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.
19 #include <linux/types.h>
20 #include <linux/string.h>
23 #include <linux/highmem.h>
24 #include <linux/module.h>
34 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
36 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
41 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
42 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
46 #define pgprintk(x...) do { } while (0)
47 #define rmap_printk(x...) do { } while (0)
51 #if defined(MMU_DEBUG) || defined(AUDIT)
56 #define ASSERT(x) do { } while (0)
60 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
61 __FILE__, __LINE__, #x); \
65 #define PT64_PT_BITS 9
66 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
67 #define PT32_PT_BITS 10
68 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
70 #define PT_WRITABLE_SHIFT 1
72 #define PT_PRESENT_MASK (1ULL << 0)
73 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
74 #define PT_USER_MASK (1ULL << 2)
75 #define PT_PWT_MASK (1ULL << 3)
76 #define PT_PCD_MASK (1ULL << 4)
77 #define PT_ACCESSED_MASK (1ULL << 5)
78 #define PT_DIRTY_MASK (1ULL << 6)
79 #define PT_PAGE_SIZE_MASK (1ULL << 7)
80 #define PT_PAT_MASK (1ULL << 7)
81 #define PT_GLOBAL_MASK (1ULL << 8)
82 #define PT64_NX_MASK (1ULL << 63)
84 #define PT_PAT_SHIFT 7
85 #define PT_DIR_PAT_SHIFT 12
86 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
88 #define PT32_DIR_PSE36_SIZE 4
89 #define PT32_DIR_PSE36_SHIFT 13
90 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
93 #define PT32_PTE_COPY_MASK \
94 (PT_PRESENT_MASK | PT_ACCESSED_MASK | PT_DIRTY_MASK | PT_GLOBAL_MASK)
96 #define PT64_PTE_COPY_MASK (PT64_NX_MASK | PT32_PTE_COPY_MASK)
98 #define PT_FIRST_AVAIL_BITS_SHIFT 9
99 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
101 #define PT_SHADOW_PS_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
102 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
104 #define PT_SHADOW_WRITABLE_SHIFT (PT_FIRST_AVAIL_BITS_SHIFT + 1)
105 #define PT_SHADOW_WRITABLE_MASK (1ULL << PT_SHADOW_WRITABLE_SHIFT)
107 #define PT_SHADOW_USER_SHIFT (PT_SHADOW_WRITABLE_SHIFT + 1)
108 #define PT_SHADOW_USER_MASK (1ULL << (PT_SHADOW_USER_SHIFT))
110 #define PT_SHADOW_BITS_OFFSET (PT_SHADOW_WRITABLE_SHIFT - PT_WRITABLE_SHIFT)
112 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
114 #define PT64_LEVEL_BITS 9
116 #define PT64_LEVEL_SHIFT(level) \
117 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
119 #define PT64_LEVEL_MASK(level) \
120 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
122 #define PT64_INDEX(address, level)\
123 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
126 #define PT32_LEVEL_BITS 10
128 #define PT32_LEVEL_SHIFT(level) \
129 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
131 #define PT32_LEVEL_MASK(level) \
132 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
134 #define PT32_INDEX(address, level)\
135 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
138 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
139 #define PT64_DIR_BASE_ADDR_MASK \
140 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
142 #define PT32_BASE_ADDR_MASK PAGE_MASK
143 #define PT32_DIR_BASE_ADDR_MASK \
144 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
147 #define PFERR_PRESENT_MASK (1U << 0)
148 #define PFERR_WRITE_MASK (1U << 1)
149 #define PFERR_USER_MASK (1U << 2)
150 #define PFERR_FETCH_MASK (1U << 4)
152 #define PT64_ROOT_LEVEL 4
153 #define PT32_ROOT_LEVEL 2
154 #define PT32E_ROOT_LEVEL 3
156 #define PT_DIRECTORY_LEVEL 2
157 #define PT_PAGE_TABLE_LEVEL 1
161 struct kvm_rmap_desc {
162 u64 *shadow_ptes[RMAP_EXT];
163 struct kvm_rmap_desc *more;
166 static struct kmem_cache *pte_chain_cache;
167 static struct kmem_cache *rmap_desc_cache;
168 static struct kmem_cache *mmu_page_cache;
169 static struct kmem_cache *mmu_page_header_cache;
171 static int is_write_protection(struct kvm_vcpu *vcpu)
173 return vcpu->cr0 & CR0_WP_MASK;
176 static int is_cpuid_PSE36(void)
181 static int is_nx(struct kvm_vcpu *vcpu)
183 return vcpu->shadow_efer & EFER_NX;
186 static int is_present_pte(unsigned long pte)
188 return pte & PT_PRESENT_MASK;
191 static int is_writeble_pte(unsigned long pte)
193 return pte & PT_WRITABLE_MASK;
196 static int is_io_pte(unsigned long pte)
198 return pte & PT_SHADOW_IO_MARK;
201 static int is_rmap_pte(u64 pte)
203 return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
204 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
207 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
208 struct kmem_cache *base_cache, int min,
213 if (cache->nobjs >= min)
215 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
216 obj = kmem_cache_zalloc(base_cache, gfp_flags);
219 cache->objects[cache->nobjs++] = obj;
224 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
227 kfree(mc->objects[--mc->nobjs]);
230 static int __mmu_topup_memory_caches(struct kvm_vcpu *vcpu, gfp_t gfp_flags)
234 r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
235 pte_chain_cache, 4, gfp_flags);
238 r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
239 rmap_desc_cache, 1, gfp_flags);
242 r = mmu_topup_memory_cache(&vcpu->mmu_page_cache,
243 mmu_page_cache, 4, gfp_flags);
246 r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
247 mmu_page_header_cache, 4, gfp_flags);
252 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
256 r = __mmu_topup_memory_caches(vcpu, GFP_NOWAIT);
258 spin_unlock(&vcpu->kvm->lock);
259 kvm_arch_ops->vcpu_put(vcpu);
260 r = __mmu_topup_memory_caches(vcpu, GFP_KERNEL);
261 kvm_arch_ops->vcpu_load(vcpu);
262 spin_lock(&vcpu->kvm->lock);
267 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
269 mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
270 mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
271 mmu_free_memory_cache(&vcpu->mmu_page_cache);
272 mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
275 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
281 p = mc->objects[--mc->nobjs];
286 static void mmu_memory_cache_free(struct kvm_mmu_memory_cache *mc, void *obj)
288 if (mc->nobjs < KVM_NR_MEM_OBJS)
289 mc->objects[mc->nobjs++] = obj;
294 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
296 return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
297 sizeof(struct kvm_pte_chain));
300 static void mmu_free_pte_chain(struct kvm_vcpu *vcpu,
301 struct kvm_pte_chain *pc)
303 mmu_memory_cache_free(&vcpu->mmu_pte_chain_cache, pc);
306 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
308 return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
309 sizeof(struct kvm_rmap_desc));
312 static void mmu_free_rmap_desc(struct kvm_vcpu *vcpu,
313 struct kvm_rmap_desc *rd)
315 mmu_memory_cache_free(&vcpu->mmu_rmap_desc_cache, rd);
319 * Reverse mapping data structures:
321 * If page->private bit zero is zero, then page->private points to the
322 * shadow page table entry that points to page_address(page).
324 * If page->private bit zero is one, (then page->private & ~1) points
325 * to a struct kvm_rmap_desc containing more mappings.
327 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
330 struct kvm_rmap_desc *desc;
333 if (!is_rmap_pte(*spte))
335 page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
336 if (!page_private(page)) {
337 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
338 set_page_private(page,(unsigned long)spte);
339 } else if (!(page_private(page) & 1)) {
340 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
341 desc = mmu_alloc_rmap_desc(vcpu);
342 desc->shadow_ptes[0] = (u64 *)page_private(page);
343 desc->shadow_ptes[1] = spte;
344 set_page_private(page,(unsigned long)desc | 1);
346 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
347 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
348 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
350 if (desc->shadow_ptes[RMAP_EXT-1]) {
351 desc->more = mmu_alloc_rmap_desc(vcpu);
354 for (i = 0; desc->shadow_ptes[i]; ++i)
356 desc->shadow_ptes[i] = spte;
360 static void rmap_desc_remove_entry(struct kvm_vcpu *vcpu,
362 struct kvm_rmap_desc *desc,
364 struct kvm_rmap_desc *prev_desc)
368 for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
370 desc->shadow_ptes[i] = desc->shadow_ptes[j];
371 desc->shadow_ptes[j] = NULL;
374 if (!prev_desc && !desc->more)
375 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
378 prev_desc->more = desc->more;
380 set_page_private(page,(unsigned long)desc->more | 1);
381 mmu_free_rmap_desc(vcpu, desc);
384 static void rmap_remove(struct kvm_vcpu *vcpu, u64 *spte)
387 struct kvm_rmap_desc *desc;
388 struct kvm_rmap_desc *prev_desc;
391 if (!is_rmap_pte(*spte))
393 page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
394 if (!page_private(page)) {
395 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
397 } else if (!(page_private(page) & 1)) {
398 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
399 if ((u64 *)page_private(page) != spte) {
400 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
404 set_page_private(page,0);
406 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
407 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
410 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
411 if (desc->shadow_ptes[i] == spte) {
412 rmap_desc_remove_entry(vcpu, page,
424 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
426 struct kvm *kvm = vcpu->kvm;
428 struct kvm_rmap_desc *desc;
431 page = gfn_to_page(kvm, gfn);
434 while (page_private(page)) {
435 if (!(page_private(page) & 1))
436 spte = (u64 *)page_private(page);
438 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
439 spte = desc->shadow_ptes[0];
442 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
443 != page_to_pfn(page));
444 BUG_ON(!(*spte & PT_PRESENT_MASK));
445 BUG_ON(!(*spte & PT_WRITABLE_MASK));
446 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
447 rmap_remove(vcpu, spte);
448 kvm_arch_ops->tlb_flush(vcpu);
449 *spte &= ~(u64)PT_WRITABLE_MASK;
454 static int is_empty_shadow_page(u64 *spt)
459 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
461 printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
469 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu,
470 struct kvm_mmu_page *page_head)
472 ASSERT(is_empty_shadow_page(page_head->spt));
473 list_del(&page_head->link);
474 mmu_memory_cache_free(&vcpu->mmu_page_cache, page_head->spt);
475 mmu_memory_cache_free(&vcpu->mmu_page_header_cache, page_head);
476 ++vcpu->kvm->n_free_mmu_pages;
479 static unsigned kvm_page_table_hashfn(gfn_t gfn)
484 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
487 struct kvm_mmu_page *page;
489 if (!vcpu->kvm->n_free_mmu_pages)
492 page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
494 page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
495 set_page_private(virt_to_page(page->spt), (unsigned long)page);
496 list_add(&page->link, &vcpu->kvm->active_mmu_pages);
497 ASSERT(is_empty_shadow_page(page->spt));
498 page->slot_bitmap = 0;
499 page->multimapped = 0;
500 page->parent_pte = parent_pte;
501 --vcpu->kvm->n_free_mmu_pages;
505 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
506 struct kvm_mmu_page *page, u64 *parent_pte)
508 struct kvm_pte_chain *pte_chain;
509 struct hlist_node *node;
514 if (!page->multimapped) {
515 u64 *old = page->parent_pte;
518 page->parent_pte = parent_pte;
521 page->multimapped = 1;
522 pte_chain = mmu_alloc_pte_chain(vcpu);
523 INIT_HLIST_HEAD(&page->parent_ptes);
524 hlist_add_head(&pte_chain->link, &page->parent_ptes);
525 pte_chain->parent_ptes[0] = old;
527 hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
528 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
530 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
531 if (!pte_chain->parent_ptes[i]) {
532 pte_chain->parent_ptes[i] = parent_pte;
536 pte_chain = mmu_alloc_pte_chain(vcpu);
538 hlist_add_head(&pte_chain->link, &page->parent_ptes);
539 pte_chain->parent_ptes[0] = parent_pte;
542 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
543 struct kvm_mmu_page *page,
546 struct kvm_pte_chain *pte_chain;
547 struct hlist_node *node;
550 if (!page->multimapped) {
551 BUG_ON(page->parent_pte != parent_pte);
552 page->parent_pte = NULL;
555 hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
556 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
557 if (!pte_chain->parent_ptes[i])
559 if (pte_chain->parent_ptes[i] != parent_pte)
561 while (i + 1 < NR_PTE_CHAIN_ENTRIES
562 && pte_chain->parent_ptes[i + 1]) {
563 pte_chain->parent_ptes[i]
564 = pte_chain->parent_ptes[i + 1];
567 pte_chain->parent_ptes[i] = NULL;
569 hlist_del(&pte_chain->link);
570 mmu_free_pte_chain(vcpu, pte_chain);
571 if (hlist_empty(&page->parent_ptes)) {
572 page->multimapped = 0;
573 page->parent_pte = NULL;
581 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
585 struct hlist_head *bucket;
586 struct kvm_mmu_page *page;
587 struct hlist_node *node;
589 pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
590 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
591 bucket = &vcpu->kvm->mmu_page_hash[index];
592 hlist_for_each_entry(page, node, bucket, hash_link)
593 if (page->gfn == gfn && !page->role.metaphysical) {
594 pgprintk("%s: found role %x\n",
595 __FUNCTION__, page->role.word);
601 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
606 unsigned hugepage_access,
609 union kvm_mmu_page_role role;
612 struct hlist_head *bucket;
613 struct kvm_mmu_page *page;
614 struct hlist_node *node;
617 role.glevels = vcpu->mmu.root_level;
619 role.metaphysical = metaphysical;
620 role.hugepage_access = hugepage_access;
621 if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
622 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
623 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
624 role.quadrant = quadrant;
626 pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
628 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
629 bucket = &vcpu->kvm->mmu_page_hash[index];
630 hlist_for_each_entry(page, node, bucket, hash_link)
631 if (page->gfn == gfn && page->role.word == role.word) {
632 mmu_page_add_parent_pte(vcpu, page, parent_pte);
633 pgprintk("%s: found\n", __FUNCTION__);
636 page = kvm_mmu_alloc_page(vcpu, parent_pte);
639 pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
642 hlist_add_head(&page->hash_link, bucket);
644 rmap_write_protect(vcpu, gfn);
648 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
649 struct kvm_mmu_page *page)
657 if (page->role.level == PT_PAGE_TABLE_LEVEL) {
658 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
659 if (pt[i] & PT_PRESENT_MASK)
660 rmap_remove(vcpu, &pt[i]);
663 kvm_arch_ops->tlb_flush(vcpu);
667 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
671 if (!(ent & PT_PRESENT_MASK))
673 ent &= PT64_BASE_ADDR_MASK;
674 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
678 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
679 struct kvm_mmu_page *page,
682 mmu_page_remove_parent_pte(vcpu, page, parent_pte);
685 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
686 struct kvm_mmu_page *page)
690 while (page->multimapped || page->parent_pte) {
691 if (!page->multimapped)
692 parent_pte = page->parent_pte;
694 struct kvm_pte_chain *chain;
696 chain = container_of(page->parent_ptes.first,
697 struct kvm_pte_chain, link);
698 parent_pte = chain->parent_ptes[0];
701 kvm_mmu_put_page(vcpu, page, parent_pte);
704 kvm_mmu_page_unlink_children(vcpu, page);
705 if (!page->root_count) {
706 hlist_del(&page->hash_link);
707 kvm_mmu_free_page(vcpu, page);
709 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
712 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
715 struct hlist_head *bucket;
716 struct kvm_mmu_page *page;
717 struct hlist_node *node, *n;
720 pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
722 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
723 bucket = &vcpu->kvm->mmu_page_hash[index];
724 hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
725 if (page->gfn == gfn && !page->role.metaphysical) {
726 pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
728 kvm_mmu_zap_page(vcpu, page);
734 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
736 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
737 struct kvm_mmu_page *page_head = page_header(__pa(pte));
739 __set_bit(slot, &page_head->slot_bitmap);
742 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
744 hpa_t hpa = gpa_to_hpa(vcpu, gpa);
746 return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
749 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
753 ASSERT((gpa & HPA_ERR_MASK) == 0);
754 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
756 return gpa | HPA_ERR_MASK;
757 return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
758 | (gpa & (PAGE_SIZE-1));
761 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
763 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
765 if (gpa == UNMAPPED_GVA)
767 return gpa_to_hpa(vcpu, gpa);
770 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
772 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
774 if (gpa == UNMAPPED_GVA)
776 return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
779 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
783 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
785 int level = PT32E_ROOT_LEVEL;
786 hpa_t table_addr = vcpu->mmu.root_hpa;
789 u32 index = PT64_INDEX(v, level);
793 ASSERT(VALID_PAGE(table_addr));
794 table = __va(table_addr);
798 if (is_present_pte(pte) && is_writeble_pte(pte))
800 mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
801 page_header_update_slot(vcpu->kvm, table, v);
802 table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
804 rmap_add(vcpu, &table[index]);
808 if (table[index] == 0) {
809 struct kvm_mmu_page *new_table;
812 pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
814 new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
816 1, 0, &table[index]);
818 pgprintk("nonpaging_map: ENOMEM\n");
822 table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
823 | PT_WRITABLE_MASK | PT_USER_MASK;
825 table_addr = table[index] & PT64_BASE_ADDR_MASK;
829 static void mmu_free_roots(struct kvm_vcpu *vcpu)
832 struct kvm_mmu_page *page;
835 if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
836 hpa_t root = vcpu->mmu.root_hpa;
838 ASSERT(VALID_PAGE(root));
839 page = page_header(root);
841 vcpu->mmu.root_hpa = INVALID_PAGE;
845 for (i = 0; i < 4; ++i) {
846 hpa_t root = vcpu->mmu.pae_root[i];
849 ASSERT(VALID_PAGE(root));
850 root &= PT64_BASE_ADDR_MASK;
851 page = page_header(root);
854 vcpu->mmu.pae_root[i] = INVALID_PAGE;
856 vcpu->mmu.root_hpa = INVALID_PAGE;
859 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
863 struct kvm_mmu_page *page;
865 root_gfn = vcpu->cr3 >> PAGE_SHIFT;
868 if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
869 hpa_t root = vcpu->mmu.root_hpa;
871 ASSERT(!VALID_PAGE(root));
872 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
873 PT64_ROOT_LEVEL, 0, 0, NULL);
874 root = __pa(page->spt);
876 vcpu->mmu.root_hpa = root;
880 for (i = 0; i < 4; ++i) {
881 hpa_t root = vcpu->mmu.pae_root[i];
883 ASSERT(!VALID_PAGE(root));
884 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
885 if (!is_present_pte(vcpu->pdptrs[i])) {
886 vcpu->mmu.pae_root[i] = 0;
889 root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
890 } else if (vcpu->mmu.root_level == 0)
892 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
893 PT32_ROOT_LEVEL, !is_paging(vcpu),
895 root = __pa(page->spt);
897 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
899 vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
902 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
907 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
914 r = mmu_topup_memory_caches(vcpu);
919 ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
922 paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
924 if (is_error_hpa(paddr))
927 return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
930 static void nonpaging_free(struct kvm_vcpu *vcpu)
932 mmu_free_roots(vcpu);
935 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
937 struct kvm_mmu *context = &vcpu->mmu;
939 context->new_cr3 = nonpaging_new_cr3;
940 context->page_fault = nonpaging_page_fault;
941 context->gva_to_gpa = nonpaging_gva_to_gpa;
942 context->free = nonpaging_free;
943 context->root_level = 0;
944 context->shadow_root_level = PT32E_ROOT_LEVEL;
945 mmu_alloc_roots(vcpu);
946 ASSERT(VALID_PAGE(context->root_hpa));
947 kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
951 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
953 ++vcpu->stat.tlb_flush;
954 kvm_arch_ops->tlb_flush(vcpu);
957 static void paging_new_cr3(struct kvm_vcpu *vcpu)
959 pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
960 mmu_free_roots(vcpu);
961 if (unlikely(vcpu->kvm->n_free_mmu_pages < KVM_MIN_FREE_MMU_PAGES))
962 kvm_mmu_free_some_pages(vcpu);
963 mmu_alloc_roots(vcpu);
964 kvm_mmu_flush_tlb(vcpu);
965 kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
968 static inline void set_pte_common(struct kvm_vcpu *vcpu,
977 *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
979 access_bits &= ~PT_WRITABLE_MASK;
981 paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
983 *shadow_pte |= access_bits;
985 if (is_error_hpa(paddr)) {
986 *shadow_pte |= gaddr;
987 *shadow_pte |= PT_SHADOW_IO_MARK;
988 *shadow_pte &= ~PT_PRESENT_MASK;
992 *shadow_pte |= paddr;
994 if (access_bits & PT_WRITABLE_MASK) {
995 struct kvm_mmu_page *shadow;
997 shadow = kvm_mmu_lookup_page(vcpu, gfn);
999 pgprintk("%s: found shadow page for %lx, marking ro\n",
1001 access_bits &= ~PT_WRITABLE_MASK;
1002 if (is_writeble_pte(*shadow_pte)) {
1003 *shadow_pte &= ~PT_WRITABLE_MASK;
1004 kvm_arch_ops->tlb_flush(vcpu);
1009 if (access_bits & PT_WRITABLE_MASK)
1010 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
1012 page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
1013 rmap_add(vcpu, shadow_pte);
1016 static void inject_page_fault(struct kvm_vcpu *vcpu,
1020 kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
1023 static inline int fix_read_pf(u64 *shadow_ent)
1025 if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
1026 !(*shadow_ent & PT_USER_MASK)) {
1028 * If supervisor write protect is disabled, we shadow kernel
1029 * pages as user pages so we can trap the write access.
1031 *shadow_ent |= PT_USER_MASK;
1032 *shadow_ent &= ~PT_WRITABLE_MASK;
1040 static void paging_free(struct kvm_vcpu *vcpu)
1042 nonpaging_free(vcpu);
1046 #include "paging_tmpl.h"
1050 #include "paging_tmpl.h"
1053 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1055 struct kvm_mmu *context = &vcpu->mmu;
1057 ASSERT(is_pae(vcpu));
1058 context->new_cr3 = paging_new_cr3;
1059 context->page_fault = paging64_page_fault;
1060 context->gva_to_gpa = paging64_gva_to_gpa;
1061 context->free = paging_free;
1062 context->root_level = level;
1063 context->shadow_root_level = level;
1064 mmu_alloc_roots(vcpu);
1065 ASSERT(VALID_PAGE(context->root_hpa));
1066 kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1067 (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1071 static int paging64_init_context(struct kvm_vcpu *vcpu)
1073 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1076 static int paging32_init_context(struct kvm_vcpu *vcpu)
1078 struct kvm_mmu *context = &vcpu->mmu;
1080 context->new_cr3 = paging_new_cr3;
1081 context->page_fault = paging32_page_fault;
1082 context->gva_to_gpa = paging32_gva_to_gpa;
1083 context->free = paging_free;
1084 context->root_level = PT32_ROOT_LEVEL;
1085 context->shadow_root_level = PT32E_ROOT_LEVEL;
1086 mmu_alloc_roots(vcpu);
1087 ASSERT(VALID_PAGE(context->root_hpa));
1088 kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1089 (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1093 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1095 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1098 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1101 ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1103 mmu_topup_memory_caches(vcpu);
1104 if (!is_paging(vcpu))
1105 return nonpaging_init_context(vcpu);
1106 else if (is_long_mode(vcpu))
1107 return paging64_init_context(vcpu);
1108 else if (is_pae(vcpu))
1109 return paging32E_init_context(vcpu);
1111 return paging32_init_context(vcpu);
1114 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1117 if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1118 vcpu->mmu.free(vcpu);
1119 vcpu->mmu.root_hpa = INVALID_PAGE;
1123 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1127 destroy_kvm_mmu(vcpu);
1128 r = init_kvm_mmu(vcpu);
1131 r = mmu_topup_memory_caches(vcpu);
1136 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1137 struct kvm_mmu_page *page,
1141 struct kvm_mmu_page *child;
1144 if (is_present_pte(pte)) {
1145 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1146 rmap_remove(vcpu, spte);
1148 child = page_header(pte & PT64_BASE_ADDR_MASK);
1149 mmu_page_remove_parent_pte(vcpu, child, spte);
1155 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1156 struct kvm_mmu_page *page,
1158 const void *new, int bytes)
1160 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1163 if (page->role.glevels == PT32_ROOT_LEVEL)
1164 paging32_update_pte(vcpu, page, spte, new, bytes);
1166 paging64_update_pte(vcpu, page, spte, new, bytes);
1169 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1170 const u8 *old, const u8 *new, int bytes)
1172 gfn_t gfn = gpa >> PAGE_SHIFT;
1173 struct kvm_mmu_page *page;
1174 struct hlist_node *node, *n;
1175 struct hlist_head *bucket;
1178 unsigned offset = offset_in_page(gpa);
1180 unsigned page_offset;
1181 unsigned misaligned;
1187 pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1188 if (gfn == vcpu->last_pt_write_gfn) {
1189 ++vcpu->last_pt_write_count;
1190 if (vcpu->last_pt_write_count >= 3)
1193 vcpu->last_pt_write_gfn = gfn;
1194 vcpu->last_pt_write_count = 1;
1196 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1197 bucket = &vcpu->kvm->mmu_page_hash[index];
1198 hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1199 if (page->gfn != gfn || page->role.metaphysical)
1201 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1202 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1203 misaligned |= bytes < 4;
1204 if (misaligned || flooded) {
1206 * Misaligned accesses are too much trouble to fix
1207 * up; also, they usually indicate a page is not used
1210 * If we're seeing too many writes to a page,
1211 * it may no longer be a page table, or we may be
1212 * forking, in which case it is better to unmap the
1215 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1216 gpa, bytes, page->role.word);
1217 kvm_mmu_zap_page(vcpu, page);
1220 page_offset = offset;
1221 level = page->role.level;
1223 if (page->role.glevels == PT32_ROOT_LEVEL) {
1224 page_offset <<= 1; /* 32->64 */
1226 * A 32-bit pde maps 4MB while the shadow pdes map
1227 * only 2MB. So we need to double the offset again
1228 * and zap two pdes instead of one.
1230 if (level == PT32_ROOT_LEVEL) {
1231 page_offset &= ~7; /* kill rounding error */
1235 quadrant = page_offset >> PAGE_SHIFT;
1236 page_offset &= ~PAGE_MASK;
1237 if (quadrant != page->role.quadrant)
1240 spte = &page->spt[page_offset / sizeof(*spte)];
1242 mmu_pte_write_zap_pte(vcpu, page, spte);
1243 mmu_pte_write_new_pte(vcpu, page, spte, new, bytes);
1249 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1251 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1253 return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1256 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1258 while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1259 struct kvm_mmu_page *page;
1261 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1262 struct kvm_mmu_page, link);
1263 kvm_mmu_zap_page(vcpu, page);
1266 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1268 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1270 struct kvm_mmu_page *page;
1272 while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1273 page = container_of(vcpu->kvm->active_mmu_pages.next,
1274 struct kvm_mmu_page, link);
1275 kvm_mmu_zap_page(vcpu, page);
1277 free_page((unsigned long)vcpu->mmu.pae_root);
1280 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1287 vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1290 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1291 * Therefore we need to allocate shadow page tables in the first
1292 * 4GB of memory, which happens to fit the DMA32 zone.
1294 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1297 vcpu->mmu.pae_root = page_address(page);
1298 for (i = 0; i < 4; ++i)
1299 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1304 free_mmu_pages(vcpu);
1308 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1311 ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1313 return alloc_mmu_pages(vcpu);
1316 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1319 ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1321 return init_kvm_mmu(vcpu);
1324 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1328 destroy_kvm_mmu(vcpu);
1329 free_mmu_pages(vcpu);
1330 mmu_free_memory_caches(vcpu);
1333 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1335 struct kvm *kvm = vcpu->kvm;
1336 struct kvm_mmu_page *page;
1338 list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1342 if (!test_bit(slot, &page->slot_bitmap))
1346 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1348 if (pt[i] & PT_WRITABLE_MASK) {
1349 rmap_remove(vcpu, &pt[i]);
1350 pt[i] &= ~PT_WRITABLE_MASK;
1355 void kvm_mmu_zap_all(struct kvm_vcpu *vcpu)
1357 destroy_kvm_mmu(vcpu);
1359 while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1360 struct kvm_mmu_page *page;
1362 page = container_of(vcpu->kvm->active_mmu_pages.next,
1363 struct kvm_mmu_page, link);
1364 kvm_mmu_zap_page(vcpu, page);
1367 mmu_free_memory_caches(vcpu);
1368 kvm_arch_ops->tlb_flush(vcpu);
1372 void kvm_mmu_module_exit(void)
1374 if (pte_chain_cache)
1375 kmem_cache_destroy(pte_chain_cache);
1376 if (rmap_desc_cache)
1377 kmem_cache_destroy(rmap_desc_cache);
1379 kmem_cache_destroy(mmu_page_cache);
1380 if (mmu_page_header_cache)
1381 kmem_cache_destroy(mmu_page_header_cache);
1384 int kvm_mmu_module_init(void)
1386 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1387 sizeof(struct kvm_pte_chain),
1389 if (!pte_chain_cache)
1391 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1392 sizeof(struct kvm_rmap_desc),
1394 if (!rmap_desc_cache)
1397 mmu_page_cache = kmem_cache_create("kvm_mmu_page",
1399 PAGE_SIZE, 0, NULL, NULL);
1400 if (!mmu_page_cache)
1403 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1404 sizeof(struct kvm_mmu_page),
1406 if (!mmu_page_header_cache)
1412 kvm_mmu_module_exit();
1418 static const char *audit_msg;
1420 static gva_t canonicalize(gva_t gva)
1422 #ifdef CONFIG_X86_64
1423 gva = (long long)(gva << 16) >> 16;
1428 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1429 gva_t va, int level)
1431 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1433 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1435 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1438 if (!(ent & PT_PRESENT_MASK))
1441 va = canonicalize(va);
1443 audit_mappings_page(vcpu, ent, va, level - 1);
1445 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1446 hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1448 if ((ent & PT_PRESENT_MASK)
1449 && (ent & PT64_BASE_ADDR_MASK) != hpa)
1450 printk(KERN_ERR "audit error: (%s) levels %d"
1451 " gva %lx gpa %llx hpa %llx ent %llx\n",
1452 audit_msg, vcpu->mmu.root_level,
1458 static void audit_mappings(struct kvm_vcpu *vcpu)
1462 if (vcpu->mmu.root_level == 4)
1463 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1465 for (i = 0; i < 4; ++i)
1466 if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1467 audit_mappings_page(vcpu,
1468 vcpu->mmu.pae_root[i],
1473 static int count_rmaps(struct kvm_vcpu *vcpu)
1478 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1479 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1480 struct kvm_rmap_desc *d;
1482 for (j = 0; j < m->npages; ++j) {
1483 struct page *page = m->phys_mem[j];
1487 if (!(page->private & 1)) {
1491 d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1493 for (k = 0; k < RMAP_EXT; ++k)
1494 if (d->shadow_ptes[k])
1505 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1508 struct kvm_mmu_page *page;
1511 list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1512 u64 *pt = page->spt;
1514 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1517 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1520 if (!(ent & PT_PRESENT_MASK))
1522 if (!(ent & PT_WRITABLE_MASK))
1530 static void audit_rmap(struct kvm_vcpu *vcpu)
1532 int n_rmap = count_rmaps(vcpu);
1533 int n_actual = count_writable_mappings(vcpu);
1535 if (n_rmap != n_actual)
1536 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1537 __FUNCTION__, audit_msg, n_rmap, n_actual);
1540 static void audit_write_protection(struct kvm_vcpu *vcpu)
1542 struct kvm_mmu_page *page;
1544 list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1548 if (page->role.metaphysical)
1551 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1553 pg = pfn_to_page(hfn);
1555 printk(KERN_ERR "%s: (%s) shadow page has writable"
1556 " mappings: gfn %lx role %x\n",
1557 __FUNCTION__, audit_msg, page->gfn,
1562 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1569 audit_write_protection(vcpu);
1570 audit_mappings(vcpu);