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_header_cache;
159 static u64 __read_mostly shadow_trap_nonpresent_pte;
160 static u64 __read_mostly shadow_notrap_nonpresent_pte;
162 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
164 shadow_trap_nonpresent_pte = trap_pte;
165 shadow_notrap_nonpresent_pte = notrap_pte;
167 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
169 static int is_write_protection(struct kvm_vcpu *vcpu)
171 return vcpu->cr0 & X86_CR0_WP;
174 static int is_cpuid_PSE36(void)
179 static int is_nx(struct kvm_vcpu *vcpu)
181 return vcpu->shadow_efer & EFER_NX;
184 static int is_present_pte(unsigned long pte)
186 return pte & PT_PRESENT_MASK;
189 static int is_shadow_present_pte(u64 pte)
191 pte &= ~PT_SHADOW_IO_MARK;
192 return pte != shadow_trap_nonpresent_pte
193 && pte != shadow_notrap_nonpresent_pte;
196 static int is_writeble_pte(unsigned long pte)
198 return pte & PT_WRITABLE_MASK;
201 static int is_io_pte(unsigned long pte)
203 return pte & PT_SHADOW_IO_MARK;
206 static int is_rmap_pte(u64 pte)
208 return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
209 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
212 static void set_shadow_pte(u64 *sptep, u64 spte)
215 set_64bit((unsigned long *)sptep, spte);
217 set_64bit((unsigned long long *)sptep, spte);
221 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
222 struct kmem_cache *base_cache, int min)
226 if (cache->nobjs >= min)
228 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
229 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
232 cache->objects[cache->nobjs++] = obj;
237 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
240 kfree(mc->objects[--mc->nobjs]);
243 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
248 if (cache->nobjs >= min)
250 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
251 page = alloc_page(GFP_KERNEL);
254 set_page_private(page, 0);
255 cache->objects[cache->nobjs++] = page_address(page);
260 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
263 free_page((unsigned long)mc->objects[--mc->nobjs]);
266 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
270 kvm_mmu_free_some_pages(vcpu);
271 r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
275 r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
279 r = mmu_topup_memory_cache_page(&vcpu->mmu_page_cache, 8);
282 r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
283 mmu_page_header_cache, 4);
288 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
290 mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
291 mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
292 mmu_free_memory_cache_page(&vcpu->mmu_page_cache);
293 mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
296 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
302 p = mc->objects[--mc->nobjs];
307 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
309 return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
310 sizeof(struct kvm_pte_chain));
313 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
318 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
320 return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
321 sizeof(struct kvm_rmap_desc));
324 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
330 * Take gfn and return the reverse mapping to it.
331 * Note: gfn must be unaliased before this function get called
334 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
336 struct kvm_memory_slot *slot;
338 slot = gfn_to_memslot(kvm, gfn);
339 return &slot->rmap[gfn - slot->base_gfn];
343 * Reverse mapping data structures:
345 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
346 * that points to page_address(page).
348 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
349 * containing more mappings.
351 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
353 struct kvm_mmu_page *page;
354 struct kvm_rmap_desc *desc;
355 unsigned long *rmapp;
358 if (!is_rmap_pte(*spte))
360 gfn = unalias_gfn(vcpu->kvm, gfn);
361 page = page_header(__pa(spte));
362 page->gfns[spte - page->spt] = gfn;
363 rmapp = gfn_to_rmap(vcpu->kvm, gfn);
365 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
366 *rmapp = (unsigned long)spte;
367 } else if (!(*rmapp & 1)) {
368 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
369 desc = mmu_alloc_rmap_desc(vcpu);
370 desc->shadow_ptes[0] = (u64 *)*rmapp;
371 desc->shadow_ptes[1] = spte;
372 *rmapp = (unsigned long)desc | 1;
374 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
375 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
376 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
378 if (desc->shadow_ptes[RMAP_EXT-1]) {
379 desc->more = mmu_alloc_rmap_desc(vcpu);
382 for (i = 0; desc->shadow_ptes[i]; ++i)
384 desc->shadow_ptes[i] = spte;
388 static void rmap_desc_remove_entry(unsigned long *rmapp,
389 struct kvm_rmap_desc *desc,
391 struct kvm_rmap_desc *prev_desc)
395 for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
397 desc->shadow_ptes[i] = desc->shadow_ptes[j];
398 desc->shadow_ptes[j] = NULL;
401 if (!prev_desc && !desc->more)
402 *rmapp = (unsigned long)desc->shadow_ptes[0];
405 prev_desc->more = desc->more;
407 *rmapp = (unsigned long)desc->more | 1;
408 mmu_free_rmap_desc(desc);
411 static void rmap_remove(struct kvm *kvm, u64 *spte)
413 struct kvm_rmap_desc *desc;
414 struct kvm_rmap_desc *prev_desc;
415 struct kvm_mmu_page *page;
416 unsigned long *rmapp;
419 if (!is_rmap_pte(*spte))
421 page = page_header(__pa(spte));
422 rmapp = gfn_to_rmap(kvm, page->gfns[spte - page->spt]);
424 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
426 } else if (!(*rmapp & 1)) {
427 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
428 if ((u64 *)*rmapp != spte) {
429 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
435 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
436 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
439 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
440 if (desc->shadow_ptes[i] == spte) {
441 rmap_desc_remove_entry(rmapp,
453 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
455 struct kvm_rmap_desc *desc;
456 unsigned long *rmapp;
459 gfn = unalias_gfn(vcpu->kvm, gfn);
460 rmapp = gfn_to_rmap(vcpu->kvm, gfn);
464 spte = (u64 *)*rmapp;
466 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
467 spte = desc->shadow_ptes[0];
470 BUG_ON(!(*spte & PT_PRESENT_MASK));
471 BUG_ON(!(*spte & PT_WRITABLE_MASK));
472 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
473 rmap_remove(vcpu->kvm, spte);
474 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
475 kvm_flush_remote_tlbs(vcpu->kvm);
480 static int is_empty_shadow_page(u64 *spt)
485 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
486 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
487 printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
495 static void kvm_mmu_free_page(struct kvm *kvm,
496 struct kvm_mmu_page *page_head)
498 ASSERT(is_empty_shadow_page(page_head->spt));
499 list_del(&page_head->link);
500 __free_page(virt_to_page(page_head->spt));
501 __free_page(virt_to_page(page_head->gfns));
503 ++kvm->n_free_mmu_pages;
506 static unsigned kvm_page_table_hashfn(gfn_t gfn)
511 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
514 struct kvm_mmu_page *page;
516 if (!vcpu->kvm->n_free_mmu_pages)
519 page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
521 page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
522 page->gfns = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
523 set_page_private(virt_to_page(page->spt), (unsigned long)page);
524 list_add(&page->link, &vcpu->kvm->active_mmu_pages);
525 ASSERT(is_empty_shadow_page(page->spt));
526 page->slot_bitmap = 0;
527 page->multimapped = 0;
528 page->parent_pte = parent_pte;
529 --vcpu->kvm->n_free_mmu_pages;
533 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
534 struct kvm_mmu_page *page, u64 *parent_pte)
536 struct kvm_pte_chain *pte_chain;
537 struct hlist_node *node;
542 if (!page->multimapped) {
543 u64 *old = page->parent_pte;
546 page->parent_pte = parent_pte;
549 page->multimapped = 1;
550 pte_chain = mmu_alloc_pte_chain(vcpu);
551 INIT_HLIST_HEAD(&page->parent_ptes);
552 hlist_add_head(&pte_chain->link, &page->parent_ptes);
553 pte_chain->parent_ptes[0] = old;
555 hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
556 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
558 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
559 if (!pte_chain->parent_ptes[i]) {
560 pte_chain->parent_ptes[i] = parent_pte;
564 pte_chain = mmu_alloc_pte_chain(vcpu);
566 hlist_add_head(&pte_chain->link, &page->parent_ptes);
567 pte_chain->parent_ptes[0] = parent_pte;
570 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
573 struct kvm_pte_chain *pte_chain;
574 struct hlist_node *node;
577 if (!page->multimapped) {
578 BUG_ON(page->parent_pte != parent_pte);
579 page->parent_pte = NULL;
582 hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
583 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
584 if (!pte_chain->parent_ptes[i])
586 if (pte_chain->parent_ptes[i] != parent_pte)
588 while (i + 1 < NR_PTE_CHAIN_ENTRIES
589 && pte_chain->parent_ptes[i + 1]) {
590 pte_chain->parent_ptes[i]
591 = pte_chain->parent_ptes[i + 1];
594 pte_chain->parent_ptes[i] = NULL;
596 hlist_del(&pte_chain->link);
597 mmu_free_pte_chain(pte_chain);
598 if (hlist_empty(&page->parent_ptes)) {
599 page->multimapped = 0;
600 page->parent_pte = NULL;
608 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
612 struct hlist_head *bucket;
613 struct kvm_mmu_page *page;
614 struct hlist_node *node;
616 pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
617 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
618 bucket = &vcpu->kvm->mmu_page_hash[index];
619 hlist_for_each_entry(page, node, bucket, hash_link)
620 if (page->gfn == gfn && !page->role.metaphysical) {
621 pgprintk("%s: found role %x\n",
622 __FUNCTION__, page->role.word);
628 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
633 unsigned hugepage_access,
636 union kvm_mmu_page_role role;
639 struct hlist_head *bucket;
640 struct kvm_mmu_page *page;
641 struct hlist_node *node;
644 role.glevels = vcpu->mmu.root_level;
646 role.metaphysical = metaphysical;
647 role.hugepage_access = hugepage_access;
648 if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
649 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
650 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
651 role.quadrant = quadrant;
653 pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
655 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
656 bucket = &vcpu->kvm->mmu_page_hash[index];
657 hlist_for_each_entry(page, node, bucket, hash_link)
658 if (page->gfn == gfn && page->role.word == role.word) {
659 mmu_page_add_parent_pte(vcpu, page, parent_pte);
660 pgprintk("%s: found\n", __FUNCTION__);
663 page = kvm_mmu_alloc_page(vcpu, parent_pte);
666 pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
669 hlist_add_head(&page->hash_link, bucket);
670 vcpu->mmu.prefetch_page(vcpu, page);
672 rmap_write_protect(vcpu, gfn);
676 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
677 struct kvm_mmu_page *page)
685 if (page->role.level == PT_PAGE_TABLE_LEVEL) {
686 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
687 if (is_shadow_present_pte(pt[i]))
688 rmap_remove(kvm, &pt[i]);
689 pt[i] = shadow_trap_nonpresent_pte;
691 kvm_flush_remote_tlbs(kvm);
695 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
698 pt[i] = shadow_trap_nonpresent_pte;
699 if (!is_shadow_present_pte(ent))
701 ent &= PT64_BASE_ADDR_MASK;
702 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
704 kvm_flush_remote_tlbs(kvm);
707 static void kvm_mmu_put_page(struct kvm_mmu_page *page,
710 mmu_page_remove_parent_pte(page, parent_pte);
713 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
717 for (i = 0; i < KVM_MAX_VCPUS; ++i)
719 kvm->vcpus[i]->last_pte_updated = NULL;
722 static void kvm_mmu_zap_page(struct kvm *kvm,
723 struct kvm_mmu_page *page)
727 while (page->multimapped || page->parent_pte) {
728 if (!page->multimapped)
729 parent_pte = page->parent_pte;
731 struct kvm_pte_chain *chain;
733 chain = container_of(page->parent_ptes.first,
734 struct kvm_pte_chain, link);
735 parent_pte = chain->parent_ptes[0];
738 kvm_mmu_put_page(page, parent_pte);
739 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
741 kvm_mmu_page_unlink_children(kvm, page);
742 if (!page->root_count) {
743 hlist_del(&page->hash_link);
744 kvm_mmu_free_page(kvm, page);
746 list_move(&page->link, &kvm->active_mmu_pages);
747 kvm_mmu_reset_last_pte_updated(kvm);
750 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
753 struct hlist_head *bucket;
754 struct kvm_mmu_page *page;
755 struct hlist_node *node, *n;
758 pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
760 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
761 bucket = &vcpu->kvm->mmu_page_hash[index];
762 hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
763 if (page->gfn == gfn && !page->role.metaphysical) {
764 pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
766 kvm_mmu_zap_page(vcpu->kvm, page);
772 static void mmu_unshadow(struct kvm_vcpu *vcpu, gfn_t gfn)
774 struct kvm_mmu_page *page;
776 while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
777 pgprintk("%s: zap %lx %x\n",
778 __FUNCTION__, gfn, page->role.word);
779 kvm_mmu_zap_page(vcpu->kvm, page);
783 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
785 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
786 struct kvm_mmu_page *page_head = page_header(__pa(pte));
788 __set_bit(slot, &page_head->slot_bitmap);
791 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
793 hpa_t hpa = gpa_to_hpa(vcpu, gpa);
795 return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
798 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
802 ASSERT((gpa & HPA_ERR_MASK) == 0);
803 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
805 return gpa | HPA_ERR_MASK;
806 return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
807 | (gpa & (PAGE_SIZE-1));
810 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
812 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
814 if (gpa == UNMAPPED_GVA)
816 return gpa_to_hpa(vcpu, gpa);
819 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
821 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
823 if (gpa == UNMAPPED_GVA)
825 return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
828 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
832 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
834 int level = PT32E_ROOT_LEVEL;
835 hpa_t table_addr = vcpu->mmu.root_hpa;
838 u32 index = PT64_INDEX(v, level);
842 ASSERT(VALID_PAGE(table_addr));
843 table = __va(table_addr);
847 if (is_shadow_present_pte(pte) && is_writeble_pte(pte))
849 mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
850 page_header_update_slot(vcpu->kvm, table, v);
851 table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
853 rmap_add(vcpu, &table[index], v >> PAGE_SHIFT);
857 if (table[index] == shadow_trap_nonpresent_pte) {
858 struct kvm_mmu_page *new_table;
861 pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
863 new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
865 1, 0, &table[index]);
867 pgprintk("nonpaging_map: ENOMEM\n");
871 table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
872 | PT_WRITABLE_MASK | PT_USER_MASK;
874 table_addr = table[index] & PT64_BASE_ADDR_MASK;
878 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
879 struct kvm_mmu_page *sp)
883 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
884 sp->spt[i] = shadow_trap_nonpresent_pte;
887 static void mmu_free_roots(struct kvm_vcpu *vcpu)
890 struct kvm_mmu_page *page;
892 if (!VALID_PAGE(vcpu->mmu.root_hpa))
895 if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
896 hpa_t root = vcpu->mmu.root_hpa;
898 page = page_header(root);
900 vcpu->mmu.root_hpa = INVALID_PAGE;
904 for (i = 0; i < 4; ++i) {
905 hpa_t root = vcpu->mmu.pae_root[i];
908 root &= PT64_BASE_ADDR_MASK;
909 page = page_header(root);
912 vcpu->mmu.pae_root[i] = INVALID_PAGE;
914 vcpu->mmu.root_hpa = INVALID_PAGE;
917 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
921 struct kvm_mmu_page *page;
923 root_gfn = vcpu->cr3 >> PAGE_SHIFT;
926 if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
927 hpa_t root = vcpu->mmu.root_hpa;
929 ASSERT(!VALID_PAGE(root));
930 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
931 PT64_ROOT_LEVEL, 0, 0, NULL);
932 root = __pa(page->spt);
934 vcpu->mmu.root_hpa = root;
938 for (i = 0; i < 4; ++i) {
939 hpa_t root = vcpu->mmu.pae_root[i];
941 ASSERT(!VALID_PAGE(root));
942 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
943 if (!is_present_pte(vcpu->pdptrs[i])) {
944 vcpu->mmu.pae_root[i] = 0;
947 root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
948 } else if (vcpu->mmu.root_level == 0)
950 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
951 PT32_ROOT_LEVEL, !is_paging(vcpu),
953 root = __pa(page->spt);
955 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
957 vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
960 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
965 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
972 r = mmu_topup_memory_caches(vcpu);
977 ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
980 paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
982 if (is_error_hpa(paddr))
985 return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
988 static void nonpaging_free(struct kvm_vcpu *vcpu)
990 mmu_free_roots(vcpu);
993 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
995 struct kvm_mmu *context = &vcpu->mmu;
997 context->new_cr3 = nonpaging_new_cr3;
998 context->page_fault = nonpaging_page_fault;
999 context->gva_to_gpa = nonpaging_gva_to_gpa;
1000 context->free = nonpaging_free;
1001 context->prefetch_page = nonpaging_prefetch_page;
1002 context->root_level = 0;
1003 context->shadow_root_level = PT32E_ROOT_LEVEL;
1004 context->root_hpa = INVALID_PAGE;
1008 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1010 ++vcpu->stat.tlb_flush;
1011 kvm_x86_ops->tlb_flush(vcpu);
1014 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1016 pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
1017 mmu_free_roots(vcpu);
1020 static void inject_page_fault(struct kvm_vcpu *vcpu,
1024 kvm_x86_ops->inject_page_fault(vcpu, addr, err_code);
1027 static void paging_free(struct kvm_vcpu *vcpu)
1029 nonpaging_free(vcpu);
1033 #include "paging_tmpl.h"
1037 #include "paging_tmpl.h"
1040 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1042 struct kvm_mmu *context = &vcpu->mmu;
1044 ASSERT(is_pae(vcpu));
1045 context->new_cr3 = paging_new_cr3;
1046 context->page_fault = paging64_page_fault;
1047 context->gva_to_gpa = paging64_gva_to_gpa;
1048 context->prefetch_page = paging64_prefetch_page;
1049 context->free = paging_free;
1050 context->root_level = level;
1051 context->shadow_root_level = level;
1052 context->root_hpa = INVALID_PAGE;
1056 static int paging64_init_context(struct kvm_vcpu *vcpu)
1058 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1061 static int paging32_init_context(struct kvm_vcpu *vcpu)
1063 struct kvm_mmu *context = &vcpu->mmu;
1065 context->new_cr3 = paging_new_cr3;
1066 context->page_fault = paging32_page_fault;
1067 context->gva_to_gpa = paging32_gva_to_gpa;
1068 context->free = paging_free;
1069 context->prefetch_page = paging32_prefetch_page;
1070 context->root_level = PT32_ROOT_LEVEL;
1071 context->shadow_root_level = PT32E_ROOT_LEVEL;
1072 context->root_hpa = INVALID_PAGE;
1076 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1078 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1081 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1084 ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1086 if (!is_paging(vcpu))
1087 return nonpaging_init_context(vcpu);
1088 else if (is_long_mode(vcpu))
1089 return paging64_init_context(vcpu);
1090 else if (is_pae(vcpu))
1091 return paging32E_init_context(vcpu);
1093 return paging32_init_context(vcpu);
1096 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1099 if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1100 vcpu->mmu.free(vcpu);
1101 vcpu->mmu.root_hpa = INVALID_PAGE;
1105 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1107 destroy_kvm_mmu(vcpu);
1108 return init_kvm_mmu(vcpu);
1110 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1112 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1116 mutex_lock(&vcpu->kvm->lock);
1117 r = mmu_topup_memory_caches(vcpu);
1120 mmu_alloc_roots(vcpu);
1121 kvm_x86_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1122 kvm_mmu_flush_tlb(vcpu);
1124 mutex_unlock(&vcpu->kvm->lock);
1127 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1129 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1131 mmu_free_roots(vcpu);
1134 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1135 struct kvm_mmu_page *page,
1139 struct kvm_mmu_page *child;
1142 if (is_shadow_present_pte(pte)) {
1143 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1144 rmap_remove(vcpu->kvm, spte);
1146 child = page_header(pte & PT64_BASE_ADDR_MASK);
1147 mmu_page_remove_parent_pte(child, spte);
1150 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1151 kvm_flush_remote_tlbs(vcpu->kvm);
1154 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1155 struct kvm_mmu_page *page,
1157 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,
1167 paging64_update_pte(vcpu, page, spte, new, bytes,
1171 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1173 u64 *spte = vcpu->last_pte_updated;
1175 return !!(spte && (*spte & PT_ACCESSED_MASK));
1178 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1179 const u8 *new, int bytes)
1181 gfn_t gfn = gpa >> PAGE_SHIFT;
1182 struct kvm_mmu_page *page;
1183 struct hlist_node *node, *n;
1184 struct hlist_head *bucket;
1187 unsigned offset = offset_in_page(gpa);
1189 unsigned page_offset;
1190 unsigned misaligned;
1196 pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1197 kvm_mmu_audit(vcpu, "pre pte write");
1198 if (gfn == vcpu->last_pt_write_gfn
1199 && !last_updated_pte_accessed(vcpu)) {
1200 ++vcpu->last_pt_write_count;
1201 if (vcpu->last_pt_write_count >= 3)
1204 vcpu->last_pt_write_gfn = gfn;
1205 vcpu->last_pt_write_count = 1;
1206 vcpu->last_pte_updated = NULL;
1208 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1209 bucket = &vcpu->kvm->mmu_page_hash[index];
1210 hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1211 if (page->gfn != gfn || page->role.metaphysical)
1213 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1214 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1215 misaligned |= bytes < 4;
1216 if (misaligned || flooded) {
1218 * Misaligned accesses are too much trouble to fix
1219 * up; also, they usually indicate a page is not used
1222 * If we're seeing too many writes to a page,
1223 * it may no longer be a page table, or we may be
1224 * forking, in which case it is better to unmap the
1227 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1228 gpa, bytes, page->role.word);
1229 kvm_mmu_zap_page(vcpu->kvm, page);
1232 page_offset = offset;
1233 level = page->role.level;
1235 if (page->role.glevels == PT32_ROOT_LEVEL) {
1236 page_offset <<= 1; /* 32->64 */
1238 * A 32-bit pde maps 4MB while the shadow pdes map
1239 * only 2MB. So we need to double the offset again
1240 * and zap two pdes instead of one.
1242 if (level == PT32_ROOT_LEVEL) {
1243 page_offset &= ~7; /* kill rounding error */
1247 quadrant = page_offset >> PAGE_SHIFT;
1248 page_offset &= ~PAGE_MASK;
1249 if (quadrant != page->role.quadrant)
1252 spte = &page->spt[page_offset / sizeof(*spte)];
1254 mmu_pte_write_zap_pte(vcpu, page, spte);
1255 mmu_pte_write_new_pte(vcpu, page, spte, new, bytes,
1256 page_offset & (pte_size - 1));
1260 kvm_mmu_audit(vcpu, "post pte write");
1263 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1265 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1267 return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1270 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1272 while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1273 struct kvm_mmu_page *page;
1275 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1276 struct kvm_mmu_page, link);
1277 kvm_mmu_zap_page(vcpu->kvm, page);
1281 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1283 struct kvm_mmu_page *page;
1285 while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1286 page = container_of(vcpu->kvm->active_mmu_pages.next,
1287 struct kvm_mmu_page, link);
1288 kvm_mmu_zap_page(vcpu->kvm, page);
1290 free_page((unsigned long)vcpu->mmu.pae_root);
1293 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1300 vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1303 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1304 * Therefore we need to allocate shadow page tables in the first
1305 * 4GB of memory, which happens to fit the DMA32 zone.
1307 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1310 vcpu->mmu.pae_root = page_address(page);
1311 for (i = 0; i < 4; ++i)
1312 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1317 free_mmu_pages(vcpu);
1321 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1324 ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1326 return alloc_mmu_pages(vcpu);
1329 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1332 ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1334 return init_kvm_mmu(vcpu);
1337 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1341 destroy_kvm_mmu(vcpu);
1342 free_mmu_pages(vcpu);
1343 mmu_free_memory_caches(vcpu);
1346 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1348 struct kvm_mmu_page *page;
1350 list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1354 if (!test_bit(slot, &page->slot_bitmap))
1358 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1360 if (pt[i] & PT_WRITABLE_MASK) {
1361 rmap_remove(kvm, &pt[i]);
1362 pt[i] &= ~PT_WRITABLE_MASK;
1367 void kvm_mmu_zap_all(struct kvm *kvm)
1369 struct kvm_mmu_page *page, *node;
1371 list_for_each_entry_safe(page, node, &kvm->active_mmu_pages, link)
1372 kvm_mmu_zap_page(kvm, page);
1374 kvm_flush_remote_tlbs(kvm);
1377 void kvm_mmu_module_exit(void)
1379 if (pte_chain_cache)
1380 kmem_cache_destroy(pte_chain_cache);
1381 if (rmap_desc_cache)
1382 kmem_cache_destroy(rmap_desc_cache);
1383 if (mmu_page_header_cache)
1384 kmem_cache_destroy(mmu_page_header_cache);
1387 int kvm_mmu_module_init(void)
1389 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1390 sizeof(struct kvm_pte_chain),
1392 if (!pte_chain_cache)
1394 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1395 sizeof(struct kvm_rmap_desc),
1397 if (!rmap_desc_cache)
1400 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1401 sizeof(struct kvm_mmu_page),
1403 if (!mmu_page_header_cache)
1409 kvm_mmu_module_exit();
1415 static const char *audit_msg;
1417 static gva_t canonicalize(gva_t gva)
1419 #ifdef CONFIG_X86_64
1420 gva = (long long)(gva << 16) >> 16;
1425 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1426 gva_t va, int level)
1428 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1430 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1432 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1435 if (ent == shadow_trap_nonpresent_pte)
1438 va = canonicalize(va);
1440 if (ent == shadow_notrap_nonpresent_pte)
1441 printk(KERN_ERR "audit: (%s) nontrapping pte"
1442 " in nonleaf level: levels %d gva %lx"
1443 " level %d pte %llx\n", audit_msg,
1444 vcpu->mmu.root_level, va, level, ent);
1446 audit_mappings_page(vcpu, ent, va, level - 1);
1448 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1449 hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1451 if (is_shadow_present_pte(ent)
1452 && (ent & PT64_BASE_ADDR_MASK) != hpa)
1453 printk(KERN_ERR "xx audit error: (%s) levels %d"
1454 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1455 audit_msg, vcpu->mmu.root_level,
1456 va, gpa, hpa, ent, is_shadow_present_pte(ent));
1457 else if (ent == shadow_notrap_nonpresent_pte
1458 && !is_error_hpa(hpa))
1459 printk(KERN_ERR "audit: (%s) notrap shadow,"
1460 " valid guest gva %lx\n", audit_msg, va);
1466 static void audit_mappings(struct kvm_vcpu *vcpu)
1470 if (vcpu->mmu.root_level == 4)
1471 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1473 for (i = 0; i < 4; ++i)
1474 if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1475 audit_mappings_page(vcpu,
1476 vcpu->mmu.pae_root[i],
1481 static int count_rmaps(struct kvm_vcpu *vcpu)
1486 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1487 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1488 struct kvm_rmap_desc *d;
1490 for (j = 0; j < m->npages; ++j) {
1491 unsigned long *rmapp = &m->rmap[j];
1495 if (!(*rmapp & 1)) {
1499 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1501 for (k = 0; k < RMAP_EXT; ++k)
1502 if (d->shadow_ptes[k])
1513 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1516 struct kvm_mmu_page *page;
1519 list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1520 u64 *pt = page->spt;
1522 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1525 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1528 if (!(ent & PT_PRESENT_MASK))
1530 if (!(ent & PT_WRITABLE_MASK))
1538 static void audit_rmap(struct kvm_vcpu *vcpu)
1540 int n_rmap = count_rmaps(vcpu);
1541 int n_actual = count_writable_mappings(vcpu);
1543 if (n_rmap != n_actual)
1544 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1545 __FUNCTION__, audit_msg, n_rmap, n_actual);
1548 static void audit_write_protection(struct kvm_vcpu *vcpu)
1550 struct kvm_mmu_page *page;
1551 struct kvm_memory_slot *slot;
1552 unsigned long *rmapp;
1555 list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1556 if (page->role.metaphysical)
1559 slot = gfn_to_memslot(vcpu->kvm, page->gfn);
1560 gfn = unalias_gfn(vcpu->kvm, page->gfn);
1561 rmapp = &slot->rmap[gfn - slot->base_gfn];
1563 printk(KERN_ERR "%s: (%s) shadow page has writable"
1564 " mappings: gfn %lx role %x\n",
1565 __FUNCTION__, audit_msg, page->gfn,
1570 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1577 audit_write_protection(vcpu);
1578 audit_mappings(vcpu);