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.
10 * Copyright 2010 Red Hat, Inc. and/or its affilates.
13 * Yaniv Kamay <yaniv@qumranet.com>
14 * Avi Kivity <avi@qumranet.com>
16 * This work is licensed under the terms of the GNU GPL, version 2. See
17 * the COPYING file in the top-level directory.
23 #include "kvm_cache_regs.h"
25 #include <linux/kvm_host.h>
26 #include <linux/types.h>
27 #include <linux/string.h>
29 #include <linux/highmem.h>
30 #include <linux/module.h>
31 #include <linux/swap.h>
32 #include <linux/hugetlb.h>
33 #include <linux/compiler.h>
34 #include <linux/srcu.h>
35 #include <linux/slab.h>
36 #include <linux/uaccess.h>
39 #include <asm/cmpxchg.h>
44 * When setting this variable to true it enables Two-Dimensional-Paging
45 * where the hardware walks 2 page tables:
46 * 1. the guest-virtual to guest-physical
47 * 2. while doing 1. it walks guest-physical to host-physical
48 * If the hardware supports that we don't need to do shadow paging.
50 bool tdp_enabled = false;
54 AUDIT_POST_PAGE_FAULT,
59 char *audit_point_name[] = {
70 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
71 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
75 #define pgprintk(x...) do { } while (0)
76 #define rmap_printk(x...) do { } while (0)
82 module_param(dbg, bool, 0644);
85 static int oos_shadow = 1;
86 module_param(oos_shadow, bool, 0644);
89 #define ASSERT(x) do { } while (0)
93 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
94 __FILE__, __LINE__, #x); \
98 #define PTE_PREFETCH_NUM 8
100 #define PT_FIRST_AVAIL_BITS_SHIFT 9
101 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
103 #define PT64_LEVEL_BITS 9
105 #define PT64_LEVEL_SHIFT(level) \
106 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
108 #define PT64_LEVEL_MASK(level) \
109 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
111 #define PT64_INDEX(address, level)\
112 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
115 #define PT32_LEVEL_BITS 10
117 #define PT32_LEVEL_SHIFT(level) \
118 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
120 #define PT32_LEVEL_MASK(level) \
121 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
122 #define PT32_LVL_OFFSET_MASK(level) \
123 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
124 * PT32_LEVEL_BITS))) - 1))
126 #define PT32_INDEX(address, level)\
127 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
130 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
131 #define PT64_DIR_BASE_ADDR_MASK \
132 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
133 #define PT64_LVL_ADDR_MASK(level) \
134 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
135 * PT64_LEVEL_BITS))) - 1))
136 #define PT64_LVL_OFFSET_MASK(level) \
137 (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
138 * PT64_LEVEL_BITS))) - 1))
140 #define PT32_BASE_ADDR_MASK PAGE_MASK
141 #define PT32_DIR_BASE_ADDR_MASK \
142 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
143 #define PT32_LVL_ADDR_MASK(level) \
144 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
145 * PT32_LEVEL_BITS))) - 1))
147 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
152 #define ACC_EXEC_MASK 1
153 #define ACC_WRITE_MASK PT_WRITABLE_MASK
154 #define ACC_USER_MASK PT_USER_MASK
155 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
157 #include <trace/events/kvm.h>
159 #define CREATE_TRACE_POINTS
160 #include "mmutrace.h"
162 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
164 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
166 struct kvm_rmap_desc {
167 u64 *sptes[RMAP_EXT];
168 struct kvm_rmap_desc *more;
171 struct kvm_shadow_walk_iterator {
179 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
180 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
181 shadow_walk_okay(&(_walker)); \
182 shadow_walk_next(&(_walker)))
184 typedef void (*mmu_parent_walk_fn) (struct kvm_mmu_page *sp, u64 *spte);
186 static struct kmem_cache *pte_chain_cache;
187 static struct kmem_cache *rmap_desc_cache;
188 static struct kmem_cache *mmu_page_header_cache;
189 static struct percpu_counter kvm_total_used_mmu_pages;
191 static u64 __read_mostly shadow_trap_nonpresent_pte;
192 static u64 __read_mostly shadow_notrap_nonpresent_pte;
193 static u64 __read_mostly shadow_base_present_pte;
194 static u64 __read_mostly shadow_nx_mask;
195 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
196 static u64 __read_mostly shadow_user_mask;
197 static u64 __read_mostly shadow_accessed_mask;
198 static u64 __read_mostly shadow_dirty_mask;
200 static inline u64 rsvd_bits(int s, int e)
202 return ((1ULL << (e - s + 1)) - 1) << s;
205 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
207 shadow_trap_nonpresent_pte = trap_pte;
208 shadow_notrap_nonpresent_pte = notrap_pte;
210 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
212 void kvm_mmu_set_base_ptes(u64 base_pte)
214 shadow_base_present_pte = base_pte;
216 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
218 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
219 u64 dirty_mask, u64 nx_mask, u64 x_mask)
221 shadow_user_mask = user_mask;
222 shadow_accessed_mask = accessed_mask;
223 shadow_dirty_mask = dirty_mask;
224 shadow_nx_mask = nx_mask;
225 shadow_x_mask = x_mask;
227 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
229 static bool is_write_protection(struct kvm_vcpu *vcpu)
231 return kvm_read_cr0_bits(vcpu, X86_CR0_WP);
234 static int is_cpuid_PSE36(void)
239 static int is_nx(struct kvm_vcpu *vcpu)
241 return vcpu->arch.efer & EFER_NX;
244 static int is_shadow_present_pte(u64 pte)
246 return pte != shadow_trap_nonpresent_pte
247 && pte != shadow_notrap_nonpresent_pte;
250 static int is_large_pte(u64 pte)
252 return pte & PT_PAGE_SIZE_MASK;
255 static int is_writable_pte(unsigned long pte)
257 return pte & PT_WRITABLE_MASK;
260 static int is_dirty_gpte(unsigned long pte)
262 return pte & PT_DIRTY_MASK;
265 static int is_rmap_spte(u64 pte)
267 return is_shadow_present_pte(pte);
270 static int is_last_spte(u64 pte, int level)
272 if (level == PT_PAGE_TABLE_LEVEL)
274 if (is_large_pte(pte))
279 static pfn_t spte_to_pfn(u64 pte)
281 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
284 static gfn_t pse36_gfn_delta(u32 gpte)
286 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
288 return (gpte & PT32_DIR_PSE36_MASK) << shift;
291 static void __set_spte(u64 *sptep, u64 spte)
293 set_64bit(sptep, spte);
296 static u64 __xchg_spte(u64 *sptep, u64 new_spte)
299 return xchg(sptep, new_spte);
305 } while (cmpxchg64(sptep, old_spte, new_spte) != old_spte);
311 static bool spte_has_volatile_bits(u64 spte)
313 if (!shadow_accessed_mask)
316 if (!is_shadow_present_pte(spte))
319 if ((spte & shadow_accessed_mask) &&
320 (!is_writable_pte(spte) || (spte & shadow_dirty_mask)))
326 static bool spte_is_bit_cleared(u64 old_spte, u64 new_spte, u64 bit_mask)
328 return (old_spte & bit_mask) && !(new_spte & bit_mask);
331 static void update_spte(u64 *sptep, u64 new_spte)
333 u64 mask, old_spte = *sptep;
335 WARN_ON(!is_rmap_spte(new_spte));
337 new_spte |= old_spte & shadow_dirty_mask;
339 mask = shadow_accessed_mask;
340 if (is_writable_pte(old_spte))
341 mask |= shadow_dirty_mask;
343 if (!spte_has_volatile_bits(old_spte) || (new_spte & mask) == mask)
344 __set_spte(sptep, new_spte);
346 old_spte = __xchg_spte(sptep, new_spte);
348 if (!shadow_accessed_mask)
351 if (spte_is_bit_cleared(old_spte, new_spte, shadow_accessed_mask))
352 kvm_set_pfn_accessed(spte_to_pfn(old_spte));
353 if (spte_is_bit_cleared(old_spte, new_spte, shadow_dirty_mask))
354 kvm_set_pfn_dirty(spte_to_pfn(old_spte));
357 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
358 struct kmem_cache *base_cache, int min)
362 if (cache->nobjs >= min)
364 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
365 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
368 cache->objects[cache->nobjs++] = obj;
373 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc,
374 struct kmem_cache *cache)
377 kmem_cache_free(cache, mc->objects[--mc->nobjs]);
380 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
385 if (cache->nobjs >= min)
387 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
388 page = alloc_page(GFP_KERNEL);
391 cache->objects[cache->nobjs++] = page_address(page);
396 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
399 free_page((unsigned long)mc->objects[--mc->nobjs]);
402 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
406 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
410 r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
411 rmap_desc_cache, 4 + PTE_PREFETCH_NUM);
414 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
417 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
418 mmu_page_header_cache, 4);
423 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
425 mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache, pte_chain_cache);
426 mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache, rmap_desc_cache);
427 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
428 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache,
429 mmu_page_header_cache);
432 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
438 p = mc->objects[--mc->nobjs];
442 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
444 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
445 sizeof(struct kvm_pte_chain));
448 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
450 kmem_cache_free(pte_chain_cache, pc);
453 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
455 return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
456 sizeof(struct kvm_rmap_desc));
459 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
461 kmem_cache_free(rmap_desc_cache, rd);
464 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
466 if (!sp->role.direct)
467 return sp->gfns[index];
469 return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
472 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
475 BUG_ON(gfn != kvm_mmu_page_get_gfn(sp, index));
477 sp->gfns[index] = gfn;
481 * Return the pointer to the largepage write count for a given
482 * gfn, handling slots that are not large page aligned.
484 static int *slot_largepage_idx(gfn_t gfn,
485 struct kvm_memory_slot *slot,
490 idx = (gfn >> KVM_HPAGE_GFN_SHIFT(level)) -
491 (slot->base_gfn >> KVM_HPAGE_GFN_SHIFT(level));
492 return &slot->lpage_info[level - 2][idx].write_count;
495 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
497 struct kvm_memory_slot *slot;
501 slot = gfn_to_memslot(kvm, gfn);
502 for (i = PT_DIRECTORY_LEVEL;
503 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
504 write_count = slot_largepage_idx(gfn, slot, i);
509 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
511 struct kvm_memory_slot *slot;
515 slot = gfn_to_memslot(kvm, gfn);
516 for (i = PT_DIRECTORY_LEVEL;
517 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
518 write_count = slot_largepage_idx(gfn, slot, i);
520 WARN_ON(*write_count < 0);
524 static int has_wrprotected_page(struct kvm *kvm,
528 struct kvm_memory_slot *slot;
531 slot = gfn_to_memslot(kvm, gfn);
533 largepage_idx = slot_largepage_idx(gfn, slot, level);
534 return *largepage_idx;
540 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
542 unsigned long page_size;
545 page_size = kvm_host_page_size(kvm, gfn);
547 for (i = PT_PAGE_TABLE_LEVEL;
548 i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
549 if (page_size >= KVM_HPAGE_SIZE(i))
558 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
560 struct kvm_memory_slot *slot;
561 int host_level, level, max_level;
563 slot = gfn_to_memslot(vcpu->kvm, large_gfn);
564 if (slot && slot->dirty_bitmap)
565 return PT_PAGE_TABLE_LEVEL;
567 host_level = host_mapping_level(vcpu->kvm, large_gfn);
569 if (host_level == PT_PAGE_TABLE_LEVEL)
572 max_level = kvm_x86_ops->get_lpage_level() < host_level ?
573 kvm_x86_ops->get_lpage_level() : host_level;
575 for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
576 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
583 * Take gfn and return the reverse mapping to it.
586 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
588 struct kvm_memory_slot *slot;
591 slot = gfn_to_memslot(kvm, gfn);
592 if (likely(level == PT_PAGE_TABLE_LEVEL))
593 return &slot->rmap[gfn - slot->base_gfn];
595 idx = (gfn >> KVM_HPAGE_GFN_SHIFT(level)) -
596 (slot->base_gfn >> KVM_HPAGE_GFN_SHIFT(level));
598 return &slot->lpage_info[level - 2][idx].rmap_pde;
602 * Reverse mapping data structures:
604 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
605 * that points to page_address(page).
607 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
608 * containing more mappings.
610 * Returns the number of rmap entries before the spte was added or zero if
611 * the spte was not added.
614 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
616 struct kvm_mmu_page *sp;
617 struct kvm_rmap_desc *desc;
618 unsigned long *rmapp;
621 if (!is_rmap_spte(*spte))
623 sp = page_header(__pa(spte));
624 kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
625 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
627 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
628 *rmapp = (unsigned long)spte;
629 } else if (!(*rmapp & 1)) {
630 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
631 desc = mmu_alloc_rmap_desc(vcpu);
632 desc->sptes[0] = (u64 *)*rmapp;
633 desc->sptes[1] = spte;
634 *rmapp = (unsigned long)desc | 1;
636 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
637 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
638 while (desc->sptes[RMAP_EXT-1] && desc->more) {
642 if (desc->sptes[RMAP_EXT-1]) {
643 desc->more = mmu_alloc_rmap_desc(vcpu);
646 for (i = 0; desc->sptes[i]; ++i)
648 desc->sptes[i] = spte;
653 static void rmap_desc_remove_entry(unsigned long *rmapp,
654 struct kvm_rmap_desc *desc,
656 struct kvm_rmap_desc *prev_desc)
660 for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
662 desc->sptes[i] = desc->sptes[j];
663 desc->sptes[j] = NULL;
666 if (!prev_desc && !desc->more)
667 *rmapp = (unsigned long)desc->sptes[0];
670 prev_desc->more = desc->more;
672 *rmapp = (unsigned long)desc->more | 1;
673 mmu_free_rmap_desc(desc);
676 static void rmap_remove(struct kvm *kvm, u64 *spte)
678 struct kvm_rmap_desc *desc;
679 struct kvm_rmap_desc *prev_desc;
680 struct kvm_mmu_page *sp;
682 unsigned long *rmapp;
685 sp = page_header(__pa(spte));
686 gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
687 rmapp = gfn_to_rmap(kvm, gfn, sp->role.level);
689 printk(KERN_ERR "rmap_remove: %p 0->BUG\n", spte);
691 } else if (!(*rmapp & 1)) {
692 rmap_printk("rmap_remove: %p 1->0\n", spte);
693 if ((u64 *)*rmapp != spte) {
694 printk(KERN_ERR "rmap_remove: %p 1->BUG\n", spte);
699 rmap_printk("rmap_remove: %p many->many\n", spte);
700 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
703 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
704 if (desc->sptes[i] == spte) {
705 rmap_desc_remove_entry(rmapp,
713 pr_err("rmap_remove: %p many->many\n", spte);
718 static void set_spte_track_bits(u64 *sptep, u64 new_spte)
721 u64 old_spte = *sptep;
723 if (!spte_has_volatile_bits(old_spte))
724 __set_spte(sptep, new_spte);
726 old_spte = __xchg_spte(sptep, new_spte);
728 if (!is_rmap_spte(old_spte))
731 pfn = spte_to_pfn(old_spte);
732 if (!shadow_accessed_mask || old_spte & shadow_accessed_mask)
733 kvm_set_pfn_accessed(pfn);
734 if (!shadow_dirty_mask || (old_spte & shadow_dirty_mask))
735 kvm_set_pfn_dirty(pfn);
738 static void drop_spte(struct kvm *kvm, u64 *sptep, u64 new_spte)
740 set_spte_track_bits(sptep, new_spte);
741 rmap_remove(kvm, sptep);
744 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
746 struct kvm_rmap_desc *desc;
752 else if (!(*rmapp & 1)) {
754 return (u64 *)*rmapp;
757 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
760 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
761 if (prev_spte == spte)
762 return desc->sptes[i];
763 prev_spte = desc->sptes[i];
770 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
772 unsigned long *rmapp;
774 int i, write_protected = 0;
776 rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
778 spte = rmap_next(kvm, rmapp, NULL);
781 BUG_ON(!(*spte & PT_PRESENT_MASK));
782 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
783 if (is_writable_pte(*spte)) {
784 update_spte(spte, *spte & ~PT_WRITABLE_MASK);
787 spte = rmap_next(kvm, rmapp, spte);
790 /* check for huge page mappings */
791 for (i = PT_DIRECTORY_LEVEL;
792 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
793 rmapp = gfn_to_rmap(kvm, gfn, i);
794 spte = rmap_next(kvm, rmapp, NULL);
797 BUG_ON(!(*spte & PT_PRESENT_MASK));
798 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
799 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
800 if (is_writable_pte(*spte)) {
802 shadow_trap_nonpresent_pte);
807 spte = rmap_next(kvm, rmapp, spte);
811 return write_protected;
814 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
818 int need_tlb_flush = 0;
820 while ((spte = rmap_next(kvm, rmapp, NULL))) {
821 BUG_ON(!(*spte & PT_PRESENT_MASK));
822 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
823 drop_spte(kvm, spte, shadow_trap_nonpresent_pte);
826 return need_tlb_flush;
829 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
834 pte_t *ptep = (pte_t *)data;
837 WARN_ON(pte_huge(*ptep));
838 new_pfn = pte_pfn(*ptep);
839 spte = rmap_next(kvm, rmapp, NULL);
841 BUG_ON(!is_shadow_present_pte(*spte));
842 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
844 if (pte_write(*ptep)) {
845 drop_spte(kvm, spte, shadow_trap_nonpresent_pte);
846 spte = rmap_next(kvm, rmapp, NULL);
848 new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
849 new_spte |= (u64)new_pfn << PAGE_SHIFT;
851 new_spte &= ~PT_WRITABLE_MASK;
852 new_spte &= ~SPTE_HOST_WRITEABLE;
853 new_spte &= ~shadow_accessed_mask;
854 set_spte_track_bits(spte, new_spte);
855 spte = rmap_next(kvm, rmapp, spte);
859 kvm_flush_remote_tlbs(kvm);
864 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
866 int (*handler)(struct kvm *kvm, unsigned long *rmapp,
872 struct kvm_memslots *slots;
874 slots = kvm_memslots(kvm);
876 for (i = 0; i < slots->nmemslots; i++) {
877 struct kvm_memory_slot *memslot = &slots->memslots[i];
878 unsigned long start = memslot->userspace_addr;
881 end = start + (memslot->npages << PAGE_SHIFT);
882 if (hva >= start && hva < end) {
883 gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
885 ret = handler(kvm, &memslot->rmap[gfn_offset], data);
887 for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
891 sh = KVM_HPAGE_GFN_SHIFT(PT_DIRECTORY_LEVEL+j);
892 idx = ((memslot->base_gfn+gfn_offset) >> sh) -
893 (memslot->base_gfn >> sh);
895 &memslot->lpage_info[j][idx].rmap_pde,
898 trace_kvm_age_page(hva, memslot, ret);
906 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
908 return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
911 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
913 kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
916 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
923 * Emulate the accessed bit for EPT, by checking if this page has
924 * an EPT mapping, and clearing it if it does. On the next access,
925 * a new EPT mapping will be established.
926 * This has some overhead, but not as much as the cost of swapping
927 * out actively used pages or breaking up actively used hugepages.
929 if (!shadow_accessed_mask)
930 return kvm_unmap_rmapp(kvm, rmapp, data);
932 spte = rmap_next(kvm, rmapp, NULL);
936 BUG_ON(!(_spte & PT_PRESENT_MASK));
937 _young = _spte & PT_ACCESSED_MASK;
940 clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
942 spte = rmap_next(kvm, rmapp, spte);
947 #define RMAP_RECYCLE_THRESHOLD 1000
949 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
951 unsigned long *rmapp;
952 struct kvm_mmu_page *sp;
954 sp = page_header(__pa(spte));
956 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
958 kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
959 kvm_flush_remote_tlbs(vcpu->kvm);
962 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
964 return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
968 static int is_empty_shadow_page(u64 *spt)
973 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
974 if (is_shadow_present_pte(*pos)) {
975 printk(KERN_ERR "%s: %p %llx\n", __func__,
984 * This value is the sum of all of the kvm instances's
985 * kvm->arch.n_used_mmu_pages values. We need a global,
986 * aggregate version in order to make the slab shrinker
989 static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, int nr)
991 kvm->arch.n_used_mmu_pages += nr;
992 percpu_counter_add(&kvm_total_used_mmu_pages, nr);
995 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
997 ASSERT(is_empty_shadow_page(sp->spt));
998 hlist_del(&sp->hash_link);
1000 __free_page(virt_to_page(sp->spt));
1001 if (!sp->role.direct)
1002 __free_page(virt_to_page(sp->gfns));
1003 kmem_cache_free(mmu_page_header_cache, sp);
1004 kvm_mod_used_mmu_pages(kvm, -1);
1007 static unsigned kvm_page_table_hashfn(gfn_t gfn)
1009 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
1012 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
1013 u64 *parent_pte, int direct)
1015 struct kvm_mmu_page *sp;
1017 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
1018 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
1020 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache,
1022 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
1023 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
1024 bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
1025 sp->multimapped = 0;
1026 sp->parent_pte = parent_pte;
1027 kvm_mod_used_mmu_pages(vcpu->kvm, +1);
1031 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
1032 struct kvm_mmu_page *sp, u64 *parent_pte)
1034 struct kvm_pte_chain *pte_chain;
1035 struct hlist_node *node;
1040 if (!sp->multimapped) {
1041 u64 *old = sp->parent_pte;
1044 sp->parent_pte = parent_pte;
1047 sp->multimapped = 1;
1048 pte_chain = mmu_alloc_pte_chain(vcpu);
1049 INIT_HLIST_HEAD(&sp->parent_ptes);
1050 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
1051 pte_chain->parent_ptes[0] = old;
1053 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
1054 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
1056 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
1057 if (!pte_chain->parent_ptes[i]) {
1058 pte_chain->parent_ptes[i] = parent_pte;
1062 pte_chain = mmu_alloc_pte_chain(vcpu);
1064 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
1065 pte_chain->parent_ptes[0] = parent_pte;
1068 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
1071 struct kvm_pte_chain *pte_chain;
1072 struct hlist_node *node;
1075 if (!sp->multimapped) {
1076 BUG_ON(sp->parent_pte != parent_pte);
1077 sp->parent_pte = NULL;
1080 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1081 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1082 if (!pte_chain->parent_ptes[i])
1084 if (pte_chain->parent_ptes[i] != parent_pte)
1086 while (i + 1 < NR_PTE_CHAIN_ENTRIES
1087 && pte_chain->parent_ptes[i + 1]) {
1088 pte_chain->parent_ptes[i]
1089 = pte_chain->parent_ptes[i + 1];
1092 pte_chain->parent_ptes[i] = NULL;
1094 hlist_del(&pte_chain->link);
1095 mmu_free_pte_chain(pte_chain);
1096 if (hlist_empty(&sp->parent_ptes)) {
1097 sp->multimapped = 0;
1098 sp->parent_pte = NULL;
1106 static void mmu_parent_walk(struct kvm_mmu_page *sp, mmu_parent_walk_fn fn)
1108 struct kvm_pte_chain *pte_chain;
1109 struct hlist_node *node;
1110 struct kvm_mmu_page *parent_sp;
1113 if (!sp->multimapped && sp->parent_pte) {
1114 parent_sp = page_header(__pa(sp->parent_pte));
1115 fn(parent_sp, sp->parent_pte);
1119 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1120 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1121 u64 *spte = pte_chain->parent_ptes[i];
1125 parent_sp = page_header(__pa(spte));
1126 fn(parent_sp, spte);
1130 static void mark_unsync(struct kvm_mmu_page *sp, u64 *spte);
1131 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1133 mmu_parent_walk(sp, mark_unsync);
1136 static void mark_unsync(struct kvm_mmu_page *sp, u64 *spte)
1140 index = spte - sp->spt;
1141 if (__test_and_set_bit(index, sp->unsync_child_bitmap))
1143 if (sp->unsync_children++)
1145 kvm_mmu_mark_parents_unsync(sp);
1148 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1149 struct kvm_mmu_page *sp)
1153 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1154 sp->spt[i] = shadow_trap_nonpresent_pte;
1157 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1158 struct kvm_mmu_page *sp, bool clear_unsync)
1163 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1167 #define KVM_PAGE_ARRAY_NR 16
1169 struct kvm_mmu_pages {
1170 struct mmu_page_and_offset {
1171 struct kvm_mmu_page *sp;
1173 } page[KVM_PAGE_ARRAY_NR];
1177 #define for_each_unsync_children(bitmap, idx) \
1178 for (idx = find_first_bit(bitmap, 512); \
1180 idx = find_next_bit(bitmap, 512, idx+1))
1182 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1188 for (i=0; i < pvec->nr; i++)
1189 if (pvec->page[i].sp == sp)
1192 pvec->page[pvec->nr].sp = sp;
1193 pvec->page[pvec->nr].idx = idx;
1195 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1198 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1199 struct kvm_mmu_pages *pvec)
1201 int i, ret, nr_unsync_leaf = 0;
1203 for_each_unsync_children(sp->unsync_child_bitmap, i) {
1204 struct kvm_mmu_page *child;
1205 u64 ent = sp->spt[i];
1207 if (!is_shadow_present_pte(ent) || is_large_pte(ent))
1208 goto clear_child_bitmap;
1210 child = page_header(ent & PT64_BASE_ADDR_MASK);
1212 if (child->unsync_children) {
1213 if (mmu_pages_add(pvec, child, i))
1216 ret = __mmu_unsync_walk(child, pvec);
1218 goto clear_child_bitmap;
1220 nr_unsync_leaf += ret;
1223 } else if (child->unsync) {
1225 if (mmu_pages_add(pvec, child, i))
1228 goto clear_child_bitmap;
1233 __clear_bit(i, sp->unsync_child_bitmap);
1234 sp->unsync_children--;
1235 WARN_ON((int)sp->unsync_children < 0);
1239 return nr_unsync_leaf;
1242 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1243 struct kvm_mmu_pages *pvec)
1245 if (!sp->unsync_children)
1248 mmu_pages_add(pvec, sp, 0);
1249 return __mmu_unsync_walk(sp, pvec);
1252 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1254 WARN_ON(!sp->unsync);
1255 trace_kvm_mmu_sync_page(sp);
1257 --kvm->stat.mmu_unsync;
1260 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1261 struct list_head *invalid_list);
1262 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1263 struct list_head *invalid_list);
1265 #define for_each_gfn_sp(kvm, sp, gfn, pos) \
1266 hlist_for_each_entry(sp, pos, \
1267 &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link) \
1268 if ((sp)->gfn != (gfn)) {} else
1270 #define for_each_gfn_indirect_valid_sp(kvm, sp, gfn, pos) \
1271 hlist_for_each_entry(sp, pos, \
1272 &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link) \
1273 if ((sp)->gfn != (gfn) || (sp)->role.direct || \
1274 (sp)->role.invalid) {} else
1276 /* @sp->gfn should be write-protected at the call site */
1277 static int __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1278 struct list_head *invalid_list, bool clear_unsync)
1280 if (sp->role.cr4_pae != !!is_pae(vcpu)) {
1281 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1286 kvm_unlink_unsync_page(vcpu->kvm, sp);
1288 if (vcpu->arch.mmu.sync_page(vcpu, sp, clear_unsync)) {
1289 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1293 kvm_mmu_flush_tlb(vcpu);
1297 static int kvm_sync_page_transient(struct kvm_vcpu *vcpu,
1298 struct kvm_mmu_page *sp)
1300 LIST_HEAD(invalid_list);
1303 ret = __kvm_sync_page(vcpu, sp, &invalid_list, false);
1305 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1310 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1311 struct list_head *invalid_list)
1313 return __kvm_sync_page(vcpu, sp, invalid_list, true);
1316 /* @gfn should be write-protected at the call site */
1317 static void kvm_sync_pages(struct kvm_vcpu *vcpu, gfn_t gfn)
1319 struct kvm_mmu_page *s;
1320 struct hlist_node *node;
1321 LIST_HEAD(invalid_list);
1324 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1328 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1329 if ((s->role.cr4_pae != !!is_pae(vcpu)) ||
1330 (vcpu->arch.mmu.sync_page(vcpu, s, true))) {
1331 kvm_mmu_prepare_zap_page(vcpu->kvm, s, &invalid_list);
1334 kvm_unlink_unsync_page(vcpu->kvm, s);
1338 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1340 kvm_mmu_flush_tlb(vcpu);
1343 struct mmu_page_path {
1344 struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1345 unsigned int idx[PT64_ROOT_LEVEL-1];
1348 #define for_each_sp(pvec, sp, parents, i) \
1349 for (i = mmu_pages_next(&pvec, &parents, -1), \
1350 sp = pvec.page[i].sp; \
1351 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1352 i = mmu_pages_next(&pvec, &parents, i))
1354 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1355 struct mmu_page_path *parents,
1360 for (n = i+1; n < pvec->nr; n++) {
1361 struct kvm_mmu_page *sp = pvec->page[n].sp;
1363 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1364 parents->idx[0] = pvec->page[n].idx;
1368 parents->parent[sp->role.level-2] = sp;
1369 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1375 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1377 struct kvm_mmu_page *sp;
1378 unsigned int level = 0;
1381 unsigned int idx = parents->idx[level];
1383 sp = parents->parent[level];
1387 --sp->unsync_children;
1388 WARN_ON((int)sp->unsync_children < 0);
1389 __clear_bit(idx, sp->unsync_child_bitmap);
1391 } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1394 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1395 struct mmu_page_path *parents,
1396 struct kvm_mmu_pages *pvec)
1398 parents->parent[parent->role.level-1] = NULL;
1402 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1403 struct kvm_mmu_page *parent)
1406 struct kvm_mmu_page *sp;
1407 struct mmu_page_path parents;
1408 struct kvm_mmu_pages pages;
1409 LIST_HEAD(invalid_list);
1411 kvm_mmu_pages_init(parent, &parents, &pages);
1412 while (mmu_unsync_walk(parent, &pages)) {
1415 for_each_sp(pages, sp, parents, i)
1416 protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1419 kvm_flush_remote_tlbs(vcpu->kvm);
1421 for_each_sp(pages, sp, parents, i) {
1422 kvm_sync_page(vcpu, sp, &invalid_list);
1423 mmu_pages_clear_parents(&parents);
1425 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1426 cond_resched_lock(&vcpu->kvm->mmu_lock);
1427 kvm_mmu_pages_init(parent, &parents, &pages);
1431 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1439 union kvm_mmu_page_role role;
1441 struct kvm_mmu_page *sp;
1442 struct hlist_node *node;
1443 bool need_sync = false;
1445 role = vcpu->arch.mmu.base_role;
1447 role.direct = direct;
1450 role.access = access;
1451 if (!vcpu->arch.mmu.direct_map
1452 && vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1453 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1454 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1455 role.quadrant = quadrant;
1457 for_each_gfn_sp(vcpu->kvm, sp, gfn, node) {
1458 if (!need_sync && sp->unsync)
1461 if (sp->role.word != role.word)
1464 if (sp->unsync && kvm_sync_page_transient(vcpu, sp))
1467 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1468 if (sp->unsync_children) {
1469 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
1470 kvm_mmu_mark_parents_unsync(sp);
1471 } else if (sp->unsync)
1472 kvm_mmu_mark_parents_unsync(sp);
1474 trace_kvm_mmu_get_page(sp, false);
1477 ++vcpu->kvm->stat.mmu_cache_miss;
1478 sp = kvm_mmu_alloc_page(vcpu, parent_pte, direct);
1483 hlist_add_head(&sp->hash_link,
1484 &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
1486 if (rmap_write_protect(vcpu->kvm, gfn))
1487 kvm_flush_remote_tlbs(vcpu->kvm);
1488 if (level > PT_PAGE_TABLE_LEVEL && need_sync)
1489 kvm_sync_pages(vcpu, gfn);
1491 account_shadowed(vcpu->kvm, gfn);
1493 if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1494 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1496 nonpaging_prefetch_page(vcpu, sp);
1497 trace_kvm_mmu_get_page(sp, true);
1501 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1502 struct kvm_vcpu *vcpu, u64 addr)
1504 iterator->addr = addr;
1505 iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1506 iterator->level = vcpu->arch.mmu.shadow_root_level;
1508 if (iterator->level == PT64_ROOT_LEVEL &&
1509 vcpu->arch.mmu.root_level < PT64_ROOT_LEVEL &&
1510 !vcpu->arch.mmu.direct_map)
1513 if (iterator->level == PT32E_ROOT_LEVEL) {
1514 iterator->shadow_addr
1515 = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1516 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1518 if (!iterator->shadow_addr)
1519 iterator->level = 0;
1523 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1525 if (iterator->level < PT_PAGE_TABLE_LEVEL)
1528 if (iterator->level == PT_PAGE_TABLE_LEVEL)
1529 if (is_large_pte(*iterator->sptep))
1532 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1533 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1537 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1539 iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1543 static void link_shadow_page(u64 *sptep, struct kvm_mmu_page *sp)
1547 spte = __pa(sp->spt)
1548 | PT_PRESENT_MASK | PT_ACCESSED_MASK
1549 | PT_WRITABLE_MASK | PT_USER_MASK;
1550 __set_spte(sptep, spte);
1553 static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1555 if (is_large_pte(*sptep)) {
1556 drop_spte(vcpu->kvm, sptep, shadow_trap_nonpresent_pte);
1557 kvm_flush_remote_tlbs(vcpu->kvm);
1561 static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1562 unsigned direct_access)
1564 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
1565 struct kvm_mmu_page *child;
1568 * For the direct sp, if the guest pte's dirty bit
1569 * changed form clean to dirty, it will corrupt the
1570 * sp's access: allow writable in the read-only sp,
1571 * so we should update the spte at this point to get
1572 * a new sp with the correct access.
1574 child = page_header(*sptep & PT64_BASE_ADDR_MASK);
1575 if (child->role.access == direct_access)
1578 mmu_page_remove_parent_pte(child, sptep);
1579 __set_spte(sptep, shadow_trap_nonpresent_pte);
1580 kvm_flush_remote_tlbs(vcpu->kvm);
1584 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1585 struct kvm_mmu_page *sp)
1593 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1596 if (is_shadow_present_pte(ent)) {
1597 if (!is_last_spte(ent, sp->role.level)) {
1598 ent &= PT64_BASE_ADDR_MASK;
1599 mmu_page_remove_parent_pte(page_header(ent),
1602 if (is_large_pte(ent))
1604 drop_spte(kvm, &pt[i],
1605 shadow_trap_nonpresent_pte);
1608 pt[i] = shadow_trap_nonpresent_pte;
1612 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1614 mmu_page_remove_parent_pte(sp, parent_pte);
1617 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1620 struct kvm_vcpu *vcpu;
1622 kvm_for_each_vcpu(i, vcpu, kvm)
1623 vcpu->arch.last_pte_updated = NULL;
1626 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1630 while (sp->multimapped || sp->parent_pte) {
1631 if (!sp->multimapped)
1632 parent_pte = sp->parent_pte;
1634 struct kvm_pte_chain *chain;
1636 chain = container_of(sp->parent_ptes.first,
1637 struct kvm_pte_chain, link);
1638 parent_pte = chain->parent_ptes[0];
1640 BUG_ON(!parent_pte);
1641 kvm_mmu_put_page(sp, parent_pte);
1642 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1646 static int mmu_zap_unsync_children(struct kvm *kvm,
1647 struct kvm_mmu_page *parent,
1648 struct list_head *invalid_list)
1651 struct mmu_page_path parents;
1652 struct kvm_mmu_pages pages;
1654 if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1657 kvm_mmu_pages_init(parent, &parents, &pages);
1658 while (mmu_unsync_walk(parent, &pages)) {
1659 struct kvm_mmu_page *sp;
1661 for_each_sp(pages, sp, parents, i) {
1662 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
1663 mmu_pages_clear_parents(&parents);
1666 kvm_mmu_pages_init(parent, &parents, &pages);
1672 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1673 struct list_head *invalid_list)
1677 trace_kvm_mmu_prepare_zap_page(sp);
1678 ++kvm->stat.mmu_shadow_zapped;
1679 ret = mmu_zap_unsync_children(kvm, sp, invalid_list);
1680 kvm_mmu_page_unlink_children(kvm, sp);
1681 kvm_mmu_unlink_parents(kvm, sp);
1682 if (!sp->role.invalid && !sp->role.direct)
1683 unaccount_shadowed(kvm, sp->gfn);
1685 kvm_unlink_unsync_page(kvm, sp);
1686 if (!sp->root_count) {
1689 list_move(&sp->link, invalid_list);
1691 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1692 kvm_reload_remote_mmus(kvm);
1695 sp->role.invalid = 1;
1696 kvm_mmu_reset_last_pte_updated(kvm);
1700 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1701 struct list_head *invalid_list)
1703 struct kvm_mmu_page *sp;
1705 if (list_empty(invalid_list))
1708 kvm_flush_remote_tlbs(kvm);
1711 sp = list_first_entry(invalid_list, struct kvm_mmu_page, link);
1712 WARN_ON(!sp->role.invalid || sp->root_count);
1713 kvm_mmu_free_page(kvm, sp);
1714 } while (!list_empty(invalid_list));
1719 * Changing the number of mmu pages allocated to the vm
1720 * Note: if goal_nr_mmu_pages is too small, you will get dead lock
1722 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int goal_nr_mmu_pages)
1724 LIST_HEAD(invalid_list);
1726 * If we set the number of mmu pages to be smaller be than the
1727 * number of actived pages , we must to free some mmu pages before we
1731 if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) {
1732 while (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages &&
1733 !list_empty(&kvm->arch.active_mmu_pages)) {
1734 struct kvm_mmu_page *page;
1736 page = container_of(kvm->arch.active_mmu_pages.prev,
1737 struct kvm_mmu_page, link);
1738 kvm_mmu_prepare_zap_page(kvm, page, &invalid_list);
1739 kvm_mmu_commit_zap_page(kvm, &invalid_list);
1741 goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages;
1744 kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages;
1747 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1749 struct kvm_mmu_page *sp;
1750 struct hlist_node *node;
1751 LIST_HEAD(invalid_list);
1754 pgprintk("%s: looking for gfn %llx\n", __func__, gfn);
1757 for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
1758 pgprintk("%s: gfn %llx role %x\n", __func__, gfn,
1761 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
1763 kvm_mmu_commit_zap_page(kvm, &invalid_list);
1767 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1769 struct kvm_mmu_page *sp;
1770 struct hlist_node *node;
1771 LIST_HEAD(invalid_list);
1773 for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
1774 pgprintk("%s: zap %llx %x\n",
1775 __func__, gfn, sp->role.word);
1776 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
1778 kvm_mmu_commit_zap_page(kvm, &invalid_list);
1781 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1783 int slot = memslot_id(kvm, gfn);
1784 struct kvm_mmu_page *sp = page_header(__pa(pte));
1786 __set_bit(slot, sp->slot_bitmap);
1789 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1794 if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1797 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1798 if (pt[i] == shadow_notrap_nonpresent_pte)
1799 __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1804 * The function is based on mtrr_type_lookup() in
1805 * arch/x86/kernel/cpu/mtrr/generic.c
1807 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1812 u8 prev_match, curr_match;
1813 int num_var_ranges = KVM_NR_VAR_MTRR;
1815 if (!mtrr_state->enabled)
1818 /* Make end inclusive end, instead of exclusive */
1821 /* Look in fixed ranges. Just return the type as per start */
1822 if (mtrr_state->have_fixed && (start < 0x100000)) {
1825 if (start < 0x80000) {
1827 idx += (start >> 16);
1828 return mtrr_state->fixed_ranges[idx];
1829 } else if (start < 0xC0000) {
1831 idx += ((start - 0x80000) >> 14);
1832 return mtrr_state->fixed_ranges[idx];
1833 } else if (start < 0x1000000) {
1835 idx += ((start - 0xC0000) >> 12);
1836 return mtrr_state->fixed_ranges[idx];
1841 * Look in variable ranges
1842 * Look of multiple ranges matching this address and pick type
1843 * as per MTRR precedence
1845 if (!(mtrr_state->enabled & 2))
1846 return mtrr_state->def_type;
1849 for (i = 0; i < num_var_ranges; ++i) {
1850 unsigned short start_state, end_state;
1852 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1855 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1856 (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1857 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1858 (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1860 start_state = ((start & mask) == (base & mask));
1861 end_state = ((end & mask) == (base & mask));
1862 if (start_state != end_state)
1865 if ((start & mask) != (base & mask))
1868 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1869 if (prev_match == 0xFF) {
1870 prev_match = curr_match;
1874 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1875 curr_match == MTRR_TYPE_UNCACHABLE)
1876 return MTRR_TYPE_UNCACHABLE;
1878 if ((prev_match == MTRR_TYPE_WRBACK &&
1879 curr_match == MTRR_TYPE_WRTHROUGH) ||
1880 (prev_match == MTRR_TYPE_WRTHROUGH &&
1881 curr_match == MTRR_TYPE_WRBACK)) {
1882 prev_match = MTRR_TYPE_WRTHROUGH;
1883 curr_match = MTRR_TYPE_WRTHROUGH;
1886 if (prev_match != curr_match)
1887 return MTRR_TYPE_UNCACHABLE;
1890 if (prev_match != 0xFF)
1893 return mtrr_state->def_type;
1896 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1900 mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1901 (gfn << PAGE_SHIFT) + PAGE_SIZE);
1902 if (mtrr == 0xfe || mtrr == 0xff)
1903 mtrr = MTRR_TYPE_WRBACK;
1906 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1908 static void __kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1910 trace_kvm_mmu_unsync_page(sp);
1911 ++vcpu->kvm->stat.mmu_unsync;
1914 kvm_mmu_mark_parents_unsync(sp);
1915 mmu_convert_notrap(sp);
1918 static void kvm_unsync_pages(struct kvm_vcpu *vcpu, gfn_t gfn)
1920 struct kvm_mmu_page *s;
1921 struct hlist_node *node;
1923 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1926 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1927 __kvm_unsync_page(vcpu, s);
1931 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1934 struct kvm_mmu_page *s;
1935 struct hlist_node *node;
1936 bool need_unsync = false;
1938 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1942 if (s->role.level != PT_PAGE_TABLE_LEVEL)
1945 if (!need_unsync && !s->unsync) {
1952 kvm_unsync_pages(vcpu, gfn);
1956 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1957 unsigned pte_access, int user_fault,
1958 int write_fault, int dirty, int level,
1959 gfn_t gfn, pfn_t pfn, bool speculative,
1960 bool can_unsync, bool reset_host_protection)
1966 * We don't set the accessed bit, since we sometimes want to see
1967 * whether the guest actually used the pte (in order to detect
1970 spte = shadow_base_present_pte;
1972 spte |= shadow_accessed_mask;
1974 pte_access &= ~ACC_WRITE_MASK;
1975 if (pte_access & ACC_EXEC_MASK)
1976 spte |= shadow_x_mask;
1978 spte |= shadow_nx_mask;
1979 if (pte_access & ACC_USER_MASK)
1980 spte |= shadow_user_mask;
1981 if (level > PT_PAGE_TABLE_LEVEL)
1982 spte |= PT_PAGE_SIZE_MASK;
1984 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1985 kvm_is_mmio_pfn(pfn));
1987 if (reset_host_protection)
1988 spte |= SPTE_HOST_WRITEABLE;
1990 spte |= (u64)pfn << PAGE_SHIFT;
1992 if ((pte_access & ACC_WRITE_MASK)
1993 || (!vcpu->arch.mmu.direct_map && write_fault
1994 && !is_write_protection(vcpu) && !user_fault)) {
1996 if (level > PT_PAGE_TABLE_LEVEL &&
1997 has_wrprotected_page(vcpu->kvm, gfn, level)) {
1999 drop_spte(vcpu->kvm, sptep, shadow_trap_nonpresent_pte);
2003 spte |= PT_WRITABLE_MASK;
2005 if (!vcpu->arch.mmu.direct_map
2006 && !(pte_access & ACC_WRITE_MASK))
2007 spte &= ~PT_USER_MASK;
2010 * Optimization: for pte sync, if spte was writable the hash
2011 * lookup is unnecessary (and expensive). Write protection
2012 * is responsibility of mmu_get_page / kvm_sync_page.
2013 * Same reasoning can be applied to dirty page accounting.
2015 if (!can_unsync && is_writable_pte(*sptep))
2018 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
2019 pgprintk("%s: found shadow page for %llx, marking ro\n",
2022 pte_access &= ~ACC_WRITE_MASK;
2023 if (is_writable_pte(spte))
2024 spte &= ~PT_WRITABLE_MASK;
2028 if (pte_access & ACC_WRITE_MASK)
2029 mark_page_dirty(vcpu->kvm, gfn);
2032 update_spte(sptep, spte);
2037 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2038 unsigned pt_access, unsigned pte_access,
2039 int user_fault, int write_fault, int dirty,
2040 int *ptwrite, int level, gfn_t gfn,
2041 pfn_t pfn, bool speculative,
2042 bool reset_host_protection)
2044 int was_rmapped = 0;
2047 pgprintk("%s: spte %llx access %x write_fault %d"
2048 " user_fault %d gfn %llx\n",
2049 __func__, *sptep, pt_access,
2050 write_fault, user_fault, gfn);
2052 if (is_rmap_spte(*sptep)) {
2054 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
2055 * the parent of the now unreachable PTE.
2057 if (level > PT_PAGE_TABLE_LEVEL &&
2058 !is_large_pte(*sptep)) {
2059 struct kvm_mmu_page *child;
2062 child = page_header(pte & PT64_BASE_ADDR_MASK);
2063 mmu_page_remove_parent_pte(child, sptep);
2064 __set_spte(sptep, shadow_trap_nonpresent_pte);
2065 kvm_flush_remote_tlbs(vcpu->kvm);
2066 } else if (pfn != spte_to_pfn(*sptep)) {
2067 pgprintk("hfn old %llx new %llx\n",
2068 spte_to_pfn(*sptep), pfn);
2069 drop_spte(vcpu->kvm, sptep, shadow_trap_nonpresent_pte);
2070 kvm_flush_remote_tlbs(vcpu->kvm);
2075 if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
2076 dirty, level, gfn, pfn, speculative, true,
2077 reset_host_protection)) {
2080 kvm_mmu_flush_tlb(vcpu);
2083 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
2084 pgprintk("instantiating %s PTE (%s) at %llx (%llx) addr %p\n",
2085 is_large_pte(*sptep)? "2MB" : "4kB",
2086 *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
2088 if (!was_rmapped && is_large_pte(*sptep))
2089 ++vcpu->kvm->stat.lpages;
2091 page_header_update_slot(vcpu->kvm, sptep, gfn);
2093 rmap_count = rmap_add(vcpu, sptep, gfn);
2094 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
2095 rmap_recycle(vcpu, sptep, gfn);
2097 kvm_release_pfn_clean(pfn);
2099 vcpu->arch.last_pte_updated = sptep;
2100 vcpu->arch.last_pte_gfn = gfn;
2104 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
2108 static struct kvm_memory_slot *
2109 pte_prefetch_gfn_to_memslot(struct kvm_vcpu *vcpu, gfn_t gfn, bool no_dirty_log)
2111 struct kvm_memory_slot *slot;
2113 slot = gfn_to_memslot(vcpu->kvm, gfn);
2114 if (!slot || slot->flags & KVM_MEMSLOT_INVALID ||
2115 (no_dirty_log && slot->dirty_bitmap))
2121 static pfn_t pte_prefetch_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn,
2124 struct kvm_memory_slot *slot;
2127 slot = pte_prefetch_gfn_to_memslot(vcpu, gfn, no_dirty_log);
2130 return page_to_pfn(bad_page);
2133 hva = gfn_to_hva_memslot(slot, gfn);
2135 return hva_to_pfn_atomic(vcpu->kvm, hva);
2138 static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu,
2139 struct kvm_mmu_page *sp,
2140 u64 *start, u64 *end)
2142 struct page *pages[PTE_PREFETCH_NUM];
2143 unsigned access = sp->role.access;
2147 gfn = kvm_mmu_page_get_gfn(sp, start - sp->spt);
2148 if (!pte_prefetch_gfn_to_memslot(vcpu, gfn, access & ACC_WRITE_MASK))
2151 ret = gfn_to_page_many_atomic(vcpu->kvm, gfn, pages, end - start);
2155 for (i = 0; i < ret; i++, gfn++, start++)
2156 mmu_set_spte(vcpu, start, ACC_ALL,
2157 access, 0, 0, 1, NULL,
2158 sp->role.level, gfn,
2159 page_to_pfn(pages[i]), true, true);
2164 static void __direct_pte_prefetch(struct kvm_vcpu *vcpu,
2165 struct kvm_mmu_page *sp, u64 *sptep)
2167 u64 *spte, *start = NULL;
2170 WARN_ON(!sp->role.direct);
2172 i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
2175 for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
2176 if (*spte != shadow_trap_nonpresent_pte || spte == sptep) {
2179 if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0)
2187 static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
2189 struct kvm_mmu_page *sp;
2192 * Since it's no accessed bit on EPT, it's no way to
2193 * distinguish between actually accessed translations
2194 * and prefetched, so disable pte prefetch if EPT is
2197 if (!shadow_accessed_mask)
2200 sp = page_header(__pa(sptep));
2201 if (sp->role.level > PT_PAGE_TABLE_LEVEL)
2204 __direct_pte_prefetch(vcpu, sp, sptep);
2207 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
2208 int level, gfn_t gfn, pfn_t pfn)
2210 struct kvm_shadow_walk_iterator iterator;
2211 struct kvm_mmu_page *sp;
2215 for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
2216 if (iterator.level == level) {
2217 mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
2218 0, write, 1, &pt_write,
2219 level, gfn, pfn, false, true);
2220 direct_pte_prefetch(vcpu, iterator.sptep);
2221 ++vcpu->stat.pf_fixed;
2225 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
2226 u64 base_addr = iterator.addr;
2228 base_addr &= PT64_LVL_ADDR_MASK(iterator.level);
2229 pseudo_gfn = base_addr >> PAGE_SHIFT;
2230 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
2232 1, ACC_ALL, iterator.sptep);
2234 pgprintk("nonpaging_map: ENOMEM\n");
2235 kvm_release_pfn_clean(pfn);
2239 __set_spte(iterator.sptep,
2241 | PT_PRESENT_MASK | PT_WRITABLE_MASK
2242 | shadow_user_mask | shadow_x_mask);
2248 static void kvm_send_hwpoison_signal(struct kvm *kvm, gfn_t gfn)
2254 /* Touch the page, so send SIGBUS */
2255 hva = (void __user *)gfn_to_hva(kvm, gfn);
2256 r = copy_from_user(buf, hva, 1);
2259 static int kvm_handle_bad_page(struct kvm *kvm, gfn_t gfn, pfn_t pfn)
2261 kvm_release_pfn_clean(pfn);
2262 if (is_hwpoison_pfn(pfn)) {
2263 kvm_send_hwpoison_signal(kvm, gfn);
2265 } else if (is_fault_pfn(pfn))
2271 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
2276 unsigned long mmu_seq;
2278 level = mapping_level(vcpu, gfn);
2281 * This path builds a PAE pagetable - so we can map 2mb pages at
2282 * maximum. Therefore check if the level is larger than that.
2284 if (level > PT_DIRECTORY_LEVEL)
2285 level = PT_DIRECTORY_LEVEL;
2287 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2289 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2291 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2294 if (is_error_pfn(pfn))
2295 return kvm_handle_bad_page(vcpu->kvm, gfn, pfn);
2297 spin_lock(&vcpu->kvm->mmu_lock);
2298 if (mmu_notifier_retry(vcpu, mmu_seq))
2300 kvm_mmu_free_some_pages(vcpu);
2301 r = __direct_map(vcpu, v, write, level, gfn, pfn);
2302 spin_unlock(&vcpu->kvm->mmu_lock);
2308 spin_unlock(&vcpu->kvm->mmu_lock);
2309 kvm_release_pfn_clean(pfn);
2314 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2317 struct kvm_mmu_page *sp;
2318 LIST_HEAD(invalid_list);
2320 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2322 spin_lock(&vcpu->kvm->mmu_lock);
2323 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL &&
2324 (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL ||
2325 vcpu->arch.mmu.direct_map)) {
2326 hpa_t root = vcpu->arch.mmu.root_hpa;
2328 sp = page_header(root);
2330 if (!sp->root_count && sp->role.invalid) {
2331 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
2332 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2334 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2335 spin_unlock(&vcpu->kvm->mmu_lock);
2338 for (i = 0; i < 4; ++i) {
2339 hpa_t root = vcpu->arch.mmu.pae_root[i];
2342 root &= PT64_BASE_ADDR_MASK;
2343 sp = page_header(root);
2345 if (!sp->root_count && sp->role.invalid)
2346 kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2349 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2351 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2352 spin_unlock(&vcpu->kvm->mmu_lock);
2353 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2356 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2360 if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2361 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
2368 static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
2370 struct kvm_mmu_page *sp;
2373 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2374 spin_lock(&vcpu->kvm->mmu_lock);
2375 kvm_mmu_free_some_pages(vcpu);
2376 sp = kvm_mmu_get_page(vcpu, 0, 0, PT64_ROOT_LEVEL,
2379 spin_unlock(&vcpu->kvm->mmu_lock);
2380 vcpu->arch.mmu.root_hpa = __pa(sp->spt);
2381 } else if (vcpu->arch.mmu.shadow_root_level == PT32E_ROOT_LEVEL) {
2382 for (i = 0; i < 4; ++i) {
2383 hpa_t root = vcpu->arch.mmu.pae_root[i];
2385 ASSERT(!VALID_PAGE(root));
2386 spin_lock(&vcpu->kvm->mmu_lock);
2387 kvm_mmu_free_some_pages(vcpu);
2388 sp = kvm_mmu_get_page(vcpu, i << 30, i << 30,
2389 PT32_ROOT_LEVEL, 1, ACC_ALL,
2391 root = __pa(sp->spt);
2393 spin_unlock(&vcpu->kvm->mmu_lock);
2394 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2395 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2403 static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
2405 struct kvm_mmu_page *sp;
2410 root_gfn = vcpu->arch.mmu.get_cr3(vcpu) >> PAGE_SHIFT;
2412 if (mmu_check_root(vcpu, root_gfn))
2416 * Do we shadow a long mode page table? If so we need to
2417 * write-protect the guests page table root.
2419 if (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL) {
2420 hpa_t root = vcpu->arch.mmu.root_hpa;
2422 ASSERT(!VALID_PAGE(root));
2424 spin_lock(&vcpu->kvm->mmu_lock);
2425 kvm_mmu_free_some_pages(vcpu);
2426 sp = kvm_mmu_get_page(vcpu, root_gfn, 0, PT64_ROOT_LEVEL,
2428 root = __pa(sp->spt);
2430 spin_unlock(&vcpu->kvm->mmu_lock);
2431 vcpu->arch.mmu.root_hpa = root;
2436 * We shadow a 32 bit page table. This may be a legacy 2-level
2437 * or a PAE 3-level page table. In either case we need to be aware that
2438 * the shadow page table may be a PAE or a long mode page table.
2440 pm_mask = PT_PRESENT_MASK;
2441 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL)
2442 pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK;
2444 for (i = 0; i < 4; ++i) {
2445 hpa_t root = vcpu->arch.mmu.pae_root[i];
2447 ASSERT(!VALID_PAGE(root));
2448 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2449 pdptr = kvm_pdptr_read_mmu(vcpu, &vcpu->arch.mmu, i);
2450 if (!is_present_gpte(pdptr)) {
2451 vcpu->arch.mmu.pae_root[i] = 0;
2454 root_gfn = pdptr >> PAGE_SHIFT;
2455 if (mmu_check_root(vcpu, root_gfn))
2458 spin_lock(&vcpu->kvm->mmu_lock);
2459 kvm_mmu_free_some_pages(vcpu);
2460 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2463 root = __pa(sp->spt);
2465 spin_unlock(&vcpu->kvm->mmu_lock);
2467 vcpu->arch.mmu.pae_root[i] = root | pm_mask;
2468 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2472 * If we shadow a 32 bit page table with a long mode page
2473 * table we enter this path.
2475 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2476 if (vcpu->arch.mmu.lm_root == NULL) {
2478 * The additional page necessary for this is only
2479 * allocated on demand.
2484 lm_root = (void*)get_zeroed_page(GFP_KERNEL);
2485 if (lm_root == NULL)
2488 lm_root[0] = __pa(vcpu->arch.mmu.pae_root) | pm_mask;
2490 vcpu->arch.mmu.lm_root = lm_root;
2493 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.lm_root);
2499 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2501 if (vcpu->arch.mmu.direct_map)
2502 return mmu_alloc_direct_roots(vcpu);
2504 return mmu_alloc_shadow_roots(vcpu);
2507 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2510 struct kvm_mmu_page *sp;
2512 if (vcpu->arch.mmu.direct_map)
2515 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2517 if (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL) {
2518 hpa_t root = vcpu->arch.mmu.root_hpa;
2519 sp = page_header(root);
2520 mmu_sync_children(vcpu, sp);
2523 for (i = 0; i < 4; ++i) {
2524 hpa_t root = vcpu->arch.mmu.pae_root[i];
2526 if (root && VALID_PAGE(root)) {
2527 root &= PT64_BASE_ADDR_MASK;
2528 sp = page_header(root);
2529 mmu_sync_children(vcpu, sp);
2534 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2536 spin_lock(&vcpu->kvm->mmu_lock);
2537 mmu_sync_roots(vcpu);
2538 spin_unlock(&vcpu->kvm->mmu_lock);
2541 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2542 u32 access, u32 *error)
2549 static gpa_t nonpaging_gva_to_gpa_nested(struct kvm_vcpu *vcpu, gva_t vaddr,
2550 u32 access, u32 *error)
2554 return vcpu->arch.nested_mmu.translate_gpa(vcpu, vaddr, access);
2557 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2563 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2564 r = mmu_topup_memory_caches(vcpu);
2569 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2571 gfn = gva >> PAGE_SHIFT;
2573 return nonpaging_map(vcpu, gva & PAGE_MASK,
2574 error_code & PFERR_WRITE_MASK, gfn);
2577 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2583 gfn_t gfn = gpa >> PAGE_SHIFT;
2584 unsigned long mmu_seq;
2587 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2589 r = mmu_topup_memory_caches(vcpu);
2593 level = mapping_level(vcpu, gfn);
2595 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2597 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2599 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2600 if (is_error_pfn(pfn))
2601 return kvm_handle_bad_page(vcpu->kvm, gfn, pfn);
2602 spin_lock(&vcpu->kvm->mmu_lock);
2603 if (mmu_notifier_retry(vcpu, mmu_seq))
2605 kvm_mmu_free_some_pages(vcpu);
2606 r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2608 spin_unlock(&vcpu->kvm->mmu_lock);
2613 spin_unlock(&vcpu->kvm->mmu_lock);
2614 kvm_release_pfn_clean(pfn);
2618 static void nonpaging_free(struct kvm_vcpu *vcpu)
2620 mmu_free_roots(vcpu);
2623 static int nonpaging_init_context(struct kvm_vcpu *vcpu,
2624 struct kvm_mmu *context)
2626 context->new_cr3 = nonpaging_new_cr3;
2627 context->page_fault = nonpaging_page_fault;
2628 context->gva_to_gpa = nonpaging_gva_to_gpa;
2629 context->free = nonpaging_free;
2630 context->prefetch_page = nonpaging_prefetch_page;
2631 context->sync_page = nonpaging_sync_page;
2632 context->invlpg = nonpaging_invlpg;
2633 context->root_level = 0;
2634 context->shadow_root_level = PT32E_ROOT_LEVEL;
2635 context->root_hpa = INVALID_PAGE;
2636 context->direct_map = true;
2637 context->nx = false;
2641 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2643 ++vcpu->stat.tlb_flush;
2644 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
2647 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2649 pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2650 mmu_free_roots(vcpu);
2653 static unsigned long get_cr3(struct kvm_vcpu *vcpu)
2655 return vcpu->arch.cr3;
2658 static void inject_page_fault(struct kvm_vcpu *vcpu)
2660 vcpu->arch.mmu.inject_page_fault(vcpu);
2663 static void paging_free(struct kvm_vcpu *vcpu)
2665 nonpaging_free(vcpu);
2668 static bool is_rsvd_bits_set(struct kvm_mmu *mmu, u64 gpte, int level)
2672 bit7 = (gpte >> 7) & 1;
2673 return (gpte & mmu->rsvd_bits_mask[bit7][level-1]) != 0;
2677 #include "paging_tmpl.h"
2681 #include "paging_tmpl.h"
2684 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
2685 struct kvm_mmu *context,
2688 int maxphyaddr = cpuid_maxphyaddr(vcpu);
2689 u64 exb_bit_rsvd = 0;
2692 exb_bit_rsvd = rsvd_bits(63, 63);
2694 case PT32_ROOT_LEVEL:
2695 /* no rsvd bits for 2 level 4K page table entries */
2696 context->rsvd_bits_mask[0][1] = 0;
2697 context->rsvd_bits_mask[0][0] = 0;
2698 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2700 if (!is_pse(vcpu)) {
2701 context->rsvd_bits_mask[1][1] = 0;
2705 if (is_cpuid_PSE36())
2706 /* 36bits PSE 4MB page */
2707 context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2709 /* 32 bits PSE 4MB page */
2710 context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2712 case PT32E_ROOT_LEVEL:
2713 context->rsvd_bits_mask[0][2] =
2714 rsvd_bits(maxphyaddr, 63) |
2715 rsvd_bits(7, 8) | rsvd_bits(1, 2); /* PDPTE */
2716 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2717 rsvd_bits(maxphyaddr, 62); /* PDE */
2718 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2719 rsvd_bits(maxphyaddr, 62); /* PTE */
2720 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2721 rsvd_bits(maxphyaddr, 62) |
2722 rsvd_bits(13, 20); /* large page */
2723 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2725 case PT64_ROOT_LEVEL:
2726 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2727 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2728 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2729 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2730 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2731 rsvd_bits(maxphyaddr, 51);
2732 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2733 rsvd_bits(maxphyaddr, 51);
2734 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2735 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2736 rsvd_bits(maxphyaddr, 51) |
2738 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2739 rsvd_bits(maxphyaddr, 51) |
2740 rsvd_bits(13, 20); /* large page */
2741 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2746 static int paging64_init_context_common(struct kvm_vcpu *vcpu,
2747 struct kvm_mmu *context,
2750 context->nx = is_nx(vcpu);
2752 reset_rsvds_bits_mask(vcpu, context, level);
2754 ASSERT(is_pae(vcpu));
2755 context->new_cr3 = paging_new_cr3;
2756 context->page_fault = paging64_page_fault;
2757 context->gva_to_gpa = paging64_gva_to_gpa;
2758 context->prefetch_page = paging64_prefetch_page;
2759 context->sync_page = paging64_sync_page;
2760 context->invlpg = paging64_invlpg;
2761 context->free = paging_free;
2762 context->root_level = level;
2763 context->shadow_root_level = level;
2764 context->root_hpa = INVALID_PAGE;
2765 context->direct_map = false;
2769 static int paging64_init_context(struct kvm_vcpu *vcpu,
2770 struct kvm_mmu *context)
2772 return paging64_init_context_common(vcpu, context, PT64_ROOT_LEVEL);
2775 static int paging32_init_context(struct kvm_vcpu *vcpu,
2776 struct kvm_mmu *context)
2778 context->nx = false;
2780 reset_rsvds_bits_mask(vcpu, context, PT32_ROOT_LEVEL);
2782 context->new_cr3 = paging_new_cr3;
2783 context->page_fault = paging32_page_fault;
2784 context->gva_to_gpa = paging32_gva_to_gpa;
2785 context->free = paging_free;
2786 context->prefetch_page = paging32_prefetch_page;
2787 context->sync_page = paging32_sync_page;
2788 context->invlpg = paging32_invlpg;
2789 context->root_level = PT32_ROOT_LEVEL;
2790 context->shadow_root_level = PT32E_ROOT_LEVEL;
2791 context->root_hpa = INVALID_PAGE;
2792 context->direct_map = false;
2796 static int paging32E_init_context(struct kvm_vcpu *vcpu,
2797 struct kvm_mmu *context)
2799 return paging64_init_context_common(vcpu, context, PT32E_ROOT_LEVEL);
2802 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2804 struct kvm_mmu *context = vcpu->arch.walk_mmu;
2806 context->new_cr3 = nonpaging_new_cr3;
2807 context->page_fault = tdp_page_fault;
2808 context->free = nonpaging_free;
2809 context->prefetch_page = nonpaging_prefetch_page;
2810 context->sync_page = nonpaging_sync_page;
2811 context->invlpg = nonpaging_invlpg;
2812 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2813 context->root_hpa = INVALID_PAGE;
2814 context->direct_map = true;
2815 context->set_cr3 = kvm_x86_ops->set_tdp_cr3;
2816 context->get_cr3 = get_cr3;
2817 context->inject_page_fault = kvm_inject_page_fault;
2818 context->nx = is_nx(vcpu);
2820 if (!is_paging(vcpu)) {
2821 context->nx = false;
2822 context->gva_to_gpa = nonpaging_gva_to_gpa;
2823 context->root_level = 0;
2824 } else if (is_long_mode(vcpu)) {
2825 context->nx = is_nx(vcpu);
2826 reset_rsvds_bits_mask(vcpu, context, PT64_ROOT_LEVEL);
2827 context->gva_to_gpa = paging64_gva_to_gpa;
2828 context->root_level = PT64_ROOT_LEVEL;
2829 } else if (is_pae(vcpu)) {
2830 context->nx = is_nx(vcpu);
2831 reset_rsvds_bits_mask(vcpu, context, PT32E_ROOT_LEVEL);
2832 context->gva_to_gpa = paging64_gva_to_gpa;
2833 context->root_level = PT32E_ROOT_LEVEL;
2835 context->nx = false;
2836 reset_rsvds_bits_mask(vcpu, context, PT32_ROOT_LEVEL);
2837 context->gva_to_gpa = paging32_gva_to_gpa;
2838 context->root_level = PT32_ROOT_LEVEL;
2844 int kvm_init_shadow_mmu(struct kvm_vcpu *vcpu, struct kvm_mmu *context)
2848 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2850 if (!is_paging(vcpu))
2851 r = nonpaging_init_context(vcpu, context);
2852 else if (is_long_mode(vcpu))
2853 r = paging64_init_context(vcpu, context);
2854 else if (is_pae(vcpu))
2855 r = paging32E_init_context(vcpu, context);
2857 r = paging32_init_context(vcpu, context);
2859 vcpu->arch.mmu.base_role.cr4_pae = !!is_pae(vcpu);
2860 vcpu->arch.mmu.base_role.cr0_wp = is_write_protection(vcpu);
2864 EXPORT_SYMBOL_GPL(kvm_init_shadow_mmu);
2866 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2868 int r = kvm_init_shadow_mmu(vcpu, vcpu->arch.walk_mmu);
2870 vcpu->arch.walk_mmu->set_cr3 = kvm_x86_ops->set_cr3;
2871 vcpu->arch.walk_mmu->get_cr3 = get_cr3;
2872 vcpu->arch.walk_mmu->inject_page_fault = kvm_inject_page_fault;
2877 static int init_kvm_nested_mmu(struct kvm_vcpu *vcpu)
2879 struct kvm_mmu *g_context = &vcpu->arch.nested_mmu;
2881 g_context->get_cr3 = get_cr3;
2882 g_context->inject_page_fault = kvm_inject_page_fault;
2885 * Note that arch.mmu.gva_to_gpa translates l2_gva to l1_gpa. The
2886 * translation of l2_gpa to l1_gpa addresses is done using the
2887 * arch.nested_mmu.gva_to_gpa function. Basically the gva_to_gpa
2888 * functions between mmu and nested_mmu are swapped.
2890 if (!is_paging(vcpu)) {
2891 g_context->nx = false;
2892 g_context->root_level = 0;
2893 g_context->gva_to_gpa = nonpaging_gva_to_gpa_nested;
2894 } else if (is_long_mode(vcpu)) {
2895 g_context->nx = is_nx(vcpu);
2896 reset_rsvds_bits_mask(vcpu, g_context, PT64_ROOT_LEVEL);
2897 g_context->root_level = PT64_ROOT_LEVEL;
2898 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
2899 } else if (is_pae(vcpu)) {
2900 g_context->nx = is_nx(vcpu);
2901 reset_rsvds_bits_mask(vcpu, g_context, PT32E_ROOT_LEVEL);
2902 g_context->root_level = PT32E_ROOT_LEVEL;
2903 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
2905 g_context->nx = false;
2906 reset_rsvds_bits_mask(vcpu, g_context, PT32_ROOT_LEVEL);
2907 g_context->root_level = PT32_ROOT_LEVEL;
2908 g_context->gva_to_gpa = paging32_gva_to_gpa_nested;
2914 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2916 vcpu->arch.update_pte.pfn = bad_pfn;
2918 if (mmu_is_nested(vcpu))
2919 return init_kvm_nested_mmu(vcpu);
2920 else if (tdp_enabled)
2921 return init_kvm_tdp_mmu(vcpu);
2923 return init_kvm_softmmu(vcpu);
2926 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2929 if (VALID_PAGE(vcpu->arch.mmu.root_hpa))
2930 /* mmu.free() should set root_hpa = INVALID_PAGE */
2931 vcpu->arch.mmu.free(vcpu);
2934 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2936 destroy_kvm_mmu(vcpu);
2937 return init_kvm_mmu(vcpu);
2939 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2941 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2945 r = mmu_topup_memory_caches(vcpu);
2948 r = mmu_alloc_roots(vcpu);
2949 spin_lock(&vcpu->kvm->mmu_lock);
2950 mmu_sync_roots(vcpu);
2951 spin_unlock(&vcpu->kvm->mmu_lock);
2954 /* set_cr3() should ensure TLB has been flushed */
2955 vcpu->arch.mmu.set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2959 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2961 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2963 mmu_free_roots(vcpu);
2965 EXPORT_SYMBOL_GPL(kvm_mmu_unload);
2967 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2968 struct kvm_mmu_page *sp,
2972 struct kvm_mmu_page *child;
2975 if (is_shadow_present_pte(pte)) {
2976 if (is_last_spte(pte, sp->role.level))
2977 drop_spte(vcpu->kvm, spte, shadow_trap_nonpresent_pte);
2979 child = page_header(pte & PT64_BASE_ADDR_MASK);
2980 mmu_page_remove_parent_pte(child, spte);
2983 __set_spte(spte, shadow_trap_nonpresent_pte);
2984 if (is_large_pte(pte))
2985 --vcpu->kvm->stat.lpages;
2988 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2989 struct kvm_mmu_page *sp,
2993 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2994 ++vcpu->kvm->stat.mmu_pde_zapped;
2998 if (is_rsvd_bits_set(&vcpu->arch.mmu, *(u64 *)new, PT_PAGE_TABLE_LEVEL))
3001 ++vcpu->kvm->stat.mmu_pte_updated;
3002 if (!sp->role.cr4_pae)
3003 paging32_update_pte(vcpu, sp, spte, new);
3005 paging64_update_pte(vcpu, sp, spte, new);
3008 static bool need_remote_flush(u64 old, u64 new)
3010 if (!is_shadow_present_pte(old))
3012 if (!is_shadow_present_pte(new))
3014 if ((old ^ new) & PT64_BASE_ADDR_MASK)
3016 old ^= PT64_NX_MASK;
3017 new ^= PT64_NX_MASK;
3018 return (old & ~new & PT64_PERM_MASK) != 0;
3021 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, bool zap_page,
3022 bool remote_flush, bool local_flush)
3028 kvm_flush_remote_tlbs(vcpu->kvm);
3029 else if (local_flush)
3030 kvm_mmu_flush_tlb(vcpu);
3033 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
3035 u64 *spte = vcpu->arch.last_pte_updated;
3037 return !!(spte && (*spte & shadow_accessed_mask));
3040 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
3046 if (!is_present_gpte(gpte))
3048 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
3050 vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
3052 pfn = gfn_to_pfn(vcpu->kvm, gfn);
3054 if (is_error_pfn(pfn)) {
3055 kvm_release_pfn_clean(pfn);
3058 vcpu->arch.update_pte.gfn = gfn;
3059 vcpu->arch.update_pte.pfn = pfn;
3062 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
3064 u64 *spte = vcpu->arch.last_pte_updated;
3067 && vcpu->arch.last_pte_gfn == gfn
3068 && shadow_accessed_mask
3069 && !(*spte & shadow_accessed_mask)
3070 && is_shadow_present_pte(*spte))
3071 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
3074 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
3075 const u8 *new, int bytes,
3076 bool guest_initiated)
3078 gfn_t gfn = gpa >> PAGE_SHIFT;
3079 union kvm_mmu_page_role mask = { .word = 0 };
3080 struct kvm_mmu_page *sp;
3081 struct hlist_node *node;
3082 LIST_HEAD(invalid_list);
3085 unsigned offset = offset_in_page(gpa);
3087 unsigned page_offset;
3088 unsigned misaligned;
3095 bool remote_flush, local_flush, zap_page;
3097 zap_page = remote_flush = local_flush = false;
3099 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
3101 invlpg_counter = atomic_read(&vcpu->kvm->arch.invlpg_counter);
3104 * Assume that the pte write on a page table of the same type
3105 * as the current vcpu paging mode. This is nearly always true
3106 * (might be false while changing modes). Note it is verified later
3109 if ((is_pae(vcpu) && bytes == 4) || !new) {
3110 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
3115 r = kvm_read_guest(vcpu->kvm, gpa, &gentry, min(bytes, 8));
3118 new = (const u8 *)&gentry;
3123 gentry = *(const u32 *)new;
3126 gentry = *(const u64 *)new;
3133 mmu_guess_page_from_pte_write(vcpu, gpa, gentry);
3134 spin_lock(&vcpu->kvm->mmu_lock);
3135 if (atomic_read(&vcpu->kvm->arch.invlpg_counter) != invlpg_counter)
3137 kvm_mmu_access_page(vcpu, gfn);
3138 kvm_mmu_free_some_pages(vcpu);
3139 ++vcpu->kvm->stat.mmu_pte_write;
3140 trace_kvm_mmu_audit(vcpu, AUDIT_PRE_PTE_WRITE);
3141 if (guest_initiated) {
3142 if (gfn == vcpu->arch.last_pt_write_gfn
3143 && !last_updated_pte_accessed(vcpu)) {
3144 ++vcpu->arch.last_pt_write_count;
3145 if (vcpu->arch.last_pt_write_count >= 3)
3148 vcpu->arch.last_pt_write_gfn = gfn;
3149 vcpu->arch.last_pt_write_count = 1;
3150 vcpu->arch.last_pte_updated = NULL;
3154 mask.cr0_wp = mask.cr4_pae = mask.nxe = 1;
3155 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn, node) {
3156 pte_size = sp->role.cr4_pae ? 8 : 4;
3157 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
3158 misaligned |= bytes < 4;
3159 if (misaligned || flooded) {
3161 * Misaligned accesses are too much trouble to fix
3162 * up; also, they usually indicate a page is not used
3165 * If we're seeing too many writes to a page,
3166 * it may no longer be a page table, or we may be
3167 * forking, in which case it is better to unmap the
3170 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
3171 gpa, bytes, sp->role.word);
3172 zap_page |= !!kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
3174 ++vcpu->kvm->stat.mmu_flooded;
3177 page_offset = offset;
3178 level = sp->role.level;
3180 if (!sp->role.cr4_pae) {
3181 page_offset <<= 1; /* 32->64 */
3183 * A 32-bit pde maps 4MB while the shadow pdes map
3184 * only 2MB. So we need to double the offset again
3185 * and zap two pdes instead of one.
3187 if (level == PT32_ROOT_LEVEL) {
3188 page_offset &= ~7; /* kill rounding error */
3192 quadrant = page_offset >> PAGE_SHIFT;
3193 page_offset &= ~PAGE_MASK;
3194 if (quadrant != sp->role.quadrant)
3198 spte = &sp->spt[page_offset / sizeof(*spte)];
3201 mmu_pte_write_zap_pte(vcpu, sp, spte);
3203 !((sp->role.word ^ vcpu->arch.mmu.base_role.word)
3205 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
3206 if (!remote_flush && need_remote_flush(entry, *spte))
3207 remote_flush = true;
3211 mmu_pte_write_flush_tlb(vcpu, zap_page, remote_flush, local_flush);
3212 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
3213 trace_kvm_mmu_audit(vcpu, AUDIT_POST_PTE_WRITE);
3214 spin_unlock(&vcpu->kvm->mmu_lock);
3215 if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
3216 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
3217 vcpu->arch.update_pte.pfn = bad_pfn;
3221 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
3226 if (vcpu->arch.mmu.direct_map)
3229 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
3231 spin_lock(&vcpu->kvm->mmu_lock);
3232 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
3233 spin_unlock(&vcpu->kvm->mmu_lock);
3236 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
3238 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
3240 LIST_HEAD(invalid_list);
3242 while (kvm_mmu_available_pages(vcpu->kvm) < KVM_REFILL_PAGES &&
3243 !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
3244 struct kvm_mmu_page *sp;
3246 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
3247 struct kvm_mmu_page, link);
3248 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
3249 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
3250 ++vcpu->kvm->stat.mmu_recycled;
3254 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
3257 enum emulation_result er;
3259 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
3268 r = mmu_topup_memory_caches(vcpu);
3272 er = emulate_instruction(vcpu, cr2, error_code, 0);
3277 case EMULATE_DO_MMIO:
3278 ++vcpu->stat.mmio_exits;
3288 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
3290 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
3292 vcpu->arch.mmu.invlpg(vcpu, gva);
3293 kvm_mmu_flush_tlb(vcpu);
3294 ++vcpu->stat.invlpg;
3296 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
3298 void kvm_enable_tdp(void)
3302 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
3304 void kvm_disable_tdp(void)
3306 tdp_enabled = false;
3308 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
3310 static void free_mmu_pages(struct kvm_vcpu *vcpu)
3312 free_page((unsigned long)vcpu->arch.mmu.pae_root);
3313 if (vcpu->arch.mmu.lm_root != NULL)
3314 free_page((unsigned long)vcpu->arch.mmu.lm_root);
3317 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
3325 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
3326 * Therefore we need to allocate shadow page tables in the first
3327 * 4GB of memory, which happens to fit the DMA32 zone.
3329 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
3333 vcpu->arch.mmu.pae_root = page_address(page);
3334 for (i = 0; i < 4; ++i)
3335 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
3340 int kvm_mmu_create(struct kvm_vcpu *vcpu)
3343 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3345 return alloc_mmu_pages(vcpu);
3348 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
3351 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3353 return init_kvm_mmu(vcpu);
3356 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
3360 destroy_kvm_mmu(vcpu);
3361 free_mmu_pages(vcpu);
3362 mmu_free_memory_caches(vcpu);
3365 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
3367 struct kvm_mmu_page *sp;
3369 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
3373 if (!test_bit(slot, sp->slot_bitmap))
3377 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
3379 if (is_writable_pte(pt[i]))
3380 pt[i] &= ~PT_WRITABLE_MASK;
3382 kvm_flush_remote_tlbs(kvm);
3385 void kvm_mmu_zap_all(struct kvm *kvm)
3387 struct kvm_mmu_page *sp, *node;
3388 LIST_HEAD(invalid_list);
3390 spin_lock(&kvm->mmu_lock);
3392 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
3393 if (kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list))
3396 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3397 spin_unlock(&kvm->mmu_lock);
3400 static int kvm_mmu_remove_some_alloc_mmu_pages(struct kvm *kvm,
3401 struct list_head *invalid_list)
3403 struct kvm_mmu_page *page;
3405 page = container_of(kvm->arch.active_mmu_pages.prev,
3406 struct kvm_mmu_page, link);
3407 return kvm_mmu_prepare_zap_page(kvm, page, invalid_list);
3410 static int mmu_shrink(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
3413 struct kvm *kvm_freed = NULL;
3415 if (nr_to_scan == 0)
3418 spin_lock(&kvm_lock);
3420 list_for_each_entry(kvm, &vm_list, vm_list) {
3421 int idx, freed_pages;
3422 LIST_HEAD(invalid_list);
3424 idx = srcu_read_lock(&kvm->srcu);
3425 spin_lock(&kvm->mmu_lock);
3426 if (!kvm_freed && nr_to_scan > 0 &&
3427 kvm->arch.n_used_mmu_pages > 0) {
3428 freed_pages = kvm_mmu_remove_some_alloc_mmu_pages(kvm,
3434 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3435 spin_unlock(&kvm->mmu_lock);
3436 srcu_read_unlock(&kvm->srcu, idx);
3439 list_move_tail(&kvm_freed->vm_list, &vm_list);
3441 spin_unlock(&kvm_lock);
3444 return percpu_counter_read_positive(&kvm_total_used_mmu_pages);
3447 static struct shrinker mmu_shrinker = {
3448 .shrink = mmu_shrink,
3449 .seeks = DEFAULT_SEEKS * 10,
3452 static void mmu_destroy_caches(void)
3454 if (pte_chain_cache)
3455 kmem_cache_destroy(pte_chain_cache);
3456 if (rmap_desc_cache)
3457 kmem_cache_destroy(rmap_desc_cache);
3458 if (mmu_page_header_cache)
3459 kmem_cache_destroy(mmu_page_header_cache);
3462 void kvm_mmu_module_exit(void)
3464 mmu_destroy_caches();
3465 percpu_counter_destroy(&kvm_total_used_mmu_pages);
3466 unregister_shrinker(&mmu_shrinker);
3469 int kvm_mmu_module_init(void)
3471 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
3472 sizeof(struct kvm_pte_chain),
3474 if (!pte_chain_cache)
3476 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
3477 sizeof(struct kvm_rmap_desc),
3479 if (!rmap_desc_cache)
3482 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
3483 sizeof(struct kvm_mmu_page),
3485 if (!mmu_page_header_cache)
3488 if (percpu_counter_init(&kvm_total_used_mmu_pages, 0))
3491 register_shrinker(&mmu_shrinker);
3496 mmu_destroy_caches();
3501 * Caculate mmu pages needed for kvm.
3503 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3506 unsigned int nr_mmu_pages;
3507 unsigned int nr_pages = 0;
3508 struct kvm_memslots *slots;
3510 slots = kvm_memslots(kvm);
3512 for (i = 0; i < slots->nmemslots; i++)
3513 nr_pages += slots->memslots[i].npages;
3515 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3516 nr_mmu_pages = max(nr_mmu_pages,
3517 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3519 return nr_mmu_pages;
3522 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3525 if (len > buffer->len)
3530 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3535 ret = pv_mmu_peek_buffer(buffer, len);
3540 buffer->processed += len;
3544 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3545 gpa_t addr, gpa_t value)
3550 if (!is_long_mode(vcpu) && !is_pae(vcpu))
3553 r = mmu_topup_memory_caches(vcpu);
3557 if (!emulator_write_phys(vcpu, addr, &value, bytes))
3563 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3565 (void)kvm_set_cr3(vcpu, vcpu->arch.cr3);
3569 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3571 spin_lock(&vcpu->kvm->mmu_lock);
3572 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3573 spin_unlock(&vcpu->kvm->mmu_lock);
3577 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3578 struct kvm_pv_mmu_op_buffer *buffer)
3580 struct kvm_mmu_op_header *header;
3582 header = pv_mmu_peek_buffer(buffer, sizeof *header);
3585 switch (header->op) {
3586 case KVM_MMU_OP_WRITE_PTE: {
3587 struct kvm_mmu_op_write_pte *wpte;
3589 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3592 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3595 case KVM_MMU_OP_FLUSH_TLB: {
3596 struct kvm_mmu_op_flush_tlb *ftlb;
3598 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3601 return kvm_pv_mmu_flush_tlb(vcpu);
3603 case KVM_MMU_OP_RELEASE_PT: {
3604 struct kvm_mmu_op_release_pt *rpt;
3606 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3609 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3615 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3616 gpa_t addr, unsigned long *ret)
3619 struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3621 buffer->ptr = buffer->buf;
3622 buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3623 buffer->processed = 0;
3625 r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3629 while (buffer->len) {
3630 r = kvm_pv_mmu_op_one(vcpu, buffer);
3639 *ret = buffer->processed;
3643 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3645 struct kvm_shadow_walk_iterator iterator;
3648 spin_lock(&vcpu->kvm->mmu_lock);
3649 for_each_shadow_entry(vcpu, addr, iterator) {
3650 sptes[iterator.level-1] = *iterator.sptep;
3652 if (!is_shadow_present_pte(*iterator.sptep))
3655 spin_unlock(&vcpu->kvm->mmu_lock);
3659 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3661 #ifdef CONFIG_KVM_MMU_AUDIT
3662 #include "mmu_audit.c"