2 * Kernel-based Virtual Machine driver for Linux
4 * This module enables machines with Intel VT-x extensions to run virtual
5 * machines without emulation or binary translation.
9 * Copyright (C) 2006 Qumranet, Inc.
10 * Copyright 2010 Red Hat, Inc. and/or its affiliates.
13 * Yaniv Kamay <yaniv@qumranet.com>
14 * Avi Kivity <avi@qumranet.com>
16 * This work is licensed under the terms of the GNU GPL, version 2. See
17 * the COPYING file in the top-level directory.
24 #include "kvm_cache_regs.h"
26 #include <linux/kvm_host.h>
27 #include <linux/types.h>
28 #include <linux/string.h>
30 #include <linux/highmem.h>
31 #include <linux/module.h>
32 #include <linux/swap.h>
33 #include <linux/hugetlb.h>
34 #include <linux/compiler.h>
35 #include <linux/srcu.h>
36 #include <linux/slab.h>
37 #include <linux/uaccess.h>
40 #include <asm/cmpxchg.h>
45 * When setting this variable to true it enables Two-Dimensional-Paging
46 * where the hardware walks 2 page tables:
47 * 1. the guest-virtual to guest-physical
48 * 2. while doing 1. it walks guest-physical to host-physical
49 * If the hardware supports that we don't need to do shadow paging.
51 bool tdp_enabled = false;
55 AUDIT_POST_PAGE_FAULT,
62 char *audit_point_name[] = {
75 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
76 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
80 #define pgprintk(x...) do { } while (0)
81 #define rmap_printk(x...) do { } while (0)
87 module_param(dbg, bool, 0644);
90 static int oos_shadow = 1;
91 module_param(oos_shadow, bool, 0644);
94 #define ASSERT(x) do { } while (0)
98 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
99 __FILE__, __LINE__, #x); \
103 #define PTE_PREFETCH_NUM 8
105 #define PT_FIRST_AVAIL_BITS_SHIFT 9
106 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
108 #define PT64_LEVEL_BITS 9
110 #define PT64_LEVEL_SHIFT(level) \
111 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
113 #define PT64_INDEX(address, level)\
114 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
117 #define PT32_LEVEL_BITS 10
119 #define PT32_LEVEL_SHIFT(level) \
120 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
122 #define PT32_LVL_OFFSET_MASK(level) \
123 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
124 * PT32_LEVEL_BITS))) - 1))
126 #define PT32_INDEX(address, level)\
127 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
130 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
131 #define PT64_DIR_BASE_ADDR_MASK \
132 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
133 #define PT64_LVL_ADDR_MASK(level) \
134 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
135 * PT64_LEVEL_BITS))) - 1))
136 #define PT64_LVL_OFFSET_MASK(level) \
137 (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
138 * PT64_LEVEL_BITS))) - 1))
140 #define PT32_BASE_ADDR_MASK PAGE_MASK
141 #define PT32_DIR_BASE_ADDR_MASK \
142 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
143 #define PT32_LVL_ADDR_MASK(level) \
144 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
145 * PT32_LEVEL_BITS))) - 1))
147 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
150 #define PTE_LIST_EXT 4
152 #define ACC_EXEC_MASK 1
153 #define ACC_WRITE_MASK PT_WRITABLE_MASK
154 #define ACC_USER_MASK PT_USER_MASK
155 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
157 #include <trace/events/kvm.h>
159 #define CREATE_TRACE_POINTS
160 #include "mmutrace.h"
162 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
164 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
166 struct pte_list_desc {
167 u64 *sptes[PTE_LIST_EXT];
168 struct pte_list_desc *more;
171 struct kvm_shadow_walk_iterator {
179 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
180 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
181 shadow_walk_okay(&(_walker)); \
182 shadow_walk_next(&(_walker)))
184 #define for_each_shadow_entry_lockless(_vcpu, _addr, _walker, spte) \
185 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
186 shadow_walk_okay(&(_walker)) && \
187 ({ spte = mmu_spte_get_lockless(_walker.sptep); 1; }); \
188 __shadow_walk_next(&(_walker), spte))
190 static struct kmem_cache *pte_list_desc_cache;
191 static struct kmem_cache *mmu_page_header_cache;
192 static struct percpu_counter kvm_total_used_mmu_pages;
194 static u64 __read_mostly shadow_nx_mask;
195 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
196 static u64 __read_mostly shadow_user_mask;
197 static u64 __read_mostly shadow_accessed_mask;
198 static u64 __read_mostly shadow_dirty_mask;
199 static u64 __read_mostly shadow_mmio_mask;
201 static void mmu_spte_set(u64 *sptep, u64 spte);
203 void kvm_mmu_set_mmio_spte_mask(u64 mmio_mask)
205 shadow_mmio_mask = mmio_mask;
207 EXPORT_SYMBOL_GPL(kvm_mmu_set_mmio_spte_mask);
209 static void mark_mmio_spte(u64 *sptep, u64 gfn, unsigned access)
211 access &= ACC_WRITE_MASK | ACC_USER_MASK;
213 trace_mark_mmio_spte(sptep, gfn, access);
214 mmu_spte_set(sptep, shadow_mmio_mask | access | gfn << PAGE_SHIFT);
217 static bool is_mmio_spte(u64 spte)
219 return (spte & shadow_mmio_mask) == shadow_mmio_mask;
222 static gfn_t get_mmio_spte_gfn(u64 spte)
224 return (spte & ~shadow_mmio_mask) >> PAGE_SHIFT;
227 static unsigned get_mmio_spte_access(u64 spte)
229 return (spte & ~shadow_mmio_mask) & ~PAGE_MASK;
232 static bool set_mmio_spte(u64 *sptep, gfn_t gfn, pfn_t pfn, unsigned access)
234 if (unlikely(is_noslot_pfn(pfn))) {
235 mark_mmio_spte(sptep, gfn, access);
242 static inline u64 rsvd_bits(int s, int e)
244 return ((1ULL << (e - s + 1)) - 1) << s;
247 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
248 u64 dirty_mask, u64 nx_mask, u64 x_mask)
250 shadow_user_mask = user_mask;
251 shadow_accessed_mask = accessed_mask;
252 shadow_dirty_mask = dirty_mask;
253 shadow_nx_mask = nx_mask;
254 shadow_x_mask = x_mask;
256 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
258 static int is_cpuid_PSE36(void)
263 static int is_nx(struct kvm_vcpu *vcpu)
265 return vcpu->arch.efer & EFER_NX;
268 static int is_shadow_present_pte(u64 pte)
270 return pte & PT_PRESENT_MASK && !is_mmio_spte(pte);
273 static int is_large_pte(u64 pte)
275 return pte & PT_PAGE_SIZE_MASK;
278 static int is_dirty_gpte(unsigned long pte)
280 return pte & PT_DIRTY_MASK;
283 static int is_rmap_spte(u64 pte)
285 return is_shadow_present_pte(pte);
288 static int is_last_spte(u64 pte, int level)
290 if (level == PT_PAGE_TABLE_LEVEL)
292 if (is_large_pte(pte))
297 static pfn_t spte_to_pfn(u64 pte)
299 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
302 static gfn_t pse36_gfn_delta(u32 gpte)
304 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
306 return (gpte & PT32_DIR_PSE36_MASK) << shift;
310 static void __set_spte(u64 *sptep, u64 spte)
315 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
320 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
322 return xchg(sptep, spte);
325 static u64 __get_spte_lockless(u64 *sptep)
327 return ACCESS_ONCE(*sptep);
330 static bool __check_direct_spte_mmio_pf(u64 spte)
332 /* It is valid if the spte is zapped. */
344 static void count_spte_clear(u64 *sptep, u64 spte)
346 struct kvm_mmu_page *sp = page_header(__pa(sptep));
348 if (is_shadow_present_pte(spte))
351 /* Ensure the spte is completely set before we increase the count */
353 sp->clear_spte_count++;
356 static void __set_spte(u64 *sptep, u64 spte)
358 union split_spte *ssptep, sspte;
360 ssptep = (union split_spte *)sptep;
361 sspte = (union split_spte)spte;
363 ssptep->spte_high = sspte.spte_high;
366 * If we map the spte from nonpresent to present, We should store
367 * the high bits firstly, then set present bit, so cpu can not
368 * fetch this spte while we are setting the spte.
372 ssptep->spte_low = sspte.spte_low;
375 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
377 union split_spte *ssptep, sspte;
379 ssptep = (union split_spte *)sptep;
380 sspte = (union split_spte)spte;
382 ssptep->spte_low = sspte.spte_low;
385 * If we map the spte from present to nonpresent, we should clear
386 * present bit firstly to avoid vcpu fetch the old high bits.
390 ssptep->spte_high = sspte.spte_high;
391 count_spte_clear(sptep, spte);
394 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
396 union split_spte *ssptep, sspte, orig;
398 ssptep = (union split_spte *)sptep;
399 sspte = (union split_spte)spte;
401 /* xchg acts as a barrier before the setting of the high bits */
402 orig.spte_low = xchg(&ssptep->spte_low, sspte.spte_low);
403 orig.spte_high = ssptep->spte_high;
404 ssptep->spte_high = sspte.spte_high;
405 count_spte_clear(sptep, spte);
411 * The idea using the light way get the spte on x86_32 guest is from
412 * gup_get_pte(arch/x86/mm/gup.c).
413 * The difference is we can not catch the spte tlb flush if we leave
414 * guest mode, so we emulate it by increase clear_spte_count when spte
417 static u64 __get_spte_lockless(u64 *sptep)
419 struct kvm_mmu_page *sp = page_header(__pa(sptep));
420 union split_spte spte, *orig = (union split_spte *)sptep;
424 count = sp->clear_spte_count;
427 spte.spte_low = orig->spte_low;
430 spte.spte_high = orig->spte_high;
433 if (unlikely(spte.spte_low != orig->spte_low ||
434 count != sp->clear_spte_count))
440 static bool __check_direct_spte_mmio_pf(u64 spte)
442 union split_spte sspte = (union split_spte)spte;
443 u32 high_mmio_mask = shadow_mmio_mask >> 32;
445 /* It is valid if the spte is zapped. */
449 /* It is valid if the spte is being zapped. */
450 if (sspte.spte_low == 0ull &&
451 (sspte.spte_high & high_mmio_mask) == high_mmio_mask)
458 static bool spte_has_volatile_bits(u64 spte)
460 if (!shadow_accessed_mask)
463 if (!is_shadow_present_pte(spte))
466 if ((spte & shadow_accessed_mask) &&
467 (!is_writable_pte(spte) || (spte & shadow_dirty_mask)))
473 static bool spte_is_bit_cleared(u64 old_spte, u64 new_spte, u64 bit_mask)
475 return (old_spte & bit_mask) && !(new_spte & bit_mask);
478 /* Rules for using mmu_spte_set:
479 * Set the sptep from nonpresent to present.
480 * Note: the sptep being assigned *must* be either not present
481 * or in a state where the hardware will not attempt to update
484 static void mmu_spte_set(u64 *sptep, u64 new_spte)
486 WARN_ON(is_shadow_present_pte(*sptep));
487 __set_spte(sptep, new_spte);
490 /* Rules for using mmu_spte_update:
491 * Update the state bits, it means the mapped pfn is not changged.
493 static void mmu_spte_update(u64 *sptep, u64 new_spte)
495 u64 mask, old_spte = *sptep;
497 WARN_ON(!is_rmap_spte(new_spte));
499 if (!is_shadow_present_pte(old_spte))
500 return mmu_spte_set(sptep, new_spte);
502 new_spte |= old_spte & shadow_dirty_mask;
504 mask = shadow_accessed_mask;
505 if (is_writable_pte(old_spte))
506 mask |= shadow_dirty_mask;
508 if (!spte_has_volatile_bits(old_spte) || (new_spte & mask) == mask)
509 __update_clear_spte_fast(sptep, new_spte);
511 old_spte = __update_clear_spte_slow(sptep, new_spte);
513 if (!shadow_accessed_mask)
516 if (spte_is_bit_cleared(old_spte, new_spte, shadow_accessed_mask))
517 kvm_set_pfn_accessed(spte_to_pfn(old_spte));
518 if (spte_is_bit_cleared(old_spte, new_spte, shadow_dirty_mask))
519 kvm_set_pfn_dirty(spte_to_pfn(old_spte));
523 * Rules for using mmu_spte_clear_track_bits:
524 * It sets the sptep from present to nonpresent, and track the
525 * state bits, it is used to clear the last level sptep.
527 static int mmu_spte_clear_track_bits(u64 *sptep)
530 u64 old_spte = *sptep;
532 if (!spte_has_volatile_bits(old_spte))
533 __update_clear_spte_fast(sptep, 0ull);
535 old_spte = __update_clear_spte_slow(sptep, 0ull);
537 if (!is_rmap_spte(old_spte))
540 pfn = spte_to_pfn(old_spte);
541 if (!shadow_accessed_mask || old_spte & shadow_accessed_mask)
542 kvm_set_pfn_accessed(pfn);
543 if (!shadow_dirty_mask || (old_spte & shadow_dirty_mask))
544 kvm_set_pfn_dirty(pfn);
549 * Rules for using mmu_spte_clear_no_track:
550 * Directly clear spte without caring the state bits of sptep,
551 * it is used to set the upper level spte.
553 static void mmu_spte_clear_no_track(u64 *sptep)
555 __update_clear_spte_fast(sptep, 0ull);
558 static u64 mmu_spte_get_lockless(u64 *sptep)
560 return __get_spte_lockless(sptep);
563 static void walk_shadow_page_lockless_begin(struct kvm_vcpu *vcpu)
566 atomic_inc(&vcpu->kvm->arch.reader_counter);
568 /* Increase the counter before walking shadow page table */
569 smp_mb__after_atomic_inc();
572 static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
574 /* Decrease the counter after walking shadow page table finished */
575 smp_mb__before_atomic_dec();
576 atomic_dec(&vcpu->kvm->arch.reader_counter);
580 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
581 struct kmem_cache *base_cache, int min)
585 if (cache->nobjs >= min)
587 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
588 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
591 cache->objects[cache->nobjs++] = obj;
596 static int mmu_memory_cache_free_objects(struct kvm_mmu_memory_cache *cache)
601 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc,
602 struct kmem_cache *cache)
605 kmem_cache_free(cache, mc->objects[--mc->nobjs]);
608 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
613 if (cache->nobjs >= min)
615 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
616 page = (void *)__get_free_page(GFP_KERNEL);
619 cache->objects[cache->nobjs++] = page;
624 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
627 free_page((unsigned long)mc->objects[--mc->nobjs]);
630 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
634 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
635 pte_list_desc_cache, 8 + PTE_PREFETCH_NUM);
638 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
641 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
642 mmu_page_header_cache, 4);
647 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
649 mmu_free_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
650 pte_list_desc_cache);
651 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
652 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache,
653 mmu_page_header_cache);
656 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
662 p = mc->objects[--mc->nobjs];
666 static struct pte_list_desc *mmu_alloc_pte_list_desc(struct kvm_vcpu *vcpu)
668 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_list_desc_cache,
669 sizeof(struct pte_list_desc));
672 static void mmu_free_pte_list_desc(struct pte_list_desc *pte_list_desc)
674 kmem_cache_free(pte_list_desc_cache, pte_list_desc);
677 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
679 if (!sp->role.direct)
680 return sp->gfns[index];
682 return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
685 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
688 BUG_ON(gfn != kvm_mmu_page_get_gfn(sp, index));
690 sp->gfns[index] = gfn;
694 * Return the pointer to the large page information for a given gfn,
695 * handling slots that are not large page aligned.
697 static struct kvm_lpage_info *lpage_info_slot(gfn_t gfn,
698 struct kvm_memory_slot *slot,
703 idx = (gfn >> KVM_HPAGE_GFN_SHIFT(level)) -
704 (slot->base_gfn >> KVM_HPAGE_GFN_SHIFT(level));
705 return &slot->lpage_info[level - 2][idx];
708 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
710 struct kvm_memory_slot *slot;
711 struct kvm_lpage_info *linfo;
714 slot = gfn_to_memslot(kvm, gfn);
715 for (i = PT_DIRECTORY_LEVEL;
716 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
717 linfo = lpage_info_slot(gfn, slot, i);
718 linfo->write_count += 1;
720 kvm->arch.indirect_shadow_pages++;
723 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
725 struct kvm_memory_slot *slot;
726 struct kvm_lpage_info *linfo;
729 slot = gfn_to_memslot(kvm, gfn);
730 for (i = PT_DIRECTORY_LEVEL;
731 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
732 linfo = lpage_info_slot(gfn, slot, i);
733 linfo->write_count -= 1;
734 WARN_ON(linfo->write_count < 0);
736 kvm->arch.indirect_shadow_pages--;
739 static int has_wrprotected_page(struct kvm *kvm,
743 struct kvm_memory_slot *slot;
744 struct kvm_lpage_info *linfo;
746 slot = gfn_to_memslot(kvm, gfn);
748 linfo = lpage_info_slot(gfn, slot, level);
749 return linfo->write_count;
755 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
757 unsigned long page_size;
760 page_size = kvm_host_page_size(kvm, gfn);
762 for (i = PT_PAGE_TABLE_LEVEL;
763 i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
764 if (page_size >= KVM_HPAGE_SIZE(i))
773 static struct kvm_memory_slot *
774 gfn_to_memslot_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t gfn,
777 struct kvm_memory_slot *slot;
779 slot = gfn_to_memslot(vcpu->kvm, gfn);
780 if (!slot || slot->flags & KVM_MEMSLOT_INVALID ||
781 (no_dirty_log && slot->dirty_bitmap))
787 static bool mapping_level_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t large_gfn)
789 return !gfn_to_memslot_dirty_bitmap(vcpu, large_gfn, true);
792 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
794 int host_level, level, max_level;
796 host_level = host_mapping_level(vcpu->kvm, large_gfn);
798 if (host_level == PT_PAGE_TABLE_LEVEL)
801 max_level = kvm_x86_ops->get_lpage_level() < host_level ?
802 kvm_x86_ops->get_lpage_level() : host_level;
804 for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
805 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
812 * Pte mapping structures:
814 * If pte_list bit zero is zero, then pte_list point to the spte.
816 * If pte_list bit zero is one, (then pte_list & ~1) points to a struct
817 * pte_list_desc containing more mappings.
819 * Returns the number of pte entries before the spte was added or zero if
820 * the spte was not added.
823 static int pte_list_add(struct kvm_vcpu *vcpu, u64 *spte,
824 unsigned long *pte_list)
826 struct pte_list_desc *desc;
830 rmap_printk("pte_list_add: %p %llx 0->1\n", spte, *spte);
831 *pte_list = (unsigned long)spte;
832 } else if (!(*pte_list & 1)) {
833 rmap_printk("pte_list_add: %p %llx 1->many\n", spte, *spte);
834 desc = mmu_alloc_pte_list_desc(vcpu);
835 desc->sptes[0] = (u64 *)*pte_list;
836 desc->sptes[1] = spte;
837 *pte_list = (unsigned long)desc | 1;
840 rmap_printk("pte_list_add: %p %llx many->many\n", spte, *spte);
841 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
842 while (desc->sptes[PTE_LIST_EXT-1] && desc->more) {
844 count += PTE_LIST_EXT;
846 if (desc->sptes[PTE_LIST_EXT-1]) {
847 desc->more = mmu_alloc_pte_list_desc(vcpu);
850 for (i = 0; desc->sptes[i]; ++i)
852 desc->sptes[i] = spte;
857 static u64 *pte_list_next(unsigned long *pte_list, u64 *spte)
859 struct pte_list_desc *desc;
865 else if (!(*pte_list & 1)) {
867 return (u64 *)*pte_list;
870 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
873 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i) {
874 if (prev_spte == spte)
875 return desc->sptes[i];
876 prev_spte = desc->sptes[i];
884 pte_list_desc_remove_entry(unsigned long *pte_list, struct pte_list_desc *desc,
885 int i, struct pte_list_desc *prev_desc)
889 for (j = PTE_LIST_EXT - 1; !desc->sptes[j] && j > i; --j)
891 desc->sptes[i] = desc->sptes[j];
892 desc->sptes[j] = NULL;
895 if (!prev_desc && !desc->more)
896 *pte_list = (unsigned long)desc->sptes[0];
899 prev_desc->more = desc->more;
901 *pte_list = (unsigned long)desc->more | 1;
902 mmu_free_pte_list_desc(desc);
905 static void pte_list_remove(u64 *spte, unsigned long *pte_list)
907 struct pte_list_desc *desc;
908 struct pte_list_desc *prev_desc;
912 printk(KERN_ERR "pte_list_remove: %p 0->BUG\n", spte);
914 } else if (!(*pte_list & 1)) {
915 rmap_printk("pte_list_remove: %p 1->0\n", spte);
916 if ((u64 *)*pte_list != spte) {
917 printk(KERN_ERR "pte_list_remove: %p 1->BUG\n", spte);
922 rmap_printk("pte_list_remove: %p many->many\n", spte);
923 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
926 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i)
927 if (desc->sptes[i] == spte) {
928 pte_list_desc_remove_entry(pte_list,
936 pr_err("pte_list_remove: %p many->many\n", spte);
941 typedef void (*pte_list_walk_fn) (u64 *spte);
942 static void pte_list_walk(unsigned long *pte_list, pte_list_walk_fn fn)
944 struct pte_list_desc *desc;
950 if (!(*pte_list & 1))
951 return fn((u64 *)*pte_list);
953 desc = (struct pte_list_desc *)(*pte_list & ~1ul);
955 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i)
962 * Take gfn and return the reverse mapping to it.
964 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
966 struct kvm_memory_slot *slot;
967 struct kvm_lpage_info *linfo;
969 slot = gfn_to_memslot(kvm, gfn);
970 if (likely(level == PT_PAGE_TABLE_LEVEL))
971 return &slot->rmap[gfn - slot->base_gfn];
973 linfo = lpage_info_slot(gfn, slot, level);
975 return &linfo->rmap_pde;
978 static bool rmap_can_add(struct kvm_vcpu *vcpu)
980 struct kvm_mmu_memory_cache *cache;
982 cache = &vcpu->arch.mmu_pte_list_desc_cache;
983 return mmu_memory_cache_free_objects(cache);
986 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
988 struct kvm_mmu_page *sp;
989 unsigned long *rmapp;
991 sp = page_header(__pa(spte));
992 kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
993 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
994 return pte_list_add(vcpu, spte, rmapp);
997 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
999 return pte_list_next(rmapp, spte);
1002 static void rmap_remove(struct kvm *kvm, u64 *spte)
1004 struct kvm_mmu_page *sp;
1006 unsigned long *rmapp;
1008 sp = page_header(__pa(spte));
1009 gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
1010 rmapp = gfn_to_rmap(kvm, gfn, sp->role.level);
1011 pte_list_remove(spte, rmapp);
1014 static void drop_spte(struct kvm *kvm, u64 *sptep)
1016 if (mmu_spte_clear_track_bits(sptep))
1017 rmap_remove(kvm, sptep);
1020 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
1022 unsigned long *rmapp;
1024 int i, write_protected = 0;
1026 rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
1028 spte = rmap_next(kvm, rmapp, NULL);
1031 BUG_ON(!(*spte & PT_PRESENT_MASK));
1032 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
1033 if (is_writable_pte(*spte)) {
1034 mmu_spte_update(spte, *spte & ~PT_WRITABLE_MASK);
1035 write_protected = 1;
1037 spte = rmap_next(kvm, rmapp, spte);
1040 /* check for huge page mappings */
1041 for (i = PT_DIRECTORY_LEVEL;
1042 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
1043 rmapp = gfn_to_rmap(kvm, gfn, i);
1044 spte = rmap_next(kvm, rmapp, NULL);
1047 BUG_ON(!(*spte & PT_PRESENT_MASK));
1048 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
1049 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
1050 if (is_writable_pte(*spte)) {
1051 drop_spte(kvm, spte);
1054 write_protected = 1;
1056 spte = rmap_next(kvm, rmapp, spte);
1060 return write_protected;
1063 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
1067 int need_tlb_flush = 0;
1069 while ((spte = rmap_next(kvm, rmapp, NULL))) {
1070 BUG_ON(!(*spte & PT_PRESENT_MASK));
1071 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
1072 drop_spte(kvm, spte);
1075 return need_tlb_flush;
1078 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
1082 u64 *spte, new_spte;
1083 pte_t *ptep = (pte_t *)data;
1086 WARN_ON(pte_huge(*ptep));
1087 new_pfn = pte_pfn(*ptep);
1088 spte = rmap_next(kvm, rmapp, NULL);
1090 BUG_ON(!is_shadow_present_pte(*spte));
1091 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
1093 if (pte_write(*ptep)) {
1094 drop_spte(kvm, spte);
1095 spte = rmap_next(kvm, rmapp, NULL);
1097 new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
1098 new_spte |= (u64)new_pfn << PAGE_SHIFT;
1100 new_spte &= ~PT_WRITABLE_MASK;
1101 new_spte &= ~SPTE_HOST_WRITEABLE;
1102 new_spte &= ~shadow_accessed_mask;
1103 mmu_spte_clear_track_bits(spte);
1104 mmu_spte_set(spte, new_spte);
1105 spte = rmap_next(kvm, rmapp, spte);
1109 kvm_flush_remote_tlbs(kvm);
1114 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
1116 int (*handler)(struct kvm *kvm, unsigned long *rmapp,
1117 unsigned long data))
1122 struct kvm_memslots *slots;
1124 slots = kvm_memslots(kvm);
1126 for (i = 0; i < slots->nmemslots; i++) {
1127 struct kvm_memory_slot *memslot = &slots->memslots[i];
1128 unsigned long start = memslot->userspace_addr;
1131 end = start + (memslot->npages << PAGE_SHIFT);
1132 if (hva >= start && hva < end) {
1133 gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
1134 gfn_t gfn = memslot->base_gfn + gfn_offset;
1136 ret = handler(kvm, &memslot->rmap[gfn_offset], data);
1138 for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
1139 struct kvm_lpage_info *linfo;
1141 linfo = lpage_info_slot(gfn, memslot,
1142 PT_DIRECTORY_LEVEL + j);
1143 ret |= handler(kvm, &linfo->rmap_pde, data);
1145 trace_kvm_age_page(hva, memslot, ret);
1153 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1155 return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
1158 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1160 kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
1163 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1170 * Emulate the accessed bit for EPT, by checking if this page has
1171 * an EPT mapping, and clearing it if it does. On the next access,
1172 * a new EPT mapping will be established.
1173 * This has some overhead, but not as much as the cost of swapping
1174 * out actively used pages or breaking up actively used hugepages.
1176 if (!shadow_accessed_mask)
1177 return kvm_unmap_rmapp(kvm, rmapp, data);
1179 spte = rmap_next(kvm, rmapp, NULL);
1183 BUG_ON(!(_spte & PT_PRESENT_MASK));
1184 _young = _spte & PT_ACCESSED_MASK;
1187 clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
1189 spte = rmap_next(kvm, rmapp, spte);
1194 static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1201 * If there's no access bit in the secondary pte set by the
1202 * hardware it's up to gup-fast/gup to set the access bit in
1203 * the primary pte or in the page structure.
1205 if (!shadow_accessed_mask)
1208 spte = rmap_next(kvm, rmapp, NULL);
1211 BUG_ON(!(_spte & PT_PRESENT_MASK));
1212 young = _spte & PT_ACCESSED_MASK;
1217 spte = rmap_next(kvm, rmapp, spte);
1223 #define RMAP_RECYCLE_THRESHOLD 1000
1225 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
1227 unsigned long *rmapp;
1228 struct kvm_mmu_page *sp;
1230 sp = page_header(__pa(spte));
1232 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
1234 kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
1235 kvm_flush_remote_tlbs(vcpu->kvm);
1238 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
1240 return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
1243 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1245 return kvm_handle_hva(kvm, hva, 0, kvm_test_age_rmapp);
1249 static int is_empty_shadow_page(u64 *spt)
1254 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
1255 if (is_shadow_present_pte(*pos)) {
1256 printk(KERN_ERR "%s: %p %llx\n", __func__,
1265 * This value is the sum of all of the kvm instances's
1266 * kvm->arch.n_used_mmu_pages values. We need a global,
1267 * aggregate version in order to make the slab shrinker
1270 static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, int nr)
1272 kvm->arch.n_used_mmu_pages += nr;
1273 percpu_counter_add(&kvm_total_used_mmu_pages, nr);
1277 * Remove the sp from shadow page cache, after call it,
1278 * we can not find this sp from the cache, and the shadow
1279 * page table is still valid.
1280 * It should be under the protection of mmu lock.
1282 static void kvm_mmu_isolate_page(struct kvm_mmu_page *sp)
1284 ASSERT(is_empty_shadow_page(sp->spt));
1285 hlist_del(&sp->hash_link);
1286 if (!sp->role.direct)
1287 free_page((unsigned long)sp->gfns);
1291 * Free the shadow page table and the sp, we can do it
1292 * out of the protection of mmu lock.
1294 static void kvm_mmu_free_page(struct kvm_mmu_page *sp)
1296 list_del(&sp->link);
1297 free_page((unsigned long)sp->spt);
1298 kmem_cache_free(mmu_page_header_cache, sp);
1301 static unsigned kvm_page_table_hashfn(gfn_t gfn)
1303 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
1306 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
1307 struct kvm_mmu_page *sp, u64 *parent_pte)
1312 pte_list_add(vcpu, parent_pte, &sp->parent_ptes);
1315 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
1318 pte_list_remove(parent_pte, &sp->parent_ptes);
1321 static void drop_parent_pte(struct kvm_mmu_page *sp,
1324 mmu_page_remove_parent_pte(sp, parent_pte);
1325 mmu_spte_clear_no_track(parent_pte);
1328 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
1329 u64 *parent_pte, int direct)
1331 struct kvm_mmu_page *sp;
1332 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache,
1334 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
1336 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache,
1338 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
1339 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
1340 bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
1341 sp->parent_ptes = 0;
1342 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1343 kvm_mod_used_mmu_pages(vcpu->kvm, +1);
1347 static void mark_unsync(u64 *spte);
1348 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1350 pte_list_walk(&sp->parent_ptes, mark_unsync);
1353 static void mark_unsync(u64 *spte)
1355 struct kvm_mmu_page *sp;
1358 sp = page_header(__pa(spte));
1359 index = spte - sp->spt;
1360 if (__test_and_set_bit(index, sp->unsync_child_bitmap))
1362 if (sp->unsync_children++)
1364 kvm_mmu_mark_parents_unsync(sp);
1367 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1368 struct kvm_mmu_page *sp)
1373 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1377 static void nonpaging_update_pte(struct kvm_vcpu *vcpu,
1378 struct kvm_mmu_page *sp, u64 *spte,
1384 #define KVM_PAGE_ARRAY_NR 16
1386 struct kvm_mmu_pages {
1387 struct mmu_page_and_offset {
1388 struct kvm_mmu_page *sp;
1390 } page[KVM_PAGE_ARRAY_NR];
1394 #define for_each_unsync_children(bitmap, idx) \
1395 for (idx = find_first_bit(bitmap, 512); \
1397 idx = find_next_bit(bitmap, 512, idx+1))
1399 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1405 for (i=0; i < pvec->nr; i++)
1406 if (pvec->page[i].sp == sp)
1409 pvec->page[pvec->nr].sp = sp;
1410 pvec->page[pvec->nr].idx = idx;
1412 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1415 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1416 struct kvm_mmu_pages *pvec)
1418 int i, ret, nr_unsync_leaf = 0;
1420 for_each_unsync_children(sp->unsync_child_bitmap, i) {
1421 struct kvm_mmu_page *child;
1422 u64 ent = sp->spt[i];
1424 if (!is_shadow_present_pte(ent) || is_large_pte(ent))
1425 goto clear_child_bitmap;
1427 child = page_header(ent & PT64_BASE_ADDR_MASK);
1429 if (child->unsync_children) {
1430 if (mmu_pages_add(pvec, child, i))
1433 ret = __mmu_unsync_walk(child, pvec);
1435 goto clear_child_bitmap;
1437 nr_unsync_leaf += ret;
1440 } else if (child->unsync) {
1442 if (mmu_pages_add(pvec, child, i))
1445 goto clear_child_bitmap;
1450 __clear_bit(i, sp->unsync_child_bitmap);
1451 sp->unsync_children--;
1452 WARN_ON((int)sp->unsync_children < 0);
1456 return nr_unsync_leaf;
1459 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1460 struct kvm_mmu_pages *pvec)
1462 if (!sp->unsync_children)
1465 mmu_pages_add(pvec, sp, 0);
1466 return __mmu_unsync_walk(sp, pvec);
1469 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1471 WARN_ON(!sp->unsync);
1472 trace_kvm_mmu_sync_page(sp);
1474 --kvm->stat.mmu_unsync;
1477 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1478 struct list_head *invalid_list);
1479 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1480 struct list_head *invalid_list);
1482 #define for_each_gfn_sp(kvm, sp, gfn, pos) \
1483 hlist_for_each_entry(sp, pos, \
1484 &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link) \
1485 if ((sp)->gfn != (gfn)) {} else
1487 #define for_each_gfn_indirect_valid_sp(kvm, sp, gfn, pos) \
1488 hlist_for_each_entry(sp, pos, \
1489 &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link) \
1490 if ((sp)->gfn != (gfn) || (sp)->role.direct || \
1491 (sp)->role.invalid) {} else
1493 /* @sp->gfn should be write-protected at the call site */
1494 static int __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1495 struct list_head *invalid_list, bool clear_unsync)
1497 if (sp->role.cr4_pae != !!is_pae(vcpu)) {
1498 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1503 kvm_unlink_unsync_page(vcpu->kvm, sp);
1505 if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1506 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1510 kvm_mmu_flush_tlb(vcpu);
1514 static int kvm_sync_page_transient(struct kvm_vcpu *vcpu,
1515 struct kvm_mmu_page *sp)
1517 LIST_HEAD(invalid_list);
1520 ret = __kvm_sync_page(vcpu, sp, &invalid_list, false);
1522 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1527 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1528 struct list_head *invalid_list)
1530 return __kvm_sync_page(vcpu, sp, invalid_list, true);
1533 /* @gfn should be write-protected at the call site */
1534 static void kvm_sync_pages(struct kvm_vcpu *vcpu, gfn_t gfn)
1536 struct kvm_mmu_page *s;
1537 struct hlist_node *node;
1538 LIST_HEAD(invalid_list);
1541 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1545 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1546 kvm_unlink_unsync_page(vcpu->kvm, s);
1547 if ((s->role.cr4_pae != !!is_pae(vcpu)) ||
1548 (vcpu->arch.mmu.sync_page(vcpu, s))) {
1549 kvm_mmu_prepare_zap_page(vcpu->kvm, s, &invalid_list);
1555 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1557 kvm_mmu_flush_tlb(vcpu);
1560 struct mmu_page_path {
1561 struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1562 unsigned int idx[PT64_ROOT_LEVEL-1];
1565 #define for_each_sp(pvec, sp, parents, i) \
1566 for (i = mmu_pages_next(&pvec, &parents, -1), \
1567 sp = pvec.page[i].sp; \
1568 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1569 i = mmu_pages_next(&pvec, &parents, i))
1571 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1572 struct mmu_page_path *parents,
1577 for (n = i+1; n < pvec->nr; n++) {
1578 struct kvm_mmu_page *sp = pvec->page[n].sp;
1580 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1581 parents->idx[0] = pvec->page[n].idx;
1585 parents->parent[sp->role.level-2] = sp;
1586 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1592 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1594 struct kvm_mmu_page *sp;
1595 unsigned int level = 0;
1598 unsigned int idx = parents->idx[level];
1600 sp = parents->parent[level];
1604 --sp->unsync_children;
1605 WARN_ON((int)sp->unsync_children < 0);
1606 __clear_bit(idx, sp->unsync_child_bitmap);
1608 } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1611 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1612 struct mmu_page_path *parents,
1613 struct kvm_mmu_pages *pvec)
1615 parents->parent[parent->role.level-1] = NULL;
1619 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1620 struct kvm_mmu_page *parent)
1623 struct kvm_mmu_page *sp;
1624 struct mmu_page_path parents;
1625 struct kvm_mmu_pages pages;
1626 LIST_HEAD(invalid_list);
1628 kvm_mmu_pages_init(parent, &parents, &pages);
1629 while (mmu_unsync_walk(parent, &pages)) {
1632 for_each_sp(pages, sp, parents, i)
1633 protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1636 kvm_flush_remote_tlbs(vcpu->kvm);
1638 for_each_sp(pages, sp, parents, i) {
1639 kvm_sync_page(vcpu, sp, &invalid_list);
1640 mmu_pages_clear_parents(&parents);
1642 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1643 cond_resched_lock(&vcpu->kvm->mmu_lock);
1644 kvm_mmu_pages_init(parent, &parents, &pages);
1648 static void init_shadow_page_table(struct kvm_mmu_page *sp)
1652 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1656 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1664 union kvm_mmu_page_role role;
1666 struct kvm_mmu_page *sp;
1667 struct hlist_node *node;
1668 bool need_sync = false;
1670 role = vcpu->arch.mmu.base_role;
1672 role.direct = direct;
1675 role.access = access;
1676 if (!vcpu->arch.mmu.direct_map
1677 && vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1678 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1679 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1680 role.quadrant = quadrant;
1682 for_each_gfn_sp(vcpu->kvm, sp, gfn, node) {
1683 if (!need_sync && sp->unsync)
1686 if (sp->role.word != role.word)
1689 if (sp->unsync && kvm_sync_page_transient(vcpu, sp))
1692 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1693 if (sp->unsync_children) {
1694 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
1695 kvm_mmu_mark_parents_unsync(sp);
1696 } else if (sp->unsync)
1697 kvm_mmu_mark_parents_unsync(sp);
1699 trace_kvm_mmu_get_page(sp, false);
1702 ++vcpu->kvm->stat.mmu_cache_miss;
1703 sp = kvm_mmu_alloc_page(vcpu, parent_pte, direct);
1708 hlist_add_head(&sp->hash_link,
1709 &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
1711 if (rmap_write_protect(vcpu->kvm, gfn))
1712 kvm_flush_remote_tlbs(vcpu->kvm);
1713 if (level > PT_PAGE_TABLE_LEVEL && need_sync)
1714 kvm_sync_pages(vcpu, gfn);
1716 account_shadowed(vcpu->kvm, gfn);
1718 init_shadow_page_table(sp);
1719 trace_kvm_mmu_get_page(sp, true);
1723 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1724 struct kvm_vcpu *vcpu, u64 addr)
1726 iterator->addr = addr;
1727 iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1728 iterator->level = vcpu->arch.mmu.shadow_root_level;
1730 if (iterator->level == PT64_ROOT_LEVEL &&
1731 vcpu->arch.mmu.root_level < PT64_ROOT_LEVEL &&
1732 !vcpu->arch.mmu.direct_map)
1735 if (iterator->level == PT32E_ROOT_LEVEL) {
1736 iterator->shadow_addr
1737 = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1738 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1740 if (!iterator->shadow_addr)
1741 iterator->level = 0;
1745 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1747 if (iterator->level < PT_PAGE_TABLE_LEVEL)
1750 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1751 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1755 static void __shadow_walk_next(struct kvm_shadow_walk_iterator *iterator,
1758 if (is_last_spte(spte, iterator->level)) {
1759 iterator->level = 0;
1763 iterator->shadow_addr = spte & PT64_BASE_ADDR_MASK;
1767 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1769 return __shadow_walk_next(iterator, *iterator->sptep);
1772 static void link_shadow_page(u64 *sptep, struct kvm_mmu_page *sp)
1776 spte = __pa(sp->spt)
1777 | PT_PRESENT_MASK | PT_ACCESSED_MASK
1778 | PT_WRITABLE_MASK | PT_USER_MASK;
1779 mmu_spte_set(sptep, spte);
1782 static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1784 if (is_large_pte(*sptep)) {
1785 drop_spte(vcpu->kvm, sptep);
1786 kvm_flush_remote_tlbs(vcpu->kvm);
1790 static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1791 unsigned direct_access)
1793 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
1794 struct kvm_mmu_page *child;
1797 * For the direct sp, if the guest pte's dirty bit
1798 * changed form clean to dirty, it will corrupt the
1799 * sp's access: allow writable in the read-only sp,
1800 * so we should update the spte at this point to get
1801 * a new sp with the correct access.
1803 child = page_header(*sptep & PT64_BASE_ADDR_MASK);
1804 if (child->role.access == direct_access)
1807 drop_parent_pte(child, sptep);
1808 kvm_flush_remote_tlbs(vcpu->kvm);
1812 static bool mmu_page_zap_pte(struct kvm *kvm, struct kvm_mmu_page *sp,
1816 struct kvm_mmu_page *child;
1819 if (is_shadow_present_pte(pte)) {
1820 if (is_last_spte(pte, sp->role.level)) {
1821 drop_spte(kvm, spte);
1822 if (is_large_pte(pte))
1825 child = page_header(pte & PT64_BASE_ADDR_MASK);
1826 drop_parent_pte(child, spte);
1831 if (is_mmio_spte(pte))
1832 mmu_spte_clear_no_track(spte);
1837 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1838 struct kvm_mmu_page *sp)
1842 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1843 mmu_page_zap_pte(kvm, sp, sp->spt + i);
1846 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1848 mmu_page_remove_parent_pte(sp, parent_pte);
1851 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1854 struct kvm_vcpu *vcpu;
1856 kvm_for_each_vcpu(i, vcpu, kvm)
1857 vcpu->arch.last_pte_updated = NULL;
1860 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1864 while ((parent_pte = pte_list_next(&sp->parent_ptes, NULL)))
1865 drop_parent_pte(sp, parent_pte);
1868 static int mmu_zap_unsync_children(struct kvm *kvm,
1869 struct kvm_mmu_page *parent,
1870 struct list_head *invalid_list)
1873 struct mmu_page_path parents;
1874 struct kvm_mmu_pages pages;
1876 if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1879 kvm_mmu_pages_init(parent, &parents, &pages);
1880 while (mmu_unsync_walk(parent, &pages)) {
1881 struct kvm_mmu_page *sp;
1883 for_each_sp(pages, sp, parents, i) {
1884 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
1885 mmu_pages_clear_parents(&parents);
1888 kvm_mmu_pages_init(parent, &parents, &pages);
1894 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1895 struct list_head *invalid_list)
1899 trace_kvm_mmu_prepare_zap_page(sp);
1900 ++kvm->stat.mmu_shadow_zapped;
1901 ret = mmu_zap_unsync_children(kvm, sp, invalid_list);
1902 kvm_mmu_page_unlink_children(kvm, sp);
1903 kvm_mmu_unlink_parents(kvm, sp);
1904 if (!sp->role.invalid && !sp->role.direct)
1905 unaccount_shadowed(kvm, sp->gfn);
1907 kvm_unlink_unsync_page(kvm, sp);
1908 if (!sp->root_count) {
1911 list_move(&sp->link, invalid_list);
1912 kvm_mod_used_mmu_pages(kvm, -1);
1914 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1915 kvm_reload_remote_mmus(kvm);
1918 sp->role.invalid = 1;
1919 kvm_mmu_reset_last_pte_updated(kvm);
1923 static void kvm_mmu_isolate_pages(struct list_head *invalid_list)
1925 struct kvm_mmu_page *sp;
1927 list_for_each_entry(sp, invalid_list, link)
1928 kvm_mmu_isolate_page(sp);
1931 static void free_pages_rcu(struct rcu_head *head)
1933 struct kvm_mmu_page *next, *sp;
1935 sp = container_of(head, struct kvm_mmu_page, rcu);
1937 if (!list_empty(&sp->link))
1938 next = list_first_entry(&sp->link,
1939 struct kvm_mmu_page, link);
1942 kvm_mmu_free_page(sp);
1947 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1948 struct list_head *invalid_list)
1950 struct kvm_mmu_page *sp;
1952 if (list_empty(invalid_list))
1955 kvm_flush_remote_tlbs(kvm);
1957 if (atomic_read(&kvm->arch.reader_counter)) {
1958 kvm_mmu_isolate_pages(invalid_list);
1959 sp = list_first_entry(invalid_list, struct kvm_mmu_page, link);
1960 list_del_init(invalid_list);
1962 trace_kvm_mmu_delay_free_pages(sp);
1963 call_rcu(&sp->rcu, free_pages_rcu);
1968 sp = list_first_entry(invalid_list, struct kvm_mmu_page, link);
1969 WARN_ON(!sp->role.invalid || sp->root_count);
1970 kvm_mmu_isolate_page(sp);
1971 kvm_mmu_free_page(sp);
1972 } while (!list_empty(invalid_list));
1977 * Changing the number of mmu pages allocated to the vm
1978 * Note: if goal_nr_mmu_pages is too small, you will get dead lock
1980 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int goal_nr_mmu_pages)
1982 LIST_HEAD(invalid_list);
1984 * If we set the number of mmu pages to be smaller be than the
1985 * number of actived pages , we must to free some mmu pages before we
1989 if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) {
1990 while (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages &&
1991 !list_empty(&kvm->arch.active_mmu_pages)) {
1992 struct kvm_mmu_page *page;
1994 page = container_of(kvm->arch.active_mmu_pages.prev,
1995 struct kvm_mmu_page, link);
1996 kvm_mmu_prepare_zap_page(kvm, page, &invalid_list);
1998 kvm_mmu_commit_zap_page(kvm, &invalid_list);
1999 goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages;
2002 kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages;
2005 int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
2007 struct kvm_mmu_page *sp;
2008 struct hlist_node *node;
2009 LIST_HEAD(invalid_list);
2012 pgprintk("%s: looking for gfn %llx\n", __func__, gfn);
2014 spin_lock(&kvm->mmu_lock);
2015 for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
2016 pgprintk("%s: gfn %llx role %x\n", __func__, gfn,
2019 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
2021 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2022 spin_unlock(&kvm->mmu_lock);
2026 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page);
2028 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
2030 struct kvm_mmu_page *sp;
2031 struct hlist_node *node;
2032 LIST_HEAD(invalid_list);
2034 for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
2035 pgprintk("%s: zap %llx %x\n",
2036 __func__, gfn, sp->role.word);
2037 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
2039 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2042 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
2044 int slot = memslot_id(kvm, gfn);
2045 struct kvm_mmu_page *sp = page_header(__pa(pte));
2047 __set_bit(slot, sp->slot_bitmap);
2051 * The function is based on mtrr_type_lookup() in
2052 * arch/x86/kernel/cpu/mtrr/generic.c
2054 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
2059 u8 prev_match, curr_match;
2060 int num_var_ranges = KVM_NR_VAR_MTRR;
2062 if (!mtrr_state->enabled)
2065 /* Make end inclusive end, instead of exclusive */
2068 /* Look in fixed ranges. Just return the type as per start */
2069 if (mtrr_state->have_fixed && (start < 0x100000)) {
2072 if (start < 0x80000) {
2074 idx += (start >> 16);
2075 return mtrr_state->fixed_ranges[idx];
2076 } else if (start < 0xC0000) {
2078 idx += ((start - 0x80000) >> 14);
2079 return mtrr_state->fixed_ranges[idx];
2080 } else if (start < 0x1000000) {
2082 idx += ((start - 0xC0000) >> 12);
2083 return mtrr_state->fixed_ranges[idx];
2088 * Look in variable ranges
2089 * Look of multiple ranges matching this address and pick type
2090 * as per MTRR precedence
2092 if (!(mtrr_state->enabled & 2))
2093 return mtrr_state->def_type;
2096 for (i = 0; i < num_var_ranges; ++i) {
2097 unsigned short start_state, end_state;
2099 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
2102 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
2103 (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
2104 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
2105 (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
2107 start_state = ((start & mask) == (base & mask));
2108 end_state = ((end & mask) == (base & mask));
2109 if (start_state != end_state)
2112 if ((start & mask) != (base & mask))
2115 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
2116 if (prev_match == 0xFF) {
2117 prev_match = curr_match;
2121 if (prev_match == MTRR_TYPE_UNCACHABLE ||
2122 curr_match == MTRR_TYPE_UNCACHABLE)
2123 return MTRR_TYPE_UNCACHABLE;
2125 if ((prev_match == MTRR_TYPE_WRBACK &&
2126 curr_match == MTRR_TYPE_WRTHROUGH) ||
2127 (prev_match == MTRR_TYPE_WRTHROUGH &&
2128 curr_match == MTRR_TYPE_WRBACK)) {
2129 prev_match = MTRR_TYPE_WRTHROUGH;
2130 curr_match = MTRR_TYPE_WRTHROUGH;
2133 if (prev_match != curr_match)
2134 return MTRR_TYPE_UNCACHABLE;
2137 if (prev_match != 0xFF)
2140 return mtrr_state->def_type;
2143 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
2147 mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
2148 (gfn << PAGE_SHIFT) + PAGE_SIZE);
2149 if (mtrr == 0xfe || mtrr == 0xff)
2150 mtrr = MTRR_TYPE_WRBACK;
2153 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
2155 static void __kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
2157 trace_kvm_mmu_unsync_page(sp);
2158 ++vcpu->kvm->stat.mmu_unsync;
2161 kvm_mmu_mark_parents_unsync(sp);
2164 static void kvm_unsync_pages(struct kvm_vcpu *vcpu, gfn_t gfn)
2166 struct kvm_mmu_page *s;
2167 struct hlist_node *node;
2169 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
2172 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
2173 __kvm_unsync_page(vcpu, s);
2177 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
2180 struct kvm_mmu_page *s;
2181 struct hlist_node *node;
2182 bool need_unsync = false;
2184 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
2188 if (s->role.level != PT_PAGE_TABLE_LEVEL)
2191 if (!need_unsync && !s->unsync) {
2198 kvm_unsync_pages(vcpu, gfn);
2202 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2203 unsigned pte_access, int user_fault,
2204 int write_fault, int level,
2205 gfn_t gfn, pfn_t pfn, bool speculative,
2206 bool can_unsync, bool host_writable)
2208 u64 spte, entry = *sptep;
2211 if (set_mmio_spte(sptep, gfn, pfn, pte_access))
2214 spte = PT_PRESENT_MASK;
2216 spte |= shadow_accessed_mask;
2218 if (pte_access & ACC_EXEC_MASK)
2219 spte |= shadow_x_mask;
2221 spte |= shadow_nx_mask;
2222 if (pte_access & ACC_USER_MASK)
2223 spte |= shadow_user_mask;
2224 if (level > PT_PAGE_TABLE_LEVEL)
2225 spte |= PT_PAGE_SIZE_MASK;
2227 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
2228 kvm_is_mmio_pfn(pfn));
2231 spte |= SPTE_HOST_WRITEABLE;
2233 pte_access &= ~ACC_WRITE_MASK;
2235 spte |= (u64)pfn << PAGE_SHIFT;
2237 if ((pte_access & ACC_WRITE_MASK)
2238 || (!vcpu->arch.mmu.direct_map && write_fault
2239 && !is_write_protection(vcpu) && !user_fault)) {
2241 if (level > PT_PAGE_TABLE_LEVEL &&
2242 has_wrprotected_page(vcpu->kvm, gfn, level)) {
2244 drop_spte(vcpu->kvm, sptep);
2248 spte |= PT_WRITABLE_MASK;
2250 if (!vcpu->arch.mmu.direct_map
2251 && !(pte_access & ACC_WRITE_MASK)) {
2252 spte &= ~PT_USER_MASK;
2254 * If we converted a user page to a kernel page,
2255 * so that the kernel can write to it when cr0.wp=0,
2256 * then we should prevent the kernel from executing it
2257 * if SMEP is enabled.
2259 if (kvm_read_cr4_bits(vcpu, X86_CR4_SMEP))
2260 spte |= PT64_NX_MASK;
2264 * Optimization: for pte sync, if spte was writable the hash
2265 * lookup is unnecessary (and expensive). Write protection
2266 * is responsibility of mmu_get_page / kvm_sync_page.
2267 * Same reasoning can be applied to dirty page accounting.
2269 if (!can_unsync && is_writable_pte(*sptep))
2272 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
2273 pgprintk("%s: found shadow page for %llx, marking ro\n",
2276 pte_access &= ~ACC_WRITE_MASK;
2277 if (is_writable_pte(spte))
2278 spte &= ~PT_WRITABLE_MASK;
2282 if (pte_access & ACC_WRITE_MASK)
2283 mark_page_dirty(vcpu->kvm, gfn);
2286 mmu_spte_update(sptep, spte);
2288 * If we overwrite a writable spte with a read-only one we
2289 * should flush remote TLBs. Otherwise rmap_write_protect
2290 * will find a read-only spte, even though the writable spte
2291 * might be cached on a CPU's TLB.
2293 if (is_writable_pte(entry) && !is_writable_pte(*sptep))
2294 kvm_flush_remote_tlbs(vcpu->kvm);
2299 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2300 unsigned pt_access, unsigned pte_access,
2301 int user_fault, int write_fault,
2302 int *emulate, int level, gfn_t gfn,
2303 pfn_t pfn, bool speculative,
2306 int was_rmapped = 0;
2309 pgprintk("%s: spte %llx access %x write_fault %d"
2310 " user_fault %d gfn %llx\n",
2311 __func__, *sptep, pt_access,
2312 write_fault, user_fault, gfn);
2314 if (is_rmap_spte(*sptep)) {
2316 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
2317 * the parent of the now unreachable PTE.
2319 if (level > PT_PAGE_TABLE_LEVEL &&
2320 !is_large_pte(*sptep)) {
2321 struct kvm_mmu_page *child;
2324 child = page_header(pte & PT64_BASE_ADDR_MASK);
2325 drop_parent_pte(child, sptep);
2326 kvm_flush_remote_tlbs(vcpu->kvm);
2327 } else if (pfn != spte_to_pfn(*sptep)) {
2328 pgprintk("hfn old %llx new %llx\n",
2329 spte_to_pfn(*sptep), pfn);
2330 drop_spte(vcpu->kvm, sptep);
2331 kvm_flush_remote_tlbs(vcpu->kvm);
2336 if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
2337 level, gfn, pfn, speculative, true,
2341 kvm_mmu_flush_tlb(vcpu);
2344 if (unlikely(is_mmio_spte(*sptep) && emulate))
2347 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
2348 pgprintk("instantiating %s PTE (%s) at %llx (%llx) addr %p\n",
2349 is_large_pte(*sptep)? "2MB" : "4kB",
2350 *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
2352 if (!was_rmapped && is_large_pte(*sptep))
2353 ++vcpu->kvm->stat.lpages;
2355 if (is_shadow_present_pte(*sptep)) {
2356 page_header_update_slot(vcpu->kvm, sptep, gfn);
2358 rmap_count = rmap_add(vcpu, sptep, gfn);
2359 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
2360 rmap_recycle(vcpu, sptep, gfn);
2363 kvm_release_pfn_clean(pfn);
2365 vcpu->arch.last_pte_updated = sptep;
2368 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
2372 static pfn_t pte_prefetch_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn,
2375 struct kvm_memory_slot *slot;
2378 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, no_dirty_log);
2380 get_page(fault_page);
2381 return page_to_pfn(fault_page);
2384 hva = gfn_to_hva_memslot(slot, gfn);
2386 return hva_to_pfn_atomic(vcpu->kvm, hva);
2389 static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu,
2390 struct kvm_mmu_page *sp,
2391 u64 *start, u64 *end)
2393 struct page *pages[PTE_PREFETCH_NUM];
2394 unsigned access = sp->role.access;
2398 gfn = kvm_mmu_page_get_gfn(sp, start - sp->spt);
2399 if (!gfn_to_memslot_dirty_bitmap(vcpu, gfn, access & ACC_WRITE_MASK))
2402 ret = gfn_to_page_many_atomic(vcpu->kvm, gfn, pages, end - start);
2406 for (i = 0; i < ret; i++, gfn++, start++)
2407 mmu_set_spte(vcpu, start, ACC_ALL,
2409 sp->role.level, gfn,
2410 page_to_pfn(pages[i]), true, true);
2415 static void __direct_pte_prefetch(struct kvm_vcpu *vcpu,
2416 struct kvm_mmu_page *sp, u64 *sptep)
2418 u64 *spte, *start = NULL;
2421 WARN_ON(!sp->role.direct);
2423 i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
2426 for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
2427 if (is_shadow_present_pte(*spte) || spte == sptep) {
2430 if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0)
2438 static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
2440 struct kvm_mmu_page *sp;
2443 * Since it's no accessed bit on EPT, it's no way to
2444 * distinguish between actually accessed translations
2445 * and prefetched, so disable pte prefetch if EPT is
2448 if (!shadow_accessed_mask)
2451 sp = page_header(__pa(sptep));
2452 if (sp->role.level > PT_PAGE_TABLE_LEVEL)
2455 __direct_pte_prefetch(vcpu, sp, sptep);
2458 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
2459 int map_writable, int level, gfn_t gfn, pfn_t pfn,
2462 struct kvm_shadow_walk_iterator iterator;
2463 struct kvm_mmu_page *sp;
2467 for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
2468 if (iterator.level == level) {
2469 unsigned pte_access = ACC_ALL;
2471 mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, pte_access,
2473 level, gfn, pfn, prefault, map_writable);
2474 direct_pte_prefetch(vcpu, iterator.sptep);
2475 ++vcpu->stat.pf_fixed;
2479 if (!is_shadow_present_pte(*iterator.sptep)) {
2480 u64 base_addr = iterator.addr;
2482 base_addr &= PT64_LVL_ADDR_MASK(iterator.level);
2483 pseudo_gfn = base_addr >> PAGE_SHIFT;
2484 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
2486 1, ACC_ALL, iterator.sptep);
2488 pgprintk("nonpaging_map: ENOMEM\n");
2489 kvm_release_pfn_clean(pfn);
2493 mmu_spte_set(iterator.sptep,
2495 | PT_PRESENT_MASK | PT_WRITABLE_MASK
2496 | shadow_user_mask | shadow_x_mask
2497 | shadow_accessed_mask);
2503 static void kvm_send_hwpoison_signal(unsigned long address, struct task_struct *tsk)
2507 info.si_signo = SIGBUS;
2509 info.si_code = BUS_MCEERR_AR;
2510 info.si_addr = (void __user *)address;
2511 info.si_addr_lsb = PAGE_SHIFT;
2513 send_sig_info(SIGBUS, &info, tsk);
2516 static int kvm_handle_bad_page(struct kvm_vcpu *vcpu, gfn_t gfn, pfn_t pfn)
2518 kvm_release_pfn_clean(pfn);
2519 if (is_hwpoison_pfn(pfn)) {
2520 kvm_send_hwpoison_signal(gfn_to_hva(vcpu->kvm, gfn), current);
2527 static void transparent_hugepage_adjust(struct kvm_vcpu *vcpu,
2528 gfn_t *gfnp, pfn_t *pfnp, int *levelp)
2532 int level = *levelp;
2535 * Check if it's a transparent hugepage. If this would be an
2536 * hugetlbfs page, level wouldn't be set to
2537 * PT_PAGE_TABLE_LEVEL and there would be no adjustment done
2540 if (!is_error_pfn(pfn) && !kvm_is_mmio_pfn(pfn) &&
2541 level == PT_PAGE_TABLE_LEVEL &&
2542 PageTransCompound(pfn_to_page(pfn)) &&
2543 !has_wrprotected_page(vcpu->kvm, gfn, PT_DIRECTORY_LEVEL)) {
2546 * mmu_notifier_retry was successful and we hold the
2547 * mmu_lock here, so the pmd can't become splitting
2548 * from under us, and in turn
2549 * __split_huge_page_refcount() can't run from under
2550 * us and we can safely transfer the refcount from
2551 * PG_tail to PG_head as we switch the pfn to tail to
2554 *levelp = level = PT_DIRECTORY_LEVEL;
2555 mask = KVM_PAGES_PER_HPAGE(level) - 1;
2556 VM_BUG_ON((gfn & mask) != (pfn & mask));
2560 kvm_release_pfn_clean(pfn);
2562 if (!get_page_unless_zero(pfn_to_page(pfn)))
2569 static bool mmu_invalid_pfn(pfn_t pfn)
2571 return unlikely(is_invalid_pfn(pfn));
2574 static bool handle_abnormal_pfn(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn,
2575 pfn_t pfn, unsigned access, int *ret_val)
2579 /* The pfn is invalid, report the error! */
2580 if (unlikely(is_invalid_pfn(pfn))) {
2581 *ret_val = kvm_handle_bad_page(vcpu, gfn, pfn);
2585 if (unlikely(is_noslot_pfn(pfn)))
2586 vcpu_cache_mmio_info(vcpu, gva, gfn, access);
2593 static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
2594 gva_t gva, pfn_t *pfn, bool write, bool *writable);
2596 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn,
2603 unsigned long mmu_seq;
2606 force_pt_level = mapping_level_dirty_bitmap(vcpu, gfn);
2607 if (likely(!force_pt_level)) {
2608 level = mapping_level(vcpu, gfn);
2610 * This path builds a PAE pagetable - so we can map
2611 * 2mb pages at maximum. Therefore check if the level
2612 * is larger than that.
2614 if (level > PT_DIRECTORY_LEVEL)
2615 level = PT_DIRECTORY_LEVEL;
2617 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2619 level = PT_PAGE_TABLE_LEVEL;
2621 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2624 if (try_async_pf(vcpu, prefault, gfn, v, &pfn, write, &map_writable))
2627 if (handle_abnormal_pfn(vcpu, v, gfn, pfn, ACC_ALL, &r))
2630 spin_lock(&vcpu->kvm->mmu_lock);
2631 if (mmu_notifier_retry(vcpu, mmu_seq))
2633 kvm_mmu_free_some_pages(vcpu);
2634 if (likely(!force_pt_level))
2635 transparent_hugepage_adjust(vcpu, &gfn, &pfn, &level);
2636 r = __direct_map(vcpu, v, write, map_writable, level, gfn, pfn,
2638 spin_unlock(&vcpu->kvm->mmu_lock);
2644 spin_unlock(&vcpu->kvm->mmu_lock);
2645 kvm_release_pfn_clean(pfn);
2650 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2653 struct kvm_mmu_page *sp;
2654 LIST_HEAD(invalid_list);
2656 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2658 spin_lock(&vcpu->kvm->mmu_lock);
2659 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL &&
2660 (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL ||
2661 vcpu->arch.mmu.direct_map)) {
2662 hpa_t root = vcpu->arch.mmu.root_hpa;
2664 sp = page_header(root);
2666 if (!sp->root_count && sp->role.invalid) {
2667 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
2668 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2670 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2671 spin_unlock(&vcpu->kvm->mmu_lock);
2674 for (i = 0; i < 4; ++i) {
2675 hpa_t root = vcpu->arch.mmu.pae_root[i];
2678 root &= PT64_BASE_ADDR_MASK;
2679 sp = page_header(root);
2681 if (!sp->root_count && sp->role.invalid)
2682 kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2685 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2687 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2688 spin_unlock(&vcpu->kvm->mmu_lock);
2689 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2692 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2696 if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2697 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
2704 static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
2706 struct kvm_mmu_page *sp;
2709 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2710 spin_lock(&vcpu->kvm->mmu_lock);
2711 kvm_mmu_free_some_pages(vcpu);
2712 sp = kvm_mmu_get_page(vcpu, 0, 0, PT64_ROOT_LEVEL,
2715 spin_unlock(&vcpu->kvm->mmu_lock);
2716 vcpu->arch.mmu.root_hpa = __pa(sp->spt);
2717 } else if (vcpu->arch.mmu.shadow_root_level == PT32E_ROOT_LEVEL) {
2718 for (i = 0; i < 4; ++i) {
2719 hpa_t root = vcpu->arch.mmu.pae_root[i];
2721 ASSERT(!VALID_PAGE(root));
2722 spin_lock(&vcpu->kvm->mmu_lock);
2723 kvm_mmu_free_some_pages(vcpu);
2724 sp = kvm_mmu_get_page(vcpu, i << (30 - PAGE_SHIFT),
2726 PT32_ROOT_LEVEL, 1, ACC_ALL,
2728 root = __pa(sp->spt);
2730 spin_unlock(&vcpu->kvm->mmu_lock);
2731 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2733 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2740 static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
2742 struct kvm_mmu_page *sp;
2747 root_gfn = vcpu->arch.mmu.get_cr3(vcpu) >> PAGE_SHIFT;
2749 if (mmu_check_root(vcpu, root_gfn))
2753 * Do we shadow a long mode page table? If so we need to
2754 * write-protect the guests page table root.
2756 if (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL) {
2757 hpa_t root = vcpu->arch.mmu.root_hpa;
2759 ASSERT(!VALID_PAGE(root));
2761 spin_lock(&vcpu->kvm->mmu_lock);
2762 kvm_mmu_free_some_pages(vcpu);
2763 sp = kvm_mmu_get_page(vcpu, root_gfn, 0, PT64_ROOT_LEVEL,
2765 root = __pa(sp->spt);
2767 spin_unlock(&vcpu->kvm->mmu_lock);
2768 vcpu->arch.mmu.root_hpa = root;
2773 * We shadow a 32 bit page table. This may be a legacy 2-level
2774 * or a PAE 3-level page table. In either case we need to be aware that
2775 * the shadow page table may be a PAE or a long mode page table.
2777 pm_mask = PT_PRESENT_MASK;
2778 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL)
2779 pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK;
2781 for (i = 0; i < 4; ++i) {
2782 hpa_t root = vcpu->arch.mmu.pae_root[i];
2784 ASSERT(!VALID_PAGE(root));
2785 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2786 pdptr = vcpu->arch.mmu.get_pdptr(vcpu, i);
2787 if (!is_present_gpte(pdptr)) {
2788 vcpu->arch.mmu.pae_root[i] = 0;
2791 root_gfn = pdptr >> PAGE_SHIFT;
2792 if (mmu_check_root(vcpu, root_gfn))
2795 spin_lock(&vcpu->kvm->mmu_lock);
2796 kvm_mmu_free_some_pages(vcpu);
2797 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2800 root = __pa(sp->spt);
2802 spin_unlock(&vcpu->kvm->mmu_lock);
2804 vcpu->arch.mmu.pae_root[i] = root | pm_mask;
2806 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2809 * If we shadow a 32 bit page table with a long mode page
2810 * table we enter this path.
2812 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2813 if (vcpu->arch.mmu.lm_root == NULL) {
2815 * The additional page necessary for this is only
2816 * allocated on demand.
2821 lm_root = (void*)get_zeroed_page(GFP_KERNEL);
2822 if (lm_root == NULL)
2825 lm_root[0] = __pa(vcpu->arch.mmu.pae_root) | pm_mask;
2827 vcpu->arch.mmu.lm_root = lm_root;
2830 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.lm_root);
2836 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2838 if (vcpu->arch.mmu.direct_map)
2839 return mmu_alloc_direct_roots(vcpu);
2841 return mmu_alloc_shadow_roots(vcpu);
2844 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2847 struct kvm_mmu_page *sp;
2849 if (vcpu->arch.mmu.direct_map)
2852 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2855 vcpu_clear_mmio_info(vcpu, ~0ul);
2856 trace_kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
2857 if (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL) {
2858 hpa_t root = vcpu->arch.mmu.root_hpa;
2859 sp = page_header(root);
2860 mmu_sync_children(vcpu, sp);
2861 trace_kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
2864 for (i = 0; i < 4; ++i) {
2865 hpa_t root = vcpu->arch.mmu.pae_root[i];
2867 if (root && VALID_PAGE(root)) {
2868 root &= PT64_BASE_ADDR_MASK;
2869 sp = page_header(root);
2870 mmu_sync_children(vcpu, sp);
2873 trace_kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
2876 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2878 spin_lock(&vcpu->kvm->mmu_lock);
2879 mmu_sync_roots(vcpu);
2880 spin_unlock(&vcpu->kvm->mmu_lock);
2883 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2884 u32 access, struct x86_exception *exception)
2887 exception->error_code = 0;
2891 static gpa_t nonpaging_gva_to_gpa_nested(struct kvm_vcpu *vcpu, gva_t vaddr,
2893 struct x86_exception *exception)
2896 exception->error_code = 0;
2897 return vcpu->arch.nested_mmu.translate_gpa(vcpu, vaddr, access);
2900 static bool quickly_check_mmio_pf(struct kvm_vcpu *vcpu, u64 addr, bool direct)
2903 return vcpu_match_mmio_gpa(vcpu, addr);
2905 return vcpu_match_mmio_gva(vcpu, addr);
2910 * On direct hosts, the last spte is only allows two states
2911 * for mmio page fault:
2912 * - It is the mmio spte
2913 * - It is zapped or it is being zapped.
2915 * This function completely checks the spte when the last spte
2916 * is not the mmio spte.
2918 static bool check_direct_spte_mmio_pf(u64 spte)
2920 return __check_direct_spte_mmio_pf(spte);
2923 static u64 walk_shadow_page_get_mmio_spte(struct kvm_vcpu *vcpu, u64 addr)
2925 struct kvm_shadow_walk_iterator iterator;
2928 walk_shadow_page_lockless_begin(vcpu);
2929 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte)
2930 if (!is_shadow_present_pte(spte))
2932 walk_shadow_page_lockless_end(vcpu);
2938 * If it is a real mmio page fault, return 1 and emulat the instruction
2939 * directly, return 0 to let CPU fault again on the address, -1 is
2940 * returned if bug is detected.
2942 int handle_mmio_page_fault_common(struct kvm_vcpu *vcpu, u64 addr, bool direct)
2946 if (quickly_check_mmio_pf(vcpu, addr, direct))
2949 spte = walk_shadow_page_get_mmio_spte(vcpu, addr);
2951 if (is_mmio_spte(spte)) {
2952 gfn_t gfn = get_mmio_spte_gfn(spte);
2953 unsigned access = get_mmio_spte_access(spte);
2958 trace_handle_mmio_page_fault(addr, gfn, access);
2959 vcpu_cache_mmio_info(vcpu, addr, gfn, access);
2964 * It's ok if the gva is remapped by other cpus on shadow guest,
2965 * it's a BUG if the gfn is not a mmio page.
2967 if (direct && !check_direct_spte_mmio_pf(spte))
2971 * If the page table is zapped by other cpus, let CPU fault again on
2976 EXPORT_SYMBOL_GPL(handle_mmio_page_fault_common);
2978 static int handle_mmio_page_fault(struct kvm_vcpu *vcpu, u64 addr,
2979 u32 error_code, bool direct)
2983 ret = handle_mmio_page_fault_common(vcpu, addr, direct);
2988 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2989 u32 error_code, bool prefault)
2994 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2996 if (unlikely(error_code & PFERR_RSVD_MASK))
2997 return handle_mmio_page_fault(vcpu, gva, error_code, true);
2999 r = mmu_topup_memory_caches(vcpu);
3004 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
3006 gfn = gva >> PAGE_SHIFT;
3008 return nonpaging_map(vcpu, gva & PAGE_MASK,
3009 error_code & PFERR_WRITE_MASK, gfn, prefault);
3012 static int kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn)
3014 struct kvm_arch_async_pf arch;
3016 arch.token = (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id;
3018 arch.direct_map = vcpu->arch.mmu.direct_map;
3019 arch.cr3 = vcpu->arch.mmu.get_cr3(vcpu);
3021 return kvm_setup_async_pf(vcpu, gva, gfn, &arch);
3024 static bool can_do_async_pf(struct kvm_vcpu *vcpu)
3026 if (unlikely(!irqchip_in_kernel(vcpu->kvm) ||
3027 kvm_event_needs_reinjection(vcpu)))
3030 return kvm_x86_ops->interrupt_allowed(vcpu);
3033 static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
3034 gva_t gva, pfn_t *pfn, bool write, bool *writable)
3038 *pfn = gfn_to_pfn_async(vcpu->kvm, gfn, &async, write, writable);
3041 return false; /* *pfn has correct page already */
3043 put_page(pfn_to_page(*pfn));
3045 if (!prefault && can_do_async_pf(vcpu)) {
3046 trace_kvm_try_async_get_page(gva, gfn);
3047 if (kvm_find_async_pf_gfn(vcpu, gfn)) {
3048 trace_kvm_async_pf_doublefault(gva, gfn);
3049 kvm_make_request(KVM_REQ_APF_HALT, vcpu);
3051 } else if (kvm_arch_setup_async_pf(vcpu, gva, gfn))
3055 *pfn = gfn_to_pfn_prot(vcpu->kvm, gfn, write, writable);
3060 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa, u32 error_code,
3067 gfn_t gfn = gpa >> PAGE_SHIFT;
3068 unsigned long mmu_seq;
3069 int write = error_code & PFERR_WRITE_MASK;
3073 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
3075 if (unlikely(error_code & PFERR_RSVD_MASK))
3076 return handle_mmio_page_fault(vcpu, gpa, error_code, true);
3078 r = mmu_topup_memory_caches(vcpu);
3082 force_pt_level = mapping_level_dirty_bitmap(vcpu, gfn);
3083 if (likely(!force_pt_level)) {
3084 level = mapping_level(vcpu, gfn);
3085 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
3087 level = PT_PAGE_TABLE_LEVEL;
3089 mmu_seq = vcpu->kvm->mmu_notifier_seq;
3092 if (try_async_pf(vcpu, prefault, gfn, gpa, &pfn, write, &map_writable))
3095 if (handle_abnormal_pfn(vcpu, 0, gfn, pfn, ACC_ALL, &r))
3098 spin_lock(&vcpu->kvm->mmu_lock);
3099 if (mmu_notifier_retry(vcpu, mmu_seq))
3101 kvm_mmu_free_some_pages(vcpu);
3102 if (likely(!force_pt_level))
3103 transparent_hugepage_adjust(vcpu, &gfn, &pfn, &level);
3104 r = __direct_map(vcpu, gpa, write, map_writable,
3105 level, gfn, pfn, prefault);
3106 spin_unlock(&vcpu->kvm->mmu_lock);
3111 spin_unlock(&vcpu->kvm->mmu_lock);
3112 kvm_release_pfn_clean(pfn);
3116 static void nonpaging_free(struct kvm_vcpu *vcpu)
3118 mmu_free_roots(vcpu);
3121 static int nonpaging_init_context(struct kvm_vcpu *vcpu,
3122 struct kvm_mmu *context)
3124 context->new_cr3 = nonpaging_new_cr3;
3125 context->page_fault = nonpaging_page_fault;
3126 context->gva_to_gpa = nonpaging_gva_to_gpa;
3127 context->free = nonpaging_free;
3128 context->sync_page = nonpaging_sync_page;
3129 context->invlpg = nonpaging_invlpg;
3130 context->update_pte = nonpaging_update_pte;
3131 context->root_level = 0;
3132 context->shadow_root_level = PT32E_ROOT_LEVEL;
3133 context->root_hpa = INVALID_PAGE;
3134 context->direct_map = true;
3135 context->nx = false;
3139 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3141 ++vcpu->stat.tlb_flush;
3142 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
3145 static void paging_new_cr3(struct kvm_vcpu *vcpu)
3147 pgprintk("%s: cr3 %lx\n", __func__, kvm_read_cr3(vcpu));
3148 mmu_free_roots(vcpu);
3151 static unsigned long get_cr3(struct kvm_vcpu *vcpu)
3153 return kvm_read_cr3(vcpu);
3156 static void inject_page_fault(struct kvm_vcpu *vcpu,
3157 struct x86_exception *fault)
3159 vcpu->arch.mmu.inject_page_fault(vcpu, fault);
3162 static void paging_free(struct kvm_vcpu *vcpu)
3164 nonpaging_free(vcpu);
3167 static bool is_rsvd_bits_set(struct kvm_mmu *mmu, u64 gpte, int level)
3171 bit7 = (gpte >> 7) & 1;
3172 return (gpte & mmu->rsvd_bits_mask[bit7][level-1]) != 0;
3175 static bool sync_mmio_spte(u64 *sptep, gfn_t gfn, unsigned access,
3178 if (unlikely(is_mmio_spte(*sptep))) {
3179 if (gfn != get_mmio_spte_gfn(*sptep)) {
3180 mmu_spte_clear_no_track(sptep);
3185 mark_mmio_spte(sptep, gfn, access);
3193 #include "paging_tmpl.h"
3197 #include "paging_tmpl.h"
3200 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
3201 struct kvm_mmu *context,
3204 int maxphyaddr = cpuid_maxphyaddr(vcpu);
3205 u64 exb_bit_rsvd = 0;
3208 exb_bit_rsvd = rsvd_bits(63, 63);
3210 case PT32_ROOT_LEVEL:
3211 /* no rsvd bits for 2 level 4K page table entries */
3212 context->rsvd_bits_mask[0][1] = 0;
3213 context->rsvd_bits_mask[0][0] = 0;
3214 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3216 if (!is_pse(vcpu)) {
3217 context->rsvd_bits_mask[1][1] = 0;
3221 if (is_cpuid_PSE36())
3222 /* 36bits PSE 4MB page */
3223 context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
3225 /* 32 bits PSE 4MB page */
3226 context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
3228 case PT32E_ROOT_LEVEL:
3229 context->rsvd_bits_mask[0][2] =
3230 rsvd_bits(maxphyaddr, 63) |
3231 rsvd_bits(7, 8) | rsvd_bits(1, 2); /* PDPTE */
3232 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
3233 rsvd_bits(maxphyaddr, 62); /* PDE */
3234 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
3235 rsvd_bits(maxphyaddr, 62); /* PTE */
3236 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
3237 rsvd_bits(maxphyaddr, 62) |
3238 rsvd_bits(13, 20); /* large page */
3239 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3241 case PT64_ROOT_LEVEL:
3242 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
3243 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
3244 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
3245 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
3246 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
3247 rsvd_bits(maxphyaddr, 51);
3248 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
3249 rsvd_bits(maxphyaddr, 51);
3250 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
3251 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
3252 rsvd_bits(maxphyaddr, 51) |
3254 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
3255 rsvd_bits(maxphyaddr, 51) |
3256 rsvd_bits(13, 20); /* large page */
3257 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
3262 static int paging64_init_context_common(struct kvm_vcpu *vcpu,
3263 struct kvm_mmu *context,
3266 context->nx = is_nx(vcpu);
3268 reset_rsvds_bits_mask(vcpu, context, level);
3270 ASSERT(is_pae(vcpu));
3271 context->new_cr3 = paging_new_cr3;
3272 context->page_fault = paging64_page_fault;
3273 context->gva_to_gpa = paging64_gva_to_gpa;
3274 context->sync_page = paging64_sync_page;
3275 context->invlpg = paging64_invlpg;
3276 context->update_pte = paging64_update_pte;
3277 context->free = paging_free;
3278 context->root_level = level;
3279 context->shadow_root_level = level;
3280 context->root_hpa = INVALID_PAGE;
3281 context->direct_map = false;
3285 static int paging64_init_context(struct kvm_vcpu *vcpu,
3286 struct kvm_mmu *context)
3288 return paging64_init_context_common(vcpu, context, PT64_ROOT_LEVEL);
3291 static int paging32_init_context(struct kvm_vcpu *vcpu,
3292 struct kvm_mmu *context)
3294 context->nx = false;
3296 reset_rsvds_bits_mask(vcpu, context, PT32_ROOT_LEVEL);
3298 context->new_cr3 = paging_new_cr3;
3299 context->page_fault = paging32_page_fault;
3300 context->gva_to_gpa = paging32_gva_to_gpa;
3301 context->free = paging_free;
3302 context->sync_page = paging32_sync_page;
3303 context->invlpg = paging32_invlpg;
3304 context->update_pte = paging32_update_pte;
3305 context->root_level = PT32_ROOT_LEVEL;
3306 context->shadow_root_level = PT32E_ROOT_LEVEL;
3307 context->root_hpa = INVALID_PAGE;
3308 context->direct_map = false;
3312 static int paging32E_init_context(struct kvm_vcpu *vcpu,
3313 struct kvm_mmu *context)
3315 return paging64_init_context_common(vcpu, context, PT32E_ROOT_LEVEL);
3318 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
3320 struct kvm_mmu *context = vcpu->arch.walk_mmu;
3322 context->base_role.word = 0;
3323 context->new_cr3 = nonpaging_new_cr3;
3324 context->page_fault = tdp_page_fault;
3325 context->free = nonpaging_free;
3326 context->sync_page = nonpaging_sync_page;
3327 context->invlpg = nonpaging_invlpg;
3328 context->update_pte = nonpaging_update_pte;
3329 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
3330 context->root_hpa = INVALID_PAGE;
3331 context->direct_map = true;
3332 context->set_cr3 = kvm_x86_ops->set_tdp_cr3;
3333 context->get_cr3 = get_cr3;
3334 context->get_pdptr = kvm_pdptr_read;
3335 context->inject_page_fault = kvm_inject_page_fault;
3336 context->nx = is_nx(vcpu);
3338 if (!is_paging(vcpu)) {
3339 context->nx = false;
3340 context->gva_to_gpa = nonpaging_gva_to_gpa;
3341 context->root_level = 0;
3342 } else if (is_long_mode(vcpu)) {
3343 context->nx = is_nx(vcpu);
3344 reset_rsvds_bits_mask(vcpu, context, PT64_ROOT_LEVEL);
3345 context->gva_to_gpa = paging64_gva_to_gpa;
3346 context->root_level = PT64_ROOT_LEVEL;
3347 } else if (is_pae(vcpu)) {
3348 context->nx = is_nx(vcpu);
3349 reset_rsvds_bits_mask(vcpu, context, PT32E_ROOT_LEVEL);
3350 context->gva_to_gpa = paging64_gva_to_gpa;
3351 context->root_level = PT32E_ROOT_LEVEL;
3353 context->nx = false;
3354 reset_rsvds_bits_mask(vcpu, context, PT32_ROOT_LEVEL);
3355 context->gva_to_gpa = paging32_gva_to_gpa;
3356 context->root_level = PT32_ROOT_LEVEL;
3362 int kvm_init_shadow_mmu(struct kvm_vcpu *vcpu, struct kvm_mmu *context)
3365 bool smep = kvm_read_cr4_bits(vcpu, X86_CR4_SMEP);
3367 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3369 if (!is_paging(vcpu))
3370 r = nonpaging_init_context(vcpu, context);
3371 else if (is_long_mode(vcpu))
3372 r = paging64_init_context(vcpu, context);
3373 else if (is_pae(vcpu))
3374 r = paging32E_init_context(vcpu, context);
3376 r = paging32_init_context(vcpu, context);
3378 vcpu->arch.mmu.base_role.cr4_pae = !!is_pae(vcpu);
3379 vcpu->arch.mmu.base_role.cr0_wp = is_write_protection(vcpu);
3380 vcpu->arch.mmu.base_role.smep_andnot_wp
3381 = smep && !is_write_protection(vcpu);
3385 EXPORT_SYMBOL_GPL(kvm_init_shadow_mmu);
3387 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
3389 int r = kvm_init_shadow_mmu(vcpu, vcpu->arch.walk_mmu);
3391 vcpu->arch.walk_mmu->set_cr3 = kvm_x86_ops->set_cr3;
3392 vcpu->arch.walk_mmu->get_cr3 = get_cr3;
3393 vcpu->arch.walk_mmu->get_pdptr = kvm_pdptr_read;
3394 vcpu->arch.walk_mmu->inject_page_fault = kvm_inject_page_fault;
3399 static int init_kvm_nested_mmu(struct kvm_vcpu *vcpu)
3401 struct kvm_mmu *g_context = &vcpu->arch.nested_mmu;
3403 g_context->get_cr3 = get_cr3;
3404 g_context->get_pdptr = kvm_pdptr_read;
3405 g_context->inject_page_fault = kvm_inject_page_fault;
3408 * Note that arch.mmu.gva_to_gpa translates l2_gva to l1_gpa. The
3409 * translation of l2_gpa to l1_gpa addresses is done using the
3410 * arch.nested_mmu.gva_to_gpa function. Basically the gva_to_gpa
3411 * functions between mmu and nested_mmu are swapped.
3413 if (!is_paging(vcpu)) {
3414 g_context->nx = false;
3415 g_context->root_level = 0;
3416 g_context->gva_to_gpa = nonpaging_gva_to_gpa_nested;
3417 } else if (is_long_mode(vcpu)) {
3418 g_context->nx = is_nx(vcpu);
3419 reset_rsvds_bits_mask(vcpu, g_context, PT64_ROOT_LEVEL);
3420 g_context->root_level = PT64_ROOT_LEVEL;
3421 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
3422 } else if (is_pae(vcpu)) {
3423 g_context->nx = is_nx(vcpu);
3424 reset_rsvds_bits_mask(vcpu, g_context, PT32E_ROOT_LEVEL);
3425 g_context->root_level = PT32E_ROOT_LEVEL;
3426 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
3428 g_context->nx = false;
3429 reset_rsvds_bits_mask(vcpu, g_context, PT32_ROOT_LEVEL);
3430 g_context->root_level = PT32_ROOT_LEVEL;
3431 g_context->gva_to_gpa = paging32_gva_to_gpa_nested;
3437 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
3439 if (mmu_is_nested(vcpu))
3440 return init_kvm_nested_mmu(vcpu);
3441 else if (tdp_enabled)
3442 return init_kvm_tdp_mmu(vcpu);
3444 return init_kvm_softmmu(vcpu);
3447 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
3450 if (VALID_PAGE(vcpu->arch.mmu.root_hpa))
3451 /* mmu.free() should set root_hpa = INVALID_PAGE */
3452 vcpu->arch.mmu.free(vcpu);
3455 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
3457 destroy_kvm_mmu(vcpu);
3458 return init_kvm_mmu(vcpu);
3460 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
3462 int kvm_mmu_load(struct kvm_vcpu *vcpu)
3466 r = mmu_topup_memory_caches(vcpu);
3469 r = mmu_alloc_roots(vcpu);
3470 spin_lock(&vcpu->kvm->mmu_lock);
3471 mmu_sync_roots(vcpu);
3472 spin_unlock(&vcpu->kvm->mmu_lock);
3475 /* set_cr3() should ensure TLB has been flushed */
3476 vcpu->arch.mmu.set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
3480 EXPORT_SYMBOL_GPL(kvm_mmu_load);
3482 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
3484 mmu_free_roots(vcpu);
3486 EXPORT_SYMBOL_GPL(kvm_mmu_unload);
3488 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
3489 struct kvm_mmu_page *sp, u64 *spte,
3492 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
3493 ++vcpu->kvm->stat.mmu_pde_zapped;
3497 ++vcpu->kvm->stat.mmu_pte_updated;
3498 vcpu->arch.mmu.update_pte(vcpu, sp, spte, new);
3501 static bool need_remote_flush(u64 old, u64 new)
3503 if (!is_shadow_present_pte(old))
3505 if (!is_shadow_present_pte(new))
3507 if ((old ^ new) & PT64_BASE_ADDR_MASK)
3509 old ^= PT64_NX_MASK;
3510 new ^= PT64_NX_MASK;
3511 return (old & ~new & PT64_PERM_MASK) != 0;
3514 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, bool zap_page,
3515 bool remote_flush, bool local_flush)
3521 kvm_flush_remote_tlbs(vcpu->kvm);
3522 else if (local_flush)
3523 kvm_mmu_flush_tlb(vcpu);
3526 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
3528 u64 *spte = vcpu->arch.last_pte_updated;
3530 return !!(spte && (*spte & shadow_accessed_mask));
3533 static u64 mmu_pte_write_fetch_gpte(struct kvm_vcpu *vcpu, gpa_t *gpa,
3534 const u8 *new, int *bytes)
3540 * Assume that the pte write on a page table of the same type
3541 * as the current vcpu paging mode since we update the sptes only
3542 * when they have the same mode.
3544 if (is_pae(vcpu) && *bytes == 4) {
3545 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
3548 r = kvm_read_guest(vcpu->kvm, *gpa, &gentry, min(*bytes, 8));
3551 new = (const u8 *)&gentry;
3556 gentry = *(const u32 *)new;
3559 gentry = *(const u64 *)new;
3570 * If we're seeing too many writes to a page, it may no longer be a page table,
3571 * or we may be forking, in which case it is better to unmap the page.
3573 static bool detect_write_flooding(struct kvm_vcpu *vcpu, gfn_t gfn)
3575 bool flooded = false;
3577 if (gfn == vcpu->arch.last_pt_write_gfn
3578 && !last_updated_pte_accessed(vcpu)) {
3579 ++vcpu->arch.last_pt_write_count;
3580 if (vcpu->arch.last_pt_write_count >= 3)
3583 vcpu->arch.last_pt_write_gfn = gfn;
3584 vcpu->arch.last_pt_write_count = 1;
3585 vcpu->arch.last_pte_updated = NULL;
3592 * Misaligned accesses are too much trouble to fix up; also, they usually
3593 * indicate a page is not used as a page table.
3595 static bool detect_write_misaligned(struct kvm_mmu_page *sp, gpa_t gpa,
3598 unsigned offset, pte_size, misaligned;
3600 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
3601 gpa, bytes, sp->role.word);
3603 offset = offset_in_page(gpa);
3604 pte_size = sp->role.cr4_pae ? 8 : 4;
3605 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
3606 misaligned |= bytes < 4;
3611 static u64 *get_written_sptes(struct kvm_mmu_page *sp, gpa_t gpa, int *nspte)
3613 unsigned page_offset, quadrant;
3617 page_offset = offset_in_page(gpa);
3618 level = sp->role.level;
3620 if (!sp->role.cr4_pae) {
3621 page_offset <<= 1; /* 32->64 */
3623 * A 32-bit pde maps 4MB while the shadow pdes map
3624 * only 2MB. So we need to double the offset again
3625 * and zap two pdes instead of one.
3627 if (level == PT32_ROOT_LEVEL) {
3628 page_offset &= ~7; /* kill rounding error */
3632 quadrant = page_offset >> PAGE_SHIFT;
3633 page_offset &= ~PAGE_MASK;
3634 if (quadrant != sp->role.quadrant)
3638 spte = &sp->spt[page_offset / sizeof(*spte)];
3642 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
3643 const u8 *new, int bytes)
3645 gfn_t gfn = gpa >> PAGE_SHIFT;
3646 union kvm_mmu_page_role mask = { .word = 0 };
3647 struct kvm_mmu_page *sp;
3648 struct hlist_node *node;
3649 LIST_HEAD(invalid_list);
3650 u64 entry, gentry, *spte;
3652 bool remote_flush, local_flush, zap_page, flooded, misaligned;
3655 * If we don't have indirect shadow pages, it means no page is
3656 * write-protected, so we can exit simply.
3658 if (!ACCESS_ONCE(vcpu->kvm->arch.indirect_shadow_pages))
3661 zap_page = remote_flush = local_flush = false;
3663 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
3665 gentry = mmu_pte_write_fetch_gpte(vcpu, &gpa, new, &bytes);
3668 * No need to care whether allocation memory is successful
3669 * or not since pte prefetch is skiped if it does not have
3670 * enough objects in the cache.
3672 mmu_topup_memory_caches(vcpu);
3674 spin_lock(&vcpu->kvm->mmu_lock);
3675 ++vcpu->kvm->stat.mmu_pte_write;
3676 trace_kvm_mmu_audit(vcpu, AUDIT_PRE_PTE_WRITE);
3678 flooded = detect_write_flooding(vcpu, gfn);
3679 mask.cr0_wp = mask.cr4_pae = mask.nxe = 1;
3680 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn, node) {
3681 misaligned = detect_write_misaligned(sp, gpa, bytes);
3683 if (misaligned || flooded) {
3684 zap_page |= !!kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
3686 ++vcpu->kvm->stat.mmu_flooded;
3690 spte = get_written_sptes(sp, gpa, &npte);
3697 mmu_page_zap_pte(vcpu->kvm, sp, spte);
3699 !((sp->role.word ^ vcpu->arch.mmu.base_role.word)
3700 & mask.word) && rmap_can_add(vcpu))
3701 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
3702 if (!remote_flush && need_remote_flush(entry, *spte))
3703 remote_flush = true;
3707 mmu_pte_write_flush_tlb(vcpu, zap_page, remote_flush, local_flush);
3708 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
3709 trace_kvm_mmu_audit(vcpu, AUDIT_POST_PTE_WRITE);
3710 spin_unlock(&vcpu->kvm->mmu_lock);
3713 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
3718 if (vcpu->arch.mmu.direct_map)
3721 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
3723 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
3727 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
3729 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
3731 LIST_HEAD(invalid_list);
3733 while (kvm_mmu_available_pages(vcpu->kvm) < KVM_REFILL_PAGES &&
3734 !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
3735 struct kvm_mmu_page *sp;
3737 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
3738 struct kvm_mmu_page, link);
3739 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
3740 ++vcpu->kvm->stat.mmu_recycled;
3742 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
3745 static bool is_mmio_page_fault(struct kvm_vcpu *vcpu, gva_t addr)
3747 if (vcpu->arch.mmu.direct_map || mmu_is_nested(vcpu))
3748 return vcpu_match_mmio_gpa(vcpu, addr);
3750 return vcpu_match_mmio_gva(vcpu, addr);
3753 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code,
3754 void *insn, int insn_len)
3756 int r, emulation_type = EMULTYPE_RETRY;
3757 enum emulation_result er;
3759 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code, false);
3768 if (is_mmio_page_fault(vcpu, cr2))
3771 er = x86_emulate_instruction(vcpu, cr2, emulation_type, insn, insn_len);
3776 case EMULATE_DO_MMIO:
3777 ++vcpu->stat.mmio_exits;
3787 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
3789 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
3791 vcpu->arch.mmu.invlpg(vcpu, gva);
3792 kvm_mmu_flush_tlb(vcpu);
3793 ++vcpu->stat.invlpg;
3795 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
3797 void kvm_enable_tdp(void)
3801 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
3803 void kvm_disable_tdp(void)
3805 tdp_enabled = false;
3807 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
3809 static void free_mmu_pages(struct kvm_vcpu *vcpu)
3811 free_page((unsigned long)vcpu->arch.mmu.pae_root);
3812 if (vcpu->arch.mmu.lm_root != NULL)
3813 free_page((unsigned long)vcpu->arch.mmu.lm_root);
3816 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
3824 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
3825 * Therefore we need to allocate shadow page tables in the first
3826 * 4GB of memory, which happens to fit the DMA32 zone.
3828 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
3832 vcpu->arch.mmu.pae_root = page_address(page);
3833 for (i = 0; i < 4; ++i)
3834 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
3839 int kvm_mmu_create(struct kvm_vcpu *vcpu)
3842 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3844 return alloc_mmu_pages(vcpu);
3847 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
3850 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3852 return init_kvm_mmu(vcpu);
3855 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
3857 struct kvm_mmu_page *sp;
3859 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
3863 if (!test_bit(slot, sp->slot_bitmap))
3867 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3868 if (!is_shadow_present_pte(pt[i]) ||
3869 !is_last_spte(pt[i], sp->role.level))
3872 if (is_large_pte(pt[i])) {
3873 drop_spte(kvm, &pt[i]);
3879 if (is_writable_pte(pt[i]))
3880 mmu_spte_update(&pt[i],
3881 pt[i] & ~PT_WRITABLE_MASK);
3884 kvm_flush_remote_tlbs(kvm);
3887 void kvm_mmu_zap_all(struct kvm *kvm)
3889 struct kvm_mmu_page *sp, *node;
3890 LIST_HEAD(invalid_list);
3892 spin_lock(&kvm->mmu_lock);
3894 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
3895 if (kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list))
3898 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3899 spin_unlock(&kvm->mmu_lock);
3902 static int kvm_mmu_remove_some_alloc_mmu_pages(struct kvm *kvm,
3903 struct list_head *invalid_list)
3905 struct kvm_mmu_page *page;
3907 page = container_of(kvm->arch.active_mmu_pages.prev,
3908 struct kvm_mmu_page, link);
3909 return kvm_mmu_prepare_zap_page(kvm, page, invalid_list);
3912 static int mmu_shrink(struct shrinker *shrink, struct shrink_control *sc)
3915 struct kvm *kvm_freed = NULL;
3916 int nr_to_scan = sc->nr_to_scan;
3918 if (nr_to_scan == 0)
3921 raw_spin_lock(&kvm_lock);
3923 list_for_each_entry(kvm, &vm_list, vm_list) {
3924 int idx, freed_pages;
3925 LIST_HEAD(invalid_list);
3927 idx = srcu_read_lock(&kvm->srcu);
3928 spin_lock(&kvm->mmu_lock);
3929 if (!kvm_freed && nr_to_scan > 0 &&
3930 kvm->arch.n_used_mmu_pages > 0) {
3931 freed_pages = kvm_mmu_remove_some_alloc_mmu_pages(kvm,
3937 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3938 spin_unlock(&kvm->mmu_lock);
3939 srcu_read_unlock(&kvm->srcu, idx);
3942 list_move_tail(&kvm_freed->vm_list, &vm_list);
3944 raw_spin_unlock(&kvm_lock);
3947 return percpu_counter_read_positive(&kvm_total_used_mmu_pages);
3950 static struct shrinker mmu_shrinker = {
3951 .shrink = mmu_shrink,
3952 .seeks = DEFAULT_SEEKS * 10,
3955 static void mmu_destroy_caches(void)
3957 if (pte_list_desc_cache)
3958 kmem_cache_destroy(pte_list_desc_cache);
3959 if (mmu_page_header_cache)
3960 kmem_cache_destroy(mmu_page_header_cache);
3963 int kvm_mmu_module_init(void)
3965 pte_list_desc_cache = kmem_cache_create("pte_list_desc",
3966 sizeof(struct pte_list_desc),
3968 if (!pte_list_desc_cache)
3971 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
3972 sizeof(struct kvm_mmu_page),
3974 if (!mmu_page_header_cache)
3977 if (percpu_counter_init(&kvm_total_used_mmu_pages, 0))
3980 register_shrinker(&mmu_shrinker);
3985 mmu_destroy_caches();
3990 * Caculate mmu pages needed for kvm.
3992 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3995 unsigned int nr_mmu_pages;
3996 unsigned int nr_pages = 0;
3997 struct kvm_memslots *slots;
3999 slots = kvm_memslots(kvm);
4001 for (i = 0; i < slots->nmemslots; i++)
4002 nr_pages += slots->memslots[i].npages;
4004 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
4005 nr_mmu_pages = max(nr_mmu_pages,
4006 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
4008 return nr_mmu_pages;
4011 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
4014 if (len > buffer->len)
4019 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
4024 ret = pv_mmu_peek_buffer(buffer, len);
4029 buffer->processed += len;
4033 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
4034 gpa_t addr, gpa_t value)
4039 if (!is_long_mode(vcpu) && !is_pae(vcpu))
4042 r = mmu_topup_memory_caches(vcpu);
4046 if (!emulator_write_phys(vcpu, addr, &value, bytes))
4052 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
4054 (void)kvm_set_cr3(vcpu, kvm_read_cr3(vcpu));
4058 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
4060 spin_lock(&vcpu->kvm->mmu_lock);
4061 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
4062 spin_unlock(&vcpu->kvm->mmu_lock);
4066 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
4067 struct kvm_pv_mmu_op_buffer *buffer)
4069 struct kvm_mmu_op_header *header;
4071 header = pv_mmu_peek_buffer(buffer, sizeof *header);
4074 switch (header->op) {
4075 case KVM_MMU_OP_WRITE_PTE: {
4076 struct kvm_mmu_op_write_pte *wpte;
4078 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
4081 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
4084 case KVM_MMU_OP_FLUSH_TLB: {
4085 struct kvm_mmu_op_flush_tlb *ftlb;
4087 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
4090 return kvm_pv_mmu_flush_tlb(vcpu);
4092 case KVM_MMU_OP_RELEASE_PT: {
4093 struct kvm_mmu_op_release_pt *rpt;
4095 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
4098 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
4104 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
4105 gpa_t addr, unsigned long *ret)
4108 struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
4110 buffer->ptr = buffer->buf;
4111 buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
4112 buffer->processed = 0;
4114 r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
4118 while (buffer->len) {
4119 r = kvm_pv_mmu_op_one(vcpu, buffer);
4128 *ret = buffer->processed;
4132 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
4134 struct kvm_shadow_walk_iterator iterator;
4138 walk_shadow_page_lockless_begin(vcpu);
4139 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte) {
4140 sptes[iterator.level-1] = spte;
4142 if (!is_shadow_present_pte(spte))
4145 walk_shadow_page_lockless_end(vcpu);
4149 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
4151 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
4155 destroy_kvm_mmu(vcpu);
4156 free_mmu_pages(vcpu);
4157 mmu_free_memory_caches(vcpu);
4160 #ifdef CONFIG_KVM_MMU_AUDIT
4161 #include "mmu_audit.c"
4163 static void mmu_audit_disable(void) { }
4166 void kvm_mmu_module_exit(void)
4168 mmu_destroy_caches();
4169 percpu_counter_destroy(&kvm_total_used_mmu_pages);
4170 unregister_shrinker(&mmu_shrinker);
4171 mmu_audit_disable();