2 * Kernel-based Virtual Machine driver for Linux
4 * This module enables machines with Intel VT-x extensions to run virtual
5 * machines without emulation or binary translation.
9 * Copyright (C) 2006 Qumranet, Inc.
12 * Yaniv Kamay <yaniv@qumranet.com>
13 * Avi Kivity <avi@qumranet.com>
15 * This work is licensed under the terms of the GNU GPL, version 2. See
16 * the COPYING file in the top-level directory.
22 #include <linux/kvm_host.h>
23 #include <linux/types.h>
24 #include <linux/string.h>
26 #include <linux/highmem.h>
27 #include <linux/module.h>
28 #include <linux/swap.h>
29 #include <linux/hugetlb.h>
30 #include <linux/compiler.h>
33 #include <asm/cmpxchg.h>
38 * When setting this variable to true it enables Two-Dimensional-Paging
39 * where the hardware walks 2 page tables:
40 * 1. the guest-virtual to guest-physical
41 * 2. while doing 1. it walks guest-physical to host-physical
42 * If the hardware supports that we don't need to do shadow paging.
44 bool tdp_enabled = false;
51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
63 #define pgprintk(x...) do { } while (0)
64 #define rmap_printk(x...) do { } while (0)
68 #if defined(MMU_DEBUG) || defined(AUDIT)
70 module_param(dbg, bool, 0644);
73 static int oos_shadow = 1;
74 module_param(oos_shadow, bool, 0644);
77 #define ASSERT(x) do { } while (0)
81 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
82 __FILE__, __LINE__, #x); \
86 #define PT_FIRST_AVAIL_BITS_SHIFT 9
87 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
89 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
91 #define PT64_LEVEL_BITS 9
93 #define PT64_LEVEL_SHIFT(level) \
94 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
96 #define PT64_LEVEL_MASK(level) \
97 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
99 #define PT64_INDEX(address, level)\
100 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
103 #define PT32_LEVEL_BITS 10
105 #define PT32_LEVEL_SHIFT(level) \
106 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
108 #define PT32_LEVEL_MASK(level) \
109 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
111 #define PT32_INDEX(address, level)\
112 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
115 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
116 #define PT64_DIR_BASE_ADDR_MASK \
117 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
119 #define PT32_BASE_ADDR_MASK PAGE_MASK
120 #define PT32_DIR_BASE_ADDR_MASK \
121 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
123 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
126 #define PFERR_PRESENT_MASK (1U << 0)
127 #define PFERR_WRITE_MASK (1U << 1)
128 #define PFERR_USER_MASK (1U << 2)
129 #define PFERR_FETCH_MASK (1U << 4)
131 #define PT_DIRECTORY_LEVEL 2
132 #define PT_PAGE_TABLE_LEVEL 1
136 #define ACC_EXEC_MASK 1
137 #define ACC_WRITE_MASK PT_WRITABLE_MASK
138 #define ACC_USER_MASK PT_USER_MASK
139 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
141 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
143 struct kvm_rmap_desc {
144 u64 *shadow_ptes[RMAP_EXT];
145 struct kvm_rmap_desc *more;
148 struct kvm_shadow_walk {
149 int (*entry)(struct kvm_shadow_walk *walk, struct kvm_vcpu *vcpu,
150 u64 addr, u64 *spte, int level);
153 struct kvm_unsync_walk {
154 int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
157 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
159 static struct kmem_cache *pte_chain_cache;
160 static struct kmem_cache *rmap_desc_cache;
161 static struct kmem_cache *mmu_page_header_cache;
163 static u64 __read_mostly shadow_trap_nonpresent_pte;
164 static u64 __read_mostly shadow_notrap_nonpresent_pte;
165 static u64 __read_mostly shadow_base_present_pte;
166 static u64 __read_mostly shadow_nx_mask;
167 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
168 static u64 __read_mostly shadow_user_mask;
169 static u64 __read_mostly shadow_accessed_mask;
170 static u64 __read_mostly shadow_dirty_mask;
171 static u64 __read_mostly shadow_mt_mask;
173 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
175 shadow_trap_nonpresent_pte = trap_pte;
176 shadow_notrap_nonpresent_pte = notrap_pte;
178 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
180 void kvm_mmu_set_base_ptes(u64 base_pte)
182 shadow_base_present_pte = base_pte;
184 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
186 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
187 u64 dirty_mask, u64 nx_mask, u64 x_mask, u64 mt_mask)
189 shadow_user_mask = user_mask;
190 shadow_accessed_mask = accessed_mask;
191 shadow_dirty_mask = dirty_mask;
192 shadow_nx_mask = nx_mask;
193 shadow_x_mask = x_mask;
194 shadow_mt_mask = mt_mask;
196 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
198 static int is_write_protection(struct kvm_vcpu *vcpu)
200 return vcpu->arch.cr0 & X86_CR0_WP;
203 static int is_cpuid_PSE36(void)
208 static int is_nx(struct kvm_vcpu *vcpu)
210 return vcpu->arch.shadow_efer & EFER_NX;
213 static int is_present_pte(unsigned long pte)
215 return pte & PT_PRESENT_MASK;
218 static int is_shadow_present_pte(u64 pte)
220 return pte != shadow_trap_nonpresent_pte
221 && pte != shadow_notrap_nonpresent_pte;
224 static int is_large_pte(u64 pte)
226 return pte & PT_PAGE_SIZE_MASK;
229 static int is_writeble_pte(unsigned long pte)
231 return pte & PT_WRITABLE_MASK;
234 static int is_dirty_pte(unsigned long pte)
236 return pte & shadow_dirty_mask;
239 static int is_rmap_pte(u64 pte)
241 return is_shadow_present_pte(pte);
244 static pfn_t spte_to_pfn(u64 pte)
246 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
249 static gfn_t pse36_gfn_delta(u32 gpte)
251 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
253 return (gpte & PT32_DIR_PSE36_MASK) << shift;
256 static void set_shadow_pte(u64 *sptep, u64 spte)
259 set_64bit((unsigned long *)sptep, spte);
261 set_64bit((unsigned long long *)sptep, spte);
265 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
266 struct kmem_cache *base_cache, int min)
270 if (cache->nobjs >= min)
272 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
273 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
276 cache->objects[cache->nobjs++] = obj;
281 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
284 kfree(mc->objects[--mc->nobjs]);
287 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
292 if (cache->nobjs >= min)
294 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
295 page = alloc_page(GFP_KERNEL);
298 set_page_private(page, 0);
299 cache->objects[cache->nobjs++] = page_address(page);
304 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
307 free_page((unsigned long)mc->objects[--mc->nobjs]);
310 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
314 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
318 r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
322 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
325 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
326 mmu_page_header_cache, 4);
331 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
333 mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
334 mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
335 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
336 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
339 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
345 p = mc->objects[--mc->nobjs];
350 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
352 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
353 sizeof(struct kvm_pte_chain));
356 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
361 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
363 return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
364 sizeof(struct kvm_rmap_desc));
367 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
373 * Return the pointer to the largepage write count for a given
374 * gfn, handling slots that are not large page aligned.
376 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
380 idx = (gfn / KVM_PAGES_PER_HPAGE) -
381 (slot->base_gfn / KVM_PAGES_PER_HPAGE);
382 return &slot->lpage_info[idx].write_count;
385 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
389 gfn = unalias_gfn(kvm, gfn);
390 write_count = slot_largepage_idx(gfn,
391 gfn_to_memslot_unaliased(kvm, gfn));
395 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
399 gfn = unalias_gfn(kvm, gfn);
400 write_count = slot_largepage_idx(gfn,
401 gfn_to_memslot_unaliased(kvm, gfn));
403 WARN_ON(*write_count < 0);
406 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
408 struct kvm_memory_slot *slot;
411 gfn = unalias_gfn(kvm, gfn);
412 slot = gfn_to_memslot_unaliased(kvm, gfn);
414 largepage_idx = slot_largepage_idx(gfn, slot);
415 return *largepage_idx;
421 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
423 struct vm_area_struct *vma;
427 addr = gfn_to_hva(kvm, gfn);
428 if (kvm_is_error_hva(addr))
431 down_read(¤t->mm->mmap_sem);
432 vma = find_vma(current->mm, addr);
433 if (vma && is_vm_hugetlb_page(vma))
435 up_read(¤t->mm->mmap_sem);
440 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
442 struct kvm_memory_slot *slot;
444 if (has_wrprotected_page(vcpu->kvm, large_gfn))
447 if (!host_largepage_backed(vcpu->kvm, large_gfn))
450 slot = gfn_to_memslot(vcpu->kvm, large_gfn);
451 if (slot && slot->dirty_bitmap)
458 * Take gfn and return the reverse mapping to it.
459 * Note: gfn must be unaliased before this function get called
462 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
464 struct kvm_memory_slot *slot;
467 slot = gfn_to_memslot(kvm, gfn);
469 return &slot->rmap[gfn - slot->base_gfn];
471 idx = (gfn / KVM_PAGES_PER_HPAGE) -
472 (slot->base_gfn / KVM_PAGES_PER_HPAGE);
474 return &slot->lpage_info[idx].rmap_pde;
478 * Reverse mapping data structures:
480 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
481 * that points to page_address(page).
483 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
484 * containing more mappings.
486 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
488 struct kvm_mmu_page *sp;
489 struct kvm_rmap_desc *desc;
490 unsigned long *rmapp;
493 if (!is_rmap_pte(*spte))
495 gfn = unalias_gfn(vcpu->kvm, gfn);
496 sp = page_header(__pa(spte));
497 sp->gfns[spte - sp->spt] = gfn;
498 rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
500 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
501 *rmapp = (unsigned long)spte;
502 } else if (!(*rmapp & 1)) {
503 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
504 desc = mmu_alloc_rmap_desc(vcpu);
505 desc->shadow_ptes[0] = (u64 *)*rmapp;
506 desc->shadow_ptes[1] = spte;
507 *rmapp = (unsigned long)desc | 1;
509 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
510 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
511 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
513 if (desc->shadow_ptes[RMAP_EXT-1]) {
514 desc->more = mmu_alloc_rmap_desc(vcpu);
517 for (i = 0; desc->shadow_ptes[i]; ++i)
519 desc->shadow_ptes[i] = spte;
523 static void rmap_desc_remove_entry(unsigned long *rmapp,
524 struct kvm_rmap_desc *desc,
526 struct kvm_rmap_desc *prev_desc)
530 for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
532 desc->shadow_ptes[i] = desc->shadow_ptes[j];
533 desc->shadow_ptes[j] = NULL;
536 if (!prev_desc && !desc->more)
537 *rmapp = (unsigned long)desc->shadow_ptes[0];
540 prev_desc->more = desc->more;
542 *rmapp = (unsigned long)desc->more | 1;
543 mmu_free_rmap_desc(desc);
546 static void rmap_remove(struct kvm *kvm, u64 *spte)
548 struct kvm_rmap_desc *desc;
549 struct kvm_rmap_desc *prev_desc;
550 struct kvm_mmu_page *sp;
552 unsigned long *rmapp;
555 if (!is_rmap_pte(*spte))
557 sp = page_header(__pa(spte));
558 pfn = spte_to_pfn(*spte);
559 if (*spte & shadow_accessed_mask)
560 kvm_set_pfn_accessed(pfn);
561 if (is_writeble_pte(*spte))
562 kvm_release_pfn_dirty(pfn);
564 kvm_release_pfn_clean(pfn);
565 rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
567 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
569 } else if (!(*rmapp & 1)) {
570 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
571 if ((u64 *)*rmapp != spte) {
572 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
578 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
579 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
582 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
583 if (desc->shadow_ptes[i] == spte) {
584 rmap_desc_remove_entry(rmapp,
596 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
598 struct kvm_rmap_desc *desc;
599 struct kvm_rmap_desc *prev_desc;
605 else if (!(*rmapp & 1)) {
607 return (u64 *)*rmapp;
610 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
614 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
615 if (prev_spte == spte)
616 return desc->shadow_ptes[i];
617 prev_spte = desc->shadow_ptes[i];
624 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
626 unsigned long *rmapp;
628 int write_protected = 0;
630 gfn = unalias_gfn(kvm, gfn);
631 rmapp = gfn_to_rmap(kvm, gfn, 0);
633 spte = rmap_next(kvm, rmapp, NULL);
636 BUG_ON(!(*spte & PT_PRESENT_MASK));
637 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
638 if (is_writeble_pte(*spte)) {
639 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
642 spte = rmap_next(kvm, rmapp, spte);
644 if (write_protected) {
647 spte = rmap_next(kvm, rmapp, NULL);
648 pfn = spte_to_pfn(*spte);
649 kvm_set_pfn_dirty(pfn);
652 /* check for huge page mappings */
653 rmapp = gfn_to_rmap(kvm, gfn, 1);
654 spte = rmap_next(kvm, rmapp, NULL);
657 BUG_ON(!(*spte & PT_PRESENT_MASK));
658 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
659 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
660 if (is_writeble_pte(*spte)) {
661 rmap_remove(kvm, spte);
663 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
667 spte = rmap_next(kvm, rmapp, spte);
671 kvm_flush_remote_tlbs(kvm);
674 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
677 int need_tlb_flush = 0;
679 while ((spte = rmap_next(kvm, rmapp, NULL))) {
680 BUG_ON(!(*spte & PT_PRESENT_MASK));
681 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
682 rmap_remove(kvm, spte);
683 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
686 return need_tlb_flush;
689 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
690 int (*handler)(struct kvm *kvm, unsigned long *rmapp))
696 * If mmap_sem isn't taken, we can look the memslots with only
697 * the mmu_lock by skipping over the slots with userspace_addr == 0.
699 for (i = 0; i < kvm->nmemslots; i++) {
700 struct kvm_memory_slot *memslot = &kvm->memslots[i];
701 unsigned long start = memslot->userspace_addr;
704 /* mmu_lock protects userspace_addr */
708 end = start + (memslot->npages << PAGE_SHIFT);
709 if (hva >= start && hva < end) {
710 gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
711 retval |= handler(kvm, &memslot->rmap[gfn_offset]);
712 retval |= handler(kvm,
713 &memslot->lpage_info[
715 KVM_PAGES_PER_HPAGE].rmap_pde);
722 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
724 return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
727 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
732 /* always return old for EPT */
733 if (!shadow_accessed_mask)
736 spte = rmap_next(kvm, rmapp, NULL);
740 BUG_ON(!(_spte & PT_PRESENT_MASK));
741 _young = _spte & PT_ACCESSED_MASK;
744 clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
746 spte = rmap_next(kvm, rmapp, spte);
751 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
753 return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
757 static int is_empty_shadow_page(u64 *spt)
762 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
763 if (is_shadow_present_pte(*pos)) {
764 printk(KERN_ERR "%s: %p %llx\n", __func__,
772 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
774 ASSERT(is_empty_shadow_page(sp->spt));
776 __free_page(virt_to_page(sp->spt));
777 __free_page(virt_to_page(sp->gfns));
779 ++kvm->arch.n_free_mmu_pages;
782 static unsigned kvm_page_table_hashfn(gfn_t gfn)
784 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
787 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
790 struct kvm_mmu_page *sp;
792 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
793 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
794 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
795 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
796 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
797 ASSERT(is_empty_shadow_page(sp->spt));
798 bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
800 sp->parent_pte = parent_pte;
801 --vcpu->kvm->arch.n_free_mmu_pages;
805 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
806 struct kvm_mmu_page *sp, u64 *parent_pte)
808 struct kvm_pte_chain *pte_chain;
809 struct hlist_node *node;
814 if (!sp->multimapped) {
815 u64 *old = sp->parent_pte;
818 sp->parent_pte = parent_pte;
822 pte_chain = mmu_alloc_pte_chain(vcpu);
823 INIT_HLIST_HEAD(&sp->parent_ptes);
824 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
825 pte_chain->parent_ptes[0] = old;
827 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
828 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
830 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
831 if (!pte_chain->parent_ptes[i]) {
832 pte_chain->parent_ptes[i] = parent_pte;
836 pte_chain = mmu_alloc_pte_chain(vcpu);
838 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
839 pte_chain->parent_ptes[0] = parent_pte;
842 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
845 struct kvm_pte_chain *pte_chain;
846 struct hlist_node *node;
849 if (!sp->multimapped) {
850 BUG_ON(sp->parent_pte != parent_pte);
851 sp->parent_pte = NULL;
854 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
855 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
856 if (!pte_chain->parent_ptes[i])
858 if (pte_chain->parent_ptes[i] != parent_pte)
860 while (i + 1 < NR_PTE_CHAIN_ENTRIES
861 && pte_chain->parent_ptes[i + 1]) {
862 pte_chain->parent_ptes[i]
863 = pte_chain->parent_ptes[i + 1];
866 pte_chain->parent_ptes[i] = NULL;
868 hlist_del(&pte_chain->link);
869 mmu_free_pte_chain(pte_chain);
870 if (hlist_empty(&sp->parent_ptes)) {
872 sp->parent_pte = NULL;
881 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
882 mmu_parent_walk_fn fn)
884 struct kvm_pte_chain *pte_chain;
885 struct hlist_node *node;
886 struct kvm_mmu_page *parent_sp;
889 if (!sp->multimapped && sp->parent_pte) {
890 parent_sp = page_header(__pa(sp->parent_pte));
892 mmu_parent_walk(vcpu, parent_sp, fn);
895 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
896 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
897 if (!pte_chain->parent_ptes[i])
899 parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
901 mmu_parent_walk(vcpu, parent_sp, fn);
905 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
908 struct kvm_mmu_page *sp = page_header(__pa(spte));
910 index = spte - sp->spt;
911 if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
912 sp->unsync_children++;
913 WARN_ON(!sp->unsync_children);
916 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
918 struct kvm_pte_chain *pte_chain;
919 struct hlist_node *node;
925 if (!sp->multimapped) {
926 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
930 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
931 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
932 if (!pte_chain->parent_ptes[i])
934 kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
938 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
940 kvm_mmu_update_parents_unsync(sp);
944 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
945 struct kvm_mmu_page *sp)
947 mmu_parent_walk(vcpu, sp, unsync_walk_fn);
948 kvm_mmu_update_parents_unsync(sp);
951 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
952 struct kvm_mmu_page *sp)
956 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
957 sp->spt[i] = shadow_trap_nonpresent_pte;
960 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
961 struct kvm_mmu_page *sp)
966 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
970 #define KVM_PAGE_ARRAY_NR 16
972 struct kvm_mmu_pages {
973 struct mmu_page_and_offset {
974 struct kvm_mmu_page *sp;
976 } page[KVM_PAGE_ARRAY_NR];
980 #define for_each_unsync_children(bitmap, idx) \
981 for (idx = find_first_bit(bitmap, 512); \
983 idx = find_next_bit(bitmap, 512, idx+1))
985 int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
991 for (i=0; i < pvec->nr; i++)
992 if (pvec->page[i].sp == sp)
995 pvec->page[pvec->nr].sp = sp;
996 pvec->page[pvec->nr].idx = idx;
998 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1001 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1002 struct kvm_mmu_pages *pvec)
1004 int i, ret, nr_unsync_leaf = 0;
1006 for_each_unsync_children(sp->unsync_child_bitmap, i) {
1007 u64 ent = sp->spt[i];
1009 if (is_shadow_present_pte(ent)) {
1010 struct kvm_mmu_page *child;
1011 child = page_header(ent & PT64_BASE_ADDR_MASK);
1013 if (child->unsync_children) {
1014 if (mmu_pages_add(pvec, child, i))
1017 ret = __mmu_unsync_walk(child, pvec);
1019 __clear_bit(i, sp->unsync_child_bitmap);
1021 nr_unsync_leaf += ret;
1026 if (child->unsync) {
1028 if (mmu_pages_add(pvec, child, i))
1034 if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1035 sp->unsync_children = 0;
1037 return nr_unsync_leaf;
1040 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1041 struct kvm_mmu_pages *pvec)
1043 if (!sp->unsync_children)
1046 mmu_pages_add(pvec, sp, 0);
1047 return __mmu_unsync_walk(sp, pvec);
1050 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1053 struct hlist_head *bucket;
1054 struct kvm_mmu_page *sp;
1055 struct hlist_node *node;
1057 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1058 index = kvm_page_table_hashfn(gfn);
1059 bucket = &kvm->arch.mmu_page_hash[index];
1060 hlist_for_each_entry(sp, node, bucket, hash_link)
1061 if (sp->gfn == gfn && !sp->role.metaphysical
1062 && !sp->role.invalid) {
1063 pgprintk("%s: found role %x\n",
1064 __func__, sp->role.word);
1070 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1072 WARN_ON(!sp->unsync);
1074 --kvm->stat.mmu_unsync;
1077 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1079 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1081 if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1082 kvm_mmu_zap_page(vcpu->kvm, sp);
1086 rmap_write_protect(vcpu->kvm, sp->gfn);
1087 kvm_unlink_unsync_page(vcpu->kvm, sp);
1088 if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1089 kvm_mmu_zap_page(vcpu->kvm, sp);
1093 kvm_mmu_flush_tlb(vcpu);
1097 struct mmu_page_path {
1098 struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1099 unsigned int idx[PT64_ROOT_LEVEL-1];
1102 #define for_each_sp(pvec, sp, parents, i) \
1103 for (i = mmu_pages_next(&pvec, &parents, -1), \
1104 sp = pvec.page[i].sp; \
1105 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1106 i = mmu_pages_next(&pvec, &parents, i))
1108 int mmu_pages_next(struct kvm_mmu_pages *pvec, struct mmu_page_path *parents,
1113 for (n = i+1; n < pvec->nr; n++) {
1114 struct kvm_mmu_page *sp = pvec->page[n].sp;
1116 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1117 parents->idx[0] = pvec->page[n].idx;
1121 parents->parent[sp->role.level-2] = sp;
1122 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1128 void mmu_pages_clear_parents(struct mmu_page_path *parents)
1130 struct kvm_mmu_page *sp;
1131 unsigned int level = 0;
1134 unsigned int idx = parents->idx[level];
1136 sp = parents->parent[level];
1140 --sp->unsync_children;
1141 WARN_ON((int)sp->unsync_children < 0);
1142 __clear_bit(idx, sp->unsync_child_bitmap);
1144 } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1147 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1148 struct mmu_page_path *parents,
1149 struct kvm_mmu_pages *pvec)
1151 parents->parent[parent->role.level-1] = NULL;
1155 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1156 struct kvm_mmu_page *parent)
1159 struct kvm_mmu_page *sp;
1160 struct mmu_page_path parents;
1161 struct kvm_mmu_pages pages;
1163 kvm_mmu_pages_init(parent, &parents, &pages);
1164 while (mmu_unsync_walk(parent, &pages)) {
1165 for_each_sp(pages, sp, parents, i) {
1166 kvm_sync_page(vcpu, sp);
1167 mmu_pages_clear_parents(&parents);
1169 cond_resched_lock(&vcpu->kvm->mmu_lock);
1170 kvm_mmu_pages_init(parent, &parents, &pages);
1174 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1182 union kvm_mmu_page_role role;
1185 struct hlist_head *bucket;
1186 struct kvm_mmu_page *sp;
1187 struct hlist_node *node, *tmp;
1190 role.glevels = vcpu->arch.mmu.root_level;
1192 role.metaphysical = metaphysical;
1193 role.access = access;
1194 if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1195 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1196 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1197 role.quadrant = quadrant;
1199 pgprintk("%s: looking gfn %lx role %x\n", __func__,
1201 index = kvm_page_table_hashfn(gfn);
1202 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1203 hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1204 if (sp->gfn == gfn) {
1206 if (kvm_sync_page(vcpu, sp))
1209 if (sp->role.word != role.word)
1212 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1213 if (sp->unsync_children) {
1214 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1215 kvm_mmu_mark_parents_unsync(vcpu, sp);
1217 pgprintk("%s: found\n", __func__);
1220 ++vcpu->kvm->stat.mmu_cache_miss;
1221 sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1224 pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1227 hlist_add_head(&sp->hash_link, bucket);
1228 if (!metaphysical) {
1229 rmap_write_protect(vcpu->kvm, gfn);
1230 account_shadowed(vcpu->kvm, gfn);
1232 if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1233 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1235 nonpaging_prefetch_page(vcpu, sp);
1239 static int walk_shadow(struct kvm_shadow_walk *walker,
1240 struct kvm_vcpu *vcpu, u64 addr)
1248 shadow_addr = vcpu->arch.mmu.root_hpa;
1249 level = vcpu->arch.mmu.shadow_root_level;
1250 if (level == PT32E_ROOT_LEVEL) {
1251 shadow_addr = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1252 shadow_addr &= PT64_BASE_ADDR_MASK;
1256 while (level >= PT_PAGE_TABLE_LEVEL) {
1257 index = SHADOW_PT_INDEX(addr, level);
1258 sptep = ((u64 *)__va(shadow_addr)) + index;
1259 r = walker->entry(walker, vcpu, addr, sptep, level);
1262 shadow_addr = *sptep & PT64_BASE_ADDR_MASK;
1268 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1269 struct kvm_mmu_page *sp)
1277 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1278 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1279 if (is_shadow_present_pte(pt[i]))
1280 rmap_remove(kvm, &pt[i]);
1281 pt[i] = shadow_trap_nonpresent_pte;
1286 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1289 if (is_shadow_present_pte(ent)) {
1290 if (!is_large_pte(ent)) {
1291 ent &= PT64_BASE_ADDR_MASK;
1292 mmu_page_remove_parent_pte(page_header(ent),
1296 rmap_remove(kvm, &pt[i]);
1299 pt[i] = shadow_trap_nonpresent_pte;
1303 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1305 mmu_page_remove_parent_pte(sp, parent_pte);
1308 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1312 for (i = 0; i < KVM_MAX_VCPUS; ++i)
1314 kvm->vcpus[i]->arch.last_pte_updated = NULL;
1317 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1321 while (sp->multimapped || sp->parent_pte) {
1322 if (!sp->multimapped)
1323 parent_pte = sp->parent_pte;
1325 struct kvm_pte_chain *chain;
1327 chain = container_of(sp->parent_ptes.first,
1328 struct kvm_pte_chain, link);
1329 parent_pte = chain->parent_ptes[0];
1331 BUG_ON(!parent_pte);
1332 kvm_mmu_put_page(sp, parent_pte);
1333 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1337 static int mmu_zap_unsync_children(struct kvm *kvm,
1338 struct kvm_mmu_page *parent)
1341 struct mmu_page_path parents;
1342 struct kvm_mmu_pages pages;
1344 if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1347 kvm_mmu_pages_init(parent, &parents, &pages);
1348 while (mmu_unsync_walk(parent, &pages)) {
1349 struct kvm_mmu_page *sp;
1351 for_each_sp(pages, sp, parents, i) {
1352 kvm_mmu_zap_page(kvm, sp);
1353 mmu_pages_clear_parents(&parents);
1356 kvm_mmu_pages_init(parent, &parents, &pages);
1362 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1365 ++kvm->stat.mmu_shadow_zapped;
1366 ret = mmu_zap_unsync_children(kvm, sp);
1367 kvm_mmu_page_unlink_children(kvm, sp);
1368 kvm_mmu_unlink_parents(kvm, sp);
1369 kvm_flush_remote_tlbs(kvm);
1370 if (!sp->role.invalid && !sp->role.metaphysical)
1371 unaccount_shadowed(kvm, sp->gfn);
1373 kvm_unlink_unsync_page(kvm, sp);
1374 if (!sp->root_count) {
1375 hlist_del(&sp->hash_link);
1376 kvm_mmu_free_page(kvm, sp);
1378 sp->role.invalid = 1;
1379 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1380 kvm_reload_remote_mmus(kvm);
1382 kvm_mmu_reset_last_pte_updated(kvm);
1387 * Changing the number of mmu pages allocated to the vm
1388 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1390 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1393 * If we set the number of mmu pages to be smaller be than the
1394 * number of actived pages , we must to free some mmu pages before we
1398 if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1400 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1401 - kvm->arch.n_free_mmu_pages;
1403 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1404 struct kvm_mmu_page *page;
1406 page = container_of(kvm->arch.active_mmu_pages.prev,
1407 struct kvm_mmu_page, link);
1408 kvm_mmu_zap_page(kvm, page);
1411 kvm->arch.n_free_mmu_pages = 0;
1414 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1415 - kvm->arch.n_alloc_mmu_pages;
1417 kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1420 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1423 struct hlist_head *bucket;
1424 struct kvm_mmu_page *sp;
1425 struct hlist_node *node, *n;
1428 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1430 index = kvm_page_table_hashfn(gfn);
1431 bucket = &kvm->arch.mmu_page_hash[index];
1432 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1433 if (sp->gfn == gfn && !sp->role.metaphysical) {
1434 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1437 if (kvm_mmu_zap_page(kvm, sp))
1443 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1445 struct kvm_mmu_page *sp;
1447 while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1448 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1449 kvm_mmu_zap_page(kvm, sp);
1453 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1455 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1456 struct kvm_mmu_page *sp = page_header(__pa(pte));
1458 __set_bit(slot, sp->slot_bitmap);
1461 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1466 if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1469 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1470 if (pt[i] == shadow_notrap_nonpresent_pte)
1471 set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1475 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1479 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1481 if (gpa == UNMAPPED_GVA)
1484 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1490 * The function is based on mtrr_type_lookup() in
1491 * arch/x86/kernel/cpu/mtrr/generic.c
1493 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1498 u8 prev_match, curr_match;
1499 int num_var_ranges = KVM_NR_VAR_MTRR;
1501 if (!mtrr_state->enabled)
1504 /* Make end inclusive end, instead of exclusive */
1507 /* Look in fixed ranges. Just return the type as per start */
1508 if (mtrr_state->have_fixed && (start < 0x100000)) {
1511 if (start < 0x80000) {
1513 idx += (start >> 16);
1514 return mtrr_state->fixed_ranges[idx];
1515 } else if (start < 0xC0000) {
1517 idx += ((start - 0x80000) >> 14);
1518 return mtrr_state->fixed_ranges[idx];
1519 } else if (start < 0x1000000) {
1521 idx += ((start - 0xC0000) >> 12);
1522 return mtrr_state->fixed_ranges[idx];
1527 * Look in variable ranges
1528 * Look of multiple ranges matching this address and pick type
1529 * as per MTRR precedence
1531 if (!(mtrr_state->enabled & 2))
1532 return mtrr_state->def_type;
1535 for (i = 0; i < num_var_ranges; ++i) {
1536 unsigned short start_state, end_state;
1538 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1541 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1542 (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1543 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1544 (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1546 start_state = ((start & mask) == (base & mask));
1547 end_state = ((end & mask) == (base & mask));
1548 if (start_state != end_state)
1551 if ((start & mask) != (base & mask))
1554 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1555 if (prev_match == 0xFF) {
1556 prev_match = curr_match;
1560 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1561 curr_match == MTRR_TYPE_UNCACHABLE)
1562 return MTRR_TYPE_UNCACHABLE;
1564 if ((prev_match == MTRR_TYPE_WRBACK &&
1565 curr_match == MTRR_TYPE_WRTHROUGH) ||
1566 (prev_match == MTRR_TYPE_WRTHROUGH &&
1567 curr_match == MTRR_TYPE_WRBACK)) {
1568 prev_match = MTRR_TYPE_WRTHROUGH;
1569 curr_match = MTRR_TYPE_WRTHROUGH;
1572 if (prev_match != curr_match)
1573 return MTRR_TYPE_UNCACHABLE;
1576 if (prev_match != 0xFF)
1579 return mtrr_state->def_type;
1582 static u8 get_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1586 mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1587 (gfn << PAGE_SHIFT) + PAGE_SIZE);
1588 if (mtrr == 0xfe || mtrr == 0xff)
1589 mtrr = MTRR_TYPE_WRBACK;
1593 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1596 struct hlist_head *bucket;
1597 struct kvm_mmu_page *s;
1598 struct hlist_node *node, *n;
1600 index = kvm_page_table_hashfn(sp->gfn);
1601 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1602 /* don't unsync if pagetable is shadowed with multiple roles */
1603 hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1604 if (s->gfn != sp->gfn || s->role.metaphysical)
1606 if (s->role.word != sp->role.word)
1609 kvm_mmu_mark_parents_unsync(vcpu, sp);
1610 ++vcpu->kvm->stat.mmu_unsync;
1612 mmu_convert_notrap(sp);
1616 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1619 struct kvm_mmu_page *shadow;
1621 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1623 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1627 if (can_unsync && oos_shadow)
1628 return kvm_unsync_page(vcpu, shadow);
1634 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1635 unsigned pte_access, int user_fault,
1636 int write_fault, int dirty, int largepage,
1637 gfn_t gfn, pfn_t pfn, bool speculative,
1642 u64 mt_mask = shadow_mt_mask;
1645 * We don't set the accessed bit, since we sometimes want to see
1646 * whether the guest actually used the pte (in order to detect
1649 spte = shadow_base_present_pte | shadow_dirty_mask;
1651 spte |= shadow_accessed_mask;
1653 pte_access &= ~ACC_WRITE_MASK;
1654 if (pte_access & ACC_EXEC_MASK)
1655 spte |= shadow_x_mask;
1657 spte |= shadow_nx_mask;
1658 if (pte_access & ACC_USER_MASK)
1659 spte |= shadow_user_mask;
1661 spte |= PT_PAGE_SIZE_MASK;
1663 mt_mask = get_memory_type(vcpu, gfn) <<
1664 kvm_x86_ops->get_mt_mask_shift();
1668 spte |= (u64)pfn << PAGE_SHIFT;
1670 if ((pte_access & ACC_WRITE_MASK)
1671 || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1673 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1675 spte = shadow_trap_nonpresent_pte;
1679 spte |= PT_WRITABLE_MASK;
1682 * Optimization: for pte sync, if spte was writable the hash
1683 * lookup is unnecessary (and expensive). Write protection
1684 * is responsibility of mmu_get_page / kvm_sync_page.
1685 * Same reasoning can be applied to dirty page accounting.
1687 if (!can_unsync && is_writeble_pte(*shadow_pte))
1690 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1691 pgprintk("%s: found shadow page for %lx, marking ro\n",
1694 pte_access &= ~ACC_WRITE_MASK;
1695 if (is_writeble_pte(spte))
1696 spte &= ~PT_WRITABLE_MASK;
1700 if (pte_access & ACC_WRITE_MASK)
1701 mark_page_dirty(vcpu->kvm, gfn);
1704 set_shadow_pte(shadow_pte, spte);
1708 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1709 unsigned pt_access, unsigned pte_access,
1710 int user_fault, int write_fault, int dirty,
1711 int *ptwrite, int largepage, gfn_t gfn,
1712 pfn_t pfn, bool speculative)
1714 int was_rmapped = 0;
1715 int was_writeble = is_writeble_pte(*shadow_pte);
1717 pgprintk("%s: spte %llx access %x write_fault %d"
1718 " user_fault %d gfn %lx\n",
1719 __func__, *shadow_pte, pt_access,
1720 write_fault, user_fault, gfn);
1722 if (is_rmap_pte(*shadow_pte)) {
1724 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1725 * the parent of the now unreachable PTE.
1727 if (largepage && !is_large_pte(*shadow_pte)) {
1728 struct kvm_mmu_page *child;
1729 u64 pte = *shadow_pte;
1731 child = page_header(pte & PT64_BASE_ADDR_MASK);
1732 mmu_page_remove_parent_pte(child, shadow_pte);
1733 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1734 pgprintk("hfn old %lx new %lx\n",
1735 spte_to_pfn(*shadow_pte), pfn);
1736 rmap_remove(vcpu->kvm, shadow_pte);
1739 was_rmapped = is_large_pte(*shadow_pte);
1744 if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1745 dirty, largepage, gfn, pfn, speculative, true)) {
1748 kvm_x86_ops->tlb_flush(vcpu);
1751 pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1752 pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1753 is_large_pte(*shadow_pte)? "2MB" : "4kB",
1754 is_present_pte(*shadow_pte)?"RW":"R", gfn,
1755 *shadow_pte, shadow_pte);
1756 if (!was_rmapped && is_large_pte(*shadow_pte))
1757 ++vcpu->kvm->stat.lpages;
1759 page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1761 rmap_add(vcpu, shadow_pte, gfn, largepage);
1762 if (!is_rmap_pte(*shadow_pte))
1763 kvm_release_pfn_clean(pfn);
1766 kvm_release_pfn_dirty(pfn);
1768 kvm_release_pfn_clean(pfn);
1771 vcpu->arch.last_pte_updated = shadow_pte;
1772 vcpu->arch.last_pte_gfn = gfn;
1776 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1780 struct direct_shadow_walk {
1781 struct kvm_shadow_walk walker;
1788 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1789 struct kvm_vcpu *vcpu,
1790 u64 addr, u64 *sptep, int level)
1792 struct direct_shadow_walk *walk =
1793 container_of(_walk, struct direct_shadow_walk, walker);
1794 struct kvm_mmu_page *sp;
1796 gfn_t gfn = addr >> PAGE_SHIFT;
1798 if (level == PT_PAGE_TABLE_LEVEL
1799 || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1800 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1801 0, walk->write, 1, &walk->pt_write,
1802 walk->largepage, gfn, walk->pfn, false);
1803 ++vcpu->stat.pf_fixed;
1807 if (*sptep == shadow_trap_nonpresent_pte) {
1808 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1809 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1812 pgprintk("nonpaging_map: ENOMEM\n");
1813 kvm_release_pfn_clean(walk->pfn);
1817 set_shadow_pte(sptep,
1819 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1820 | shadow_user_mask | shadow_x_mask);
1825 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1826 int largepage, gfn_t gfn, pfn_t pfn)
1829 struct direct_shadow_walk walker = {
1830 .walker = { .entry = direct_map_entry, },
1832 .largepage = largepage,
1837 r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1840 return walker.pt_write;
1843 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1848 unsigned long mmu_seq;
1850 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1851 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1855 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1857 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1860 if (is_error_pfn(pfn)) {
1861 kvm_release_pfn_clean(pfn);
1865 spin_lock(&vcpu->kvm->mmu_lock);
1866 if (mmu_notifier_retry(vcpu, mmu_seq))
1868 kvm_mmu_free_some_pages(vcpu);
1869 r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1870 spin_unlock(&vcpu->kvm->mmu_lock);
1876 spin_unlock(&vcpu->kvm->mmu_lock);
1877 kvm_release_pfn_clean(pfn);
1882 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1885 struct kvm_mmu_page *sp;
1887 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1889 spin_lock(&vcpu->kvm->mmu_lock);
1890 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1891 hpa_t root = vcpu->arch.mmu.root_hpa;
1893 sp = page_header(root);
1895 if (!sp->root_count && sp->role.invalid)
1896 kvm_mmu_zap_page(vcpu->kvm, sp);
1897 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1898 spin_unlock(&vcpu->kvm->mmu_lock);
1901 for (i = 0; i < 4; ++i) {
1902 hpa_t root = vcpu->arch.mmu.pae_root[i];
1905 root &= PT64_BASE_ADDR_MASK;
1906 sp = page_header(root);
1908 if (!sp->root_count && sp->role.invalid)
1909 kvm_mmu_zap_page(vcpu->kvm, sp);
1911 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1913 spin_unlock(&vcpu->kvm->mmu_lock);
1914 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1917 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1921 struct kvm_mmu_page *sp;
1922 int metaphysical = 0;
1924 root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1926 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1927 hpa_t root = vcpu->arch.mmu.root_hpa;
1929 ASSERT(!VALID_PAGE(root));
1932 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1933 PT64_ROOT_LEVEL, metaphysical,
1935 root = __pa(sp->spt);
1937 vcpu->arch.mmu.root_hpa = root;
1940 metaphysical = !is_paging(vcpu);
1943 for (i = 0; i < 4; ++i) {
1944 hpa_t root = vcpu->arch.mmu.pae_root[i];
1946 ASSERT(!VALID_PAGE(root));
1947 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1948 if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1949 vcpu->arch.mmu.pae_root[i] = 0;
1952 root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1953 } else if (vcpu->arch.mmu.root_level == 0)
1955 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1956 PT32_ROOT_LEVEL, metaphysical,
1958 root = __pa(sp->spt);
1960 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1962 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1965 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
1968 struct kvm_mmu_page *sp;
1970 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1972 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1973 hpa_t root = vcpu->arch.mmu.root_hpa;
1974 sp = page_header(root);
1975 mmu_sync_children(vcpu, sp);
1978 for (i = 0; i < 4; ++i) {
1979 hpa_t root = vcpu->arch.mmu.pae_root[i];
1982 root &= PT64_BASE_ADDR_MASK;
1983 sp = page_header(root);
1984 mmu_sync_children(vcpu, sp);
1989 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
1991 spin_lock(&vcpu->kvm->mmu_lock);
1992 mmu_sync_roots(vcpu);
1993 spin_unlock(&vcpu->kvm->mmu_lock);
1996 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2001 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2007 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2008 r = mmu_topup_memory_caches(vcpu);
2013 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2015 gfn = gva >> PAGE_SHIFT;
2017 return nonpaging_map(vcpu, gva & PAGE_MASK,
2018 error_code & PFERR_WRITE_MASK, gfn);
2021 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2027 gfn_t gfn = gpa >> PAGE_SHIFT;
2028 unsigned long mmu_seq;
2031 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2033 r = mmu_topup_memory_caches(vcpu);
2037 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
2038 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2041 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2043 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2044 if (is_error_pfn(pfn)) {
2045 kvm_release_pfn_clean(pfn);
2048 spin_lock(&vcpu->kvm->mmu_lock);
2049 if (mmu_notifier_retry(vcpu, mmu_seq))
2051 kvm_mmu_free_some_pages(vcpu);
2052 r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2053 largepage, gfn, pfn);
2054 spin_unlock(&vcpu->kvm->mmu_lock);
2059 spin_unlock(&vcpu->kvm->mmu_lock);
2060 kvm_release_pfn_clean(pfn);
2064 static void nonpaging_free(struct kvm_vcpu *vcpu)
2066 mmu_free_roots(vcpu);
2069 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2071 struct kvm_mmu *context = &vcpu->arch.mmu;
2073 context->new_cr3 = nonpaging_new_cr3;
2074 context->page_fault = nonpaging_page_fault;
2075 context->gva_to_gpa = nonpaging_gva_to_gpa;
2076 context->free = nonpaging_free;
2077 context->prefetch_page = nonpaging_prefetch_page;
2078 context->sync_page = nonpaging_sync_page;
2079 context->invlpg = nonpaging_invlpg;
2080 context->root_level = 0;
2081 context->shadow_root_level = PT32E_ROOT_LEVEL;
2082 context->root_hpa = INVALID_PAGE;
2086 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2088 ++vcpu->stat.tlb_flush;
2089 kvm_x86_ops->tlb_flush(vcpu);
2092 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2094 pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2095 mmu_free_roots(vcpu);
2098 static void inject_page_fault(struct kvm_vcpu *vcpu,
2102 kvm_inject_page_fault(vcpu, addr, err_code);
2105 static void paging_free(struct kvm_vcpu *vcpu)
2107 nonpaging_free(vcpu);
2111 #include "paging_tmpl.h"
2115 #include "paging_tmpl.h"
2118 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2120 struct kvm_mmu *context = &vcpu->arch.mmu;
2122 ASSERT(is_pae(vcpu));
2123 context->new_cr3 = paging_new_cr3;
2124 context->page_fault = paging64_page_fault;
2125 context->gva_to_gpa = paging64_gva_to_gpa;
2126 context->prefetch_page = paging64_prefetch_page;
2127 context->sync_page = paging64_sync_page;
2128 context->invlpg = paging64_invlpg;
2129 context->free = paging_free;
2130 context->root_level = level;
2131 context->shadow_root_level = level;
2132 context->root_hpa = INVALID_PAGE;
2136 static int paging64_init_context(struct kvm_vcpu *vcpu)
2138 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2141 static int paging32_init_context(struct kvm_vcpu *vcpu)
2143 struct kvm_mmu *context = &vcpu->arch.mmu;
2145 context->new_cr3 = paging_new_cr3;
2146 context->page_fault = paging32_page_fault;
2147 context->gva_to_gpa = paging32_gva_to_gpa;
2148 context->free = paging_free;
2149 context->prefetch_page = paging32_prefetch_page;
2150 context->sync_page = paging32_sync_page;
2151 context->invlpg = paging32_invlpg;
2152 context->root_level = PT32_ROOT_LEVEL;
2153 context->shadow_root_level = PT32E_ROOT_LEVEL;
2154 context->root_hpa = INVALID_PAGE;
2158 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2160 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2163 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2165 struct kvm_mmu *context = &vcpu->arch.mmu;
2167 context->new_cr3 = nonpaging_new_cr3;
2168 context->page_fault = tdp_page_fault;
2169 context->free = nonpaging_free;
2170 context->prefetch_page = nonpaging_prefetch_page;
2171 context->sync_page = nonpaging_sync_page;
2172 context->invlpg = nonpaging_invlpg;
2173 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2174 context->root_hpa = INVALID_PAGE;
2176 if (!is_paging(vcpu)) {
2177 context->gva_to_gpa = nonpaging_gva_to_gpa;
2178 context->root_level = 0;
2179 } else if (is_long_mode(vcpu)) {
2180 context->gva_to_gpa = paging64_gva_to_gpa;
2181 context->root_level = PT64_ROOT_LEVEL;
2182 } else if (is_pae(vcpu)) {
2183 context->gva_to_gpa = paging64_gva_to_gpa;
2184 context->root_level = PT32E_ROOT_LEVEL;
2186 context->gva_to_gpa = paging32_gva_to_gpa;
2187 context->root_level = PT32_ROOT_LEVEL;
2193 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2196 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2198 if (!is_paging(vcpu))
2199 return nonpaging_init_context(vcpu);
2200 else if (is_long_mode(vcpu))
2201 return paging64_init_context(vcpu);
2202 else if (is_pae(vcpu))
2203 return paging32E_init_context(vcpu);
2205 return paging32_init_context(vcpu);
2208 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2210 vcpu->arch.update_pte.pfn = bad_pfn;
2213 return init_kvm_tdp_mmu(vcpu);
2215 return init_kvm_softmmu(vcpu);
2218 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2221 if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2222 vcpu->arch.mmu.free(vcpu);
2223 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2227 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2229 destroy_kvm_mmu(vcpu);
2230 return init_kvm_mmu(vcpu);
2232 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2234 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2238 r = mmu_topup_memory_caches(vcpu);
2241 spin_lock(&vcpu->kvm->mmu_lock);
2242 kvm_mmu_free_some_pages(vcpu);
2243 mmu_alloc_roots(vcpu);
2244 mmu_sync_roots(vcpu);
2245 spin_unlock(&vcpu->kvm->mmu_lock);
2246 kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2247 kvm_mmu_flush_tlb(vcpu);
2251 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2253 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2255 mmu_free_roots(vcpu);
2258 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2259 struct kvm_mmu_page *sp,
2263 struct kvm_mmu_page *child;
2266 if (is_shadow_present_pte(pte)) {
2267 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2269 rmap_remove(vcpu->kvm, spte);
2271 child = page_header(pte & PT64_BASE_ADDR_MASK);
2272 mmu_page_remove_parent_pte(child, spte);
2275 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2276 if (is_large_pte(pte))
2277 --vcpu->kvm->stat.lpages;
2280 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2281 struct kvm_mmu_page *sp,
2285 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2286 if (!vcpu->arch.update_pte.largepage ||
2287 sp->role.glevels == PT32_ROOT_LEVEL) {
2288 ++vcpu->kvm->stat.mmu_pde_zapped;
2293 ++vcpu->kvm->stat.mmu_pte_updated;
2294 if (sp->role.glevels == PT32_ROOT_LEVEL)
2295 paging32_update_pte(vcpu, sp, spte, new);
2297 paging64_update_pte(vcpu, sp, spte, new);
2300 static bool need_remote_flush(u64 old, u64 new)
2302 if (!is_shadow_present_pte(old))
2304 if (!is_shadow_present_pte(new))
2306 if ((old ^ new) & PT64_BASE_ADDR_MASK)
2308 old ^= PT64_NX_MASK;
2309 new ^= PT64_NX_MASK;
2310 return (old & ~new & PT64_PERM_MASK) != 0;
2313 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2315 if (need_remote_flush(old, new))
2316 kvm_flush_remote_tlbs(vcpu->kvm);
2318 kvm_mmu_flush_tlb(vcpu);
2321 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2323 u64 *spte = vcpu->arch.last_pte_updated;
2325 return !!(spte && (*spte & shadow_accessed_mask));
2328 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2329 const u8 *new, int bytes)
2336 vcpu->arch.update_pte.largepage = 0;
2338 if (bytes != 4 && bytes != 8)
2342 * Assume that the pte write on a page table of the same type
2343 * as the current vcpu paging mode. This is nearly always true
2344 * (might be false while changing modes). Note it is verified later
2348 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2349 if ((bytes == 4) && (gpa % 4 == 0)) {
2350 r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2353 memcpy((void *)&gpte + (gpa % 8), new, 4);
2354 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2355 memcpy((void *)&gpte, new, 8);
2358 if ((bytes == 4) && (gpa % 4 == 0))
2359 memcpy((void *)&gpte, new, 4);
2361 if (!is_present_pte(gpte))
2363 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2365 if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2366 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2367 vcpu->arch.update_pte.largepage = 1;
2369 vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2371 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2373 if (is_error_pfn(pfn)) {
2374 kvm_release_pfn_clean(pfn);
2377 vcpu->arch.update_pte.gfn = gfn;
2378 vcpu->arch.update_pte.pfn = pfn;
2381 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2383 u64 *spte = vcpu->arch.last_pte_updated;
2386 && vcpu->arch.last_pte_gfn == gfn
2387 && shadow_accessed_mask
2388 && !(*spte & shadow_accessed_mask)
2389 && is_shadow_present_pte(*spte))
2390 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2393 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2394 const u8 *new, int bytes)
2396 gfn_t gfn = gpa >> PAGE_SHIFT;
2397 struct kvm_mmu_page *sp;
2398 struct hlist_node *node, *n;
2399 struct hlist_head *bucket;
2403 unsigned offset = offset_in_page(gpa);
2405 unsigned page_offset;
2406 unsigned misaligned;
2413 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2414 mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2415 spin_lock(&vcpu->kvm->mmu_lock);
2416 kvm_mmu_access_page(vcpu, gfn);
2417 kvm_mmu_free_some_pages(vcpu);
2418 ++vcpu->kvm->stat.mmu_pte_write;
2419 kvm_mmu_audit(vcpu, "pre pte write");
2420 if (gfn == vcpu->arch.last_pt_write_gfn
2421 && !last_updated_pte_accessed(vcpu)) {
2422 ++vcpu->arch.last_pt_write_count;
2423 if (vcpu->arch.last_pt_write_count >= 3)
2426 vcpu->arch.last_pt_write_gfn = gfn;
2427 vcpu->arch.last_pt_write_count = 1;
2428 vcpu->arch.last_pte_updated = NULL;
2430 index = kvm_page_table_hashfn(gfn);
2431 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2432 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2433 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
2435 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2436 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2437 misaligned |= bytes < 4;
2438 if (misaligned || flooded) {
2440 * Misaligned accesses are too much trouble to fix
2441 * up; also, they usually indicate a page is not used
2444 * If we're seeing too many writes to a page,
2445 * it may no longer be a page table, or we may be
2446 * forking, in which case it is better to unmap the
2449 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2450 gpa, bytes, sp->role.word);
2451 if (kvm_mmu_zap_page(vcpu->kvm, sp))
2453 ++vcpu->kvm->stat.mmu_flooded;
2456 page_offset = offset;
2457 level = sp->role.level;
2459 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2460 page_offset <<= 1; /* 32->64 */
2462 * A 32-bit pde maps 4MB while the shadow pdes map
2463 * only 2MB. So we need to double the offset again
2464 * and zap two pdes instead of one.
2466 if (level == PT32_ROOT_LEVEL) {
2467 page_offset &= ~7; /* kill rounding error */
2471 quadrant = page_offset >> PAGE_SHIFT;
2472 page_offset &= ~PAGE_MASK;
2473 if (quadrant != sp->role.quadrant)
2476 spte = &sp->spt[page_offset / sizeof(*spte)];
2477 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2479 r = kvm_read_guest_atomic(vcpu->kvm,
2480 gpa & ~(u64)(pte_size - 1),
2482 new = (const void *)&gentry;
2488 mmu_pte_write_zap_pte(vcpu, sp, spte);
2490 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2491 mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2495 kvm_mmu_audit(vcpu, "post pte write");
2496 spin_unlock(&vcpu->kvm->mmu_lock);
2497 if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2498 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2499 vcpu->arch.update_pte.pfn = bad_pfn;
2503 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2508 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2510 spin_lock(&vcpu->kvm->mmu_lock);
2511 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2512 spin_unlock(&vcpu->kvm->mmu_lock);
2515 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2517 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2519 while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2520 struct kvm_mmu_page *sp;
2522 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2523 struct kvm_mmu_page, link);
2524 kvm_mmu_zap_page(vcpu->kvm, sp);
2525 ++vcpu->kvm->stat.mmu_recycled;
2529 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2532 enum emulation_result er;
2534 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2543 r = mmu_topup_memory_caches(vcpu);
2547 er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2552 case EMULATE_DO_MMIO:
2553 ++vcpu->stat.mmio_exits;
2556 kvm_report_emulation_failure(vcpu, "pagetable");
2564 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2566 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2568 spin_lock(&vcpu->kvm->mmu_lock);
2569 vcpu->arch.mmu.invlpg(vcpu, gva);
2570 spin_unlock(&vcpu->kvm->mmu_lock);
2571 kvm_mmu_flush_tlb(vcpu);
2572 ++vcpu->stat.invlpg;
2574 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2576 void kvm_enable_tdp(void)
2580 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2582 void kvm_disable_tdp(void)
2584 tdp_enabled = false;
2586 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2588 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2590 struct kvm_mmu_page *sp;
2592 while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2593 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2594 struct kvm_mmu_page, link);
2595 kvm_mmu_zap_page(vcpu->kvm, sp);
2598 free_page((unsigned long)vcpu->arch.mmu.pae_root);
2601 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2608 if (vcpu->kvm->arch.n_requested_mmu_pages)
2609 vcpu->kvm->arch.n_free_mmu_pages =
2610 vcpu->kvm->arch.n_requested_mmu_pages;
2612 vcpu->kvm->arch.n_free_mmu_pages =
2613 vcpu->kvm->arch.n_alloc_mmu_pages;
2615 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2616 * Therefore we need to allocate shadow page tables in the first
2617 * 4GB of memory, which happens to fit the DMA32 zone.
2619 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2622 vcpu->arch.mmu.pae_root = page_address(page);
2623 for (i = 0; i < 4; ++i)
2624 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2629 free_mmu_pages(vcpu);
2633 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2636 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2638 return alloc_mmu_pages(vcpu);
2641 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2644 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2646 return init_kvm_mmu(vcpu);
2649 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2653 destroy_kvm_mmu(vcpu);
2654 free_mmu_pages(vcpu);
2655 mmu_free_memory_caches(vcpu);
2658 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2660 struct kvm_mmu_page *sp;
2662 spin_lock(&kvm->mmu_lock);
2663 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2667 if (!test_bit(slot, sp->slot_bitmap))
2671 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2673 if (pt[i] & PT_WRITABLE_MASK)
2674 pt[i] &= ~PT_WRITABLE_MASK;
2676 kvm_flush_remote_tlbs(kvm);
2677 spin_unlock(&kvm->mmu_lock);
2680 void kvm_mmu_zap_all(struct kvm *kvm)
2682 struct kvm_mmu_page *sp, *node;
2684 spin_lock(&kvm->mmu_lock);
2685 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2686 if (kvm_mmu_zap_page(kvm, sp))
2687 node = container_of(kvm->arch.active_mmu_pages.next,
2688 struct kvm_mmu_page, link);
2689 spin_unlock(&kvm->mmu_lock);
2691 kvm_flush_remote_tlbs(kvm);
2694 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2696 struct kvm_mmu_page *page;
2698 page = container_of(kvm->arch.active_mmu_pages.prev,
2699 struct kvm_mmu_page, link);
2700 kvm_mmu_zap_page(kvm, page);
2703 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2706 struct kvm *kvm_freed = NULL;
2707 int cache_count = 0;
2709 spin_lock(&kvm_lock);
2711 list_for_each_entry(kvm, &vm_list, vm_list) {
2714 if (!down_read_trylock(&kvm->slots_lock))
2716 spin_lock(&kvm->mmu_lock);
2717 npages = kvm->arch.n_alloc_mmu_pages -
2718 kvm->arch.n_free_mmu_pages;
2719 cache_count += npages;
2720 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2721 kvm_mmu_remove_one_alloc_mmu_page(kvm);
2727 spin_unlock(&kvm->mmu_lock);
2728 up_read(&kvm->slots_lock);
2731 list_move_tail(&kvm_freed->vm_list, &vm_list);
2733 spin_unlock(&kvm_lock);
2738 static struct shrinker mmu_shrinker = {
2739 .shrink = mmu_shrink,
2740 .seeks = DEFAULT_SEEKS * 10,
2743 static void mmu_destroy_caches(void)
2745 if (pte_chain_cache)
2746 kmem_cache_destroy(pte_chain_cache);
2747 if (rmap_desc_cache)
2748 kmem_cache_destroy(rmap_desc_cache);
2749 if (mmu_page_header_cache)
2750 kmem_cache_destroy(mmu_page_header_cache);
2753 void kvm_mmu_module_exit(void)
2755 mmu_destroy_caches();
2756 unregister_shrinker(&mmu_shrinker);
2759 int kvm_mmu_module_init(void)
2761 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2762 sizeof(struct kvm_pte_chain),
2764 if (!pte_chain_cache)
2766 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2767 sizeof(struct kvm_rmap_desc),
2769 if (!rmap_desc_cache)
2772 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2773 sizeof(struct kvm_mmu_page),
2775 if (!mmu_page_header_cache)
2778 register_shrinker(&mmu_shrinker);
2783 mmu_destroy_caches();
2788 * Caculate mmu pages needed for kvm.
2790 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2793 unsigned int nr_mmu_pages;
2794 unsigned int nr_pages = 0;
2796 for (i = 0; i < kvm->nmemslots; i++)
2797 nr_pages += kvm->memslots[i].npages;
2799 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2800 nr_mmu_pages = max(nr_mmu_pages,
2801 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2803 return nr_mmu_pages;
2806 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2809 if (len > buffer->len)
2814 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2819 ret = pv_mmu_peek_buffer(buffer, len);
2824 buffer->processed += len;
2828 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2829 gpa_t addr, gpa_t value)
2834 if (!is_long_mode(vcpu) && !is_pae(vcpu))
2837 r = mmu_topup_memory_caches(vcpu);
2841 if (!emulator_write_phys(vcpu, addr, &value, bytes))
2847 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2849 kvm_x86_ops->tlb_flush(vcpu);
2850 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
2854 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2856 spin_lock(&vcpu->kvm->mmu_lock);
2857 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2858 spin_unlock(&vcpu->kvm->mmu_lock);
2862 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2863 struct kvm_pv_mmu_op_buffer *buffer)
2865 struct kvm_mmu_op_header *header;
2867 header = pv_mmu_peek_buffer(buffer, sizeof *header);
2870 switch (header->op) {
2871 case KVM_MMU_OP_WRITE_PTE: {
2872 struct kvm_mmu_op_write_pte *wpte;
2874 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2877 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2880 case KVM_MMU_OP_FLUSH_TLB: {
2881 struct kvm_mmu_op_flush_tlb *ftlb;
2883 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2886 return kvm_pv_mmu_flush_tlb(vcpu);
2888 case KVM_MMU_OP_RELEASE_PT: {
2889 struct kvm_mmu_op_release_pt *rpt;
2891 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2894 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2900 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2901 gpa_t addr, unsigned long *ret)
2904 struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2906 buffer->ptr = buffer->buf;
2907 buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
2908 buffer->processed = 0;
2910 r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
2914 while (buffer->len) {
2915 r = kvm_pv_mmu_op_one(vcpu, buffer);
2924 *ret = buffer->processed;
2930 static const char *audit_msg;
2932 static gva_t canonicalize(gva_t gva)
2934 #ifdef CONFIG_X86_64
2935 gva = (long long)(gva << 16) >> 16;
2940 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2941 gva_t va, int level)
2943 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2945 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2947 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2950 if (ent == shadow_trap_nonpresent_pte)
2953 va = canonicalize(va);
2955 if (ent == shadow_notrap_nonpresent_pte)
2956 printk(KERN_ERR "audit: (%s) nontrapping pte"
2957 " in nonleaf level: levels %d gva %lx"
2958 " level %d pte %llx\n", audit_msg,
2959 vcpu->arch.mmu.root_level, va, level, ent);
2961 audit_mappings_page(vcpu, ent, va, level - 1);
2963 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2964 hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2966 if (is_shadow_present_pte(ent)
2967 && (ent & PT64_BASE_ADDR_MASK) != hpa)
2968 printk(KERN_ERR "xx audit error: (%s) levels %d"
2969 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2970 audit_msg, vcpu->arch.mmu.root_level,
2972 is_shadow_present_pte(ent));
2973 else if (ent == shadow_notrap_nonpresent_pte
2974 && !is_error_hpa(hpa))
2975 printk(KERN_ERR "audit: (%s) notrap shadow,"
2976 " valid guest gva %lx\n", audit_msg, va);
2977 kvm_release_pfn_clean(pfn);
2983 static void audit_mappings(struct kvm_vcpu *vcpu)
2987 if (vcpu->arch.mmu.root_level == 4)
2988 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2990 for (i = 0; i < 4; ++i)
2991 if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2992 audit_mappings_page(vcpu,
2993 vcpu->arch.mmu.pae_root[i],
2998 static int count_rmaps(struct kvm_vcpu *vcpu)
3003 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3004 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3005 struct kvm_rmap_desc *d;
3007 for (j = 0; j < m->npages; ++j) {
3008 unsigned long *rmapp = &m->rmap[j];
3012 if (!(*rmapp & 1)) {
3016 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3018 for (k = 0; k < RMAP_EXT; ++k)
3019 if (d->shadow_ptes[k])
3030 static int count_writable_mappings(struct kvm_vcpu *vcpu)
3033 struct kvm_mmu_page *sp;
3036 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3039 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3042 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3045 if (!(ent & PT_PRESENT_MASK))
3047 if (!(ent & PT_WRITABLE_MASK))
3055 static void audit_rmap(struct kvm_vcpu *vcpu)
3057 int n_rmap = count_rmaps(vcpu);
3058 int n_actual = count_writable_mappings(vcpu);
3060 if (n_rmap != n_actual)
3061 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
3062 __func__, audit_msg, n_rmap, n_actual);
3065 static void audit_write_protection(struct kvm_vcpu *vcpu)
3067 struct kvm_mmu_page *sp;
3068 struct kvm_memory_slot *slot;
3069 unsigned long *rmapp;
3072 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3073 if (sp->role.metaphysical)
3076 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3077 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3078 rmapp = &slot->rmap[gfn - slot->base_gfn];
3080 printk(KERN_ERR "%s: (%s) shadow page has writable"
3081 " mappings: gfn %lx role %x\n",
3082 __func__, audit_msg, sp->gfn,
3087 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3094 audit_write_protection(vcpu);
3095 audit_mappings(vcpu);