsigned int nr_pages = 0;
4845 struct kvm_memslots *slots;
4846 struct kvm_memory_slot *memslot;
4849 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
4850 slots = __kvm_memslots(kvm, i);
4852 kvm_for_each_memslot(memslot, slots)
4853 nr_pages += memslot->npages;
4856 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
4857 nr_mmu_pages = max(nr_mmu_pages,
4858 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
4860 return nr_mmu_pages;
4863 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
4865 struct kvm_shadow_walk_iterator iterator;
4869 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
4872 walk_shadow_page_lockless_begin(vcpu);
4873 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte) {
4874 sptes[iterator.level-1] = spte;
4876 if (!is_shadow_present_pte(spte))
4879 walk_shadow_page_lockless_end(vcpu);
4883 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
4885 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
4887 kvm_mmu_unload(vcpu);
4888 free_mmu_pages(vcpu);
4889 mmu_free_memory_caches(vcpu);
4892 void kvm_mmu_module_exit(void)
4894 mmu_destroy_caches();
4895 percpu_counter_destroy(&kvm_total_used_mmu_pages);
4896 unregister_shrinker(&mmu_shrinker);
4897 mmu_audit_disable();