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 __set_bit(index, sp->unsync_child_bitmap);
912 sp->unsync_children = 1;
915 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
917 struct kvm_pte_chain *pte_chain;
918 struct hlist_node *node;
924 if (!sp->multimapped) {
925 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
929 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
930 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
931 if (!pte_chain->parent_ptes[i])
933 kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
937 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
939 sp->unsync_children = 1;
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 for_each_unsync_children(bitmap, idx) \
971 for (idx = find_first_bit(bitmap, 512); \
973 idx = find_next_bit(bitmap, 512, idx+1))
975 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
976 struct kvm_unsync_walk *walker)
980 if (!sp->unsync_children)
983 for_each_unsync_children(sp->unsync_child_bitmap, i) {
984 u64 ent = sp->spt[i];
986 if (is_shadow_present_pte(ent)) {
987 struct kvm_mmu_page *child;
988 child = page_header(ent & PT64_BASE_ADDR_MASK);
990 if (child->unsync_children) {
991 ret = mmu_unsync_walk(child, walker);
994 __clear_bit(i, sp->unsync_child_bitmap);
998 ret = walker->entry(child, walker);
999 __clear_bit(i, sp->unsync_child_bitmap);
1006 if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1007 sp->unsync_children = 0;
1012 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1015 struct hlist_head *bucket;
1016 struct kvm_mmu_page *sp;
1017 struct hlist_node *node;
1019 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1020 index = kvm_page_table_hashfn(gfn);
1021 bucket = &kvm->arch.mmu_page_hash[index];
1022 hlist_for_each_entry(sp, node, bucket, hash_link)
1023 if (sp->gfn == gfn && !sp->role.metaphysical
1024 && !sp->role.invalid) {
1025 pgprintk("%s: found role %x\n",
1026 __func__, sp->role.word);
1032 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1034 WARN_ON(!sp->unsync);
1036 --kvm->stat.mmu_unsync;
1039 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1041 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1043 if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1044 kvm_mmu_zap_page(vcpu->kvm, sp);
1048 rmap_write_protect(vcpu->kvm, sp->gfn);
1049 kvm_unlink_unsync_page(vcpu->kvm, sp);
1050 if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1051 kvm_mmu_zap_page(vcpu->kvm, sp);
1055 kvm_mmu_flush_tlb(vcpu);
1059 struct sync_walker {
1060 struct kvm_vcpu *vcpu;
1061 struct kvm_unsync_walk walker;
1064 static int mmu_sync_fn(struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk)
1066 struct sync_walker *sync_walk = container_of(walk, struct sync_walker,
1068 struct kvm_vcpu *vcpu = sync_walk->vcpu;
1070 kvm_sync_page(vcpu, sp);
1071 return (need_resched() || spin_needbreak(&vcpu->kvm->mmu_lock));
1074 static void mmu_sync_children(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1076 struct sync_walker walker = {
1077 .walker = { .entry = mmu_sync_fn, },
1081 while (mmu_unsync_walk(sp, &walker.walker))
1082 cond_resched_lock(&vcpu->kvm->mmu_lock);
1085 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1093 union kvm_mmu_page_role role;
1096 struct hlist_head *bucket;
1097 struct kvm_mmu_page *sp;
1098 struct hlist_node *node, *tmp;
1101 role.glevels = vcpu->arch.mmu.root_level;
1103 role.metaphysical = metaphysical;
1104 role.access = access;
1105 if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1106 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1107 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1108 role.quadrant = quadrant;
1110 pgprintk("%s: looking gfn %lx role %x\n", __func__,
1112 index = kvm_page_table_hashfn(gfn);
1113 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1114 hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1115 if (sp->gfn == gfn) {
1117 if (kvm_sync_page(vcpu, sp))
1120 if (sp->role.word != role.word)
1123 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1124 if (sp->unsync_children) {
1125 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1126 kvm_mmu_mark_parents_unsync(vcpu, sp);
1128 pgprintk("%s: found\n", __func__);
1131 ++vcpu->kvm->stat.mmu_cache_miss;
1132 sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1135 pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1138 hlist_add_head(&sp->hash_link, bucket);
1139 if (!metaphysical) {
1140 rmap_write_protect(vcpu->kvm, gfn);
1141 account_shadowed(vcpu->kvm, gfn);
1143 if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1144 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1146 nonpaging_prefetch_page(vcpu, sp);
1150 static int walk_shadow(struct kvm_shadow_walk *walker,
1151 struct kvm_vcpu *vcpu, u64 addr)
1159 shadow_addr = vcpu->arch.mmu.root_hpa;
1160 level = vcpu->arch.mmu.shadow_root_level;
1161 if (level == PT32E_ROOT_LEVEL) {
1162 shadow_addr = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1163 shadow_addr &= PT64_BASE_ADDR_MASK;
1167 while (level >= PT_PAGE_TABLE_LEVEL) {
1168 index = SHADOW_PT_INDEX(addr, level);
1169 sptep = ((u64 *)__va(shadow_addr)) + index;
1170 r = walker->entry(walker, vcpu, addr, sptep, level);
1173 shadow_addr = *sptep & PT64_BASE_ADDR_MASK;
1179 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1180 struct kvm_mmu_page *sp)
1188 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1189 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1190 if (is_shadow_present_pte(pt[i]))
1191 rmap_remove(kvm, &pt[i]);
1192 pt[i] = shadow_trap_nonpresent_pte;
1197 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1200 if (is_shadow_present_pte(ent)) {
1201 if (!is_large_pte(ent)) {
1202 ent &= PT64_BASE_ADDR_MASK;
1203 mmu_page_remove_parent_pte(page_header(ent),
1207 rmap_remove(kvm, &pt[i]);
1210 pt[i] = shadow_trap_nonpresent_pte;
1214 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1216 mmu_page_remove_parent_pte(sp, parent_pte);
1219 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1223 for (i = 0; i < KVM_MAX_VCPUS; ++i)
1225 kvm->vcpus[i]->arch.last_pte_updated = NULL;
1228 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1232 while (sp->multimapped || sp->parent_pte) {
1233 if (!sp->multimapped)
1234 parent_pte = sp->parent_pte;
1236 struct kvm_pte_chain *chain;
1238 chain = container_of(sp->parent_ptes.first,
1239 struct kvm_pte_chain, link);
1240 parent_pte = chain->parent_ptes[0];
1242 BUG_ON(!parent_pte);
1243 kvm_mmu_put_page(sp, parent_pte);
1244 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1249 struct kvm_unsync_walk walker;
1254 static int mmu_zap_fn(struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk)
1256 struct zap_walker *zap_walk = container_of(walk, struct zap_walker,
1258 kvm_mmu_zap_page(zap_walk->kvm, sp);
1259 zap_walk->zapped = 1;
1263 static int mmu_zap_unsync_children(struct kvm *kvm, struct kvm_mmu_page *sp)
1265 struct zap_walker walker = {
1266 .walker = { .entry = mmu_zap_fn, },
1271 if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1273 mmu_unsync_walk(sp, &walker.walker);
1274 return walker.zapped;
1277 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1280 ++kvm->stat.mmu_shadow_zapped;
1281 ret = mmu_zap_unsync_children(kvm, sp);
1282 kvm_mmu_page_unlink_children(kvm, sp);
1283 kvm_mmu_unlink_parents(kvm, sp);
1284 kvm_flush_remote_tlbs(kvm);
1285 if (!sp->role.invalid && !sp->role.metaphysical)
1286 unaccount_shadowed(kvm, sp->gfn);
1288 kvm_unlink_unsync_page(kvm, sp);
1289 if (!sp->root_count) {
1290 hlist_del(&sp->hash_link);
1291 kvm_mmu_free_page(kvm, sp);
1293 sp->role.invalid = 1;
1294 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1295 kvm_reload_remote_mmus(kvm);
1297 kvm_mmu_reset_last_pte_updated(kvm);
1302 * Changing the number of mmu pages allocated to the vm
1303 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1305 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1308 * If we set the number of mmu pages to be smaller be than the
1309 * number of actived pages , we must to free some mmu pages before we
1313 if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1315 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1316 - kvm->arch.n_free_mmu_pages;
1318 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1319 struct kvm_mmu_page *page;
1321 page = container_of(kvm->arch.active_mmu_pages.prev,
1322 struct kvm_mmu_page, link);
1323 kvm_mmu_zap_page(kvm, page);
1326 kvm->arch.n_free_mmu_pages = 0;
1329 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1330 - kvm->arch.n_alloc_mmu_pages;
1332 kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1335 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1338 struct hlist_head *bucket;
1339 struct kvm_mmu_page *sp;
1340 struct hlist_node *node, *n;
1343 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1345 index = kvm_page_table_hashfn(gfn);
1346 bucket = &kvm->arch.mmu_page_hash[index];
1347 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1348 if (sp->gfn == gfn && !sp->role.metaphysical) {
1349 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1352 if (kvm_mmu_zap_page(kvm, sp))
1358 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1360 struct kvm_mmu_page *sp;
1362 while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1363 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1364 kvm_mmu_zap_page(kvm, sp);
1368 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1370 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1371 struct kvm_mmu_page *sp = page_header(__pa(pte));
1373 __set_bit(slot, sp->slot_bitmap);
1376 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1381 if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1384 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1385 if (pt[i] == shadow_notrap_nonpresent_pte)
1386 set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1390 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1394 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1396 if (gpa == UNMAPPED_GVA)
1399 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1405 * The function is based on mtrr_type_lookup() in
1406 * arch/x86/kernel/cpu/mtrr/generic.c
1408 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1413 u8 prev_match, curr_match;
1414 int num_var_ranges = KVM_NR_VAR_MTRR;
1416 if (!mtrr_state->enabled)
1419 /* Make end inclusive end, instead of exclusive */
1422 /* Look in fixed ranges. Just return the type as per start */
1423 if (mtrr_state->have_fixed && (start < 0x100000)) {
1426 if (start < 0x80000) {
1428 idx += (start >> 16);
1429 return mtrr_state->fixed_ranges[idx];
1430 } else if (start < 0xC0000) {
1432 idx += ((start - 0x80000) >> 14);
1433 return mtrr_state->fixed_ranges[idx];
1434 } else if (start < 0x1000000) {
1436 idx += ((start - 0xC0000) >> 12);
1437 return mtrr_state->fixed_ranges[idx];
1442 * Look in variable ranges
1443 * Look of multiple ranges matching this address and pick type
1444 * as per MTRR precedence
1446 if (!(mtrr_state->enabled & 2))
1447 return mtrr_state->def_type;
1450 for (i = 0; i < num_var_ranges; ++i) {
1451 unsigned short start_state, end_state;
1453 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1456 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1457 (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1458 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1459 (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1461 start_state = ((start & mask) == (base & mask));
1462 end_state = ((end & mask) == (base & mask));
1463 if (start_state != end_state)
1466 if ((start & mask) != (base & mask))
1469 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1470 if (prev_match == 0xFF) {
1471 prev_match = curr_match;
1475 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1476 curr_match == MTRR_TYPE_UNCACHABLE)
1477 return MTRR_TYPE_UNCACHABLE;
1479 if ((prev_match == MTRR_TYPE_WRBACK &&
1480 curr_match == MTRR_TYPE_WRTHROUGH) ||
1481 (prev_match == MTRR_TYPE_WRTHROUGH &&
1482 curr_match == MTRR_TYPE_WRBACK)) {
1483 prev_match = MTRR_TYPE_WRTHROUGH;
1484 curr_match = MTRR_TYPE_WRTHROUGH;
1487 if (prev_match != curr_match)
1488 return MTRR_TYPE_UNCACHABLE;
1491 if (prev_match != 0xFF)
1494 return mtrr_state->def_type;
1497 static u8 get_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1501 mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1502 (gfn << PAGE_SHIFT) + PAGE_SIZE);
1503 if (mtrr == 0xfe || mtrr == 0xff)
1504 mtrr = MTRR_TYPE_WRBACK;
1508 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1511 struct hlist_head *bucket;
1512 struct kvm_mmu_page *s;
1513 struct hlist_node *node, *n;
1515 index = kvm_page_table_hashfn(sp->gfn);
1516 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1517 /* don't unsync if pagetable is shadowed with multiple roles */
1518 hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1519 if (s->gfn != sp->gfn || s->role.metaphysical)
1521 if (s->role.word != sp->role.word)
1524 kvm_mmu_mark_parents_unsync(vcpu, sp);
1525 ++vcpu->kvm->stat.mmu_unsync;
1527 mmu_convert_notrap(sp);
1531 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1534 struct kvm_mmu_page *shadow;
1536 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1538 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1542 if (can_unsync && oos_shadow)
1543 return kvm_unsync_page(vcpu, shadow);
1549 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1550 unsigned pte_access, int user_fault,
1551 int write_fault, int dirty, int largepage,
1552 gfn_t gfn, pfn_t pfn, bool speculative,
1557 u64 mt_mask = shadow_mt_mask;
1560 * We don't set the accessed bit, since we sometimes want to see
1561 * whether the guest actually used the pte (in order to detect
1564 spte = shadow_base_present_pte | shadow_dirty_mask;
1566 spte |= shadow_accessed_mask;
1568 pte_access &= ~ACC_WRITE_MASK;
1569 if (pte_access & ACC_EXEC_MASK)
1570 spte |= shadow_x_mask;
1572 spte |= shadow_nx_mask;
1573 if (pte_access & ACC_USER_MASK)
1574 spte |= shadow_user_mask;
1576 spte |= PT_PAGE_SIZE_MASK;
1578 mt_mask = get_memory_type(vcpu, gfn) <<
1579 kvm_x86_ops->get_mt_mask_shift();
1583 spte |= (u64)pfn << PAGE_SHIFT;
1585 if ((pte_access & ACC_WRITE_MASK)
1586 || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1588 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1590 spte = shadow_trap_nonpresent_pte;
1594 spte |= PT_WRITABLE_MASK;
1596 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1597 pgprintk("%s: found shadow page for %lx, marking ro\n",
1600 pte_access &= ~ACC_WRITE_MASK;
1601 if (is_writeble_pte(spte))
1602 spte &= ~PT_WRITABLE_MASK;
1606 if (pte_access & ACC_WRITE_MASK)
1607 mark_page_dirty(vcpu->kvm, gfn);
1610 set_shadow_pte(shadow_pte, spte);
1614 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1615 unsigned pt_access, unsigned pte_access,
1616 int user_fault, int write_fault, int dirty,
1617 int *ptwrite, int largepage, gfn_t gfn,
1618 pfn_t pfn, bool speculative)
1620 int was_rmapped = 0;
1621 int was_writeble = is_writeble_pte(*shadow_pte);
1623 pgprintk("%s: spte %llx access %x write_fault %d"
1624 " user_fault %d gfn %lx\n",
1625 __func__, *shadow_pte, pt_access,
1626 write_fault, user_fault, gfn);
1628 if (is_rmap_pte(*shadow_pte)) {
1630 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1631 * the parent of the now unreachable PTE.
1633 if (largepage && !is_large_pte(*shadow_pte)) {
1634 struct kvm_mmu_page *child;
1635 u64 pte = *shadow_pte;
1637 child = page_header(pte & PT64_BASE_ADDR_MASK);
1638 mmu_page_remove_parent_pte(child, shadow_pte);
1639 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1640 pgprintk("hfn old %lx new %lx\n",
1641 spte_to_pfn(*shadow_pte), pfn);
1642 rmap_remove(vcpu->kvm, shadow_pte);
1645 was_rmapped = is_large_pte(*shadow_pte);
1650 if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1651 dirty, largepage, gfn, pfn, speculative, true)) {
1654 kvm_x86_ops->tlb_flush(vcpu);
1657 pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1658 pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1659 is_large_pte(*shadow_pte)? "2MB" : "4kB",
1660 is_present_pte(*shadow_pte)?"RW":"R", gfn,
1661 *shadow_pte, shadow_pte);
1662 if (!was_rmapped && is_large_pte(*shadow_pte))
1663 ++vcpu->kvm->stat.lpages;
1665 page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1667 rmap_add(vcpu, shadow_pte, gfn, largepage);
1668 if (!is_rmap_pte(*shadow_pte))
1669 kvm_release_pfn_clean(pfn);
1672 kvm_release_pfn_dirty(pfn);
1674 kvm_release_pfn_clean(pfn);
1677 vcpu->arch.last_pte_updated = shadow_pte;
1678 vcpu->arch.last_pte_gfn = gfn;
1682 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1686 struct direct_shadow_walk {
1687 struct kvm_shadow_walk walker;
1694 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1695 struct kvm_vcpu *vcpu,
1696 u64 addr, u64 *sptep, int level)
1698 struct direct_shadow_walk *walk =
1699 container_of(_walk, struct direct_shadow_walk, walker);
1700 struct kvm_mmu_page *sp;
1702 gfn_t gfn = addr >> PAGE_SHIFT;
1704 if (level == PT_PAGE_TABLE_LEVEL
1705 || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1706 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1707 0, walk->write, 1, &walk->pt_write,
1708 walk->largepage, gfn, walk->pfn, false);
1709 ++vcpu->stat.pf_fixed;
1713 if (*sptep == shadow_trap_nonpresent_pte) {
1714 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1715 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1718 pgprintk("nonpaging_map: ENOMEM\n");
1719 kvm_release_pfn_clean(walk->pfn);
1723 set_shadow_pte(sptep,
1725 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1726 | shadow_user_mask | shadow_x_mask);
1731 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1732 int largepage, gfn_t gfn, pfn_t pfn)
1735 struct direct_shadow_walk walker = {
1736 .walker = { .entry = direct_map_entry, },
1738 .largepage = largepage,
1743 r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1746 return walker.pt_write;
1749 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1754 unsigned long mmu_seq;
1756 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1757 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1761 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1763 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1766 if (is_error_pfn(pfn)) {
1767 kvm_release_pfn_clean(pfn);
1771 spin_lock(&vcpu->kvm->mmu_lock);
1772 if (mmu_notifier_retry(vcpu, mmu_seq))
1774 kvm_mmu_free_some_pages(vcpu);
1775 r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1776 spin_unlock(&vcpu->kvm->mmu_lock);
1782 spin_unlock(&vcpu->kvm->mmu_lock);
1783 kvm_release_pfn_clean(pfn);
1788 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1791 struct kvm_mmu_page *sp;
1793 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1795 spin_lock(&vcpu->kvm->mmu_lock);
1796 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1797 hpa_t root = vcpu->arch.mmu.root_hpa;
1799 sp = page_header(root);
1801 if (!sp->root_count && sp->role.invalid)
1802 kvm_mmu_zap_page(vcpu->kvm, sp);
1803 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1804 spin_unlock(&vcpu->kvm->mmu_lock);
1807 for (i = 0; i < 4; ++i) {
1808 hpa_t root = vcpu->arch.mmu.pae_root[i];
1811 root &= PT64_BASE_ADDR_MASK;
1812 sp = page_header(root);
1814 if (!sp->root_count && sp->role.invalid)
1815 kvm_mmu_zap_page(vcpu->kvm, sp);
1817 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1819 spin_unlock(&vcpu->kvm->mmu_lock);
1820 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1823 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1827 struct kvm_mmu_page *sp;
1828 int metaphysical = 0;
1830 root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1832 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1833 hpa_t root = vcpu->arch.mmu.root_hpa;
1835 ASSERT(!VALID_PAGE(root));
1838 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1839 PT64_ROOT_LEVEL, metaphysical,
1841 root = __pa(sp->spt);
1843 vcpu->arch.mmu.root_hpa = root;
1846 metaphysical = !is_paging(vcpu);
1849 for (i = 0; i < 4; ++i) {
1850 hpa_t root = vcpu->arch.mmu.pae_root[i];
1852 ASSERT(!VALID_PAGE(root));
1853 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1854 if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1855 vcpu->arch.mmu.pae_root[i] = 0;
1858 root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1859 } else if (vcpu->arch.mmu.root_level == 0)
1861 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1862 PT32_ROOT_LEVEL, metaphysical,
1864 root = __pa(sp->spt);
1866 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1868 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1871 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
1874 struct kvm_mmu_page *sp;
1876 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1878 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1879 hpa_t root = vcpu->arch.mmu.root_hpa;
1880 sp = page_header(root);
1881 mmu_sync_children(vcpu, sp);
1884 for (i = 0; i < 4; ++i) {
1885 hpa_t root = vcpu->arch.mmu.pae_root[i];
1888 root &= PT64_BASE_ADDR_MASK;
1889 sp = page_header(root);
1890 mmu_sync_children(vcpu, sp);
1895 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
1897 spin_lock(&vcpu->kvm->mmu_lock);
1898 mmu_sync_roots(vcpu);
1899 spin_unlock(&vcpu->kvm->mmu_lock);
1902 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1907 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1913 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1914 r = mmu_topup_memory_caches(vcpu);
1919 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1921 gfn = gva >> PAGE_SHIFT;
1923 return nonpaging_map(vcpu, gva & PAGE_MASK,
1924 error_code & PFERR_WRITE_MASK, gfn);
1927 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1933 gfn_t gfn = gpa >> PAGE_SHIFT;
1934 unsigned long mmu_seq;
1937 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1939 r = mmu_topup_memory_caches(vcpu);
1943 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1944 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1947 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1949 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1950 if (is_error_pfn(pfn)) {
1951 kvm_release_pfn_clean(pfn);
1954 spin_lock(&vcpu->kvm->mmu_lock);
1955 if (mmu_notifier_retry(vcpu, mmu_seq))
1957 kvm_mmu_free_some_pages(vcpu);
1958 r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1959 largepage, gfn, pfn);
1960 spin_unlock(&vcpu->kvm->mmu_lock);
1965 spin_unlock(&vcpu->kvm->mmu_lock);
1966 kvm_release_pfn_clean(pfn);
1970 static void nonpaging_free(struct kvm_vcpu *vcpu)
1972 mmu_free_roots(vcpu);
1975 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1977 struct kvm_mmu *context = &vcpu->arch.mmu;
1979 context->new_cr3 = nonpaging_new_cr3;
1980 context->page_fault = nonpaging_page_fault;
1981 context->gva_to_gpa = nonpaging_gva_to_gpa;
1982 context->free = nonpaging_free;
1983 context->prefetch_page = nonpaging_prefetch_page;
1984 context->sync_page = nonpaging_sync_page;
1985 context->invlpg = nonpaging_invlpg;
1986 context->root_level = 0;
1987 context->shadow_root_level = PT32E_ROOT_LEVEL;
1988 context->root_hpa = INVALID_PAGE;
1992 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1994 ++vcpu->stat.tlb_flush;
1995 kvm_x86_ops->tlb_flush(vcpu);
1998 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2000 pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2001 mmu_free_roots(vcpu);
2004 static void inject_page_fault(struct kvm_vcpu *vcpu,
2008 kvm_inject_page_fault(vcpu, addr, err_code);
2011 static void paging_free(struct kvm_vcpu *vcpu)
2013 nonpaging_free(vcpu);
2017 #include "paging_tmpl.h"
2021 #include "paging_tmpl.h"
2024 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2026 struct kvm_mmu *context = &vcpu->arch.mmu;
2028 ASSERT(is_pae(vcpu));
2029 context->new_cr3 = paging_new_cr3;
2030 context->page_fault = paging64_page_fault;
2031 context->gva_to_gpa = paging64_gva_to_gpa;
2032 context->prefetch_page = paging64_prefetch_page;
2033 context->sync_page = paging64_sync_page;
2034 context->invlpg = paging64_invlpg;
2035 context->free = paging_free;
2036 context->root_level = level;
2037 context->shadow_root_level = level;
2038 context->root_hpa = INVALID_PAGE;
2042 static int paging64_init_context(struct kvm_vcpu *vcpu)
2044 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2047 static int paging32_init_context(struct kvm_vcpu *vcpu)
2049 struct kvm_mmu *context = &vcpu->arch.mmu;
2051 context->new_cr3 = paging_new_cr3;
2052 context->page_fault = paging32_page_fault;
2053 context->gva_to_gpa = paging32_gva_to_gpa;
2054 context->free = paging_free;
2055 context->prefetch_page = paging32_prefetch_page;
2056 context->sync_page = paging32_sync_page;
2057 context->invlpg = paging32_invlpg;
2058 context->root_level = PT32_ROOT_LEVEL;
2059 context->shadow_root_level = PT32E_ROOT_LEVEL;
2060 context->root_hpa = INVALID_PAGE;
2064 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2066 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2069 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2071 struct kvm_mmu *context = &vcpu->arch.mmu;
2073 context->new_cr3 = nonpaging_new_cr3;
2074 context->page_fault = tdp_page_fault;
2075 context->free = nonpaging_free;
2076 context->prefetch_page = nonpaging_prefetch_page;
2077 context->sync_page = nonpaging_sync_page;
2078 context->invlpg = nonpaging_invlpg;
2079 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2080 context->root_hpa = INVALID_PAGE;
2082 if (!is_paging(vcpu)) {
2083 context->gva_to_gpa = nonpaging_gva_to_gpa;
2084 context->root_level = 0;
2085 } else if (is_long_mode(vcpu)) {
2086 context->gva_to_gpa = paging64_gva_to_gpa;
2087 context->root_level = PT64_ROOT_LEVEL;
2088 } else if (is_pae(vcpu)) {
2089 context->gva_to_gpa = paging64_gva_to_gpa;
2090 context->root_level = PT32E_ROOT_LEVEL;
2092 context->gva_to_gpa = paging32_gva_to_gpa;
2093 context->root_level = PT32_ROOT_LEVEL;
2099 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2102 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2104 if (!is_paging(vcpu))
2105 return nonpaging_init_context(vcpu);
2106 else if (is_long_mode(vcpu))
2107 return paging64_init_context(vcpu);
2108 else if (is_pae(vcpu))
2109 return paging32E_init_context(vcpu);
2111 return paging32_init_context(vcpu);
2114 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2116 vcpu->arch.update_pte.pfn = bad_pfn;
2119 return init_kvm_tdp_mmu(vcpu);
2121 return init_kvm_softmmu(vcpu);
2124 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2127 if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2128 vcpu->arch.mmu.free(vcpu);
2129 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2133 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2135 destroy_kvm_mmu(vcpu);
2136 return init_kvm_mmu(vcpu);
2138 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2140 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2144 r = mmu_topup_memory_caches(vcpu);
2147 spin_lock(&vcpu->kvm->mmu_lock);
2148 kvm_mmu_free_some_pages(vcpu);
2149 mmu_alloc_roots(vcpu);
2150 mmu_sync_roots(vcpu);
2151 spin_unlock(&vcpu->kvm->mmu_lock);
2152 kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2153 kvm_mmu_flush_tlb(vcpu);
2157 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2159 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2161 mmu_free_roots(vcpu);
2164 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2165 struct kvm_mmu_page *sp,
2169 struct kvm_mmu_page *child;
2172 if (is_shadow_present_pte(pte)) {
2173 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2175 rmap_remove(vcpu->kvm, spte);
2177 child = page_header(pte & PT64_BASE_ADDR_MASK);
2178 mmu_page_remove_parent_pte(child, spte);
2181 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2182 if (is_large_pte(pte))
2183 --vcpu->kvm->stat.lpages;
2186 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2187 struct kvm_mmu_page *sp,
2191 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2192 if (!vcpu->arch.update_pte.largepage ||
2193 sp->role.glevels == PT32_ROOT_LEVEL) {
2194 ++vcpu->kvm->stat.mmu_pde_zapped;
2199 ++vcpu->kvm->stat.mmu_pte_updated;
2200 if (sp->role.glevels == PT32_ROOT_LEVEL)
2201 paging32_update_pte(vcpu, sp, spte, new);
2203 paging64_update_pte(vcpu, sp, spte, new);
2206 static bool need_remote_flush(u64 old, u64 new)
2208 if (!is_shadow_present_pte(old))
2210 if (!is_shadow_present_pte(new))
2212 if ((old ^ new) & PT64_BASE_ADDR_MASK)
2214 old ^= PT64_NX_MASK;
2215 new ^= PT64_NX_MASK;
2216 return (old & ~new & PT64_PERM_MASK) != 0;
2219 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2221 if (need_remote_flush(old, new))
2222 kvm_flush_remote_tlbs(vcpu->kvm);
2224 kvm_mmu_flush_tlb(vcpu);
2227 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2229 u64 *spte = vcpu->arch.last_pte_updated;
2231 return !!(spte && (*spte & shadow_accessed_mask));
2234 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2235 const u8 *new, int bytes)
2242 vcpu->arch.update_pte.largepage = 0;
2244 if (bytes != 4 && bytes != 8)
2248 * Assume that the pte write on a page table of the same type
2249 * as the current vcpu paging mode. This is nearly always true
2250 * (might be false while changing modes). Note it is verified later
2254 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2255 if ((bytes == 4) && (gpa % 4 == 0)) {
2256 r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2259 memcpy((void *)&gpte + (gpa % 8), new, 4);
2260 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2261 memcpy((void *)&gpte, new, 8);
2264 if ((bytes == 4) && (gpa % 4 == 0))
2265 memcpy((void *)&gpte, new, 4);
2267 if (!is_present_pte(gpte))
2269 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2271 if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2272 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2273 vcpu->arch.update_pte.largepage = 1;
2275 vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2277 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2279 if (is_error_pfn(pfn)) {
2280 kvm_release_pfn_clean(pfn);
2283 vcpu->arch.update_pte.gfn = gfn;
2284 vcpu->arch.update_pte.pfn = pfn;
2287 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2289 u64 *spte = vcpu->arch.last_pte_updated;
2292 && vcpu->arch.last_pte_gfn == gfn
2293 && shadow_accessed_mask
2294 && !(*spte & shadow_accessed_mask)
2295 && is_shadow_present_pte(*spte))
2296 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2299 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2300 const u8 *new, int bytes)
2302 gfn_t gfn = gpa >> PAGE_SHIFT;
2303 struct kvm_mmu_page *sp;
2304 struct hlist_node *node, *n;
2305 struct hlist_head *bucket;
2309 unsigned offset = offset_in_page(gpa);
2311 unsigned page_offset;
2312 unsigned misaligned;
2319 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2320 mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2321 spin_lock(&vcpu->kvm->mmu_lock);
2322 kvm_mmu_access_page(vcpu, gfn);
2323 kvm_mmu_free_some_pages(vcpu);
2324 ++vcpu->kvm->stat.mmu_pte_write;
2325 kvm_mmu_audit(vcpu, "pre pte write");
2326 if (gfn == vcpu->arch.last_pt_write_gfn
2327 && !last_updated_pte_accessed(vcpu)) {
2328 ++vcpu->arch.last_pt_write_count;
2329 if (vcpu->arch.last_pt_write_count >= 3)
2332 vcpu->arch.last_pt_write_gfn = gfn;
2333 vcpu->arch.last_pt_write_count = 1;
2334 vcpu->arch.last_pte_updated = NULL;
2336 index = kvm_page_table_hashfn(gfn);
2337 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2338 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2339 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
2341 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2342 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2343 misaligned |= bytes < 4;
2344 if (misaligned || flooded) {
2346 * Misaligned accesses are too much trouble to fix
2347 * up; also, they usually indicate a page is not used
2350 * If we're seeing too many writes to a page,
2351 * it may no longer be a page table, or we may be
2352 * forking, in which case it is better to unmap the
2355 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2356 gpa, bytes, sp->role.word);
2357 if (kvm_mmu_zap_page(vcpu->kvm, sp))
2359 ++vcpu->kvm->stat.mmu_flooded;
2362 page_offset = offset;
2363 level = sp->role.level;
2365 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2366 page_offset <<= 1; /* 32->64 */
2368 * A 32-bit pde maps 4MB while the shadow pdes map
2369 * only 2MB. So we need to double the offset again
2370 * and zap two pdes instead of one.
2372 if (level == PT32_ROOT_LEVEL) {
2373 page_offset &= ~7; /* kill rounding error */
2377 quadrant = page_offset >> PAGE_SHIFT;
2378 page_offset &= ~PAGE_MASK;
2379 if (quadrant != sp->role.quadrant)
2382 spte = &sp->spt[page_offset / sizeof(*spte)];
2383 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2385 r = kvm_read_guest_atomic(vcpu->kvm,
2386 gpa & ~(u64)(pte_size - 1),
2388 new = (const void *)&gentry;
2394 mmu_pte_write_zap_pte(vcpu, sp, spte);
2396 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2397 mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2401 kvm_mmu_audit(vcpu, "post pte write");
2402 spin_unlock(&vcpu->kvm->mmu_lock);
2403 if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2404 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2405 vcpu->arch.update_pte.pfn = bad_pfn;
2409 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2414 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2416 spin_lock(&vcpu->kvm->mmu_lock);
2417 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2418 spin_unlock(&vcpu->kvm->mmu_lock);
2421 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2423 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2425 while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2426 struct kvm_mmu_page *sp;
2428 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2429 struct kvm_mmu_page, link);
2430 kvm_mmu_zap_page(vcpu->kvm, sp);
2431 ++vcpu->kvm->stat.mmu_recycled;
2435 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2438 enum emulation_result er;
2440 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2449 r = mmu_topup_memory_caches(vcpu);
2453 er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2458 case EMULATE_DO_MMIO:
2459 ++vcpu->stat.mmio_exits;
2462 kvm_report_emulation_failure(vcpu, "pagetable");
2470 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2472 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2474 spin_lock(&vcpu->kvm->mmu_lock);
2475 vcpu->arch.mmu.invlpg(vcpu, gva);
2476 spin_unlock(&vcpu->kvm->mmu_lock);
2477 kvm_mmu_flush_tlb(vcpu);
2478 ++vcpu->stat.invlpg;
2480 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2482 void kvm_enable_tdp(void)
2486 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2488 void kvm_disable_tdp(void)
2490 tdp_enabled = false;
2492 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2494 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2496 struct kvm_mmu_page *sp;
2498 while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2499 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2500 struct kvm_mmu_page, link);
2501 kvm_mmu_zap_page(vcpu->kvm, sp);
2504 free_page((unsigned long)vcpu->arch.mmu.pae_root);
2507 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2514 if (vcpu->kvm->arch.n_requested_mmu_pages)
2515 vcpu->kvm->arch.n_free_mmu_pages =
2516 vcpu->kvm->arch.n_requested_mmu_pages;
2518 vcpu->kvm->arch.n_free_mmu_pages =
2519 vcpu->kvm->arch.n_alloc_mmu_pages;
2521 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2522 * Therefore we need to allocate shadow page tables in the first
2523 * 4GB of memory, which happens to fit the DMA32 zone.
2525 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2528 vcpu->arch.mmu.pae_root = page_address(page);
2529 for (i = 0; i < 4; ++i)
2530 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2535 free_mmu_pages(vcpu);
2539 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2542 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2544 return alloc_mmu_pages(vcpu);
2547 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2550 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2552 return init_kvm_mmu(vcpu);
2555 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2559 destroy_kvm_mmu(vcpu);
2560 free_mmu_pages(vcpu);
2561 mmu_free_memory_caches(vcpu);
2564 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2566 struct kvm_mmu_page *sp;
2568 spin_lock(&kvm->mmu_lock);
2569 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2573 if (!test_bit(slot, sp->slot_bitmap))
2577 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2579 if (pt[i] & PT_WRITABLE_MASK)
2580 pt[i] &= ~PT_WRITABLE_MASK;
2582 kvm_flush_remote_tlbs(kvm);
2583 spin_unlock(&kvm->mmu_lock);
2586 void kvm_mmu_zap_all(struct kvm *kvm)
2588 struct kvm_mmu_page *sp, *node;
2590 spin_lock(&kvm->mmu_lock);
2591 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2592 if (kvm_mmu_zap_page(kvm, sp))
2593 node = container_of(kvm->arch.active_mmu_pages.next,
2594 struct kvm_mmu_page, link);
2595 spin_unlock(&kvm->mmu_lock);
2597 kvm_flush_remote_tlbs(kvm);
2600 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2602 struct kvm_mmu_page *page;
2604 page = container_of(kvm->arch.active_mmu_pages.prev,
2605 struct kvm_mmu_page, link);
2606 kvm_mmu_zap_page(kvm, page);
2609 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2612 struct kvm *kvm_freed = NULL;
2613 int cache_count = 0;
2615 spin_lock(&kvm_lock);
2617 list_for_each_entry(kvm, &vm_list, vm_list) {
2620 if (!down_read_trylock(&kvm->slots_lock))
2622 spin_lock(&kvm->mmu_lock);
2623 npages = kvm->arch.n_alloc_mmu_pages -
2624 kvm->arch.n_free_mmu_pages;
2625 cache_count += npages;
2626 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2627 kvm_mmu_remove_one_alloc_mmu_page(kvm);
2633 spin_unlock(&kvm->mmu_lock);
2634 up_read(&kvm->slots_lock);
2637 list_move_tail(&kvm_freed->vm_list, &vm_list);
2639 spin_unlock(&kvm_lock);
2644 static struct shrinker mmu_shrinker = {
2645 .shrink = mmu_shrink,
2646 .seeks = DEFAULT_SEEKS * 10,
2649 static void mmu_destroy_caches(void)
2651 if (pte_chain_cache)
2652 kmem_cache_destroy(pte_chain_cache);
2653 if (rmap_desc_cache)
2654 kmem_cache_destroy(rmap_desc_cache);
2655 if (mmu_page_header_cache)
2656 kmem_cache_destroy(mmu_page_header_cache);
2659 void kvm_mmu_module_exit(void)
2661 mmu_destroy_caches();
2662 unregister_shrinker(&mmu_shrinker);
2665 int kvm_mmu_module_init(void)
2667 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2668 sizeof(struct kvm_pte_chain),
2670 if (!pte_chain_cache)
2672 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2673 sizeof(struct kvm_rmap_desc),
2675 if (!rmap_desc_cache)
2678 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2679 sizeof(struct kvm_mmu_page),
2681 if (!mmu_page_header_cache)
2684 register_shrinker(&mmu_shrinker);
2689 mmu_destroy_caches();
2694 * Caculate mmu pages needed for kvm.
2696 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2699 unsigned int nr_mmu_pages;
2700 unsigned int nr_pages = 0;
2702 for (i = 0; i < kvm->nmemslots; i++)
2703 nr_pages += kvm->memslots[i].npages;
2705 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2706 nr_mmu_pages = max(nr_mmu_pages,
2707 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2709 return nr_mmu_pages;
2712 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2715 if (len > buffer->len)
2720 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2725 ret = pv_mmu_peek_buffer(buffer, len);
2730 buffer->processed += len;
2734 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2735 gpa_t addr, gpa_t value)
2740 if (!is_long_mode(vcpu) && !is_pae(vcpu))
2743 r = mmu_topup_memory_caches(vcpu);
2747 if (!emulator_write_phys(vcpu, addr, &value, bytes))
2753 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2755 kvm_x86_ops->tlb_flush(vcpu);
2756 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
2760 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2762 spin_lock(&vcpu->kvm->mmu_lock);
2763 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2764 spin_unlock(&vcpu->kvm->mmu_lock);
2768 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2769 struct kvm_pv_mmu_op_buffer *buffer)
2771 struct kvm_mmu_op_header *header;
2773 header = pv_mmu_peek_buffer(buffer, sizeof *header);
2776 switch (header->op) {
2777 case KVM_MMU_OP_WRITE_PTE: {
2778 struct kvm_mmu_op_write_pte *wpte;
2780 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2783 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2786 case KVM_MMU_OP_FLUSH_TLB: {
2787 struct kvm_mmu_op_flush_tlb *ftlb;
2789 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2792 return kvm_pv_mmu_flush_tlb(vcpu);
2794 case KVM_MMU_OP_RELEASE_PT: {
2795 struct kvm_mmu_op_release_pt *rpt;
2797 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2800 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2806 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2807 gpa_t addr, unsigned long *ret)
2810 struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2812 buffer->ptr = buffer->buf;
2813 buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
2814 buffer->processed = 0;
2816 r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
2820 while (buffer->len) {
2821 r = kvm_pv_mmu_op_one(vcpu, buffer);
2830 *ret = buffer->processed;
2836 static const char *audit_msg;
2838 static gva_t canonicalize(gva_t gva)
2840 #ifdef CONFIG_X86_64
2841 gva = (long long)(gva << 16) >> 16;
2846 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2847 gva_t va, int level)
2849 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2851 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2853 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2856 if (ent == shadow_trap_nonpresent_pte)
2859 va = canonicalize(va);
2861 if (ent == shadow_notrap_nonpresent_pte)
2862 printk(KERN_ERR "audit: (%s) nontrapping pte"
2863 " in nonleaf level: levels %d gva %lx"
2864 " level %d pte %llx\n", audit_msg,
2865 vcpu->arch.mmu.root_level, va, level, ent);
2867 audit_mappings_page(vcpu, ent, va, level - 1);
2869 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2870 hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2872 if (is_shadow_present_pte(ent)
2873 && (ent & PT64_BASE_ADDR_MASK) != hpa)
2874 printk(KERN_ERR "xx audit error: (%s) levels %d"
2875 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2876 audit_msg, vcpu->arch.mmu.root_level,
2878 is_shadow_present_pte(ent));
2879 else if (ent == shadow_notrap_nonpresent_pte
2880 && !is_error_hpa(hpa))
2881 printk(KERN_ERR "audit: (%s) notrap shadow,"
2882 " valid guest gva %lx\n", audit_msg, va);
2883 kvm_release_pfn_clean(pfn);
2889 static void audit_mappings(struct kvm_vcpu *vcpu)
2893 if (vcpu->arch.mmu.root_level == 4)
2894 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2896 for (i = 0; i < 4; ++i)
2897 if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2898 audit_mappings_page(vcpu,
2899 vcpu->arch.mmu.pae_root[i],
2904 static int count_rmaps(struct kvm_vcpu *vcpu)
2909 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2910 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2911 struct kvm_rmap_desc *d;
2913 for (j = 0; j < m->npages; ++j) {
2914 unsigned long *rmapp = &m->rmap[j];
2918 if (!(*rmapp & 1)) {
2922 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2924 for (k = 0; k < RMAP_EXT; ++k)
2925 if (d->shadow_ptes[k])
2936 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2939 struct kvm_mmu_page *sp;
2942 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2945 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2948 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2951 if (!(ent & PT_PRESENT_MASK))
2953 if (!(ent & PT_WRITABLE_MASK))
2961 static void audit_rmap(struct kvm_vcpu *vcpu)
2963 int n_rmap = count_rmaps(vcpu);
2964 int n_actual = count_writable_mappings(vcpu);
2966 if (n_rmap != n_actual)
2967 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2968 __func__, audit_msg, n_rmap, n_actual);
2971 static void audit_write_protection(struct kvm_vcpu *vcpu)
2973 struct kvm_mmu_page *sp;
2974 struct kvm_memory_slot *slot;
2975 unsigned long *rmapp;
2978 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2979 if (sp->role.metaphysical)
2982 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2983 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
2984 rmapp = &slot->rmap[gfn - slot->base_gfn];
2986 printk(KERN_ERR "%s: (%s) shadow page has writable"
2987 " mappings: gfn %lx role %x\n",
2988 __func__, audit_msg, sp->gfn,
2993 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3000 audit_write_protection(vcpu);
3001 audit_mappings(vcpu);