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.
22 #include <linux/kvm_host.h>
23 #include <linux/types.h>
24 #include <linux/string.h>
26 #include <linux/highmem.h>
27 #include <linux/module.h>
28 #include <linux/swap.h>
29 #include <linux/hugetlb.h>
30 #include <linux/compiler.h>
33 #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);
73 static int oos_shadow = 1;
74 module_param(oos_shadow, bool, 0644);
77 #define ASSERT(x) do { } while (0)
81 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
82 __FILE__, __LINE__, #x); \
86 #define PT_FIRST_AVAIL_BITS_SHIFT 9
87 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
89 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
91 #define PT64_LEVEL_BITS 9
93 #define PT64_LEVEL_SHIFT(level) \
94 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
96 #define PT64_LEVEL_MASK(level) \
97 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
99 #define PT64_INDEX(address, level)\
100 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
103 #define PT32_LEVEL_BITS 10
105 #define PT32_LEVEL_SHIFT(level) \
106 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
108 #define PT32_LEVEL_MASK(level) \
109 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
111 #define PT32_INDEX(address, level)\
112 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
115 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
116 #define PT64_DIR_BASE_ADDR_MASK \
117 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
119 #define PT32_BASE_ADDR_MASK PAGE_MASK
120 #define PT32_DIR_BASE_ADDR_MASK \
121 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
123 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
126 #define PFERR_PRESENT_MASK (1U << 0)
127 #define PFERR_WRITE_MASK (1U << 1)
128 #define PFERR_USER_MASK (1U << 2)
129 #define PFERR_FETCH_MASK (1U << 4)
131 #define PT_DIRECTORY_LEVEL 2
132 #define PT_PAGE_TABLE_LEVEL 1
136 #define ACC_EXEC_MASK 1
137 #define ACC_WRITE_MASK PT_WRITABLE_MASK
138 #define ACC_USER_MASK PT_USER_MASK
139 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
141 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
143 struct kvm_rmap_desc {
144 u64 *shadow_ptes[RMAP_EXT];
145 struct kvm_rmap_desc *more;
148 struct kvm_shadow_walk {
149 int (*entry)(struct kvm_shadow_walk *walk, struct kvm_vcpu *vcpu,
150 u64 addr, u64 *spte, int level);
153 struct kvm_shadow_walk_iterator {
161 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
162 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
163 shadow_walk_okay(&(_walker)); \
164 shadow_walk_next(&(_walker)))
167 struct kvm_unsync_walk {
168 int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
171 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
173 static struct kmem_cache *pte_chain_cache;
174 static struct kmem_cache *rmap_desc_cache;
175 static struct kmem_cache *mmu_page_header_cache;
177 static u64 __read_mostly shadow_trap_nonpresent_pte;
178 static u64 __read_mostly shadow_notrap_nonpresent_pte;
179 static u64 __read_mostly shadow_base_present_pte;
180 static u64 __read_mostly shadow_nx_mask;
181 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
182 static u64 __read_mostly shadow_user_mask;
183 static u64 __read_mostly shadow_accessed_mask;
184 static u64 __read_mostly shadow_dirty_mask;
185 static u64 __read_mostly shadow_mt_mask;
187 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
189 shadow_trap_nonpresent_pte = trap_pte;
190 shadow_notrap_nonpresent_pte = notrap_pte;
192 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
194 void kvm_mmu_set_base_ptes(u64 base_pte)
196 shadow_base_present_pte = base_pte;
198 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
200 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
201 u64 dirty_mask, u64 nx_mask, u64 x_mask, u64 mt_mask)
203 shadow_user_mask = user_mask;
204 shadow_accessed_mask = accessed_mask;
205 shadow_dirty_mask = dirty_mask;
206 shadow_nx_mask = nx_mask;
207 shadow_x_mask = x_mask;
208 shadow_mt_mask = mt_mask;
210 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
212 static int is_write_protection(struct kvm_vcpu *vcpu)
214 return vcpu->arch.cr0 & X86_CR0_WP;
217 static int is_cpuid_PSE36(void)
222 static int is_nx(struct kvm_vcpu *vcpu)
224 return vcpu->arch.shadow_efer & EFER_NX;
227 static int is_present_pte(unsigned long pte)
229 return pte & PT_PRESENT_MASK;
232 static int is_shadow_present_pte(u64 pte)
234 return pte != shadow_trap_nonpresent_pte
235 && pte != shadow_notrap_nonpresent_pte;
238 static int is_large_pte(u64 pte)
240 return pte & PT_PAGE_SIZE_MASK;
243 static int is_writeble_pte(unsigned long pte)
245 return pte & PT_WRITABLE_MASK;
248 static int is_dirty_pte(unsigned long pte)
250 return pte & shadow_dirty_mask;
253 static int is_rmap_pte(u64 pte)
255 return is_shadow_present_pte(pte);
258 static pfn_t spte_to_pfn(u64 pte)
260 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
263 static gfn_t pse36_gfn_delta(u32 gpte)
265 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
267 return (gpte & PT32_DIR_PSE36_MASK) << shift;
270 static void set_shadow_pte(u64 *sptep, u64 spte)
273 set_64bit((unsigned long *)sptep, spte);
275 set_64bit((unsigned long long *)sptep, spte);
279 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
280 struct kmem_cache *base_cache, int min)
284 if (cache->nobjs >= min)
286 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
287 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
290 cache->objects[cache->nobjs++] = obj;
295 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
298 kfree(mc->objects[--mc->nobjs]);
301 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
306 if (cache->nobjs >= min)
308 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
309 page = alloc_page(GFP_KERNEL);
312 set_page_private(page, 0);
313 cache->objects[cache->nobjs++] = page_address(page);
318 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
321 free_page((unsigned long)mc->objects[--mc->nobjs]);
324 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
328 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
332 r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
336 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
339 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
340 mmu_page_header_cache, 4);
345 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
347 mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
348 mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
349 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
350 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
353 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
359 p = mc->objects[--mc->nobjs];
364 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
366 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
367 sizeof(struct kvm_pte_chain));
370 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
375 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
377 return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
378 sizeof(struct kvm_rmap_desc));
381 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
387 * Return the pointer to the largepage write count for a given
388 * gfn, handling slots that are not large page aligned.
390 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
394 idx = (gfn / KVM_PAGES_PER_HPAGE) -
395 (slot->base_gfn / KVM_PAGES_PER_HPAGE);
396 return &slot->lpage_info[idx].write_count;
399 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
403 gfn = unalias_gfn(kvm, gfn);
404 write_count = slot_largepage_idx(gfn,
405 gfn_to_memslot_unaliased(kvm, gfn));
409 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
413 gfn = unalias_gfn(kvm, gfn);
414 write_count = slot_largepage_idx(gfn,
415 gfn_to_memslot_unaliased(kvm, gfn));
417 WARN_ON(*write_count < 0);
420 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
422 struct kvm_memory_slot *slot;
425 gfn = unalias_gfn(kvm, gfn);
426 slot = gfn_to_memslot_unaliased(kvm, gfn);
428 largepage_idx = slot_largepage_idx(gfn, slot);
429 return *largepage_idx;
435 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
437 struct vm_area_struct *vma;
441 addr = gfn_to_hva(kvm, gfn);
442 if (kvm_is_error_hva(addr))
445 down_read(¤t->mm->mmap_sem);
446 vma = find_vma(current->mm, addr);
447 if (vma && is_vm_hugetlb_page(vma))
449 up_read(¤t->mm->mmap_sem);
454 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
456 struct kvm_memory_slot *slot;
458 if (has_wrprotected_page(vcpu->kvm, large_gfn))
461 if (!host_largepage_backed(vcpu->kvm, large_gfn))
464 slot = gfn_to_memslot(vcpu->kvm, large_gfn);
465 if (slot && slot->dirty_bitmap)
472 * Take gfn and return the reverse mapping to it.
473 * Note: gfn must be unaliased before this function get called
476 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
478 struct kvm_memory_slot *slot;
481 slot = gfn_to_memslot(kvm, gfn);
483 return &slot->rmap[gfn - slot->base_gfn];
485 idx = (gfn / KVM_PAGES_PER_HPAGE) -
486 (slot->base_gfn / KVM_PAGES_PER_HPAGE);
488 return &slot->lpage_info[idx].rmap_pde;
492 * Reverse mapping data structures:
494 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
495 * that points to page_address(page).
497 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
498 * containing more mappings.
500 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
502 struct kvm_mmu_page *sp;
503 struct kvm_rmap_desc *desc;
504 unsigned long *rmapp;
507 if (!is_rmap_pte(*spte))
509 gfn = unalias_gfn(vcpu->kvm, gfn);
510 sp = page_header(__pa(spte));
511 sp->gfns[spte - sp->spt] = gfn;
512 rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
514 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
515 *rmapp = (unsigned long)spte;
516 } else if (!(*rmapp & 1)) {
517 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
518 desc = mmu_alloc_rmap_desc(vcpu);
519 desc->shadow_ptes[0] = (u64 *)*rmapp;
520 desc->shadow_ptes[1] = spte;
521 *rmapp = (unsigned long)desc | 1;
523 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
524 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
525 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
527 if (desc->shadow_ptes[RMAP_EXT-1]) {
528 desc->more = mmu_alloc_rmap_desc(vcpu);
531 for (i = 0; desc->shadow_ptes[i]; ++i)
533 desc->shadow_ptes[i] = spte;
537 static void rmap_desc_remove_entry(unsigned long *rmapp,
538 struct kvm_rmap_desc *desc,
540 struct kvm_rmap_desc *prev_desc)
544 for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
546 desc->shadow_ptes[i] = desc->shadow_ptes[j];
547 desc->shadow_ptes[j] = NULL;
550 if (!prev_desc && !desc->more)
551 *rmapp = (unsigned long)desc->shadow_ptes[0];
554 prev_desc->more = desc->more;
556 *rmapp = (unsigned long)desc->more | 1;
557 mmu_free_rmap_desc(desc);
560 static void rmap_remove(struct kvm *kvm, u64 *spte)
562 struct kvm_rmap_desc *desc;
563 struct kvm_rmap_desc *prev_desc;
564 struct kvm_mmu_page *sp;
566 unsigned long *rmapp;
569 if (!is_rmap_pte(*spte))
571 sp = page_header(__pa(spte));
572 pfn = spte_to_pfn(*spte);
573 if (*spte & shadow_accessed_mask)
574 kvm_set_pfn_accessed(pfn);
575 if (is_writeble_pte(*spte))
576 kvm_release_pfn_dirty(pfn);
578 kvm_release_pfn_clean(pfn);
579 rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
581 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
583 } else if (!(*rmapp & 1)) {
584 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
585 if ((u64 *)*rmapp != spte) {
586 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
592 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
593 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
596 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
597 if (desc->shadow_ptes[i] == spte) {
598 rmap_desc_remove_entry(rmapp,
610 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
612 struct kvm_rmap_desc *desc;
613 struct kvm_rmap_desc *prev_desc;
619 else if (!(*rmapp & 1)) {
621 return (u64 *)*rmapp;
624 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
628 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
629 if (prev_spte == spte)
630 return desc->shadow_ptes[i];
631 prev_spte = desc->shadow_ptes[i];
638 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
640 unsigned long *rmapp;
642 int write_protected = 0;
644 gfn = unalias_gfn(kvm, gfn);
645 rmapp = gfn_to_rmap(kvm, gfn, 0);
647 spte = rmap_next(kvm, rmapp, NULL);
650 BUG_ON(!(*spte & PT_PRESENT_MASK));
651 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
652 if (is_writeble_pte(*spte)) {
653 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
656 spte = rmap_next(kvm, rmapp, spte);
658 if (write_protected) {
661 spte = rmap_next(kvm, rmapp, NULL);
662 pfn = spte_to_pfn(*spte);
663 kvm_set_pfn_dirty(pfn);
666 /* check for huge page mappings */
667 rmapp = gfn_to_rmap(kvm, gfn, 1);
668 spte = rmap_next(kvm, rmapp, NULL);
671 BUG_ON(!(*spte & PT_PRESENT_MASK));
672 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
673 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
674 if (is_writeble_pte(*spte)) {
675 rmap_remove(kvm, spte);
677 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
681 spte = rmap_next(kvm, rmapp, spte);
684 return write_protected;
687 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
690 int need_tlb_flush = 0;
692 while ((spte = rmap_next(kvm, rmapp, NULL))) {
693 BUG_ON(!(*spte & PT_PRESENT_MASK));
694 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
695 rmap_remove(kvm, spte);
696 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
699 return need_tlb_flush;
702 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
703 int (*handler)(struct kvm *kvm, unsigned long *rmapp))
709 * If mmap_sem isn't taken, we can look the memslots with only
710 * the mmu_lock by skipping over the slots with userspace_addr == 0.
712 for (i = 0; i < kvm->nmemslots; i++) {
713 struct kvm_memory_slot *memslot = &kvm->memslots[i];
714 unsigned long start = memslot->userspace_addr;
717 /* mmu_lock protects userspace_addr */
721 end = start + (memslot->npages << PAGE_SHIFT);
722 if (hva >= start && hva < end) {
723 gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
724 retval |= handler(kvm, &memslot->rmap[gfn_offset]);
725 retval |= handler(kvm,
726 &memslot->lpage_info[
728 KVM_PAGES_PER_HPAGE].rmap_pde);
735 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
737 return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
740 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
745 /* always return old for EPT */
746 if (!shadow_accessed_mask)
749 spte = rmap_next(kvm, rmapp, NULL);
753 BUG_ON(!(_spte & PT_PRESENT_MASK));
754 _young = _spte & PT_ACCESSED_MASK;
757 clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
759 spte = rmap_next(kvm, rmapp, spte);
764 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
766 return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
770 static int is_empty_shadow_page(u64 *spt)
775 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
776 if (is_shadow_present_pte(*pos)) {
777 printk(KERN_ERR "%s: %p %llx\n", __func__,
785 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
787 ASSERT(is_empty_shadow_page(sp->spt));
789 __free_page(virt_to_page(sp->spt));
790 __free_page(virt_to_page(sp->gfns));
792 ++kvm->arch.n_free_mmu_pages;
795 static unsigned kvm_page_table_hashfn(gfn_t gfn)
797 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
800 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
803 struct kvm_mmu_page *sp;
805 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
806 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
807 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
808 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
809 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
810 INIT_LIST_HEAD(&sp->oos_link);
811 ASSERT(is_empty_shadow_page(sp->spt));
812 bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
814 sp->parent_pte = parent_pte;
815 --vcpu->kvm->arch.n_free_mmu_pages;
819 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
820 struct kvm_mmu_page *sp, u64 *parent_pte)
822 struct kvm_pte_chain *pte_chain;
823 struct hlist_node *node;
828 if (!sp->multimapped) {
829 u64 *old = sp->parent_pte;
832 sp->parent_pte = parent_pte;
836 pte_chain = mmu_alloc_pte_chain(vcpu);
837 INIT_HLIST_HEAD(&sp->parent_ptes);
838 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
839 pte_chain->parent_ptes[0] = old;
841 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
842 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
844 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
845 if (!pte_chain->parent_ptes[i]) {
846 pte_chain->parent_ptes[i] = parent_pte;
850 pte_chain = mmu_alloc_pte_chain(vcpu);
852 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
853 pte_chain->parent_ptes[0] = parent_pte;
856 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
859 struct kvm_pte_chain *pte_chain;
860 struct hlist_node *node;
863 if (!sp->multimapped) {
864 BUG_ON(sp->parent_pte != parent_pte);
865 sp->parent_pte = NULL;
868 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
869 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
870 if (!pte_chain->parent_ptes[i])
872 if (pte_chain->parent_ptes[i] != parent_pte)
874 while (i + 1 < NR_PTE_CHAIN_ENTRIES
875 && pte_chain->parent_ptes[i + 1]) {
876 pte_chain->parent_ptes[i]
877 = pte_chain->parent_ptes[i + 1];
880 pte_chain->parent_ptes[i] = NULL;
882 hlist_del(&pte_chain->link);
883 mmu_free_pte_chain(pte_chain);
884 if (hlist_empty(&sp->parent_ptes)) {
886 sp->parent_pte = NULL;
895 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
896 mmu_parent_walk_fn fn)
898 struct kvm_pte_chain *pte_chain;
899 struct hlist_node *node;
900 struct kvm_mmu_page *parent_sp;
903 if (!sp->multimapped && sp->parent_pte) {
904 parent_sp = page_header(__pa(sp->parent_pte));
906 mmu_parent_walk(vcpu, parent_sp, fn);
909 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
910 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
911 if (!pte_chain->parent_ptes[i])
913 parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
915 mmu_parent_walk(vcpu, parent_sp, fn);
919 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
922 struct kvm_mmu_page *sp = page_header(__pa(spte));
924 index = spte - sp->spt;
925 if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
926 sp->unsync_children++;
927 WARN_ON(!sp->unsync_children);
930 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
932 struct kvm_pte_chain *pte_chain;
933 struct hlist_node *node;
939 if (!sp->multimapped) {
940 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
944 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
945 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
946 if (!pte_chain->parent_ptes[i])
948 kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
952 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
954 kvm_mmu_update_parents_unsync(sp);
958 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
959 struct kvm_mmu_page *sp)
961 mmu_parent_walk(vcpu, sp, unsync_walk_fn);
962 kvm_mmu_update_parents_unsync(sp);
965 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
966 struct kvm_mmu_page *sp)
970 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
971 sp->spt[i] = shadow_trap_nonpresent_pte;
974 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
975 struct kvm_mmu_page *sp)
980 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
984 #define KVM_PAGE_ARRAY_NR 16
986 struct kvm_mmu_pages {
987 struct mmu_page_and_offset {
988 struct kvm_mmu_page *sp;
990 } page[KVM_PAGE_ARRAY_NR];
994 #define for_each_unsync_children(bitmap, idx) \
995 for (idx = find_first_bit(bitmap, 512); \
997 idx = find_next_bit(bitmap, 512, idx+1))
999 int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1005 for (i=0; i < pvec->nr; i++)
1006 if (pvec->page[i].sp == sp)
1009 pvec->page[pvec->nr].sp = sp;
1010 pvec->page[pvec->nr].idx = idx;
1012 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1015 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1016 struct kvm_mmu_pages *pvec)
1018 int i, ret, nr_unsync_leaf = 0;
1020 for_each_unsync_children(sp->unsync_child_bitmap, i) {
1021 u64 ent = sp->spt[i];
1023 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1024 struct kvm_mmu_page *child;
1025 child = page_header(ent & PT64_BASE_ADDR_MASK);
1027 if (child->unsync_children) {
1028 if (mmu_pages_add(pvec, child, i))
1031 ret = __mmu_unsync_walk(child, pvec);
1033 __clear_bit(i, sp->unsync_child_bitmap);
1035 nr_unsync_leaf += ret;
1040 if (child->unsync) {
1042 if (mmu_pages_add(pvec, child, i))
1048 if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1049 sp->unsync_children = 0;
1051 return nr_unsync_leaf;
1054 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1055 struct kvm_mmu_pages *pvec)
1057 if (!sp->unsync_children)
1060 mmu_pages_add(pvec, sp, 0);
1061 return __mmu_unsync_walk(sp, pvec);
1064 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1067 struct hlist_head *bucket;
1068 struct kvm_mmu_page *sp;
1069 struct hlist_node *node;
1071 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1072 index = kvm_page_table_hashfn(gfn);
1073 bucket = &kvm->arch.mmu_page_hash[index];
1074 hlist_for_each_entry(sp, node, bucket, hash_link)
1075 if (sp->gfn == gfn && !sp->role.metaphysical
1076 && !sp->role.invalid) {
1077 pgprintk("%s: found role %x\n",
1078 __func__, sp->role.word);
1084 static void kvm_unlink_unsync_global(struct kvm *kvm, struct kvm_mmu_page *sp)
1086 list_del(&sp->oos_link);
1087 --kvm->stat.mmu_unsync_global;
1090 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1092 WARN_ON(!sp->unsync);
1095 kvm_unlink_unsync_global(kvm, sp);
1096 --kvm->stat.mmu_unsync;
1099 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1101 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1103 if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1104 kvm_mmu_zap_page(vcpu->kvm, sp);
1108 if (rmap_write_protect(vcpu->kvm, sp->gfn))
1109 kvm_flush_remote_tlbs(vcpu->kvm);
1110 kvm_unlink_unsync_page(vcpu->kvm, sp);
1111 if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1112 kvm_mmu_zap_page(vcpu->kvm, sp);
1116 kvm_mmu_flush_tlb(vcpu);
1120 struct mmu_page_path {
1121 struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1122 unsigned int idx[PT64_ROOT_LEVEL-1];
1125 #define for_each_sp(pvec, sp, parents, i) \
1126 for (i = mmu_pages_next(&pvec, &parents, -1), \
1127 sp = pvec.page[i].sp; \
1128 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1129 i = mmu_pages_next(&pvec, &parents, i))
1131 int mmu_pages_next(struct kvm_mmu_pages *pvec, struct mmu_page_path *parents,
1136 for (n = i+1; n < pvec->nr; n++) {
1137 struct kvm_mmu_page *sp = pvec->page[n].sp;
1139 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1140 parents->idx[0] = pvec->page[n].idx;
1144 parents->parent[sp->role.level-2] = sp;
1145 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1151 void mmu_pages_clear_parents(struct mmu_page_path *parents)
1153 struct kvm_mmu_page *sp;
1154 unsigned int level = 0;
1157 unsigned int idx = parents->idx[level];
1159 sp = parents->parent[level];
1163 --sp->unsync_children;
1164 WARN_ON((int)sp->unsync_children < 0);
1165 __clear_bit(idx, sp->unsync_child_bitmap);
1167 } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1170 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1171 struct mmu_page_path *parents,
1172 struct kvm_mmu_pages *pvec)
1174 parents->parent[parent->role.level-1] = NULL;
1178 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1179 struct kvm_mmu_page *parent)
1182 struct kvm_mmu_page *sp;
1183 struct mmu_page_path parents;
1184 struct kvm_mmu_pages pages;
1186 kvm_mmu_pages_init(parent, &parents, &pages);
1187 while (mmu_unsync_walk(parent, &pages)) {
1190 for_each_sp(pages, sp, parents, i)
1191 protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1194 kvm_flush_remote_tlbs(vcpu->kvm);
1196 for_each_sp(pages, sp, parents, i) {
1197 kvm_sync_page(vcpu, sp);
1198 mmu_pages_clear_parents(&parents);
1200 cond_resched_lock(&vcpu->kvm->mmu_lock);
1201 kvm_mmu_pages_init(parent, &parents, &pages);
1205 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1213 union kvm_mmu_page_role role;
1216 struct hlist_head *bucket;
1217 struct kvm_mmu_page *sp;
1218 struct hlist_node *node, *tmp;
1220 role = vcpu->arch.mmu.base_role;
1222 role.metaphysical = metaphysical;
1223 role.access = access;
1224 if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1225 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1226 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1227 role.quadrant = quadrant;
1229 pgprintk("%s: looking gfn %lx role %x\n", __func__,
1231 index = kvm_page_table_hashfn(gfn);
1232 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1233 hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1234 if (sp->gfn == gfn) {
1236 if (kvm_sync_page(vcpu, sp))
1239 if (sp->role.word != role.word)
1242 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1243 if (sp->unsync_children) {
1244 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1245 kvm_mmu_mark_parents_unsync(vcpu, sp);
1247 pgprintk("%s: found\n", __func__);
1250 ++vcpu->kvm->stat.mmu_cache_miss;
1251 sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1254 pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1257 sp->global = role.cr4_pge;
1258 hlist_add_head(&sp->hash_link, bucket);
1259 if (!metaphysical) {
1260 if (rmap_write_protect(vcpu->kvm, gfn))
1261 kvm_flush_remote_tlbs(vcpu->kvm);
1262 account_shadowed(vcpu->kvm, gfn);
1264 if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1265 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1267 nonpaging_prefetch_page(vcpu, sp);
1271 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1272 struct kvm_vcpu *vcpu, u64 addr)
1274 iterator->addr = addr;
1275 iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1276 iterator->level = vcpu->arch.mmu.shadow_root_level;
1277 if (iterator->level == PT32E_ROOT_LEVEL) {
1278 iterator->shadow_addr
1279 = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1280 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1282 if (!iterator->shadow_addr)
1283 iterator->level = 0;
1287 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1289 if (iterator->level < PT_PAGE_TABLE_LEVEL)
1291 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1292 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1296 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1298 iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1302 static int walk_shadow(struct kvm_shadow_walk *walker,
1303 struct kvm_vcpu *vcpu, u64 addr)
1305 struct kvm_shadow_walk_iterator iterator;
1308 for_each_shadow_entry(vcpu, addr, iterator) {
1309 r = walker->entry(walker, vcpu, addr,
1310 iterator.sptep, iterator.level);
1317 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1318 struct kvm_mmu_page *sp)
1326 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1327 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1328 if (is_shadow_present_pte(pt[i]))
1329 rmap_remove(kvm, &pt[i]);
1330 pt[i] = shadow_trap_nonpresent_pte;
1335 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1338 if (is_shadow_present_pte(ent)) {
1339 if (!is_large_pte(ent)) {
1340 ent &= PT64_BASE_ADDR_MASK;
1341 mmu_page_remove_parent_pte(page_header(ent),
1345 rmap_remove(kvm, &pt[i]);
1348 pt[i] = shadow_trap_nonpresent_pte;
1352 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1354 mmu_page_remove_parent_pte(sp, parent_pte);
1357 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1361 for (i = 0; i < KVM_MAX_VCPUS; ++i)
1363 kvm->vcpus[i]->arch.last_pte_updated = NULL;
1366 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1370 while (sp->multimapped || sp->parent_pte) {
1371 if (!sp->multimapped)
1372 parent_pte = sp->parent_pte;
1374 struct kvm_pte_chain *chain;
1376 chain = container_of(sp->parent_ptes.first,
1377 struct kvm_pte_chain, link);
1378 parent_pte = chain->parent_ptes[0];
1380 BUG_ON(!parent_pte);
1381 kvm_mmu_put_page(sp, parent_pte);
1382 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1386 static int mmu_zap_unsync_children(struct kvm *kvm,
1387 struct kvm_mmu_page *parent)
1390 struct mmu_page_path parents;
1391 struct kvm_mmu_pages pages;
1393 if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1396 kvm_mmu_pages_init(parent, &parents, &pages);
1397 while (mmu_unsync_walk(parent, &pages)) {
1398 struct kvm_mmu_page *sp;
1400 for_each_sp(pages, sp, parents, i) {
1401 kvm_mmu_zap_page(kvm, sp);
1402 mmu_pages_clear_parents(&parents);
1405 kvm_mmu_pages_init(parent, &parents, &pages);
1411 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1414 ++kvm->stat.mmu_shadow_zapped;
1415 ret = mmu_zap_unsync_children(kvm, sp);
1416 kvm_mmu_page_unlink_children(kvm, sp);
1417 kvm_mmu_unlink_parents(kvm, sp);
1418 kvm_flush_remote_tlbs(kvm);
1419 if (!sp->role.invalid && !sp->role.metaphysical)
1420 unaccount_shadowed(kvm, sp->gfn);
1422 kvm_unlink_unsync_page(kvm, sp);
1423 if (!sp->root_count) {
1424 hlist_del(&sp->hash_link);
1425 kvm_mmu_free_page(kvm, sp);
1427 sp->role.invalid = 1;
1428 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1429 kvm_reload_remote_mmus(kvm);
1431 kvm_mmu_reset_last_pte_updated(kvm);
1436 * Changing the number of mmu pages allocated to the vm
1437 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1439 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1442 * If we set the number of mmu pages to be smaller be than the
1443 * number of actived pages , we must to free some mmu pages before we
1447 if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1449 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1450 - kvm->arch.n_free_mmu_pages;
1452 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1453 struct kvm_mmu_page *page;
1455 page = container_of(kvm->arch.active_mmu_pages.prev,
1456 struct kvm_mmu_page, link);
1457 kvm_mmu_zap_page(kvm, page);
1460 kvm->arch.n_free_mmu_pages = 0;
1463 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1464 - kvm->arch.n_alloc_mmu_pages;
1466 kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1469 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1472 struct hlist_head *bucket;
1473 struct kvm_mmu_page *sp;
1474 struct hlist_node *node, *n;
1477 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1479 index = kvm_page_table_hashfn(gfn);
1480 bucket = &kvm->arch.mmu_page_hash[index];
1481 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1482 if (sp->gfn == gfn && !sp->role.metaphysical) {
1483 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1486 if (kvm_mmu_zap_page(kvm, sp))
1492 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1494 struct kvm_mmu_page *sp;
1496 while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1497 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1498 kvm_mmu_zap_page(kvm, sp);
1502 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1504 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1505 struct kvm_mmu_page *sp = page_header(__pa(pte));
1507 __set_bit(slot, sp->slot_bitmap);
1510 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1515 if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1518 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1519 if (pt[i] == shadow_notrap_nonpresent_pte)
1520 set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1524 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1528 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1530 if (gpa == UNMAPPED_GVA)
1533 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1539 * The function is based on mtrr_type_lookup() in
1540 * arch/x86/kernel/cpu/mtrr/generic.c
1542 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1547 u8 prev_match, curr_match;
1548 int num_var_ranges = KVM_NR_VAR_MTRR;
1550 if (!mtrr_state->enabled)
1553 /* Make end inclusive end, instead of exclusive */
1556 /* Look in fixed ranges. Just return the type as per start */
1557 if (mtrr_state->have_fixed && (start < 0x100000)) {
1560 if (start < 0x80000) {
1562 idx += (start >> 16);
1563 return mtrr_state->fixed_ranges[idx];
1564 } else if (start < 0xC0000) {
1566 idx += ((start - 0x80000) >> 14);
1567 return mtrr_state->fixed_ranges[idx];
1568 } else if (start < 0x1000000) {
1570 idx += ((start - 0xC0000) >> 12);
1571 return mtrr_state->fixed_ranges[idx];
1576 * Look in variable ranges
1577 * Look of multiple ranges matching this address and pick type
1578 * as per MTRR precedence
1580 if (!(mtrr_state->enabled & 2))
1581 return mtrr_state->def_type;
1584 for (i = 0; i < num_var_ranges; ++i) {
1585 unsigned short start_state, end_state;
1587 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1590 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1591 (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1592 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1593 (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1595 start_state = ((start & mask) == (base & mask));
1596 end_state = ((end & mask) == (base & mask));
1597 if (start_state != end_state)
1600 if ((start & mask) != (base & mask))
1603 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1604 if (prev_match == 0xFF) {
1605 prev_match = curr_match;
1609 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1610 curr_match == MTRR_TYPE_UNCACHABLE)
1611 return MTRR_TYPE_UNCACHABLE;
1613 if ((prev_match == MTRR_TYPE_WRBACK &&
1614 curr_match == MTRR_TYPE_WRTHROUGH) ||
1615 (prev_match == MTRR_TYPE_WRTHROUGH &&
1616 curr_match == MTRR_TYPE_WRBACK)) {
1617 prev_match = MTRR_TYPE_WRTHROUGH;
1618 curr_match = MTRR_TYPE_WRTHROUGH;
1621 if (prev_match != curr_match)
1622 return MTRR_TYPE_UNCACHABLE;
1625 if (prev_match != 0xFF)
1628 return mtrr_state->def_type;
1631 static u8 get_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1635 mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1636 (gfn << PAGE_SHIFT) + PAGE_SIZE);
1637 if (mtrr == 0xfe || mtrr == 0xff)
1638 mtrr = MTRR_TYPE_WRBACK;
1642 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1645 struct hlist_head *bucket;
1646 struct kvm_mmu_page *s;
1647 struct hlist_node *node, *n;
1649 index = kvm_page_table_hashfn(sp->gfn);
1650 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1651 /* don't unsync if pagetable is shadowed with multiple roles */
1652 hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1653 if (s->gfn != sp->gfn || s->role.metaphysical)
1655 if (s->role.word != sp->role.word)
1658 ++vcpu->kvm->stat.mmu_unsync;
1662 list_add(&sp->oos_link, &vcpu->kvm->arch.oos_global_pages);
1663 ++vcpu->kvm->stat.mmu_unsync_global;
1665 kvm_mmu_mark_parents_unsync(vcpu, sp);
1667 mmu_convert_notrap(sp);
1671 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1674 struct kvm_mmu_page *shadow;
1676 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1678 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1682 if (can_unsync && oos_shadow)
1683 return kvm_unsync_page(vcpu, shadow);
1689 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1690 unsigned pte_access, int user_fault,
1691 int write_fault, int dirty, int largepage,
1692 int global, gfn_t gfn, pfn_t pfn, bool speculative,
1697 u64 mt_mask = shadow_mt_mask;
1698 struct kvm_mmu_page *sp = page_header(__pa(shadow_pte));
1700 if (!global && sp->global) {
1703 kvm_unlink_unsync_global(vcpu->kvm, sp);
1704 kvm_mmu_mark_parents_unsync(vcpu, sp);
1709 * We don't set the accessed bit, since we sometimes want to see
1710 * whether the guest actually used the pte (in order to detect
1713 spte = shadow_base_present_pte | shadow_dirty_mask;
1715 spte |= shadow_accessed_mask;
1717 pte_access &= ~ACC_WRITE_MASK;
1718 if (pte_access & ACC_EXEC_MASK)
1719 spte |= shadow_x_mask;
1721 spte |= shadow_nx_mask;
1722 if (pte_access & ACC_USER_MASK)
1723 spte |= shadow_user_mask;
1725 spte |= PT_PAGE_SIZE_MASK;
1727 if (!kvm_is_mmio_pfn(pfn)) {
1728 mt_mask = get_memory_type(vcpu, gfn) <<
1729 kvm_x86_ops->get_mt_mask_shift();
1730 mt_mask |= VMX_EPT_IGMT_BIT;
1732 mt_mask = MTRR_TYPE_UNCACHABLE <<
1733 kvm_x86_ops->get_mt_mask_shift();
1737 spte |= (u64)pfn << PAGE_SHIFT;
1739 if ((pte_access & ACC_WRITE_MASK)
1740 || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1742 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1744 spte = shadow_trap_nonpresent_pte;
1748 spte |= PT_WRITABLE_MASK;
1751 * Optimization: for pte sync, if spte was writable the hash
1752 * lookup is unnecessary (and expensive). Write protection
1753 * is responsibility of mmu_get_page / kvm_sync_page.
1754 * Same reasoning can be applied to dirty page accounting.
1756 if (!can_unsync && is_writeble_pte(*shadow_pte))
1759 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1760 pgprintk("%s: found shadow page for %lx, marking ro\n",
1763 pte_access &= ~ACC_WRITE_MASK;
1764 if (is_writeble_pte(spte))
1765 spte &= ~PT_WRITABLE_MASK;
1769 if (pte_access & ACC_WRITE_MASK)
1770 mark_page_dirty(vcpu->kvm, gfn);
1773 set_shadow_pte(shadow_pte, spte);
1777 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1778 unsigned pt_access, unsigned pte_access,
1779 int user_fault, int write_fault, int dirty,
1780 int *ptwrite, int largepage, int global,
1781 gfn_t gfn, pfn_t pfn, bool speculative)
1783 int was_rmapped = 0;
1784 int was_writeble = is_writeble_pte(*shadow_pte);
1786 pgprintk("%s: spte %llx access %x write_fault %d"
1787 " user_fault %d gfn %lx\n",
1788 __func__, *shadow_pte, pt_access,
1789 write_fault, user_fault, gfn);
1791 if (is_rmap_pte(*shadow_pte)) {
1793 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1794 * the parent of the now unreachable PTE.
1796 if (largepage && !is_large_pte(*shadow_pte)) {
1797 struct kvm_mmu_page *child;
1798 u64 pte = *shadow_pte;
1800 child = page_header(pte & PT64_BASE_ADDR_MASK);
1801 mmu_page_remove_parent_pte(child, shadow_pte);
1802 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1803 pgprintk("hfn old %lx new %lx\n",
1804 spte_to_pfn(*shadow_pte), pfn);
1805 rmap_remove(vcpu->kvm, shadow_pte);
1808 was_rmapped = is_large_pte(*shadow_pte);
1813 if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1814 dirty, largepage, global, gfn, pfn, speculative, true)) {
1817 kvm_x86_ops->tlb_flush(vcpu);
1820 pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1821 pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1822 is_large_pte(*shadow_pte)? "2MB" : "4kB",
1823 is_present_pte(*shadow_pte)?"RW":"R", gfn,
1824 *shadow_pte, shadow_pte);
1825 if (!was_rmapped && is_large_pte(*shadow_pte))
1826 ++vcpu->kvm->stat.lpages;
1828 page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1830 rmap_add(vcpu, shadow_pte, gfn, largepage);
1831 if (!is_rmap_pte(*shadow_pte))
1832 kvm_release_pfn_clean(pfn);
1835 kvm_release_pfn_dirty(pfn);
1837 kvm_release_pfn_clean(pfn);
1840 vcpu->arch.last_pte_updated = shadow_pte;
1841 vcpu->arch.last_pte_gfn = gfn;
1845 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1849 struct direct_shadow_walk {
1850 struct kvm_shadow_walk walker;
1857 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1858 struct kvm_vcpu *vcpu,
1859 u64 addr, u64 *sptep, int level)
1861 struct direct_shadow_walk *walk =
1862 container_of(_walk, struct direct_shadow_walk, walker);
1863 struct kvm_mmu_page *sp;
1865 gfn_t gfn = addr >> PAGE_SHIFT;
1867 if (level == PT_PAGE_TABLE_LEVEL
1868 || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1869 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1870 0, walk->write, 1, &walk->pt_write,
1871 walk->largepage, 0, gfn, walk->pfn, false);
1872 ++vcpu->stat.pf_fixed;
1876 if (*sptep == shadow_trap_nonpresent_pte) {
1877 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1878 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1881 pgprintk("nonpaging_map: ENOMEM\n");
1882 kvm_release_pfn_clean(walk->pfn);
1886 set_shadow_pte(sptep,
1888 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1889 | shadow_user_mask | shadow_x_mask);
1894 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1895 int largepage, gfn_t gfn, pfn_t pfn)
1898 struct direct_shadow_walk walker = {
1899 .walker = { .entry = direct_map_entry, },
1901 .largepage = largepage,
1906 r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1909 return walker.pt_write;
1912 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1917 unsigned long mmu_seq;
1919 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1920 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1924 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1926 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1929 if (is_error_pfn(pfn)) {
1930 kvm_release_pfn_clean(pfn);
1934 spin_lock(&vcpu->kvm->mmu_lock);
1935 if (mmu_notifier_retry(vcpu, mmu_seq))
1937 kvm_mmu_free_some_pages(vcpu);
1938 r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1939 spin_unlock(&vcpu->kvm->mmu_lock);
1945 spin_unlock(&vcpu->kvm->mmu_lock);
1946 kvm_release_pfn_clean(pfn);
1951 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1954 struct kvm_mmu_page *sp;
1956 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1958 spin_lock(&vcpu->kvm->mmu_lock);
1959 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1960 hpa_t root = vcpu->arch.mmu.root_hpa;
1962 sp = page_header(root);
1964 if (!sp->root_count && sp->role.invalid)
1965 kvm_mmu_zap_page(vcpu->kvm, sp);
1966 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1967 spin_unlock(&vcpu->kvm->mmu_lock);
1970 for (i = 0; i < 4; ++i) {
1971 hpa_t root = vcpu->arch.mmu.pae_root[i];
1974 root &= PT64_BASE_ADDR_MASK;
1975 sp = page_header(root);
1977 if (!sp->root_count && sp->role.invalid)
1978 kvm_mmu_zap_page(vcpu->kvm, sp);
1980 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1982 spin_unlock(&vcpu->kvm->mmu_lock);
1983 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1986 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1990 struct kvm_mmu_page *sp;
1991 int metaphysical = 0;
1993 root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1995 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1996 hpa_t root = vcpu->arch.mmu.root_hpa;
1998 ASSERT(!VALID_PAGE(root));
2001 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2002 PT64_ROOT_LEVEL, metaphysical,
2004 root = __pa(sp->spt);
2006 vcpu->arch.mmu.root_hpa = root;
2009 metaphysical = !is_paging(vcpu);
2012 for (i = 0; i < 4; ++i) {
2013 hpa_t root = vcpu->arch.mmu.pae_root[i];
2015 ASSERT(!VALID_PAGE(root));
2016 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2017 if (!is_present_pte(vcpu->arch.pdptrs[i])) {
2018 vcpu->arch.mmu.pae_root[i] = 0;
2021 root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
2022 } else if (vcpu->arch.mmu.root_level == 0)
2024 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2025 PT32_ROOT_LEVEL, metaphysical,
2027 root = __pa(sp->spt);
2029 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2031 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2034 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2037 struct kvm_mmu_page *sp;
2039 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2041 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2042 hpa_t root = vcpu->arch.mmu.root_hpa;
2043 sp = page_header(root);
2044 mmu_sync_children(vcpu, sp);
2047 for (i = 0; i < 4; ++i) {
2048 hpa_t root = vcpu->arch.mmu.pae_root[i];
2051 root &= PT64_BASE_ADDR_MASK;
2052 sp = page_header(root);
2053 mmu_sync_children(vcpu, sp);
2058 static void mmu_sync_global(struct kvm_vcpu *vcpu)
2060 struct kvm *kvm = vcpu->kvm;
2061 struct kvm_mmu_page *sp, *n;
2063 list_for_each_entry_safe(sp, n, &kvm->arch.oos_global_pages, oos_link)
2064 kvm_sync_page(vcpu, sp);
2067 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2069 spin_lock(&vcpu->kvm->mmu_lock);
2070 mmu_sync_roots(vcpu);
2071 spin_unlock(&vcpu->kvm->mmu_lock);
2074 void kvm_mmu_sync_global(struct kvm_vcpu *vcpu)
2076 spin_lock(&vcpu->kvm->mmu_lock);
2077 mmu_sync_global(vcpu);
2078 spin_unlock(&vcpu->kvm->mmu_lock);
2081 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2086 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2092 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2093 r = mmu_topup_memory_caches(vcpu);
2098 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2100 gfn = gva >> PAGE_SHIFT;
2102 return nonpaging_map(vcpu, gva & PAGE_MASK,
2103 error_code & PFERR_WRITE_MASK, gfn);
2106 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2112 gfn_t gfn = gpa >> PAGE_SHIFT;
2113 unsigned long mmu_seq;
2116 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2118 r = mmu_topup_memory_caches(vcpu);
2122 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
2123 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2126 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2128 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2129 if (is_error_pfn(pfn)) {
2130 kvm_release_pfn_clean(pfn);
2133 spin_lock(&vcpu->kvm->mmu_lock);
2134 if (mmu_notifier_retry(vcpu, mmu_seq))
2136 kvm_mmu_free_some_pages(vcpu);
2137 r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2138 largepage, gfn, pfn);
2139 spin_unlock(&vcpu->kvm->mmu_lock);
2144 spin_unlock(&vcpu->kvm->mmu_lock);
2145 kvm_release_pfn_clean(pfn);
2149 static void nonpaging_free(struct kvm_vcpu *vcpu)
2151 mmu_free_roots(vcpu);
2154 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2156 struct kvm_mmu *context = &vcpu->arch.mmu;
2158 context->new_cr3 = nonpaging_new_cr3;
2159 context->page_fault = nonpaging_page_fault;
2160 context->gva_to_gpa = nonpaging_gva_to_gpa;
2161 context->free = nonpaging_free;
2162 context->prefetch_page = nonpaging_prefetch_page;
2163 context->sync_page = nonpaging_sync_page;
2164 context->invlpg = nonpaging_invlpg;
2165 context->root_level = 0;
2166 context->shadow_root_level = PT32E_ROOT_LEVEL;
2167 context->root_hpa = INVALID_PAGE;
2171 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2173 ++vcpu->stat.tlb_flush;
2174 kvm_x86_ops->tlb_flush(vcpu);
2177 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2179 pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2180 mmu_free_roots(vcpu);
2183 static void inject_page_fault(struct kvm_vcpu *vcpu,
2187 kvm_inject_page_fault(vcpu, addr, err_code);
2190 static void paging_free(struct kvm_vcpu *vcpu)
2192 nonpaging_free(vcpu);
2196 #include "paging_tmpl.h"
2200 #include "paging_tmpl.h"
2203 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2205 struct kvm_mmu *context = &vcpu->arch.mmu;
2207 ASSERT(is_pae(vcpu));
2208 context->new_cr3 = paging_new_cr3;
2209 context->page_fault = paging64_page_fault;
2210 context->gva_to_gpa = paging64_gva_to_gpa;
2211 context->prefetch_page = paging64_prefetch_page;
2212 context->sync_page = paging64_sync_page;
2213 context->invlpg = paging64_invlpg;
2214 context->free = paging_free;
2215 context->root_level = level;
2216 context->shadow_root_level = level;
2217 context->root_hpa = INVALID_PAGE;
2221 static int paging64_init_context(struct kvm_vcpu *vcpu)
2223 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2226 static int paging32_init_context(struct kvm_vcpu *vcpu)
2228 struct kvm_mmu *context = &vcpu->arch.mmu;
2230 context->new_cr3 = paging_new_cr3;
2231 context->page_fault = paging32_page_fault;
2232 context->gva_to_gpa = paging32_gva_to_gpa;
2233 context->free = paging_free;
2234 context->prefetch_page = paging32_prefetch_page;
2235 context->sync_page = paging32_sync_page;
2236 context->invlpg = paging32_invlpg;
2237 context->root_level = PT32_ROOT_LEVEL;
2238 context->shadow_root_level = PT32E_ROOT_LEVEL;
2239 context->root_hpa = INVALID_PAGE;
2243 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2245 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2248 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2250 struct kvm_mmu *context = &vcpu->arch.mmu;
2252 context->new_cr3 = nonpaging_new_cr3;
2253 context->page_fault = tdp_page_fault;
2254 context->free = nonpaging_free;
2255 context->prefetch_page = nonpaging_prefetch_page;
2256 context->sync_page = nonpaging_sync_page;
2257 context->invlpg = nonpaging_invlpg;
2258 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2259 context->root_hpa = INVALID_PAGE;
2261 if (!is_paging(vcpu)) {
2262 context->gva_to_gpa = nonpaging_gva_to_gpa;
2263 context->root_level = 0;
2264 } else if (is_long_mode(vcpu)) {
2265 context->gva_to_gpa = paging64_gva_to_gpa;
2266 context->root_level = PT64_ROOT_LEVEL;
2267 } else if (is_pae(vcpu)) {
2268 context->gva_to_gpa = paging64_gva_to_gpa;
2269 context->root_level = PT32E_ROOT_LEVEL;
2271 context->gva_to_gpa = paging32_gva_to_gpa;
2272 context->root_level = PT32_ROOT_LEVEL;
2278 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2283 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2285 if (!is_paging(vcpu))
2286 r = nonpaging_init_context(vcpu);
2287 else if (is_long_mode(vcpu))
2288 r = paging64_init_context(vcpu);
2289 else if (is_pae(vcpu))
2290 r = paging32E_init_context(vcpu);
2292 r = paging32_init_context(vcpu);
2294 vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2299 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2301 vcpu->arch.update_pte.pfn = bad_pfn;
2304 return init_kvm_tdp_mmu(vcpu);
2306 return init_kvm_softmmu(vcpu);
2309 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2312 if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2313 vcpu->arch.mmu.free(vcpu);
2314 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2318 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2320 destroy_kvm_mmu(vcpu);
2321 return init_kvm_mmu(vcpu);
2323 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2325 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2329 r = mmu_topup_memory_caches(vcpu);
2332 spin_lock(&vcpu->kvm->mmu_lock);
2333 kvm_mmu_free_some_pages(vcpu);
2334 mmu_alloc_roots(vcpu);
2335 mmu_sync_roots(vcpu);
2336 spin_unlock(&vcpu->kvm->mmu_lock);
2337 kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2338 kvm_mmu_flush_tlb(vcpu);
2342 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2344 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2346 mmu_free_roots(vcpu);
2349 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2350 struct kvm_mmu_page *sp,
2354 struct kvm_mmu_page *child;
2357 if (is_shadow_present_pte(pte)) {
2358 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2360 rmap_remove(vcpu->kvm, spte);
2362 child = page_header(pte & PT64_BASE_ADDR_MASK);
2363 mmu_page_remove_parent_pte(child, spte);
2366 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2367 if (is_large_pte(pte))
2368 --vcpu->kvm->stat.lpages;
2371 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2372 struct kvm_mmu_page *sp,
2376 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2377 if (!vcpu->arch.update_pte.largepage ||
2378 sp->role.glevels == PT32_ROOT_LEVEL) {
2379 ++vcpu->kvm->stat.mmu_pde_zapped;
2384 ++vcpu->kvm->stat.mmu_pte_updated;
2385 if (sp->role.glevels == PT32_ROOT_LEVEL)
2386 paging32_update_pte(vcpu, sp, spte, new);
2388 paging64_update_pte(vcpu, sp, spte, new);
2391 static bool need_remote_flush(u64 old, u64 new)
2393 if (!is_shadow_present_pte(old))
2395 if (!is_shadow_present_pte(new))
2397 if ((old ^ new) & PT64_BASE_ADDR_MASK)
2399 old ^= PT64_NX_MASK;
2400 new ^= PT64_NX_MASK;
2401 return (old & ~new & PT64_PERM_MASK) != 0;
2404 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2406 if (need_remote_flush(old, new))
2407 kvm_flush_remote_tlbs(vcpu->kvm);
2409 kvm_mmu_flush_tlb(vcpu);
2412 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2414 u64 *spte = vcpu->arch.last_pte_updated;
2416 return !!(spte && (*spte & shadow_accessed_mask));
2419 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2420 const u8 *new, int bytes)
2427 vcpu->arch.update_pte.largepage = 0;
2429 if (bytes != 4 && bytes != 8)
2433 * Assume that the pte write on a page table of the same type
2434 * as the current vcpu paging mode. This is nearly always true
2435 * (might be false while changing modes). Note it is verified later
2439 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2440 if ((bytes == 4) && (gpa % 4 == 0)) {
2441 r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2444 memcpy((void *)&gpte + (gpa % 8), new, 4);
2445 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2446 memcpy((void *)&gpte, new, 8);
2449 if ((bytes == 4) && (gpa % 4 == 0))
2450 memcpy((void *)&gpte, new, 4);
2452 if (!is_present_pte(gpte))
2454 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2456 if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2457 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2458 vcpu->arch.update_pte.largepage = 1;
2460 vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2462 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2464 if (is_error_pfn(pfn)) {
2465 kvm_release_pfn_clean(pfn);
2468 vcpu->arch.update_pte.gfn = gfn;
2469 vcpu->arch.update_pte.pfn = pfn;
2472 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2474 u64 *spte = vcpu->arch.last_pte_updated;
2477 && vcpu->arch.last_pte_gfn == gfn
2478 && shadow_accessed_mask
2479 && !(*spte & shadow_accessed_mask)
2480 && is_shadow_present_pte(*spte))
2481 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2484 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2485 const u8 *new, int bytes,
2486 bool guest_initiated)
2488 gfn_t gfn = gpa >> PAGE_SHIFT;
2489 struct kvm_mmu_page *sp;
2490 struct hlist_node *node, *n;
2491 struct hlist_head *bucket;
2495 unsigned offset = offset_in_page(gpa);
2497 unsigned page_offset;
2498 unsigned misaligned;
2505 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2506 mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2507 spin_lock(&vcpu->kvm->mmu_lock);
2508 kvm_mmu_access_page(vcpu, gfn);
2509 kvm_mmu_free_some_pages(vcpu);
2510 ++vcpu->kvm->stat.mmu_pte_write;
2511 kvm_mmu_audit(vcpu, "pre pte write");
2512 if (guest_initiated) {
2513 if (gfn == vcpu->arch.last_pt_write_gfn
2514 && !last_updated_pte_accessed(vcpu)) {
2515 ++vcpu->arch.last_pt_write_count;
2516 if (vcpu->arch.last_pt_write_count >= 3)
2519 vcpu->arch.last_pt_write_gfn = gfn;
2520 vcpu->arch.last_pt_write_count = 1;
2521 vcpu->arch.last_pte_updated = NULL;
2524 index = kvm_page_table_hashfn(gfn);
2525 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2526 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2527 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
2529 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2530 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2531 misaligned |= bytes < 4;
2532 if (misaligned || flooded) {
2534 * Misaligned accesses are too much trouble to fix
2535 * up; also, they usually indicate a page is not used
2538 * If we're seeing too many writes to a page,
2539 * it may no longer be a page table, or we may be
2540 * forking, in which case it is better to unmap the
2543 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2544 gpa, bytes, sp->role.word);
2545 if (kvm_mmu_zap_page(vcpu->kvm, sp))
2547 ++vcpu->kvm->stat.mmu_flooded;
2550 page_offset = offset;
2551 level = sp->role.level;
2553 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2554 page_offset <<= 1; /* 32->64 */
2556 * A 32-bit pde maps 4MB while the shadow pdes map
2557 * only 2MB. So we need to double the offset again
2558 * and zap two pdes instead of one.
2560 if (level == PT32_ROOT_LEVEL) {
2561 page_offset &= ~7; /* kill rounding error */
2565 quadrant = page_offset >> PAGE_SHIFT;
2566 page_offset &= ~PAGE_MASK;
2567 if (quadrant != sp->role.quadrant)
2570 spte = &sp->spt[page_offset / sizeof(*spte)];
2571 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2573 r = kvm_read_guest_atomic(vcpu->kvm,
2574 gpa & ~(u64)(pte_size - 1),
2576 new = (const void *)&gentry;
2582 mmu_pte_write_zap_pte(vcpu, sp, spte);
2584 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2585 mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2589 kvm_mmu_audit(vcpu, "post pte write");
2590 spin_unlock(&vcpu->kvm->mmu_lock);
2591 if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2592 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2593 vcpu->arch.update_pte.pfn = bad_pfn;
2597 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2602 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2604 spin_lock(&vcpu->kvm->mmu_lock);
2605 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2606 spin_unlock(&vcpu->kvm->mmu_lock);
2609 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2611 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2613 while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2614 struct kvm_mmu_page *sp;
2616 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2617 struct kvm_mmu_page, link);
2618 kvm_mmu_zap_page(vcpu->kvm, sp);
2619 ++vcpu->kvm->stat.mmu_recycled;
2623 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2626 enum emulation_result er;
2628 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2637 r = mmu_topup_memory_caches(vcpu);
2641 er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2646 case EMULATE_DO_MMIO:
2647 ++vcpu->stat.mmio_exits;
2650 kvm_report_emulation_failure(vcpu, "pagetable");
2658 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2660 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2662 vcpu->arch.mmu.invlpg(vcpu, gva);
2663 kvm_mmu_flush_tlb(vcpu);
2664 ++vcpu->stat.invlpg;
2666 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2668 void kvm_enable_tdp(void)
2672 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2674 void kvm_disable_tdp(void)
2676 tdp_enabled = false;
2678 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2680 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2682 struct kvm_mmu_page *sp;
2684 while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2685 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2686 struct kvm_mmu_page, link);
2687 kvm_mmu_zap_page(vcpu->kvm, sp);
2690 free_page((unsigned long)vcpu->arch.mmu.pae_root);
2693 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2700 if (vcpu->kvm->arch.n_requested_mmu_pages)
2701 vcpu->kvm->arch.n_free_mmu_pages =
2702 vcpu->kvm->arch.n_requested_mmu_pages;
2704 vcpu->kvm->arch.n_free_mmu_pages =
2705 vcpu->kvm->arch.n_alloc_mmu_pages;
2707 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2708 * Therefore we need to allocate shadow page tables in the first
2709 * 4GB of memory, which happens to fit the DMA32 zone.
2711 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2714 vcpu->arch.mmu.pae_root = page_address(page);
2715 for (i = 0; i < 4; ++i)
2716 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2721 free_mmu_pages(vcpu);
2725 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2728 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2730 return alloc_mmu_pages(vcpu);
2733 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2736 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2738 return init_kvm_mmu(vcpu);
2741 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2745 destroy_kvm_mmu(vcpu);
2746 free_mmu_pages(vcpu);
2747 mmu_free_memory_caches(vcpu);
2750 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2752 struct kvm_mmu_page *sp;
2754 spin_lock(&kvm->mmu_lock);
2755 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2759 if (!test_bit(slot, sp->slot_bitmap))
2763 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2765 if (pt[i] & PT_WRITABLE_MASK)
2766 pt[i] &= ~PT_WRITABLE_MASK;
2768 kvm_flush_remote_tlbs(kvm);
2769 spin_unlock(&kvm->mmu_lock);
2772 void kvm_mmu_zap_all(struct kvm *kvm)
2774 struct kvm_mmu_page *sp, *node;
2776 spin_lock(&kvm->mmu_lock);
2777 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2778 if (kvm_mmu_zap_page(kvm, sp))
2779 node = container_of(kvm->arch.active_mmu_pages.next,
2780 struct kvm_mmu_page, link);
2781 spin_unlock(&kvm->mmu_lock);
2783 kvm_flush_remote_tlbs(kvm);
2786 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2788 struct kvm_mmu_page *page;
2790 page = container_of(kvm->arch.active_mmu_pages.prev,
2791 struct kvm_mmu_page, link);
2792 kvm_mmu_zap_page(kvm, page);
2795 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2798 struct kvm *kvm_freed = NULL;
2799 int cache_count = 0;
2801 spin_lock(&kvm_lock);
2803 list_for_each_entry(kvm, &vm_list, vm_list) {
2806 if (!down_read_trylock(&kvm->slots_lock))
2808 spin_lock(&kvm->mmu_lock);
2809 npages = kvm->arch.n_alloc_mmu_pages -
2810 kvm->arch.n_free_mmu_pages;
2811 cache_count += npages;
2812 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2813 kvm_mmu_remove_one_alloc_mmu_page(kvm);
2819 spin_unlock(&kvm->mmu_lock);
2820 up_read(&kvm->slots_lock);
2823 list_move_tail(&kvm_freed->vm_list, &vm_list);
2825 spin_unlock(&kvm_lock);
2830 static struct shrinker mmu_shrinker = {
2831 .shrink = mmu_shrink,
2832 .seeks = DEFAULT_SEEKS * 10,
2835 static void mmu_destroy_caches(void)
2837 if (pte_chain_cache)
2838 kmem_cache_destroy(pte_chain_cache);
2839 if (rmap_desc_cache)
2840 kmem_cache_destroy(rmap_desc_cache);
2841 if (mmu_page_header_cache)
2842 kmem_cache_destroy(mmu_page_header_cache);
2845 void kvm_mmu_module_exit(void)
2847 mmu_destroy_caches();
2848 unregister_shrinker(&mmu_shrinker);
2851 int kvm_mmu_module_init(void)
2853 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2854 sizeof(struct kvm_pte_chain),
2856 if (!pte_chain_cache)
2858 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2859 sizeof(struct kvm_rmap_desc),
2861 if (!rmap_desc_cache)
2864 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2865 sizeof(struct kvm_mmu_page),
2867 if (!mmu_page_header_cache)
2870 register_shrinker(&mmu_shrinker);
2875 mmu_destroy_caches();
2880 * Caculate mmu pages needed for kvm.
2882 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2885 unsigned int nr_mmu_pages;
2886 unsigned int nr_pages = 0;
2888 for (i = 0; i < kvm->nmemslots; i++)
2889 nr_pages += kvm->memslots[i].npages;
2891 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2892 nr_mmu_pages = max(nr_mmu_pages,
2893 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2895 return nr_mmu_pages;
2898 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2901 if (len > buffer->len)
2906 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2911 ret = pv_mmu_peek_buffer(buffer, len);
2916 buffer->processed += len;
2920 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2921 gpa_t addr, gpa_t value)
2926 if (!is_long_mode(vcpu) && !is_pae(vcpu))
2929 r = mmu_topup_memory_caches(vcpu);
2933 if (!emulator_write_phys(vcpu, addr, &value, bytes))
2939 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2941 kvm_x86_ops->tlb_flush(vcpu);
2942 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
2946 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2948 spin_lock(&vcpu->kvm->mmu_lock);
2949 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2950 spin_unlock(&vcpu->kvm->mmu_lock);
2954 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2955 struct kvm_pv_mmu_op_buffer *buffer)
2957 struct kvm_mmu_op_header *header;
2959 header = pv_mmu_peek_buffer(buffer, sizeof *header);
2962 switch (header->op) {
2963 case KVM_MMU_OP_WRITE_PTE: {
2964 struct kvm_mmu_op_write_pte *wpte;
2966 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2969 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2972 case KVM_MMU_OP_FLUSH_TLB: {
2973 struct kvm_mmu_op_flush_tlb *ftlb;
2975 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2978 return kvm_pv_mmu_flush_tlb(vcpu);
2980 case KVM_MMU_OP_RELEASE_PT: {
2981 struct kvm_mmu_op_release_pt *rpt;
2983 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2986 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2992 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2993 gpa_t addr, unsigned long *ret)
2996 struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2998 buffer->ptr = buffer->buf;
2999 buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3000 buffer->processed = 0;
3002 r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3006 while (buffer->len) {
3007 r = kvm_pv_mmu_op_one(vcpu, buffer);
3016 *ret = buffer->processed;
3022 static const char *audit_msg;
3024 static gva_t canonicalize(gva_t gva)
3026 #ifdef CONFIG_X86_64
3027 gva = (long long)(gva << 16) >> 16;
3032 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3033 gva_t va, int level)
3035 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3037 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3039 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3042 if (ent == shadow_trap_nonpresent_pte)
3045 va = canonicalize(va);
3047 if (ent == shadow_notrap_nonpresent_pte)
3048 printk(KERN_ERR "audit: (%s) nontrapping pte"
3049 " in nonleaf level: levels %d gva %lx"
3050 " level %d pte %llx\n", audit_msg,
3051 vcpu->arch.mmu.root_level, va, level, ent);
3053 audit_mappings_page(vcpu, ent, va, level - 1);
3055 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3056 hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
3058 if (is_shadow_present_pte(ent)
3059 && (ent & PT64_BASE_ADDR_MASK) != hpa)
3060 printk(KERN_ERR "xx audit error: (%s) levels %d"
3061 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3062 audit_msg, vcpu->arch.mmu.root_level,
3064 is_shadow_present_pte(ent));
3065 else if (ent == shadow_notrap_nonpresent_pte
3066 && !is_error_hpa(hpa))
3067 printk(KERN_ERR "audit: (%s) notrap shadow,"
3068 " valid guest gva %lx\n", audit_msg, va);
3069 kvm_release_pfn_clean(pfn);
3075 static void audit_mappings(struct kvm_vcpu *vcpu)
3079 if (vcpu->arch.mmu.root_level == 4)
3080 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3082 for (i = 0; i < 4; ++i)
3083 if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3084 audit_mappings_page(vcpu,
3085 vcpu->arch.mmu.pae_root[i],
3090 static int count_rmaps(struct kvm_vcpu *vcpu)
3095 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3096 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3097 struct kvm_rmap_desc *d;
3099 for (j = 0; j < m->npages; ++j) {
3100 unsigned long *rmapp = &m->rmap[j];
3104 if (!(*rmapp & 1)) {
3108 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3110 for (k = 0; k < RMAP_EXT; ++k)
3111 if (d->shadow_ptes[k])
3122 static int count_writable_mappings(struct kvm_vcpu *vcpu)
3125 struct kvm_mmu_page *sp;
3128 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3131 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3134 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3137 if (!(ent & PT_PRESENT_MASK))
3139 if (!(ent & PT_WRITABLE_MASK))
3147 static void audit_rmap(struct kvm_vcpu *vcpu)
3149 int n_rmap = count_rmaps(vcpu);
3150 int n_actual = count_writable_mappings(vcpu);
3152 if (n_rmap != n_actual)
3153 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
3154 __func__, audit_msg, n_rmap, n_actual);
3157 static void audit_write_protection(struct kvm_vcpu *vcpu)
3159 struct kvm_mmu_page *sp;
3160 struct kvm_memory_slot *slot;
3161 unsigned long *rmapp;
3164 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3165 if (sp->role.metaphysical)
3168 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3169 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3170 rmapp = &slot->rmap[gfn - slot->base_gfn];
3172 printk(KERN_ERR "%s: (%s) shadow page has writable"
3173 " mappings: gfn %lx role %x\n",
3174 __func__, audit_msg, sp->gfn,
3179 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3186 audit_write_protection(vcpu);
3187 audit_mappings(vcpu);