2 * Kernel-based Virtual Machine driver for Linux
4 * This module enables machines with Intel VT-x extensions to run virtual
5 * machines without emulation or binary translation.
9 * Copyright (C) 2006 Qumranet, Inc.
12 * Yaniv Kamay <yaniv@qumranet.com>
13 * Avi Kivity <avi@qumranet.com>
15 * This work is licensed under the terms of the GNU GPL, version 2. See
16 * the COPYING file in the top-level directory.
23 #include <linux/types.h>
24 #include <linux/string.h>
26 #include <linux/highmem.h>
27 #include <linux/module.h>
30 #include <asm/cmpxchg.h>
37 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
39 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
44 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
45 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
49 #define pgprintk(x...) do { } while (0)
50 #define rmap_printk(x...) do { } while (0)
54 #if defined(MMU_DEBUG) || defined(AUDIT)
59 #define ASSERT(x) do { } while (0)
63 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
64 __FILE__, __LINE__, #x); \
68 #define PT64_PT_BITS 9
69 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
70 #define PT32_PT_BITS 10
71 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
73 #define PT_WRITABLE_SHIFT 1
75 #define PT_PRESENT_MASK (1ULL << 0)
76 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
77 #define PT_USER_MASK (1ULL << 2)
78 #define PT_PWT_MASK (1ULL << 3)
79 #define PT_PCD_MASK (1ULL << 4)
80 #define PT_ACCESSED_MASK (1ULL << 5)
81 #define PT_DIRTY_MASK (1ULL << 6)
82 #define PT_PAGE_SIZE_MASK (1ULL << 7)
83 #define PT_PAT_MASK (1ULL << 7)
84 #define PT_GLOBAL_MASK (1ULL << 8)
85 #define PT64_NX_MASK (1ULL << 63)
87 #define PT_PAT_SHIFT 7
88 #define PT_DIR_PAT_SHIFT 12
89 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
91 #define PT32_DIR_PSE36_SIZE 4
92 #define PT32_DIR_PSE36_SHIFT 13
93 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
96 #define PT_FIRST_AVAIL_BITS_SHIFT 9
97 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
99 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
101 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
103 #define PT64_LEVEL_BITS 9
105 #define PT64_LEVEL_SHIFT(level) \
106 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
108 #define PT64_LEVEL_MASK(level) \
109 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
111 #define PT64_INDEX(address, level)\
112 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
115 #define PT32_LEVEL_BITS 10
117 #define PT32_LEVEL_SHIFT(level) \
118 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
120 #define PT32_LEVEL_MASK(level) \
121 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
123 #define PT32_INDEX(address, level)\
124 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
127 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
128 #define PT64_DIR_BASE_ADDR_MASK \
129 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
131 #define PT32_BASE_ADDR_MASK PAGE_MASK
132 #define PT32_DIR_BASE_ADDR_MASK \
133 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
136 #define PFERR_PRESENT_MASK (1U << 0)
137 #define PFERR_WRITE_MASK (1U << 1)
138 #define PFERR_USER_MASK (1U << 2)
139 #define PFERR_FETCH_MASK (1U << 4)
141 #define PT64_ROOT_LEVEL 4
142 #define PT32_ROOT_LEVEL 2
143 #define PT32E_ROOT_LEVEL 3
145 #define PT_DIRECTORY_LEVEL 2
146 #define PT_PAGE_TABLE_LEVEL 1
150 struct kvm_rmap_desc {
151 u64 *shadow_ptes[RMAP_EXT];
152 struct kvm_rmap_desc *more;
155 static struct kmem_cache *pte_chain_cache;
156 static struct kmem_cache *rmap_desc_cache;
157 static struct kmem_cache *mmu_page_cache;
158 static struct kmem_cache *mmu_page_header_cache;
160 static int is_write_protection(struct kvm_vcpu *vcpu)
162 return vcpu->cr0 & CR0_WP_MASK;
165 static int is_cpuid_PSE36(void)
170 static int is_nx(struct kvm_vcpu *vcpu)
172 return vcpu->shadow_efer & EFER_NX;
175 static int is_present_pte(unsigned long pte)
177 return pte & PT_PRESENT_MASK;
180 static int is_writeble_pte(unsigned long pte)
182 return pte & PT_WRITABLE_MASK;
185 static int is_io_pte(unsigned long pte)
187 return pte & PT_SHADOW_IO_MARK;
190 static int is_rmap_pte(u64 pte)
192 return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
193 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
196 static void set_shadow_pte(u64 *sptep, u64 spte)
199 set_64bit((unsigned long *)sptep, spte);
201 set_64bit((unsigned long long *)sptep, spte);
205 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
206 struct kmem_cache *base_cache, int min,
211 if (cache->nobjs >= min)
213 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
214 obj = kmem_cache_zalloc(base_cache, gfp_flags);
217 cache->objects[cache->nobjs++] = obj;
222 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
225 kfree(mc->objects[--mc->nobjs]);
228 static int __mmu_topup_memory_caches(struct kvm_vcpu *vcpu, gfp_t gfp_flags)
232 r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
233 pte_chain_cache, 4, gfp_flags);
236 r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
237 rmap_desc_cache, 1, gfp_flags);
240 r = mmu_topup_memory_cache(&vcpu->mmu_page_cache,
241 mmu_page_cache, 4, gfp_flags);
244 r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
245 mmu_page_header_cache, 4, gfp_flags);
250 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
254 r = __mmu_topup_memory_caches(vcpu, GFP_NOWAIT);
256 spin_unlock(&vcpu->kvm->lock);
257 kvm_arch_ops->vcpu_put(vcpu);
258 r = __mmu_topup_memory_caches(vcpu, GFP_KERNEL);
259 kvm_arch_ops->vcpu_load(vcpu);
260 spin_lock(&vcpu->kvm->lock);
265 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
267 mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
268 mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
269 mmu_free_memory_cache(&vcpu->mmu_page_cache);
270 mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
273 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
279 p = mc->objects[--mc->nobjs];
284 static void mmu_memory_cache_free(struct kvm_mmu_memory_cache *mc, void *obj)
286 if (mc->nobjs < KVM_NR_MEM_OBJS)
287 mc->objects[mc->nobjs++] = obj;
292 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
294 return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
295 sizeof(struct kvm_pte_chain));
298 static void mmu_free_pte_chain(struct kvm_vcpu *vcpu,
299 struct kvm_pte_chain *pc)
301 mmu_memory_cache_free(&vcpu->mmu_pte_chain_cache, pc);
304 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
306 return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
307 sizeof(struct kvm_rmap_desc));
310 static void mmu_free_rmap_desc(struct kvm_vcpu *vcpu,
311 struct kvm_rmap_desc *rd)
313 mmu_memory_cache_free(&vcpu->mmu_rmap_desc_cache, rd);
317 * Reverse mapping data structures:
319 * If page->private bit zero is zero, then page->private points to the
320 * shadow page table entry that points to page_address(page).
322 * If page->private bit zero is one, (then page->private & ~1) points
323 * to a struct kvm_rmap_desc containing more mappings.
325 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
328 struct kvm_rmap_desc *desc;
331 if (!is_rmap_pte(*spte))
333 page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
334 if (!page_private(page)) {
335 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
336 set_page_private(page,(unsigned long)spte);
337 } else if (!(page_private(page) & 1)) {
338 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
339 desc = mmu_alloc_rmap_desc(vcpu);
340 desc->shadow_ptes[0] = (u64 *)page_private(page);
341 desc->shadow_ptes[1] = spte;
342 set_page_private(page,(unsigned long)desc | 1);
344 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
345 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
346 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
348 if (desc->shadow_ptes[RMAP_EXT-1]) {
349 desc->more = mmu_alloc_rmap_desc(vcpu);
352 for (i = 0; desc->shadow_ptes[i]; ++i)
354 desc->shadow_ptes[i] = spte;
358 static void rmap_desc_remove_entry(struct kvm_vcpu *vcpu,
360 struct kvm_rmap_desc *desc,
362 struct kvm_rmap_desc *prev_desc)
366 for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
368 desc->shadow_ptes[i] = desc->shadow_ptes[j];
369 desc->shadow_ptes[j] = NULL;
372 if (!prev_desc && !desc->more)
373 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
376 prev_desc->more = desc->more;
378 set_page_private(page,(unsigned long)desc->more | 1);
379 mmu_free_rmap_desc(vcpu, desc);
382 static void rmap_remove(struct kvm_vcpu *vcpu, u64 *spte)
385 struct kvm_rmap_desc *desc;
386 struct kvm_rmap_desc *prev_desc;
389 if (!is_rmap_pte(*spte))
391 page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
392 if (!page_private(page)) {
393 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
395 } else if (!(page_private(page) & 1)) {
396 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
397 if ((u64 *)page_private(page) != spte) {
398 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
402 set_page_private(page,0);
404 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
405 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
408 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
409 if (desc->shadow_ptes[i] == spte) {
410 rmap_desc_remove_entry(vcpu, page,
422 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
424 struct kvm *kvm = vcpu->kvm;
426 struct kvm_rmap_desc *desc;
429 page = gfn_to_page(kvm, gfn);
432 while (page_private(page)) {
433 if (!(page_private(page) & 1))
434 spte = (u64 *)page_private(page);
436 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
437 spte = desc->shadow_ptes[0];
440 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
441 != page_to_pfn(page));
442 BUG_ON(!(*spte & PT_PRESENT_MASK));
443 BUG_ON(!(*spte & PT_WRITABLE_MASK));
444 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
445 rmap_remove(vcpu, spte);
446 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
447 kvm_flush_remote_tlbs(vcpu->kvm);
452 static int is_empty_shadow_page(u64 *spt)
457 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
459 printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
467 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu,
468 struct kvm_mmu_page *page_head)
470 ASSERT(is_empty_shadow_page(page_head->spt));
471 list_del(&page_head->link);
472 mmu_memory_cache_free(&vcpu->mmu_page_cache, page_head->spt);
473 mmu_memory_cache_free(&vcpu->mmu_page_header_cache, page_head);
474 ++vcpu->kvm->n_free_mmu_pages;
477 static unsigned kvm_page_table_hashfn(gfn_t gfn)
482 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
485 struct kvm_mmu_page *page;
487 if (!vcpu->kvm->n_free_mmu_pages)
490 page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
492 page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
493 set_page_private(virt_to_page(page->spt), (unsigned long)page);
494 list_add(&page->link, &vcpu->kvm->active_mmu_pages);
495 ASSERT(is_empty_shadow_page(page->spt));
496 page->slot_bitmap = 0;
497 page->multimapped = 0;
498 page->parent_pte = parent_pte;
499 --vcpu->kvm->n_free_mmu_pages;
503 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
504 struct kvm_mmu_page *page, u64 *parent_pte)
506 struct kvm_pte_chain *pte_chain;
507 struct hlist_node *node;
512 if (!page->multimapped) {
513 u64 *old = page->parent_pte;
516 page->parent_pte = parent_pte;
519 page->multimapped = 1;
520 pte_chain = mmu_alloc_pte_chain(vcpu);
521 INIT_HLIST_HEAD(&page->parent_ptes);
522 hlist_add_head(&pte_chain->link, &page->parent_ptes);
523 pte_chain->parent_ptes[0] = old;
525 hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
526 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
528 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
529 if (!pte_chain->parent_ptes[i]) {
530 pte_chain->parent_ptes[i] = parent_pte;
534 pte_chain = mmu_alloc_pte_chain(vcpu);
536 hlist_add_head(&pte_chain->link, &page->parent_ptes);
537 pte_chain->parent_ptes[0] = parent_pte;
540 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
541 struct kvm_mmu_page *page,
544 struct kvm_pte_chain *pte_chain;
545 struct hlist_node *node;
548 if (!page->multimapped) {
549 BUG_ON(page->parent_pte != parent_pte);
550 page->parent_pte = NULL;
553 hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
554 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
555 if (!pte_chain->parent_ptes[i])
557 if (pte_chain->parent_ptes[i] != parent_pte)
559 while (i + 1 < NR_PTE_CHAIN_ENTRIES
560 && pte_chain->parent_ptes[i + 1]) {
561 pte_chain->parent_ptes[i]
562 = pte_chain->parent_ptes[i + 1];
565 pte_chain->parent_ptes[i] = NULL;
567 hlist_del(&pte_chain->link);
568 mmu_free_pte_chain(vcpu, pte_chain);
569 if (hlist_empty(&page->parent_ptes)) {
570 page->multimapped = 0;
571 page->parent_pte = NULL;
579 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
583 struct hlist_head *bucket;
584 struct kvm_mmu_page *page;
585 struct hlist_node *node;
587 pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
588 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
589 bucket = &vcpu->kvm->mmu_page_hash[index];
590 hlist_for_each_entry(page, node, bucket, hash_link)
591 if (page->gfn == gfn && !page->role.metaphysical) {
592 pgprintk("%s: found role %x\n",
593 __FUNCTION__, page->role.word);
599 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
604 unsigned hugepage_access,
607 union kvm_mmu_page_role role;
610 struct hlist_head *bucket;
611 struct kvm_mmu_page *page;
612 struct hlist_node *node;
615 role.glevels = vcpu->mmu.root_level;
617 role.metaphysical = metaphysical;
618 role.hugepage_access = hugepage_access;
619 if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
620 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
621 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
622 role.quadrant = quadrant;
624 pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
626 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
627 bucket = &vcpu->kvm->mmu_page_hash[index];
628 hlist_for_each_entry(page, node, bucket, hash_link)
629 if (page->gfn == gfn && page->role.word == role.word) {
630 mmu_page_add_parent_pte(vcpu, page, parent_pte);
631 pgprintk("%s: found\n", __FUNCTION__);
634 page = kvm_mmu_alloc_page(vcpu, parent_pte);
637 pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
640 hlist_add_head(&page->hash_link, bucket);
642 rmap_write_protect(vcpu, gfn);
646 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
647 struct kvm_mmu_page *page)
655 if (page->role.level == PT_PAGE_TABLE_LEVEL) {
656 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
657 if (pt[i] & PT_PRESENT_MASK)
658 rmap_remove(vcpu, &pt[i]);
661 kvm_flush_remote_tlbs(vcpu->kvm);
665 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
669 if (!(ent & PT_PRESENT_MASK))
671 ent &= PT64_BASE_ADDR_MASK;
672 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
674 kvm_flush_remote_tlbs(vcpu->kvm);
677 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
678 struct kvm_mmu_page *page,
681 mmu_page_remove_parent_pte(vcpu, page, parent_pte);
684 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
685 struct kvm_mmu_page *page)
689 while (page->multimapped || page->parent_pte) {
690 if (!page->multimapped)
691 parent_pte = page->parent_pte;
693 struct kvm_pte_chain *chain;
695 chain = container_of(page->parent_ptes.first,
696 struct kvm_pte_chain, link);
697 parent_pte = chain->parent_ptes[0];
700 kvm_mmu_put_page(vcpu, page, parent_pte);
701 set_shadow_pte(parent_pte, 0);
703 kvm_mmu_page_unlink_children(vcpu, page);
704 if (!page->root_count) {
705 hlist_del(&page->hash_link);
706 kvm_mmu_free_page(vcpu, page);
708 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
711 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
714 struct hlist_head *bucket;
715 struct kvm_mmu_page *page;
716 struct hlist_node *node, *n;
719 pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
721 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
722 bucket = &vcpu->kvm->mmu_page_hash[index];
723 hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
724 if (page->gfn == gfn && !page->role.metaphysical) {
725 pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
727 kvm_mmu_zap_page(vcpu, page);
733 static void mmu_unshadow(struct kvm_vcpu *vcpu, gfn_t gfn)
735 struct kvm_mmu_page *page;
737 while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
738 pgprintk("%s: zap %lx %x\n",
739 __FUNCTION__, gfn, page->role.word);
740 kvm_mmu_zap_page(vcpu, page);
744 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
746 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
747 struct kvm_mmu_page *page_head = page_header(__pa(pte));
749 __set_bit(slot, &page_head->slot_bitmap);
752 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
754 hpa_t hpa = gpa_to_hpa(vcpu, gpa);
756 return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
759 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
763 ASSERT((gpa & HPA_ERR_MASK) == 0);
764 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
766 return gpa | HPA_ERR_MASK;
767 return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
768 | (gpa & (PAGE_SIZE-1));
771 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
773 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
775 if (gpa == UNMAPPED_GVA)
777 return gpa_to_hpa(vcpu, gpa);
780 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
782 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
784 if (gpa == UNMAPPED_GVA)
786 return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
789 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
793 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
795 int level = PT32E_ROOT_LEVEL;
796 hpa_t table_addr = vcpu->mmu.root_hpa;
799 u32 index = PT64_INDEX(v, level);
803 ASSERT(VALID_PAGE(table_addr));
804 table = __va(table_addr);
808 if (is_present_pte(pte) && is_writeble_pte(pte))
810 mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
811 page_header_update_slot(vcpu->kvm, table, v);
812 table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
814 rmap_add(vcpu, &table[index]);
818 if (table[index] == 0) {
819 struct kvm_mmu_page *new_table;
822 pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
824 new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
826 1, 0, &table[index]);
828 pgprintk("nonpaging_map: ENOMEM\n");
832 table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
833 | PT_WRITABLE_MASK | PT_USER_MASK;
835 table_addr = table[index] & PT64_BASE_ADDR_MASK;
839 static void mmu_free_roots(struct kvm_vcpu *vcpu)
842 struct kvm_mmu_page *page;
844 if (!VALID_PAGE(vcpu->mmu.root_hpa))
847 if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
848 hpa_t root = vcpu->mmu.root_hpa;
850 page = page_header(root);
852 vcpu->mmu.root_hpa = INVALID_PAGE;
856 for (i = 0; i < 4; ++i) {
857 hpa_t root = vcpu->mmu.pae_root[i];
860 root &= PT64_BASE_ADDR_MASK;
861 page = page_header(root);
864 vcpu->mmu.pae_root[i] = INVALID_PAGE;
866 vcpu->mmu.root_hpa = INVALID_PAGE;
869 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
873 struct kvm_mmu_page *page;
875 root_gfn = vcpu->cr3 >> PAGE_SHIFT;
878 if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
879 hpa_t root = vcpu->mmu.root_hpa;
881 ASSERT(!VALID_PAGE(root));
882 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
883 PT64_ROOT_LEVEL, 0, 0, NULL);
884 root = __pa(page->spt);
886 vcpu->mmu.root_hpa = root;
890 for (i = 0; i < 4; ++i) {
891 hpa_t root = vcpu->mmu.pae_root[i];
893 ASSERT(!VALID_PAGE(root));
894 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
895 if (!is_present_pte(vcpu->pdptrs[i])) {
896 vcpu->mmu.pae_root[i] = 0;
899 root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
900 } else if (vcpu->mmu.root_level == 0)
902 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
903 PT32_ROOT_LEVEL, !is_paging(vcpu),
905 root = __pa(page->spt);
907 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
909 vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
912 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
917 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
924 r = mmu_topup_memory_caches(vcpu);
929 ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
932 paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
934 if (is_error_hpa(paddr))
937 return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
940 static void nonpaging_free(struct kvm_vcpu *vcpu)
942 mmu_free_roots(vcpu);
945 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
947 struct kvm_mmu *context = &vcpu->mmu;
949 context->new_cr3 = nonpaging_new_cr3;
950 context->page_fault = nonpaging_page_fault;
951 context->gva_to_gpa = nonpaging_gva_to_gpa;
952 context->free = nonpaging_free;
953 context->root_level = 0;
954 context->shadow_root_level = PT32E_ROOT_LEVEL;
955 context->root_hpa = INVALID_PAGE;
959 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
961 ++vcpu->stat.tlb_flush;
962 kvm_arch_ops->tlb_flush(vcpu);
965 static void paging_new_cr3(struct kvm_vcpu *vcpu)
967 pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
968 mmu_free_roots(vcpu);
971 static void inject_page_fault(struct kvm_vcpu *vcpu,
975 kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
978 static void paging_free(struct kvm_vcpu *vcpu)
980 nonpaging_free(vcpu);
984 #include "paging_tmpl.h"
988 #include "paging_tmpl.h"
991 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
993 struct kvm_mmu *context = &vcpu->mmu;
995 ASSERT(is_pae(vcpu));
996 context->new_cr3 = paging_new_cr3;
997 context->page_fault = paging64_page_fault;
998 context->gva_to_gpa = paging64_gva_to_gpa;
999 context->free = paging_free;
1000 context->root_level = level;
1001 context->shadow_root_level = level;
1002 context->root_hpa = INVALID_PAGE;
1006 static int paging64_init_context(struct kvm_vcpu *vcpu)
1008 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1011 static int paging32_init_context(struct kvm_vcpu *vcpu)
1013 struct kvm_mmu *context = &vcpu->mmu;
1015 context->new_cr3 = paging_new_cr3;
1016 context->page_fault = paging32_page_fault;
1017 context->gva_to_gpa = paging32_gva_to_gpa;
1018 context->free = paging_free;
1019 context->root_level = PT32_ROOT_LEVEL;
1020 context->shadow_root_level = PT32E_ROOT_LEVEL;
1021 context->root_hpa = INVALID_PAGE;
1025 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1027 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1030 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1033 ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1035 if (!is_paging(vcpu))
1036 return nonpaging_init_context(vcpu);
1037 else if (is_long_mode(vcpu))
1038 return paging64_init_context(vcpu);
1039 else if (is_pae(vcpu))
1040 return paging32E_init_context(vcpu);
1042 return paging32_init_context(vcpu);
1045 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1048 if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1049 vcpu->mmu.free(vcpu);
1050 vcpu->mmu.root_hpa = INVALID_PAGE;
1054 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1056 destroy_kvm_mmu(vcpu);
1057 return init_kvm_mmu(vcpu);
1060 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1064 spin_lock(&vcpu->kvm->lock);
1065 r = mmu_topup_memory_caches(vcpu);
1068 mmu_alloc_roots(vcpu);
1069 kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1070 kvm_mmu_flush_tlb(vcpu);
1072 spin_unlock(&vcpu->kvm->lock);
1075 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1077 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1079 mmu_free_roots(vcpu);
1082 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1083 struct kvm_mmu_page *page,
1087 struct kvm_mmu_page *child;
1090 if (is_present_pte(pte)) {
1091 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1092 rmap_remove(vcpu, spte);
1094 child = page_header(pte & PT64_BASE_ADDR_MASK);
1095 mmu_page_remove_parent_pte(vcpu, child, spte);
1099 kvm_flush_remote_tlbs(vcpu->kvm);
1102 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1103 struct kvm_mmu_page *page,
1105 const void *new, int bytes)
1107 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1110 if (page->role.glevels == PT32_ROOT_LEVEL)
1111 paging32_update_pte(vcpu, page, spte, new, bytes);
1113 paging64_update_pte(vcpu, page, spte, new, bytes);
1116 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1117 const u8 *old, const u8 *new, int bytes)
1119 gfn_t gfn = gpa >> PAGE_SHIFT;
1120 struct kvm_mmu_page *page;
1121 struct hlist_node *node, *n;
1122 struct hlist_head *bucket;
1125 unsigned offset = offset_in_page(gpa);
1127 unsigned page_offset;
1128 unsigned misaligned;
1134 pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1135 if (gfn == vcpu->last_pt_write_gfn) {
1136 ++vcpu->last_pt_write_count;
1137 if (vcpu->last_pt_write_count >= 3)
1140 vcpu->last_pt_write_gfn = gfn;
1141 vcpu->last_pt_write_count = 1;
1143 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1144 bucket = &vcpu->kvm->mmu_page_hash[index];
1145 hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1146 if (page->gfn != gfn || page->role.metaphysical)
1148 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1149 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1150 misaligned |= bytes < 4;
1151 if (misaligned || flooded) {
1153 * Misaligned accesses are too much trouble to fix
1154 * up; also, they usually indicate a page is not used
1157 * If we're seeing too many writes to a page,
1158 * it may no longer be a page table, or we may be
1159 * forking, in which case it is better to unmap the
1162 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1163 gpa, bytes, page->role.word);
1164 kvm_mmu_zap_page(vcpu, page);
1167 page_offset = offset;
1168 level = page->role.level;
1170 if (page->role.glevels == PT32_ROOT_LEVEL) {
1171 page_offset <<= 1; /* 32->64 */
1173 * A 32-bit pde maps 4MB while the shadow pdes map
1174 * only 2MB. So we need to double the offset again
1175 * and zap two pdes instead of one.
1177 if (level == PT32_ROOT_LEVEL) {
1178 page_offset &= ~7; /* kill rounding error */
1182 quadrant = page_offset >> PAGE_SHIFT;
1183 page_offset &= ~PAGE_MASK;
1184 if (quadrant != page->role.quadrant)
1187 spte = &page->spt[page_offset / sizeof(*spte)];
1189 mmu_pte_write_zap_pte(vcpu, page, spte);
1190 mmu_pte_write_new_pte(vcpu, page, spte, new, bytes);
1196 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1198 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1200 return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1203 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1205 while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1206 struct kvm_mmu_page *page;
1208 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1209 struct kvm_mmu_page, link);
1210 kvm_mmu_zap_page(vcpu, page);
1213 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1215 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1217 struct kvm_mmu_page *page;
1219 while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1220 page = container_of(vcpu->kvm->active_mmu_pages.next,
1221 struct kvm_mmu_page, link);
1222 kvm_mmu_zap_page(vcpu, page);
1224 free_page((unsigned long)vcpu->mmu.pae_root);
1227 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1234 vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1237 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1238 * Therefore we need to allocate shadow page tables in the first
1239 * 4GB of memory, which happens to fit the DMA32 zone.
1241 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1244 vcpu->mmu.pae_root = page_address(page);
1245 for (i = 0; i < 4; ++i)
1246 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1251 free_mmu_pages(vcpu);
1255 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1258 ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1260 return alloc_mmu_pages(vcpu);
1263 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1266 ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1268 return init_kvm_mmu(vcpu);
1271 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1275 destroy_kvm_mmu(vcpu);
1276 free_mmu_pages(vcpu);
1277 mmu_free_memory_caches(vcpu);
1280 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1282 struct kvm *kvm = vcpu->kvm;
1283 struct kvm_mmu_page *page;
1285 list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1289 if (!test_bit(slot, &page->slot_bitmap))
1293 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1295 if (pt[i] & PT_WRITABLE_MASK) {
1296 rmap_remove(vcpu, &pt[i]);
1297 pt[i] &= ~PT_WRITABLE_MASK;
1302 void kvm_mmu_zap_all(struct kvm_vcpu *vcpu)
1304 destroy_kvm_mmu(vcpu);
1306 while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1307 struct kvm_mmu_page *page;
1309 page = container_of(vcpu->kvm->active_mmu_pages.next,
1310 struct kvm_mmu_page, link);
1311 kvm_mmu_zap_page(vcpu, page);
1314 mmu_free_memory_caches(vcpu);
1315 kvm_flush_remote_tlbs(vcpu->kvm);
1319 void kvm_mmu_module_exit(void)
1321 if (pte_chain_cache)
1322 kmem_cache_destroy(pte_chain_cache);
1323 if (rmap_desc_cache)
1324 kmem_cache_destroy(rmap_desc_cache);
1326 kmem_cache_destroy(mmu_page_cache);
1327 if (mmu_page_header_cache)
1328 kmem_cache_destroy(mmu_page_header_cache);
1331 int kvm_mmu_module_init(void)
1333 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1334 sizeof(struct kvm_pte_chain),
1336 if (!pte_chain_cache)
1338 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1339 sizeof(struct kvm_rmap_desc),
1341 if (!rmap_desc_cache)
1344 mmu_page_cache = kmem_cache_create("kvm_mmu_page",
1346 PAGE_SIZE, 0, NULL);
1347 if (!mmu_page_cache)
1350 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1351 sizeof(struct kvm_mmu_page),
1353 if (!mmu_page_header_cache)
1359 kvm_mmu_module_exit();
1365 static const char *audit_msg;
1367 static gva_t canonicalize(gva_t gva)
1369 #ifdef CONFIG_X86_64
1370 gva = (long long)(gva << 16) >> 16;
1375 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1376 gva_t va, int level)
1378 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1380 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1382 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1385 if (!(ent & PT_PRESENT_MASK))
1388 va = canonicalize(va);
1390 audit_mappings_page(vcpu, ent, va, level - 1);
1392 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1393 hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1395 if ((ent & PT_PRESENT_MASK)
1396 && (ent & PT64_BASE_ADDR_MASK) != hpa)
1397 printk(KERN_ERR "audit error: (%s) levels %d"
1398 " gva %lx gpa %llx hpa %llx ent %llx\n",
1399 audit_msg, vcpu->mmu.root_level,
1405 static void audit_mappings(struct kvm_vcpu *vcpu)
1409 if (vcpu->mmu.root_level == 4)
1410 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1412 for (i = 0; i < 4; ++i)
1413 if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1414 audit_mappings_page(vcpu,
1415 vcpu->mmu.pae_root[i],
1420 static int count_rmaps(struct kvm_vcpu *vcpu)
1425 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1426 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1427 struct kvm_rmap_desc *d;
1429 for (j = 0; j < m->npages; ++j) {
1430 struct page *page = m->phys_mem[j];
1434 if (!(page->private & 1)) {
1438 d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1440 for (k = 0; k < RMAP_EXT; ++k)
1441 if (d->shadow_ptes[k])
1452 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1455 struct kvm_mmu_page *page;
1458 list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1459 u64 *pt = page->spt;
1461 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1464 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1467 if (!(ent & PT_PRESENT_MASK))
1469 if (!(ent & PT_WRITABLE_MASK))
1477 static void audit_rmap(struct kvm_vcpu *vcpu)
1479 int n_rmap = count_rmaps(vcpu);
1480 int n_actual = count_writable_mappings(vcpu);
1482 if (n_rmap != n_actual)
1483 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1484 __FUNCTION__, audit_msg, n_rmap, n_actual);
1487 static void audit_write_protection(struct kvm_vcpu *vcpu)
1489 struct kvm_mmu_page *page;
1491 list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1495 if (page->role.metaphysical)
1498 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1500 pg = pfn_to_page(hfn);
1502 printk(KERN_ERR "%s: (%s) shadow page has writable"
1503 " mappings: gfn %lx role %x\n",
1504 __FUNCTION__, audit_msg, page->gfn,
1509 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1516 audit_write_protection(vcpu);
1517 audit_mappings(vcpu);