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.
12 * Yaniv Kamay <yaniv@qumranet.com>
13 * Avi Kivity <avi@qumranet.com>
15 * This work is licensed under the terms of the GNU GPL, version 2. See
16 * the COPYING file in the top-level directory.
23 #include <linux/kvm_host.h>
24 #include <linux/types.h>
25 #include <linux/string.h>
27 #include <linux/highmem.h>
28 #include <linux/module.h>
29 #include <linux/swap.h>
30 #include <linux/hugetlb.h>
31 #include <linux/compiler.h>
34 #include <asm/cmpxchg.h>
38 * When setting this variable to true it enables Two-Dimensional-Paging
39 * where the hardware walks 2 page tables:
40 * 1. the guest-virtual to guest-physical
41 * 2. while doing 1. it walks guest-physical to host-physical
42 * If the hardware supports that we don't need to do shadow paging.
44 bool tdp_enabled = false;
51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
63 #define pgprintk(x...) do { } while (0)
64 #define rmap_printk(x...) do { } while (0)
68 #if defined(MMU_DEBUG) || defined(AUDIT)
70 module_param(dbg, bool, 0644);
74 #define ASSERT(x) do { } while (0)
78 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
79 __FILE__, __LINE__, #x); \
83 #define PT_FIRST_AVAIL_BITS_SHIFT 9
84 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
86 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
88 #define PT64_LEVEL_BITS 9
90 #define PT64_LEVEL_SHIFT(level) \
91 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
93 #define PT64_LEVEL_MASK(level) \
94 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
96 #define PT64_INDEX(address, level)\
97 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
100 #define PT32_LEVEL_BITS 10
102 #define PT32_LEVEL_SHIFT(level) \
103 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
105 #define PT32_LEVEL_MASK(level) \
106 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
108 #define PT32_INDEX(address, level)\
109 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
112 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
113 #define PT64_DIR_BASE_ADDR_MASK \
114 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
116 #define PT32_BASE_ADDR_MASK PAGE_MASK
117 #define PT32_DIR_BASE_ADDR_MASK \
118 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
120 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
123 #define PFERR_PRESENT_MASK (1U << 0)
124 #define PFERR_WRITE_MASK (1U << 1)
125 #define PFERR_USER_MASK (1U << 2)
126 #define PFERR_FETCH_MASK (1U << 4)
128 #define PT_DIRECTORY_LEVEL 2
129 #define PT_PAGE_TABLE_LEVEL 1
133 #define ACC_EXEC_MASK 1
134 #define ACC_WRITE_MASK PT_WRITABLE_MASK
135 #define ACC_USER_MASK PT_USER_MASK
136 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
138 struct kvm_pv_mmu_op_buffer {
142 char buf[512] __aligned(sizeof(long));
145 struct kvm_rmap_desc {
146 u64 *shadow_ptes[RMAP_EXT];
147 struct kvm_rmap_desc *more;
150 static struct kmem_cache *pte_chain_cache;
151 static struct kmem_cache *rmap_desc_cache;
152 static struct kmem_cache *mmu_page_header_cache;
154 static u64 __read_mostly shadow_trap_nonpresent_pte;
155 static u64 __read_mostly shadow_notrap_nonpresent_pte;
156 static u64 __read_mostly shadow_base_present_pte;
157 static u64 __read_mostly shadow_nx_mask;
158 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
159 static u64 __read_mostly shadow_user_mask;
160 static u64 __read_mostly shadow_accessed_mask;
161 static u64 __read_mostly shadow_dirty_mask;
163 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
165 shadow_trap_nonpresent_pte = trap_pte;
166 shadow_notrap_nonpresent_pte = notrap_pte;
168 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
170 void kvm_mmu_set_base_ptes(u64 base_pte)
172 shadow_base_present_pte = base_pte;
174 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
176 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
177 u64 dirty_mask, u64 nx_mask, u64 x_mask)
179 shadow_user_mask = user_mask;
180 shadow_accessed_mask = accessed_mask;
181 shadow_dirty_mask = dirty_mask;
182 shadow_nx_mask = nx_mask;
183 shadow_x_mask = x_mask;
185 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
187 static int is_write_protection(struct kvm_vcpu *vcpu)
189 return vcpu->arch.cr0 & X86_CR0_WP;
192 static int is_cpuid_PSE36(void)
197 static int is_nx(struct kvm_vcpu *vcpu)
199 return vcpu->arch.shadow_efer & EFER_NX;
202 static int is_present_pte(unsigned long pte)
204 return pte & PT_PRESENT_MASK;
207 static int is_shadow_present_pte(u64 pte)
209 return pte != shadow_trap_nonpresent_pte
210 && pte != shadow_notrap_nonpresent_pte;
213 static int is_large_pte(u64 pte)
215 return pte & PT_PAGE_SIZE_MASK;
218 static int is_writeble_pte(unsigned long pte)
220 return pte & PT_WRITABLE_MASK;
223 static int is_dirty_pte(unsigned long pte)
225 return pte & shadow_dirty_mask;
228 static int is_rmap_pte(u64 pte)
230 return is_shadow_present_pte(pte);
233 static pfn_t spte_to_pfn(u64 pte)
235 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
238 static gfn_t pse36_gfn_delta(u32 gpte)
240 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
242 return (gpte & PT32_DIR_PSE36_MASK) << shift;
245 static void set_shadow_pte(u64 *sptep, u64 spte)
248 set_64bit((unsigned long *)sptep, spte);
250 set_64bit((unsigned long long *)sptep, spte);
254 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
255 struct kmem_cache *base_cache, int min)
259 if (cache->nobjs >= min)
261 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
262 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
265 cache->objects[cache->nobjs++] = obj;
270 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
273 kfree(mc->objects[--mc->nobjs]);
276 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
281 if (cache->nobjs >= min)
283 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
284 page = alloc_page(GFP_KERNEL);
287 set_page_private(page, 0);
288 cache->objects[cache->nobjs++] = page_address(page);
293 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
296 free_page((unsigned long)mc->objects[--mc->nobjs]);
299 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
303 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
307 r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
311 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
314 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
315 mmu_page_header_cache, 4);
320 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
322 mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
323 mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
324 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
325 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
328 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
334 p = mc->objects[--mc->nobjs];
339 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
341 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
342 sizeof(struct kvm_pte_chain));
345 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
350 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
352 return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
353 sizeof(struct kvm_rmap_desc));
356 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
362 * Return the pointer to the largepage write count for a given
363 * gfn, handling slots that are not large page aligned.
365 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
369 idx = (gfn / KVM_PAGES_PER_HPAGE) -
370 (slot->base_gfn / KVM_PAGES_PER_HPAGE);
371 return &slot->lpage_info[idx].write_count;
374 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
378 write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
382 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
386 write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
388 WARN_ON(*write_count < 0);
391 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
393 struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
397 largepage_idx = slot_largepage_idx(gfn, slot);
398 return *largepage_idx;
404 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
406 struct vm_area_struct *vma;
409 addr = gfn_to_hva(kvm, gfn);
410 if (kvm_is_error_hva(addr))
413 vma = find_vma(current->mm, addr);
414 if (vma && is_vm_hugetlb_page(vma))
420 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
422 struct kvm_memory_slot *slot;
424 if (has_wrprotected_page(vcpu->kvm, large_gfn))
427 if (!host_largepage_backed(vcpu->kvm, large_gfn))
430 slot = gfn_to_memslot(vcpu->kvm, large_gfn);
431 if (slot && slot->dirty_bitmap)
438 * Take gfn and return the reverse mapping to it.
439 * Note: gfn must be unaliased before this function get called
442 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
444 struct kvm_memory_slot *slot;
447 slot = gfn_to_memslot(kvm, gfn);
449 return &slot->rmap[gfn - slot->base_gfn];
451 idx = (gfn / KVM_PAGES_PER_HPAGE) -
452 (slot->base_gfn / KVM_PAGES_PER_HPAGE);
454 return &slot->lpage_info[idx].rmap_pde;
458 * Reverse mapping data structures:
460 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
461 * that points to page_address(page).
463 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
464 * containing more mappings.
466 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
468 struct kvm_mmu_page *sp;
469 struct kvm_rmap_desc *desc;
470 unsigned long *rmapp;
473 if (!is_rmap_pte(*spte))
475 gfn = unalias_gfn(vcpu->kvm, gfn);
476 sp = page_header(__pa(spte));
477 sp->gfns[spte - sp->spt] = gfn;
478 rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
480 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
481 *rmapp = (unsigned long)spte;
482 } else if (!(*rmapp & 1)) {
483 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
484 desc = mmu_alloc_rmap_desc(vcpu);
485 desc->shadow_ptes[0] = (u64 *)*rmapp;
486 desc->shadow_ptes[1] = spte;
487 *rmapp = (unsigned long)desc | 1;
489 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
490 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
491 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
493 if (desc->shadow_ptes[RMAP_EXT-1]) {
494 desc->more = mmu_alloc_rmap_desc(vcpu);
497 for (i = 0; desc->shadow_ptes[i]; ++i)
499 desc->shadow_ptes[i] = spte;
503 static void rmap_desc_remove_entry(unsigned long *rmapp,
504 struct kvm_rmap_desc *desc,
506 struct kvm_rmap_desc *prev_desc)
510 for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
512 desc->shadow_ptes[i] = desc->shadow_ptes[j];
513 desc->shadow_ptes[j] = NULL;
516 if (!prev_desc && !desc->more)
517 *rmapp = (unsigned long)desc->shadow_ptes[0];
520 prev_desc->more = desc->more;
522 *rmapp = (unsigned long)desc->more | 1;
523 mmu_free_rmap_desc(desc);
526 static void rmap_remove(struct kvm *kvm, u64 *spte)
528 struct kvm_rmap_desc *desc;
529 struct kvm_rmap_desc *prev_desc;
530 struct kvm_mmu_page *sp;
532 unsigned long *rmapp;
535 if (!is_rmap_pte(*spte))
537 sp = page_header(__pa(spte));
538 pfn = spte_to_pfn(*spte);
539 if (*spte & shadow_accessed_mask)
540 kvm_set_pfn_accessed(pfn);
541 if (is_writeble_pte(*spte))
542 kvm_release_pfn_dirty(pfn);
544 kvm_release_pfn_clean(pfn);
545 rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
547 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
549 } else if (!(*rmapp & 1)) {
550 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
551 if ((u64 *)*rmapp != spte) {
552 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
558 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
559 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
562 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
563 if (desc->shadow_ptes[i] == spte) {
564 rmap_desc_remove_entry(rmapp,
576 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
578 struct kvm_rmap_desc *desc;
579 struct kvm_rmap_desc *prev_desc;
585 else if (!(*rmapp & 1)) {
587 return (u64 *)*rmapp;
590 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
594 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
595 if (prev_spte == spte)
596 return desc->shadow_ptes[i];
597 prev_spte = desc->shadow_ptes[i];
604 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
606 unsigned long *rmapp;
608 int write_protected = 0;
610 gfn = unalias_gfn(kvm, gfn);
611 rmapp = gfn_to_rmap(kvm, gfn, 0);
613 spte = rmap_next(kvm, rmapp, NULL);
616 BUG_ON(!(*spte & PT_PRESENT_MASK));
617 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
618 if (is_writeble_pte(*spte)) {
619 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
622 spte = rmap_next(kvm, rmapp, spte);
624 if (write_protected) {
627 spte = rmap_next(kvm, rmapp, NULL);
628 pfn = spte_to_pfn(*spte);
629 kvm_set_pfn_dirty(pfn);
632 /* check for huge page mappings */
633 rmapp = gfn_to_rmap(kvm, gfn, 1);
634 spte = rmap_next(kvm, rmapp, NULL);
637 BUG_ON(!(*spte & PT_PRESENT_MASK));
638 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
639 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
640 if (is_writeble_pte(*spte)) {
641 rmap_remove(kvm, spte);
643 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
647 spte = rmap_next(kvm, rmapp, spte);
651 kvm_flush_remote_tlbs(kvm);
653 account_shadowed(kvm, gfn);
657 static int is_empty_shadow_page(u64 *spt)
662 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
663 if (is_shadow_present_pte(*pos)) {
664 printk(KERN_ERR "%s: %p %llx\n", __func__,
672 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
674 ASSERT(is_empty_shadow_page(sp->spt));
676 __free_page(virt_to_page(sp->spt));
677 __free_page(virt_to_page(sp->gfns));
679 ++kvm->arch.n_free_mmu_pages;
682 static unsigned kvm_page_table_hashfn(gfn_t gfn)
684 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
687 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
690 struct kvm_mmu_page *sp;
692 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
693 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
694 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
695 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
696 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
697 ASSERT(is_empty_shadow_page(sp->spt));
700 sp->parent_pte = parent_pte;
701 --vcpu->kvm->arch.n_free_mmu_pages;
705 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
706 struct kvm_mmu_page *sp, u64 *parent_pte)
708 struct kvm_pte_chain *pte_chain;
709 struct hlist_node *node;
714 if (!sp->multimapped) {
715 u64 *old = sp->parent_pte;
718 sp->parent_pte = parent_pte;
722 pte_chain = mmu_alloc_pte_chain(vcpu);
723 INIT_HLIST_HEAD(&sp->parent_ptes);
724 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
725 pte_chain->parent_ptes[0] = old;
727 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
728 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
730 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
731 if (!pte_chain->parent_ptes[i]) {
732 pte_chain->parent_ptes[i] = parent_pte;
736 pte_chain = mmu_alloc_pte_chain(vcpu);
738 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
739 pte_chain->parent_ptes[0] = parent_pte;
742 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
745 struct kvm_pte_chain *pte_chain;
746 struct hlist_node *node;
749 if (!sp->multimapped) {
750 BUG_ON(sp->parent_pte != parent_pte);
751 sp->parent_pte = NULL;
754 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
755 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
756 if (!pte_chain->parent_ptes[i])
758 if (pte_chain->parent_ptes[i] != parent_pte)
760 while (i + 1 < NR_PTE_CHAIN_ENTRIES
761 && pte_chain->parent_ptes[i + 1]) {
762 pte_chain->parent_ptes[i]
763 = pte_chain->parent_ptes[i + 1];
766 pte_chain->parent_ptes[i] = NULL;
768 hlist_del(&pte_chain->link);
769 mmu_free_pte_chain(pte_chain);
770 if (hlist_empty(&sp->parent_ptes)) {
772 sp->parent_pte = NULL;
780 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
781 struct kvm_mmu_page *sp)
785 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
786 sp->spt[i] = shadow_trap_nonpresent_pte;
789 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
792 struct hlist_head *bucket;
793 struct kvm_mmu_page *sp;
794 struct hlist_node *node;
796 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
797 index = kvm_page_table_hashfn(gfn);
798 bucket = &kvm->arch.mmu_page_hash[index];
799 hlist_for_each_entry(sp, node, bucket, hash_link)
800 if (sp->gfn == gfn && !sp->role.metaphysical
801 && !sp->role.invalid) {
802 pgprintk("%s: found role %x\n",
803 __func__, sp->role.word);
809 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
817 union kvm_mmu_page_role role;
820 struct hlist_head *bucket;
821 struct kvm_mmu_page *sp;
822 struct hlist_node *node;
825 role.glevels = vcpu->arch.mmu.root_level;
827 role.metaphysical = metaphysical;
828 role.access = access;
829 if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
830 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
831 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
832 role.quadrant = quadrant;
834 pgprintk("%s: looking gfn %lx role %x\n", __func__,
836 index = kvm_page_table_hashfn(gfn);
837 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
838 hlist_for_each_entry(sp, node, bucket, hash_link)
839 if (sp->gfn == gfn && sp->role.word == role.word) {
840 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
841 pgprintk("%s: found\n", __func__);
844 ++vcpu->kvm->stat.mmu_cache_miss;
845 sp = kvm_mmu_alloc_page(vcpu, parent_pte);
848 pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
851 hlist_add_head(&sp->hash_link, bucket);
853 rmap_write_protect(vcpu->kvm, gfn);
854 if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
855 vcpu->arch.mmu.prefetch_page(vcpu, sp);
857 nonpaging_prefetch_page(vcpu, sp);
861 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
862 struct kvm_mmu_page *sp)
870 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
871 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
872 if (is_shadow_present_pte(pt[i]))
873 rmap_remove(kvm, &pt[i]);
874 pt[i] = shadow_trap_nonpresent_pte;
876 kvm_flush_remote_tlbs(kvm);
880 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
883 if (is_shadow_present_pte(ent)) {
884 if (!is_large_pte(ent)) {
885 ent &= PT64_BASE_ADDR_MASK;
886 mmu_page_remove_parent_pte(page_header(ent),
890 rmap_remove(kvm, &pt[i]);
893 pt[i] = shadow_trap_nonpresent_pte;
895 kvm_flush_remote_tlbs(kvm);
898 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
900 mmu_page_remove_parent_pte(sp, parent_pte);
903 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
907 for (i = 0; i < KVM_MAX_VCPUS; ++i)
909 kvm->vcpus[i]->arch.last_pte_updated = NULL;
912 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
916 ++kvm->stat.mmu_shadow_zapped;
917 while (sp->multimapped || sp->parent_pte) {
918 if (!sp->multimapped)
919 parent_pte = sp->parent_pte;
921 struct kvm_pte_chain *chain;
923 chain = container_of(sp->parent_ptes.first,
924 struct kvm_pte_chain, link);
925 parent_pte = chain->parent_ptes[0];
928 kvm_mmu_put_page(sp, parent_pte);
929 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
931 kvm_mmu_page_unlink_children(kvm, sp);
932 if (!sp->root_count) {
933 if (!sp->role.metaphysical && !sp->role.invalid)
934 unaccount_shadowed(kvm, sp->gfn);
935 hlist_del(&sp->hash_link);
936 kvm_mmu_free_page(kvm, sp);
938 int invalid = sp->role.invalid;
939 list_move(&sp->link, &kvm->arch.active_mmu_pages);
940 sp->role.invalid = 1;
941 kvm_reload_remote_mmus(kvm);
942 if (!sp->role.metaphysical && !invalid)
943 unaccount_shadowed(kvm, sp->gfn);
945 kvm_mmu_reset_last_pte_updated(kvm);
949 * Changing the number of mmu pages allocated to the vm
950 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
952 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
955 * If we set the number of mmu pages to be smaller be than the
956 * number of actived pages , we must to free some mmu pages before we
960 if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
962 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
963 - kvm->arch.n_free_mmu_pages;
965 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
966 struct kvm_mmu_page *page;
968 page = container_of(kvm->arch.active_mmu_pages.prev,
969 struct kvm_mmu_page, link);
970 kvm_mmu_zap_page(kvm, page);
973 kvm->arch.n_free_mmu_pages = 0;
976 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
977 - kvm->arch.n_alloc_mmu_pages;
979 kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
982 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
985 struct hlist_head *bucket;
986 struct kvm_mmu_page *sp;
987 struct hlist_node *node, *n;
990 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
992 index = kvm_page_table_hashfn(gfn);
993 bucket = &kvm->arch.mmu_page_hash[index];
994 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
995 if (sp->gfn == gfn && !sp->role.metaphysical) {
996 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
998 kvm_mmu_zap_page(kvm, sp);
1004 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1006 struct kvm_mmu_page *sp;
1008 while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1009 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1010 kvm_mmu_zap_page(kvm, sp);
1014 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1016 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1017 struct kvm_mmu_page *sp = page_header(__pa(pte));
1019 __set_bit(slot, &sp->slot_bitmap);
1022 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1026 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1028 if (gpa == UNMAPPED_GVA)
1031 down_read(¤t->mm->mmap_sem);
1032 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1033 up_read(¤t->mm->mmap_sem);
1038 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1039 unsigned pt_access, unsigned pte_access,
1040 int user_fault, int write_fault, int dirty,
1041 int *ptwrite, int largepage, gfn_t gfn,
1042 pfn_t pfn, bool speculative)
1045 int was_rmapped = 0;
1046 int was_writeble = is_writeble_pte(*shadow_pte);
1048 pgprintk("%s: spte %llx access %x write_fault %d"
1049 " user_fault %d gfn %lx\n",
1050 __func__, *shadow_pte, pt_access,
1051 write_fault, user_fault, gfn);
1053 if (is_rmap_pte(*shadow_pte)) {
1055 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1056 * the parent of the now unreachable PTE.
1058 if (largepage && !is_large_pte(*shadow_pte)) {
1059 struct kvm_mmu_page *child;
1060 u64 pte = *shadow_pte;
1062 child = page_header(pte & PT64_BASE_ADDR_MASK);
1063 mmu_page_remove_parent_pte(child, shadow_pte);
1064 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1065 pgprintk("hfn old %lx new %lx\n",
1066 spte_to_pfn(*shadow_pte), pfn);
1067 rmap_remove(vcpu->kvm, shadow_pte);
1070 was_rmapped = is_large_pte(*shadow_pte);
1077 * We don't set the accessed bit, since we sometimes want to see
1078 * whether the guest actually used the pte (in order to detect
1081 spte = shadow_base_present_pte | shadow_dirty_mask;
1083 pte_access |= PT_ACCESSED_MASK;
1085 pte_access &= ~ACC_WRITE_MASK;
1086 if (pte_access & ACC_EXEC_MASK)
1087 spte |= shadow_x_mask;
1089 spte |= shadow_nx_mask;
1090 if (pte_access & ACC_USER_MASK)
1091 spte |= shadow_user_mask;
1093 spte |= PT_PAGE_SIZE_MASK;
1095 spte |= (u64)pfn << PAGE_SHIFT;
1097 if ((pte_access & ACC_WRITE_MASK)
1098 || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1099 struct kvm_mmu_page *shadow;
1101 spte |= PT_WRITABLE_MASK;
1103 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1105 (largepage && has_wrprotected_page(vcpu->kvm, gfn))) {
1106 pgprintk("%s: found shadow page for %lx, marking ro\n",
1108 pte_access &= ~ACC_WRITE_MASK;
1109 if (is_writeble_pte(spte)) {
1110 spte &= ~PT_WRITABLE_MASK;
1111 kvm_x86_ops->tlb_flush(vcpu);
1118 if (pte_access & ACC_WRITE_MASK)
1119 mark_page_dirty(vcpu->kvm, gfn);
1121 pgprintk("%s: setting spte %llx\n", __func__, spte);
1122 pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1123 (spte&PT_PAGE_SIZE_MASK)? "2MB" : "4kB",
1124 (spte&PT_WRITABLE_MASK)?"RW":"R", gfn, spte, shadow_pte);
1125 set_shadow_pte(shadow_pte, spte);
1126 if (!was_rmapped && (spte & PT_PAGE_SIZE_MASK)
1127 && (spte & PT_PRESENT_MASK))
1128 ++vcpu->kvm->stat.lpages;
1130 page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1132 rmap_add(vcpu, shadow_pte, gfn, largepage);
1133 if (!is_rmap_pte(*shadow_pte))
1134 kvm_release_pfn_clean(pfn);
1137 kvm_release_pfn_dirty(pfn);
1139 kvm_release_pfn_clean(pfn);
1142 vcpu->arch.last_pte_updated = shadow_pte;
1143 vcpu->arch.last_pte_gfn = gfn;
1147 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1151 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1152 int largepage, gfn_t gfn, pfn_t pfn,
1155 hpa_t table_addr = vcpu->arch.mmu.root_hpa;
1159 u32 index = PT64_INDEX(v, level);
1162 ASSERT(VALID_PAGE(table_addr));
1163 table = __va(table_addr);
1166 mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1167 0, write, 1, &pt_write, 0, gfn, pfn, false);
1171 if (largepage && level == 2) {
1172 mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1173 0, write, 1, &pt_write, 1, gfn, pfn, false);
1177 if (table[index] == shadow_trap_nonpresent_pte) {
1178 struct kvm_mmu_page *new_table;
1181 pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1183 new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1185 1, ACC_ALL, &table[index]);
1187 pgprintk("nonpaging_map: ENOMEM\n");
1188 kvm_release_pfn_clean(pfn);
1192 set_shadow_pte(&table[index],
1193 __pa(new_table->spt)
1194 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1195 | shadow_user_mask | shadow_x_mask);
1197 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1201 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1207 down_read(¤t->mm->mmap_sem);
1208 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1209 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1213 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1214 up_read(¤t->mm->mmap_sem);
1217 if (is_error_pfn(pfn)) {
1218 kvm_release_pfn_clean(pfn);
1222 spin_lock(&vcpu->kvm->mmu_lock);
1223 kvm_mmu_free_some_pages(vcpu);
1224 r = __direct_map(vcpu, v, write, largepage, gfn, pfn,
1226 spin_unlock(&vcpu->kvm->mmu_lock);
1233 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1236 struct kvm_mmu_page *sp;
1238 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1240 spin_lock(&vcpu->kvm->mmu_lock);
1241 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1242 hpa_t root = vcpu->arch.mmu.root_hpa;
1244 sp = page_header(root);
1246 if (!sp->root_count && sp->role.invalid)
1247 kvm_mmu_zap_page(vcpu->kvm, sp);
1248 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1249 spin_unlock(&vcpu->kvm->mmu_lock);
1252 for (i = 0; i < 4; ++i) {
1253 hpa_t root = vcpu->arch.mmu.pae_root[i];
1256 root &= PT64_BASE_ADDR_MASK;
1257 sp = page_header(root);
1259 if (!sp->root_count && sp->role.invalid)
1260 kvm_mmu_zap_page(vcpu->kvm, sp);
1262 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1264 spin_unlock(&vcpu->kvm->mmu_lock);
1265 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1268 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1272 struct kvm_mmu_page *sp;
1273 int metaphysical = 0;
1275 root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1277 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1278 hpa_t root = vcpu->arch.mmu.root_hpa;
1280 ASSERT(!VALID_PAGE(root));
1283 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1284 PT64_ROOT_LEVEL, metaphysical,
1286 root = __pa(sp->spt);
1288 vcpu->arch.mmu.root_hpa = root;
1291 metaphysical = !is_paging(vcpu);
1294 for (i = 0; i < 4; ++i) {
1295 hpa_t root = vcpu->arch.mmu.pae_root[i];
1297 ASSERT(!VALID_PAGE(root));
1298 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1299 if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1300 vcpu->arch.mmu.pae_root[i] = 0;
1303 root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1304 } else if (vcpu->arch.mmu.root_level == 0)
1306 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1307 PT32_ROOT_LEVEL, metaphysical,
1309 root = __pa(sp->spt);
1311 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1313 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1316 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1321 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1327 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1328 r = mmu_topup_memory_caches(vcpu);
1333 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1335 gfn = gva >> PAGE_SHIFT;
1337 return nonpaging_map(vcpu, gva & PAGE_MASK,
1338 error_code & PFERR_WRITE_MASK, gfn);
1341 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1347 gfn_t gfn = gpa >> PAGE_SHIFT;
1350 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1352 r = mmu_topup_memory_caches(vcpu);
1356 down_read(¤t->mm->mmap_sem);
1357 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1358 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1361 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1362 up_read(¤t->mm->mmap_sem);
1363 if (is_error_pfn(pfn)) {
1364 kvm_release_pfn_clean(pfn);
1367 spin_lock(&vcpu->kvm->mmu_lock);
1368 kvm_mmu_free_some_pages(vcpu);
1369 r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1370 largepage, gfn, pfn, kvm_x86_ops->get_tdp_level());
1371 spin_unlock(&vcpu->kvm->mmu_lock);
1376 static void nonpaging_free(struct kvm_vcpu *vcpu)
1378 mmu_free_roots(vcpu);
1381 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1383 struct kvm_mmu *context = &vcpu->arch.mmu;
1385 context->new_cr3 = nonpaging_new_cr3;
1386 context->page_fault = nonpaging_page_fault;
1387 context->gva_to_gpa = nonpaging_gva_to_gpa;
1388 context->free = nonpaging_free;
1389 context->prefetch_page = nonpaging_prefetch_page;
1390 context->root_level = 0;
1391 context->shadow_root_level = PT32E_ROOT_LEVEL;
1392 context->root_hpa = INVALID_PAGE;
1396 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1398 ++vcpu->stat.tlb_flush;
1399 kvm_x86_ops->tlb_flush(vcpu);
1402 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1404 pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1405 mmu_free_roots(vcpu);
1408 static void inject_page_fault(struct kvm_vcpu *vcpu,
1412 kvm_inject_page_fault(vcpu, addr, err_code);
1415 static void paging_free(struct kvm_vcpu *vcpu)
1417 nonpaging_free(vcpu);
1421 #include "paging_tmpl.h"
1425 #include "paging_tmpl.h"
1428 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1430 struct kvm_mmu *context = &vcpu->arch.mmu;
1432 ASSERT(is_pae(vcpu));
1433 context->new_cr3 = paging_new_cr3;
1434 context->page_fault = paging64_page_fault;
1435 context->gva_to_gpa = paging64_gva_to_gpa;
1436 context->prefetch_page = paging64_prefetch_page;
1437 context->free = paging_free;
1438 context->root_level = level;
1439 context->shadow_root_level = level;
1440 context->root_hpa = INVALID_PAGE;
1444 static int paging64_init_context(struct kvm_vcpu *vcpu)
1446 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1449 static int paging32_init_context(struct kvm_vcpu *vcpu)
1451 struct kvm_mmu *context = &vcpu->arch.mmu;
1453 context->new_cr3 = paging_new_cr3;
1454 context->page_fault = paging32_page_fault;
1455 context->gva_to_gpa = paging32_gva_to_gpa;
1456 context->free = paging_free;
1457 context->prefetch_page = paging32_prefetch_page;
1458 context->root_level = PT32_ROOT_LEVEL;
1459 context->shadow_root_level = PT32E_ROOT_LEVEL;
1460 context->root_hpa = INVALID_PAGE;
1464 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1466 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1469 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1471 struct kvm_mmu *context = &vcpu->arch.mmu;
1473 context->new_cr3 = nonpaging_new_cr3;
1474 context->page_fault = tdp_page_fault;
1475 context->free = nonpaging_free;
1476 context->prefetch_page = nonpaging_prefetch_page;
1477 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
1478 context->root_hpa = INVALID_PAGE;
1480 if (!is_paging(vcpu)) {
1481 context->gva_to_gpa = nonpaging_gva_to_gpa;
1482 context->root_level = 0;
1483 } else if (is_long_mode(vcpu)) {
1484 context->gva_to_gpa = paging64_gva_to_gpa;
1485 context->root_level = PT64_ROOT_LEVEL;
1486 } else if (is_pae(vcpu)) {
1487 context->gva_to_gpa = paging64_gva_to_gpa;
1488 context->root_level = PT32E_ROOT_LEVEL;
1490 context->gva_to_gpa = paging32_gva_to_gpa;
1491 context->root_level = PT32_ROOT_LEVEL;
1497 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1500 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1502 if (!is_paging(vcpu))
1503 return nonpaging_init_context(vcpu);
1504 else if (is_long_mode(vcpu))
1505 return paging64_init_context(vcpu);
1506 else if (is_pae(vcpu))
1507 return paging32E_init_context(vcpu);
1509 return paging32_init_context(vcpu);
1512 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1514 vcpu->arch.update_pte.pfn = bad_pfn;
1517 return init_kvm_tdp_mmu(vcpu);
1519 return init_kvm_softmmu(vcpu);
1522 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1525 if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1526 vcpu->arch.mmu.free(vcpu);
1527 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1531 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1533 destroy_kvm_mmu(vcpu);
1534 return init_kvm_mmu(vcpu);
1536 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1538 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1542 r = mmu_topup_memory_caches(vcpu);
1545 spin_lock(&vcpu->kvm->mmu_lock);
1546 kvm_mmu_free_some_pages(vcpu);
1547 mmu_alloc_roots(vcpu);
1548 spin_unlock(&vcpu->kvm->mmu_lock);
1549 kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1550 kvm_mmu_flush_tlb(vcpu);
1554 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1556 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1558 mmu_free_roots(vcpu);
1561 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1562 struct kvm_mmu_page *sp,
1566 struct kvm_mmu_page *child;
1569 if (is_shadow_present_pte(pte)) {
1570 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
1572 rmap_remove(vcpu->kvm, spte);
1574 child = page_header(pte & PT64_BASE_ADDR_MASK);
1575 mmu_page_remove_parent_pte(child, spte);
1578 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1579 if (is_large_pte(pte))
1580 --vcpu->kvm->stat.lpages;
1583 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1584 struct kvm_mmu_page *sp,
1588 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1589 if (!vcpu->arch.update_pte.largepage ||
1590 sp->role.glevels == PT32_ROOT_LEVEL) {
1591 ++vcpu->kvm->stat.mmu_pde_zapped;
1596 ++vcpu->kvm->stat.mmu_pte_updated;
1597 if (sp->role.glevels == PT32_ROOT_LEVEL)
1598 paging32_update_pte(vcpu, sp, spte, new);
1600 paging64_update_pte(vcpu, sp, spte, new);
1603 static bool need_remote_flush(u64 old, u64 new)
1605 if (!is_shadow_present_pte(old))
1607 if (!is_shadow_present_pte(new))
1609 if ((old ^ new) & PT64_BASE_ADDR_MASK)
1611 old ^= PT64_NX_MASK;
1612 new ^= PT64_NX_MASK;
1613 return (old & ~new & PT64_PERM_MASK) != 0;
1616 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1618 if (need_remote_flush(old, new))
1619 kvm_flush_remote_tlbs(vcpu->kvm);
1621 kvm_mmu_flush_tlb(vcpu);
1624 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1626 u64 *spte = vcpu->arch.last_pte_updated;
1628 return !!(spte && (*spte & shadow_accessed_mask));
1631 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1632 const u8 *new, int bytes)
1639 vcpu->arch.update_pte.largepage = 0;
1641 if (bytes != 4 && bytes != 8)
1645 * Assume that the pte write on a page table of the same type
1646 * as the current vcpu paging mode. This is nearly always true
1647 * (might be false while changing modes). Note it is verified later
1651 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1652 if ((bytes == 4) && (gpa % 4 == 0)) {
1653 r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1656 memcpy((void *)&gpte + (gpa % 8), new, 4);
1657 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1658 memcpy((void *)&gpte, new, 8);
1661 if ((bytes == 4) && (gpa % 4 == 0))
1662 memcpy((void *)&gpte, new, 4);
1664 if (!is_present_pte(gpte))
1666 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1668 down_read(¤t->mm->mmap_sem);
1669 if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
1670 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1671 vcpu->arch.update_pte.largepage = 1;
1673 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1674 up_read(¤t->mm->mmap_sem);
1676 if (is_error_pfn(pfn)) {
1677 kvm_release_pfn_clean(pfn);
1680 vcpu->arch.update_pte.gfn = gfn;
1681 vcpu->arch.update_pte.pfn = pfn;
1684 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
1686 u64 *spte = vcpu->arch.last_pte_updated;
1689 && vcpu->arch.last_pte_gfn == gfn
1690 && shadow_accessed_mask
1691 && !(*spte & shadow_accessed_mask)
1692 && is_shadow_present_pte(*spte))
1693 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
1696 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1697 const u8 *new, int bytes)
1699 gfn_t gfn = gpa >> PAGE_SHIFT;
1700 struct kvm_mmu_page *sp;
1701 struct hlist_node *node, *n;
1702 struct hlist_head *bucket;
1706 unsigned offset = offset_in_page(gpa);
1708 unsigned page_offset;
1709 unsigned misaligned;
1716 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
1717 mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1718 spin_lock(&vcpu->kvm->mmu_lock);
1719 kvm_mmu_access_page(vcpu, gfn);
1720 kvm_mmu_free_some_pages(vcpu);
1721 ++vcpu->kvm->stat.mmu_pte_write;
1722 kvm_mmu_audit(vcpu, "pre pte write");
1723 if (gfn == vcpu->arch.last_pt_write_gfn
1724 && !last_updated_pte_accessed(vcpu)) {
1725 ++vcpu->arch.last_pt_write_count;
1726 if (vcpu->arch.last_pt_write_count >= 3)
1729 vcpu->arch.last_pt_write_gfn = gfn;
1730 vcpu->arch.last_pt_write_count = 1;
1731 vcpu->arch.last_pte_updated = NULL;
1733 index = kvm_page_table_hashfn(gfn);
1734 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1735 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1736 if (sp->gfn != gfn || sp->role.metaphysical)
1738 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1739 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1740 misaligned |= bytes < 4;
1741 if (misaligned || flooded) {
1743 * Misaligned accesses are too much trouble to fix
1744 * up; also, they usually indicate a page is not used
1747 * If we're seeing too many writes to a page,
1748 * it may no longer be a page table, or we may be
1749 * forking, in which case it is better to unmap the
1752 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1753 gpa, bytes, sp->role.word);
1754 kvm_mmu_zap_page(vcpu->kvm, sp);
1755 ++vcpu->kvm->stat.mmu_flooded;
1758 page_offset = offset;
1759 level = sp->role.level;
1761 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1762 page_offset <<= 1; /* 32->64 */
1764 * A 32-bit pde maps 4MB while the shadow pdes map
1765 * only 2MB. So we need to double the offset again
1766 * and zap two pdes instead of one.
1768 if (level == PT32_ROOT_LEVEL) {
1769 page_offset &= ~7; /* kill rounding error */
1773 quadrant = page_offset >> PAGE_SHIFT;
1774 page_offset &= ~PAGE_MASK;
1775 if (quadrant != sp->role.quadrant)
1778 spte = &sp->spt[page_offset / sizeof(*spte)];
1779 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1781 r = kvm_read_guest_atomic(vcpu->kvm,
1782 gpa & ~(u64)(pte_size - 1),
1784 new = (const void *)&gentry;
1790 mmu_pte_write_zap_pte(vcpu, sp, spte);
1792 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1793 mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1797 kvm_mmu_audit(vcpu, "post pte write");
1798 spin_unlock(&vcpu->kvm->mmu_lock);
1799 if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
1800 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
1801 vcpu->arch.update_pte.pfn = bad_pfn;
1805 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1810 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1812 spin_lock(&vcpu->kvm->mmu_lock);
1813 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1814 spin_unlock(&vcpu->kvm->mmu_lock);
1817 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
1819 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1821 while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1822 struct kvm_mmu_page *sp;
1824 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1825 struct kvm_mmu_page, link);
1826 kvm_mmu_zap_page(vcpu->kvm, sp);
1827 ++vcpu->kvm->stat.mmu_recycled;
1831 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1834 enum emulation_result er;
1836 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1845 r = mmu_topup_memory_caches(vcpu);
1849 er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1854 case EMULATE_DO_MMIO:
1855 ++vcpu->stat.mmio_exits;
1858 kvm_report_emulation_failure(vcpu, "pagetable");
1866 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1868 void kvm_enable_tdp(void)
1872 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
1874 void kvm_disable_tdp(void)
1876 tdp_enabled = false;
1878 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
1880 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1882 struct kvm_mmu_page *sp;
1884 while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1885 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1886 struct kvm_mmu_page, link);
1887 kvm_mmu_zap_page(vcpu->kvm, sp);
1890 free_page((unsigned long)vcpu->arch.mmu.pae_root);
1893 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1900 if (vcpu->kvm->arch.n_requested_mmu_pages)
1901 vcpu->kvm->arch.n_free_mmu_pages =
1902 vcpu->kvm->arch.n_requested_mmu_pages;
1904 vcpu->kvm->arch.n_free_mmu_pages =
1905 vcpu->kvm->arch.n_alloc_mmu_pages;
1907 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1908 * Therefore we need to allocate shadow page tables in the first
1909 * 4GB of memory, which happens to fit the DMA32 zone.
1911 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1914 vcpu->arch.mmu.pae_root = page_address(page);
1915 for (i = 0; i < 4; ++i)
1916 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1921 free_mmu_pages(vcpu);
1925 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1928 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1930 return alloc_mmu_pages(vcpu);
1933 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1936 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1938 return init_kvm_mmu(vcpu);
1941 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1945 destroy_kvm_mmu(vcpu);
1946 free_mmu_pages(vcpu);
1947 mmu_free_memory_caches(vcpu);
1950 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1952 struct kvm_mmu_page *sp;
1954 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1958 if (!test_bit(slot, &sp->slot_bitmap))
1962 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1964 if (pt[i] & PT_WRITABLE_MASK)
1965 pt[i] &= ~PT_WRITABLE_MASK;
1969 void kvm_mmu_zap_all(struct kvm *kvm)
1971 struct kvm_mmu_page *sp, *node;
1973 spin_lock(&kvm->mmu_lock);
1974 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1975 kvm_mmu_zap_page(kvm, sp);
1976 spin_unlock(&kvm->mmu_lock);
1978 kvm_flush_remote_tlbs(kvm);
1981 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
1983 struct kvm_mmu_page *page;
1985 page = container_of(kvm->arch.active_mmu_pages.prev,
1986 struct kvm_mmu_page, link);
1987 kvm_mmu_zap_page(kvm, page);
1990 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
1993 struct kvm *kvm_freed = NULL;
1994 int cache_count = 0;
1996 spin_lock(&kvm_lock);
1998 list_for_each_entry(kvm, &vm_list, vm_list) {
2001 if (!down_read_trylock(&kvm->slots_lock))
2003 spin_lock(&kvm->mmu_lock);
2004 npages = kvm->arch.n_alloc_mmu_pages -
2005 kvm->arch.n_free_mmu_pages;
2006 cache_count += npages;
2007 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2008 kvm_mmu_remove_one_alloc_mmu_page(kvm);
2014 spin_unlock(&kvm->mmu_lock);
2015 up_read(&kvm->slots_lock);
2018 list_move_tail(&kvm_freed->vm_list, &vm_list);
2020 spin_unlock(&kvm_lock);
2025 static struct shrinker mmu_shrinker = {
2026 .shrink = mmu_shrink,
2027 .seeks = DEFAULT_SEEKS * 10,
2030 static void mmu_destroy_caches(void)
2032 if (pte_chain_cache)
2033 kmem_cache_destroy(pte_chain_cache);
2034 if (rmap_desc_cache)
2035 kmem_cache_destroy(rmap_desc_cache);
2036 if (mmu_page_header_cache)
2037 kmem_cache_destroy(mmu_page_header_cache);
2040 void kvm_mmu_module_exit(void)
2042 mmu_destroy_caches();
2043 unregister_shrinker(&mmu_shrinker);
2046 int kvm_mmu_module_init(void)
2048 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2049 sizeof(struct kvm_pte_chain),
2051 if (!pte_chain_cache)
2053 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2054 sizeof(struct kvm_rmap_desc),
2056 if (!rmap_desc_cache)
2059 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2060 sizeof(struct kvm_mmu_page),
2062 if (!mmu_page_header_cache)
2065 register_shrinker(&mmu_shrinker);
2070 mmu_destroy_caches();
2075 * Caculate mmu pages needed for kvm.
2077 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2080 unsigned int nr_mmu_pages;
2081 unsigned int nr_pages = 0;
2083 for (i = 0; i < kvm->nmemslots; i++)
2084 nr_pages += kvm->memslots[i].npages;
2086 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2087 nr_mmu_pages = max(nr_mmu_pages,
2088 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2090 return nr_mmu_pages;
2093 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2096 if (len > buffer->len)
2101 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2106 ret = pv_mmu_peek_buffer(buffer, len);
2111 buffer->processed += len;
2115 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2116 gpa_t addr, gpa_t value)
2121 if (!is_long_mode(vcpu) && !is_pae(vcpu))
2124 r = mmu_topup_memory_caches(vcpu);
2128 if (!emulator_write_phys(vcpu, addr, &value, bytes))
2134 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2136 kvm_x86_ops->tlb_flush(vcpu);
2140 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2142 spin_lock(&vcpu->kvm->mmu_lock);
2143 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2144 spin_unlock(&vcpu->kvm->mmu_lock);
2148 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2149 struct kvm_pv_mmu_op_buffer *buffer)
2151 struct kvm_mmu_op_header *header;
2153 header = pv_mmu_peek_buffer(buffer, sizeof *header);
2156 switch (header->op) {
2157 case KVM_MMU_OP_WRITE_PTE: {
2158 struct kvm_mmu_op_write_pte *wpte;
2160 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2163 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2166 case KVM_MMU_OP_FLUSH_TLB: {
2167 struct kvm_mmu_op_flush_tlb *ftlb;
2169 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2172 return kvm_pv_mmu_flush_tlb(vcpu);
2174 case KVM_MMU_OP_RELEASE_PT: {
2175 struct kvm_mmu_op_release_pt *rpt;
2177 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2180 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2186 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2187 gpa_t addr, unsigned long *ret)
2190 struct kvm_pv_mmu_op_buffer buffer;
2192 buffer.ptr = buffer.buf;
2193 buffer.len = min_t(unsigned long, bytes, sizeof buffer.buf);
2194 buffer.processed = 0;
2196 r = kvm_read_guest(vcpu->kvm, addr, buffer.buf, buffer.len);
2200 while (buffer.len) {
2201 r = kvm_pv_mmu_op_one(vcpu, &buffer);
2210 *ret = buffer.processed;
2216 static const char *audit_msg;
2218 static gva_t canonicalize(gva_t gva)
2220 #ifdef CONFIG_X86_64
2221 gva = (long long)(gva << 16) >> 16;
2226 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2227 gva_t va, int level)
2229 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2231 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2233 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2236 if (ent == shadow_trap_nonpresent_pte)
2239 va = canonicalize(va);
2241 if (ent == shadow_notrap_nonpresent_pte)
2242 printk(KERN_ERR "audit: (%s) nontrapping pte"
2243 " in nonleaf level: levels %d gva %lx"
2244 " level %d pte %llx\n", audit_msg,
2245 vcpu->arch.mmu.root_level, va, level, ent);
2247 audit_mappings_page(vcpu, ent, va, level - 1);
2249 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2250 hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2252 if (is_shadow_present_pte(ent)
2253 && (ent & PT64_BASE_ADDR_MASK) != hpa)
2254 printk(KERN_ERR "xx audit error: (%s) levels %d"
2255 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2256 audit_msg, vcpu->arch.mmu.root_level,
2258 is_shadow_present_pte(ent));
2259 else if (ent == shadow_notrap_nonpresent_pte
2260 && !is_error_hpa(hpa))
2261 printk(KERN_ERR "audit: (%s) notrap shadow,"
2262 " valid guest gva %lx\n", audit_msg, va);
2263 kvm_release_pfn_clean(pfn);
2269 static void audit_mappings(struct kvm_vcpu *vcpu)
2273 if (vcpu->arch.mmu.root_level == 4)
2274 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2276 for (i = 0; i < 4; ++i)
2277 if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2278 audit_mappings_page(vcpu,
2279 vcpu->arch.mmu.pae_root[i],
2284 static int count_rmaps(struct kvm_vcpu *vcpu)
2289 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2290 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2291 struct kvm_rmap_desc *d;
2293 for (j = 0; j < m->npages; ++j) {
2294 unsigned long *rmapp = &m->rmap[j];
2298 if (!(*rmapp & 1)) {
2302 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2304 for (k = 0; k < RMAP_EXT; ++k)
2305 if (d->shadow_ptes[k])
2316 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2319 struct kvm_mmu_page *sp;
2322 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2325 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2328 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2331 if (!(ent & PT_PRESENT_MASK))
2333 if (!(ent & PT_WRITABLE_MASK))
2341 static void audit_rmap(struct kvm_vcpu *vcpu)
2343 int n_rmap = count_rmaps(vcpu);
2344 int n_actual = count_writable_mappings(vcpu);
2346 if (n_rmap != n_actual)
2347 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2348 __func__, audit_msg, n_rmap, n_actual);
2351 static void audit_write_protection(struct kvm_vcpu *vcpu)
2353 struct kvm_mmu_page *sp;
2354 struct kvm_memory_slot *slot;
2355 unsigned long *rmapp;
2358 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2359 if (sp->role.metaphysical)
2362 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2363 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2364 rmapp = &slot->rmap[gfn - slot->base_gfn];
2366 printk(KERN_ERR "%s: (%s) shadow page has writable"
2367 " mappings: gfn %lx role %x\n",
2368 __func__, audit_msg, sp->gfn,
2373 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2380 audit_write_protection(vcpu);
2381 audit_mappings(vcpu);