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 affiliates.
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,
61 char *audit_point_name[] = {
74 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
75 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
79 #define pgprintk(x...) do { } while (0)
80 #define rmap_printk(x...) do { } while (0)
86 module_param(dbg, bool, 0644);
89 static int oos_shadow = 1;
90 module_param(oos_shadow, bool, 0644);
93 #define ASSERT(x) do { } while (0)
97 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
98 __FILE__, __LINE__, #x); \
102 #define PTE_PREFETCH_NUM 8
104 #define PT_FIRST_AVAIL_BITS_SHIFT 9
105 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
107 #define PT64_LEVEL_BITS 9
109 #define PT64_LEVEL_SHIFT(level) \
110 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
112 #define PT64_LEVEL_MASK(level) \
113 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
115 #define PT64_INDEX(address, level)\
116 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
119 #define PT32_LEVEL_BITS 10
121 #define PT32_LEVEL_SHIFT(level) \
122 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
124 #define PT32_LEVEL_MASK(level) \
125 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
126 #define PT32_LVL_OFFSET_MASK(level) \
127 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
128 * PT32_LEVEL_BITS))) - 1))
130 #define PT32_INDEX(address, level)\
131 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
134 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
135 #define PT64_DIR_BASE_ADDR_MASK \
136 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
137 #define PT64_LVL_ADDR_MASK(level) \
138 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
139 * PT64_LEVEL_BITS))) - 1))
140 #define PT64_LVL_OFFSET_MASK(level) \
141 (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
142 * PT64_LEVEL_BITS))) - 1))
144 #define PT32_BASE_ADDR_MASK PAGE_MASK
145 #define PT32_DIR_BASE_ADDR_MASK \
146 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
147 #define PT32_LVL_ADDR_MASK(level) \
148 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
149 * PT32_LEVEL_BITS))) - 1))
151 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
156 #define ACC_EXEC_MASK 1
157 #define ACC_WRITE_MASK PT_WRITABLE_MASK
158 #define ACC_USER_MASK PT_USER_MASK
159 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
161 #include <trace/events/kvm.h>
163 #define CREATE_TRACE_POINTS
164 #include "mmutrace.h"
166 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
168 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
170 struct kvm_rmap_desc {
171 u64 *sptes[RMAP_EXT];
172 struct kvm_rmap_desc *more;
175 struct kvm_shadow_walk_iterator {
183 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
184 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
185 shadow_walk_okay(&(_walker)); \
186 shadow_walk_next(&(_walker)))
188 typedef void (*mmu_parent_walk_fn) (struct kvm_mmu_page *sp, u64 *spte);
190 static struct kmem_cache *pte_chain_cache;
191 static struct kmem_cache *rmap_desc_cache;
192 static struct kmem_cache *mmu_page_header_cache;
193 static struct percpu_counter kvm_total_used_mmu_pages;
195 static u64 __read_mostly shadow_trap_nonpresent_pte;
196 static u64 __read_mostly shadow_notrap_nonpresent_pte;
197 static u64 __read_mostly shadow_base_present_pte;
198 static u64 __read_mostly shadow_nx_mask;
199 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
200 static u64 __read_mostly shadow_user_mask;
201 static u64 __read_mostly shadow_accessed_mask;
202 static u64 __read_mostly shadow_dirty_mask;
204 static inline u64 rsvd_bits(int s, int e)
206 return ((1ULL << (e - s + 1)) - 1) << s;
209 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
211 shadow_trap_nonpresent_pte = trap_pte;
212 shadow_notrap_nonpresent_pte = notrap_pte;
214 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
216 void kvm_mmu_set_base_ptes(u64 base_pte)
218 shadow_base_present_pte = base_pte;
220 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
222 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
223 u64 dirty_mask, u64 nx_mask, u64 x_mask)
225 shadow_user_mask = user_mask;
226 shadow_accessed_mask = accessed_mask;
227 shadow_dirty_mask = dirty_mask;
228 shadow_nx_mask = nx_mask;
229 shadow_x_mask = x_mask;
231 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
233 static bool is_write_protection(struct kvm_vcpu *vcpu)
235 return kvm_read_cr0_bits(vcpu, X86_CR0_WP);
238 static int is_cpuid_PSE36(void)
243 static int is_nx(struct kvm_vcpu *vcpu)
245 return vcpu->arch.efer & EFER_NX;
248 static int is_shadow_present_pte(u64 pte)
250 return pte != shadow_trap_nonpresent_pte
251 && pte != shadow_notrap_nonpresent_pte;
254 static int is_large_pte(u64 pte)
256 return pte & PT_PAGE_SIZE_MASK;
259 static int is_writable_pte(unsigned long pte)
261 return pte & PT_WRITABLE_MASK;
264 static int is_dirty_gpte(unsigned long pte)
266 return pte & PT_DIRTY_MASK;
269 static int is_rmap_spte(u64 pte)
271 return is_shadow_present_pte(pte);
274 static int is_last_spte(u64 pte, int level)
276 if (level == PT_PAGE_TABLE_LEVEL)
278 if (is_large_pte(pte))
283 static pfn_t spte_to_pfn(u64 pte)
285 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
288 static gfn_t pse36_gfn_delta(u32 gpte)
290 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
292 return (gpte & PT32_DIR_PSE36_MASK) << shift;
295 static void __set_spte(u64 *sptep, u64 spte)
297 set_64bit(sptep, spte);
300 static u64 __xchg_spte(u64 *sptep, u64 new_spte)
303 return xchg(sptep, new_spte);
309 } while (cmpxchg64(sptep, old_spte, new_spte) != old_spte);
315 static bool spte_has_volatile_bits(u64 spte)
317 if (!shadow_accessed_mask)
320 if (!is_shadow_present_pte(spte))
323 if ((spte & shadow_accessed_mask) &&
324 (!is_writable_pte(spte) || (spte & shadow_dirty_mask)))
330 static bool spte_is_bit_cleared(u64 old_spte, u64 new_spte, u64 bit_mask)
332 return (old_spte & bit_mask) && !(new_spte & bit_mask);
335 static void update_spte(u64 *sptep, u64 new_spte)
337 u64 mask, old_spte = *sptep;
339 WARN_ON(!is_rmap_spte(new_spte));
341 new_spte |= old_spte & shadow_dirty_mask;
343 mask = shadow_accessed_mask;
344 if (is_writable_pte(old_spte))
345 mask |= shadow_dirty_mask;
347 if (!spte_has_volatile_bits(old_spte) || (new_spte & mask) == mask)
348 __set_spte(sptep, new_spte);
350 old_spte = __xchg_spte(sptep, new_spte);
352 if (!shadow_accessed_mask)
355 if (spte_is_bit_cleared(old_spte, new_spte, shadow_accessed_mask))
356 kvm_set_pfn_accessed(spte_to_pfn(old_spte));
357 if (spte_is_bit_cleared(old_spte, new_spte, shadow_dirty_mask))
358 kvm_set_pfn_dirty(spte_to_pfn(old_spte));
361 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
362 struct kmem_cache *base_cache, int min)
366 if (cache->nobjs >= min)
368 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
369 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
372 cache->objects[cache->nobjs++] = obj;
377 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc,
378 struct kmem_cache *cache)
381 kmem_cache_free(cache, mc->objects[--mc->nobjs]);
384 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
389 if (cache->nobjs >= min)
391 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
392 page = alloc_page(GFP_KERNEL);
395 cache->objects[cache->nobjs++] = page_address(page);
400 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
403 free_page((unsigned long)mc->objects[--mc->nobjs]);
406 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
410 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
414 r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
415 rmap_desc_cache, 4 + PTE_PREFETCH_NUM);
418 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
421 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
422 mmu_page_header_cache, 4);
427 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
429 mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache, pte_chain_cache);
430 mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache, rmap_desc_cache);
431 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
432 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache,
433 mmu_page_header_cache);
436 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
442 p = mc->objects[--mc->nobjs];
446 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
448 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
449 sizeof(struct kvm_pte_chain));
452 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
454 kmem_cache_free(pte_chain_cache, pc);
457 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
459 return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
460 sizeof(struct kvm_rmap_desc));
463 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
465 kmem_cache_free(rmap_desc_cache, rd);
468 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
470 if (!sp->role.direct)
471 return sp->gfns[index];
473 return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
476 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
479 BUG_ON(gfn != kvm_mmu_page_get_gfn(sp, index));
481 sp->gfns[index] = gfn;
485 * Return the pointer to the largepage write count for a given
486 * gfn, handling slots that are not large page aligned.
488 static int *slot_largepage_idx(gfn_t gfn,
489 struct kvm_memory_slot *slot,
494 idx = (gfn >> KVM_HPAGE_GFN_SHIFT(level)) -
495 (slot->base_gfn >> KVM_HPAGE_GFN_SHIFT(level));
496 return &slot->lpage_info[level - 2][idx].write_count;
499 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
501 struct kvm_memory_slot *slot;
505 slot = gfn_to_memslot(kvm, gfn);
506 for (i = PT_DIRECTORY_LEVEL;
507 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
508 write_count = slot_largepage_idx(gfn, slot, i);
513 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
515 struct kvm_memory_slot *slot;
519 slot = gfn_to_memslot(kvm, gfn);
520 for (i = PT_DIRECTORY_LEVEL;
521 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
522 write_count = slot_largepage_idx(gfn, slot, i);
524 WARN_ON(*write_count < 0);
528 static int has_wrprotected_page(struct kvm *kvm,
532 struct kvm_memory_slot *slot;
535 slot = gfn_to_memslot(kvm, gfn);
537 largepage_idx = slot_largepage_idx(gfn, slot, level);
538 return *largepage_idx;
544 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
546 unsigned long page_size;
549 page_size = kvm_host_page_size(kvm, gfn);
551 for (i = PT_PAGE_TABLE_LEVEL;
552 i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
553 if (page_size >= KVM_HPAGE_SIZE(i))
562 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
564 struct kvm_memory_slot *slot;
565 int host_level, level, max_level;
567 slot = gfn_to_memslot(vcpu->kvm, large_gfn);
568 if (slot && slot->dirty_bitmap)
569 return PT_PAGE_TABLE_LEVEL;
571 host_level = host_mapping_level(vcpu->kvm, large_gfn);
573 if (host_level == PT_PAGE_TABLE_LEVEL)
576 max_level = kvm_x86_ops->get_lpage_level() < host_level ?
577 kvm_x86_ops->get_lpage_level() : host_level;
579 for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
580 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
587 * Take gfn and return the reverse mapping to it.
590 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
592 struct kvm_memory_slot *slot;
595 slot = gfn_to_memslot(kvm, gfn);
596 if (likely(level == PT_PAGE_TABLE_LEVEL))
597 return &slot->rmap[gfn - slot->base_gfn];
599 idx = (gfn >> KVM_HPAGE_GFN_SHIFT(level)) -
600 (slot->base_gfn >> KVM_HPAGE_GFN_SHIFT(level));
602 return &slot->lpage_info[level - 2][idx].rmap_pde;
606 * Reverse mapping data structures:
608 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
609 * that points to page_address(page).
611 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
612 * containing more mappings.
614 * Returns the number of rmap entries before the spte was added or zero if
615 * the spte was not added.
618 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
620 struct kvm_mmu_page *sp;
621 struct kvm_rmap_desc *desc;
622 unsigned long *rmapp;
625 if (!is_rmap_spte(*spte))
627 sp = page_header(__pa(spte));
628 kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
629 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
631 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
632 *rmapp = (unsigned long)spte;
633 } else if (!(*rmapp & 1)) {
634 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
635 desc = mmu_alloc_rmap_desc(vcpu);
636 desc->sptes[0] = (u64 *)*rmapp;
637 desc->sptes[1] = spte;
638 *rmapp = (unsigned long)desc | 1;
641 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
642 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
643 while (desc->sptes[RMAP_EXT-1] && desc->more) {
647 if (desc->sptes[RMAP_EXT-1]) {
648 desc->more = mmu_alloc_rmap_desc(vcpu);
651 for (i = 0; desc->sptes[i]; ++i)
653 desc->sptes[i] = spte;
658 static void rmap_desc_remove_entry(unsigned long *rmapp,
659 struct kvm_rmap_desc *desc,
661 struct kvm_rmap_desc *prev_desc)
665 for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
667 desc->sptes[i] = desc->sptes[j];
668 desc->sptes[j] = NULL;
671 if (!prev_desc && !desc->more)
672 *rmapp = (unsigned long)desc->sptes[0];
675 prev_desc->more = desc->more;
677 *rmapp = (unsigned long)desc->more | 1;
678 mmu_free_rmap_desc(desc);
681 static void rmap_remove(struct kvm *kvm, u64 *spte)
683 struct kvm_rmap_desc *desc;
684 struct kvm_rmap_desc *prev_desc;
685 struct kvm_mmu_page *sp;
687 unsigned long *rmapp;
690 sp = page_header(__pa(spte));
691 gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
692 rmapp = gfn_to_rmap(kvm, gfn, sp->role.level);
694 printk(KERN_ERR "rmap_remove: %p 0->BUG\n", spte);
696 } else if (!(*rmapp & 1)) {
697 rmap_printk("rmap_remove: %p 1->0\n", spte);
698 if ((u64 *)*rmapp != spte) {
699 printk(KERN_ERR "rmap_remove: %p 1->BUG\n", spte);
704 rmap_printk("rmap_remove: %p many->many\n", spte);
705 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
708 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
709 if (desc->sptes[i] == spte) {
710 rmap_desc_remove_entry(rmapp,
718 pr_err("rmap_remove: %p many->many\n", spte);
723 static void set_spte_track_bits(u64 *sptep, u64 new_spte)
726 u64 old_spte = *sptep;
728 if (!spte_has_volatile_bits(old_spte))
729 __set_spte(sptep, new_spte);
731 old_spte = __xchg_spte(sptep, new_spte);
733 if (!is_rmap_spte(old_spte))
736 pfn = spte_to_pfn(old_spte);
737 if (!shadow_accessed_mask || old_spte & shadow_accessed_mask)
738 kvm_set_pfn_accessed(pfn);
739 if (!shadow_dirty_mask || (old_spte & shadow_dirty_mask))
740 kvm_set_pfn_dirty(pfn);
743 static void drop_spte(struct kvm *kvm, u64 *sptep, u64 new_spte)
745 set_spte_track_bits(sptep, new_spte);
746 rmap_remove(kvm, sptep);
749 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
751 struct kvm_rmap_desc *desc;
757 else if (!(*rmapp & 1)) {
759 return (u64 *)*rmapp;
762 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
765 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
766 if (prev_spte == spte)
767 return desc->sptes[i];
768 prev_spte = desc->sptes[i];
775 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
777 unsigned long *rmapp;
779 int i, write_protected = 0;
781 rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
783 spte = rmap_next(kvm, rmapp, NULL);
786 BUG_ON(!(*spte & PT_PRESENT_MASK));
787 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
788 if (is_writable_pte(*spte)) {
789 update_spte(spte, *spte & ~PT_WRITABLE_MASK);
792 spte = rmap_next(kvm, rmapp, spte);
795 /* check for huge page mappings */
796 for (i = PT_DIRECTORY_LEVEL;
797 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
798 rmapp = gfn_to_rmap(kvm, gfn, i);
799 spte = rmap_next(kvm, rmapp, NULL);
802 BUG_ON(!(*spte & PT_PRESENT_MASK));
803 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
804 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
805 if (is_writable_pte(*spte)) {
807 shadow_trap_nonpresent_pte);
812 spte = rmap_next(kvm, rmapp, spte);
816 return write_protected;
819 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
823 int need_tlb_flush = 0;
825 while ((spte = rmap_next(kvm, rmapp, NULL))) {
826 BUG_ON(!(*spte & PT_PRESENT_MASK));
827 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
828 drop_spte(kvm, spte, shadow_trap_nonpresent_pte);
831 return need_tlb_flush;
834 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
839 pte_t *ptep = (pte_t *)data;
842 WARN_ON(pte_huge(*ptep));
843 new_pfn = pte_pfn(*ptep);
844 spte = rmap_next(kvm, rmapp, NULL);
846 BUG_ON(!is_shadow_present_pte(*spte));
847 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
849 if (pte_write(*ptep)) {
850 drop_spte(kvm, spte, shadow_trap_nonpresent_pte);
851 spte = rmap_next(kvm, rmapp, NULL);
853 new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
854 new_spte |= (u64)new_pfn << PAGE_SHIFT;
856 new_spte &= ~PT_WRITABLE_MASK;
857 new_spte &= ~SPTE_HOST_WRITEABLE;
858 new_spte &= ~shadow_accessed_mask;
859 set_spte_track_bits(spte, new_spte);
860 spte = rmap_next(kvm, rmapp, spte);
864 kvm_flush_remote_tlbs(kvm);
869 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
871 int (*handler)(struct kvm *kvm, unsigned long *rmapp,
877 struct kvm_memslots *slots;
879 slots = kvm_memslots(kvm);
881 for (i = 0; i < slots->nmemslots; i++) {
882 struct kvm_memory_slot *memslot = &slots->memslots[i];
883 unsigned long start = memslot->userspace_addr;
886 end = start + (memslot->npages << PAGE_SHIFT);
887 if (hva >= start && hva < end) {
888 gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
890 ret = handler(kvm, &memslot->rmap[gfn_offset], data);
892 for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
896 sh = KVM_HPAGE_GFN_SHIFT(PT_DIRECTORY_LEVEL+j);
897 idx = ((memslot->base_gfn+gfn_offset) >> sh) -
898 (memslot->base_gfn >> sh);
900 &memslot->lpage_info[j][idx].rmap_pde,
903 trace_kvm_age_page(hva, memslot, ret);
911 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
913 return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
916 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
918 kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
921 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
928 * Emulate the accessed bit for EPT, by checking if this page has
929 * an EPT mapping, and clearing it if it does. On the next access,
930 * a new EPT mapping will be established.
931 * This has some overhead, but not as much as the cost of swapping
932 * out actively used pages or breaking up actively used hugepages.
934 if (!shadow_accessed_mask)
935 return kvm_unmap_rmapp(kvm, rmapp, data);
937 spte = rmap_next(kvm, rmapp, NULL);
941 BUG_ON(!(_spte & PT_PRESENT_MASK));
942 _young = _spte & PT_ACCESSED_MASK;
945 clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
947 spte = rmap_next(kvm, rmapp, spte);
952 #define RMAP_RECYCLE_THRESHOLD 1000
954 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
956 unsigned long *rmapp;
957 struct kvm_mmu_page *sp;
959 sp = page_header(__pa(spte));
961 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
963 kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
964 kvm_flush_remote_tlbs(vcpu->kvm);
967 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
969 return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
973 static int is_empty_shadow_page(u64 *spt)
978 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
979 if (is_shadow_present_pte(*pos)) {
980 printk(KERN_ERR "%s: %p %llx\n", __func__,
989 * This value is the sum of all of the kvm instances's
990 * kvm->arch.n_used_mmu_pages values. We need a global,
991 * aggregate version in order to make the slab shrinker
994 static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, int nr)
996 kvm->arch.n_used_mmu_pages += nr;
997 percpu_counter_add(&kvm_total_used_mmu_pages, nr);
1000 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1002 ASSERT(is_empty_shadow_page(sp->spt));
1003 hlist_del(&sp->hash_link);
1004 list_del(&sp->link);
1005 __free_page(virt_to_page(sp->spt));
1006 if (!sp->role.direct)
1007 __free_page(virt_to_page(sp->gfns));
1008 kmem_cache_free(mmu_page_header_cache, sp);
1009 kvm_mod_used_mmu_pages(kvm, -1);
1012 static unsigned kvm_page_table_hashfn(gfn_t gfn)
1014 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
1017 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
1018 u64 *parent_pte, int direct)
1020 struct kvm_mmu_page *sp;
1022 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
1023 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
1025 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache,
1027 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
1028 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
1029 bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
1030 sp->multimapped = 0;
1031 sp->parent_pte = parent_pte;
1032 kvm_mod_used_mmu_pages(vcpu->kvm, +1);
1036 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
1037 struct kvm_mmu_page *sp, u64 *parent_pte)
1039 struct kvm_pte_chain *pte_chain;
1040 struct hlist_node *node;
1045 if (!sp->multimapped) {
1046 u64 *old = sp->parent_pte;
1049 sp->parent_pte = parent_pte;
1052 sp->multimapped = 1;
1053 pte_chain = mmu_alloc_pte_chain(vcpu);
1054 INIT_HLIST_HEAD(&sp->parent_ptes);
1055 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
1056 pte_chain->parent_ptes[0] = old;
1058 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
1059 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
1061 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
1062 if (!pte_chain->parent_ptes[i]) {
1063 pte_chain->parent_ptes[i] = parent_pte;
1067 pte_chain = mmu_alloc_pte_chain(vcpu);
1069 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
1070 pte_chain->parent_ptes[0] = parent_pte;
1073 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
1076 struct kvm_pte_chain *pte_chain;
1077 struct hlist_node *node;
1080 if (!sp->multimapped) {
1081 BUG_ON(sp->parent_pte != parent_pte);
1082 sp->parent_pte = NULL;
1085 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1086 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1087 if (!pte_chain->parent_ptes[i])
1089 if (pte_chain->parent_ptes[i] != parent_pte)
1091 while (i + 1 < NR_PTE_CHAIN_ENTRIES
1092 && pte_chain->parent_ptes[i + 1]) {
1093 pte_chain->parent_ptes[i]
1094 = pte_chain->parent_ptes[i + 1];
1097 pte_chain->parent_ptes[i] = NULL;
1099 hlist_del(&pte_chain->link);
1100 mmu_free_pte_chain(pte_chain);
1101 if (hlist_empty(&sp->parent_ptes)) {
1102 sp->multimapped = 0;
1103 sp->parent_pte = NULL;
1111 static void mmu_parent_walk(struct kvm_mmu_page *sp, mmu_parent_walk_fn fn)
1113 struct kvm_pte_chain *pte_chain;
1114 struct hlist_node *node;
1115 struct kvm_mmu_page *parent_sp;
1118 if (!sp->multimapped && sp->parent_pte) {
1119 parent_sp = page_header(__pa(sp->parent_pte));
1120 fn(parent_sp, sp->parent_pte);
1124 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1125 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1126 u64 *spte = pte_chain->parent_ptes[i];
1130 parent_sp = page_header(__pa(spte));
1131 fn(parent_sp, spte);
1135 static void mark_unsync(struct kvm_mmu_page *sp, u64 *spte);
1136 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1138 mmu_parent_walk(sp, mark_unsync);
1141 static void mark_unsync(struct kvm_mmu_page *sp, u64 *spte)
1145 index = spte - sp->spt;
1146 if (__test_and_set_bit(index, sp->unsync_child_bitmap))
1148 if (sp->unsync_children++)
1150 kvm_mmu_mark_parents_unsync(sp);
1153 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1154 struct kvm_mmu_page *sp)
1158 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1159 sp->spt[i] = shadow_trap_nonpresent_pte;
1162 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1163 struct kvm_mmu_page *sp, bool clear_unsync)
1168 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1172 #define KVM_PAGE_ARRAY_NR 16
1174 struct kvm_mmu_pages {
1175 struct mmu_page_and_offset {
1176 struct kvm_mmu_page *sp;
1178 } page[KVM_PAGE_ARRAY_NR];
1182 #define for_each_unsync_children(bitmap, idx) \
1183 for (idx = find_first_bit(bitmap, 512); \
1185 idx = find_next_bit(bitmap, 512, idx+1))
1187 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1193 for (i=0; i < pvec->nr; i++)
1194 if (pvec->page[i].sp == sp)
1197 pvec->page[pvec->nr].sp = sp;
1198 pvec->page[pvec->nr].idx = idx;
1200 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1203 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1204 struct kvm_mmu_pages *pvec)
1206 int i, ret, nr_unsync_leaf = 0;
1208 for_each_unsync_children(sp->unsync_child_bitmap, i) {
1209 struct kvm_mmu_page *child;
1210 u64 ent = sp->spt[i];
1212 if (!is_shadow_present_pte(ent) || is_large_pte(ent))
1213 goto clear_child_bitmap;
1215 child = page_header(ent & PT64_BASE_ADDR_MASK);
1217 if (child->unsync_children) {
1218 if (mmu_pages_add(pvec, child, i))
1221 ret = __mmu_unsync_walk(child, pvec);
1223 goto clear_child_bitmap;
1225 nr_unsync_leaf += ret;
1228 } else if (child->unsync) {
1230 if (mmu_pages_add(pvec, child, i))
1233 goto clear_child_bitmap;
1238 __clear_bit(i, sp->unsync_child_bitmap);
1239 sp->unsync_children--;
1240 WARN_ON((int)sp->unsync_children < 0);
1244 return nr_unsync_leaf;
1247 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1248 struct kvm_mmu_pages *pvec)
1250 if (!sp->unsync_children)
1253 mmu_pages_add(pvec, sp, 0);
1254 return __mmu_unsync_walk(sp, pvec);
1257 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1259 WARN_ON(!sp->unsync);
1260 trace_kvm_mmu_sync_page(sp);
1262 --kvm->stat.mmu_unsync;
1265 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1266 struct list_head *invalid_list);
1267 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1268 struct list_head *invalid_list);
1270 #define for_each_gfn_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)) {} else
1275 #define for_each_gfn_indirect_valid_sp(kvm, sp, gfn, pos) \
1276 hlist_for_each_entry(sp, pos, \
1277 &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link) \
1278 if ((sp)->gfn != (gfn) || (sp)->role.direct || \
1279 (sp)->role.invalid) {} else
1281 /* @sp->gfn should be write-protected at the call site */
1282 static int __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1283 struct list_head *invalid_list, bool clear_unsync)
1285 if (sp->role.cr4_pae != !!is_pae(vcpu)) {
1286 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1291 kvm_unlink_unsync_page(vcpu->kvm, sp);
1293 if (vcpu->arch.mmu.sync_page(vcpu, sp, clear_unsync)) {
1294 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1298 kvm_mmu_flush_tlb(vcpu);
1302 static int kvm_sync_page_transient(struct kvm_vcpu *vcpu,
1303 struct kvm_mmu_page *sp)
1305 LIST_HEAD(invalid_list);
1308 ret = __kvm_sync_page(vcpu, sp, &invalid_list, false);
1310 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1315 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1316 struct list_head *invalid_list)
1318 return __kvm_sync_page(vcpu, sp, invalid_list, true);
1321 /* @gfn should be write-protected at the call site */
1322 static void kvm_sync_pages(struct kvm_vcpu *vcpu, gfn_t gfn)
1324 struct kvm_mmu_page *s;
1325 struct hlist_node *node;
1326 LIST_HEAD(invalid_list);
1329 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1333 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1334 if ((s->role.cr4_pae != !!is_pae(vcpu)) ||
1335 (vcpu->arch.mmu.sync_page(vcpu, s, true))) {
1336 kvm_mmu_prepare_zap_page(vcpu->kvm, s, &invalid_list);
1339 kvm_unlink_unsync_page(vcpu->kvm, s);
1343 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1345 kvm_mmu_flush_tlb(vcpu);
1348 struct mmu_page_path {
1349 struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1350 unsigned int idx[PT64_ROOT_LEVEL-1];
1353 #define for_each_sp(pvec, sp, parents, i) \
1354 for (i = mmu_pages_next(&pvec, &parents, -1), \
1355 sp = pvec.page[i].sp; \
1356 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1357 i = mmu_pages_next(&pvec, &parents, i))
1359 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1360 struct mmu_page_path *parents,
1365 for (n = i+1; n < pvec->nr; n++) {
1366 struct kvm_mmu_page *sp = pvec->page[n].sp;
1368 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1369 parents->idx[0] = pvec->page[n].idx;
1373 parents->parent[sp->role.level-2] = sp;
1374 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1380 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1382 struct kvm_mmu_page *sp;
1383 unsigned int level = 0;
1386 unsigned int idx = parents->idx[level];
1388 sp = parents->parent[level];
1392 --sp->unsync_children;
1393 WARN_ON((int)sp->unsync_children < 0);
1394 __clear_bit(idx, sp->unsync_child_bitmap);
1396 } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1399 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1400 struct mmu_page_path *parents,
1401 struct kvm_mmu_pages *pvec)
1403 parents->parent[parent->role.level-1] = NULL;
1407 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1408 struct kvm_mmu_page *parent)
1411 struct kvm_mmu_page *sp;
1412 struct mmu_page_path parents;
1413 struct kvm_mmu_pages pages;
1414 LIST_HEAD(invalid_list);
1416 kvm_mmu_pages_init(parent, &parents, &pages);
1417 while (mmu_unsync_walk(parent, &pages)) {
1420 for_each_sp(pages, sp, parents, i)
1421 protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1424 kvm_flush_remote_tlbs(vcpu->kvm);
1426 for_each_sp(pages, sp, parents, i) {
1427 kvm_sync_page(vcpu, sp, &invalid_list);
1428 mmu_pages_clear_parents(&parents);
1430 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1431 cond_resched_lock(&vcpu->kvm->mmu_lock);
1432 kvm_mmu_pages_init(parent, &parents, &pages);
1436 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1444 union kvm_mmu_page_role role;
1446 struct kvm_mmu_page *sp;
1447 struct hlist_node *node;
1448 bool need_sync = false;
1450 role = vcpu->arch.mmu.base_role;
1452 role.direct = direct;
1455 role.access = access;
1456 if (!vcpu->arch.mmu.direct_map
1457 && vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1458 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1459 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1460 role.quadrant = quadrant;
1462 for_each_gfn_sp(vcpu->kvm, sp, gfn, node) {
1463 if (!need_sync && sp->unsync)
1466 if (sp->role.word != role.word)
1469 if (sp->unsync && kvm_sync_page_transient(vcpu, sp))
1472 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1473 if (sp->unsync_children) {
1474 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
1475 kvm_mmu_mark_parents_unsync(sp);
1476 } else if (sp->unsync)
1477 kvm_mmu_mark_parents_unsync(sp);
1479 trace_kvm_mmu_get_page(sp, false);
1482 ++vcpu->kvm->stat.mmu_cache_miss;
1483 sp = kvm_mmu_alloc_page(vcpu, parent_pte, direct);
1488 hlist_add_head(&sp->hash_link,
1489 &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
1491 if (rmap_write_protect(vcpu->kvm, gfn))
1492 kvm_flush_remote_tlbs(vcpu->kvm);
1493 if (level > PT_PAGE_TABLE_LEVEL && need_sync)
1494 kvm_sync_pages(vcpu, gfn);
1496 account_shadowed(vcpu->kvm, gfn);
1498 if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1499 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1501 nonpaging_prefetch_page(vcpu, sp);
1502 trace_kvm_mmu_get_page(sp, true);
1506 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1507 struct kvm_vcpu *vcpu, u64 addr)
1509 iterator->addr = addr;
1510 iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1511 iterator->level = vcpu->arch.mmu.shadow_root_level;
1513 if (iterator->level == PT64_ROOT_LEVEL &&
1514 vcpu->arch.mmu.root_level < PT64_ROOT_LEVEL &&
1515 !vcpu->arch.mmu.direct_map)
1518 if (iterator->level == PT32E_ROOT_LEVEL) {
1519 iterator->shadow_addr
1520 = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1521 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1523 if (!iterator->shadow_addr)
1524 iterator->level = 0;
1528 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1530 if (iterator->level < PT_PAGE_TABLE_LEVEL)
1533 if (iterator->level == PT_PAGE_TABLE_LEVEL)
1534 if (is_large_pte(*iterator->sptep))
1537 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1538 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1542 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1544 iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1548 static void link_shadow_page(u64 *sptep, struct kvm_mmu_page *sp)
1552 spte = __pa(sp->spt)
1553 | PT_PRESENT_MASK | PT_ACCESSED_MASK
1554 | PT_WRITABLE_MASK | PT_USER_MASK;
1555 __set_spte(sptep, spte);
1558 static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1560 if (is_large_pte(*sptep)) {
1561 drop_spte(vcpu->kvm, sptep, shadow_trap_nonpresent_pte);
1562 kvm_flush_remote_tlbs(vcpu->kvm);
1566 static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1567 unsigned direct_access)
1569 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
1570 struct kvm_mmu_page *child;
1573 * For the direct sp, if the guest pte's dirty bit
1574 * changed form clean to dirty, it will corrupt the
1575 * sp's access: allow writable in the read-only sp,
1576 * so we should update the spte at this point to get
1577 * a new sp with the correct access.
1579 child = page_header(*sptep & PT64_BASE_ADDR_MASK);
1580 if (child->role.access == direct_access)
1583 mmu_page_remove_parent_pte(child, sptep);
1584 __set_spte(sptep, shadow_trap_nonpresent_pte);
1585 kvm_flush_remote_tlbs(vcpu->kvm);
1589 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1590 struct kvm_mmu_page *sp)
1598 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1601 if (is_shadow_present_pte(ent)) {
1602 if (!is_last_spte(ent, sp->role.level)) {
1603 ent &= PT64_BASE_ADDR_MASK;
1604 mmu_page_remove_parent_pte(page_header(ent),
1607 if (is_large_pte(ent))
1609 drop_spte(kvm, &pt[i],
1610 shadow_trap_nonpresent_pte);
1613 pt[i] = shadow_trap_nonpresent_pte;
1617 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1619 mmu_page_remove_parent_pte(sp, parent_pte);
1622 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1625 struct kvm_vcpu *vcpu;
1627 kvm_for_each_vcpu(i, vcpu, kvm)
1628 vcpu->arch.last_pte_updated = NULL;
1631 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1635 while (sp->multimapped || sp->parent_pte) {
1636 if (!sp->multimapped)
1637 parent_pte = sp->parent_pte;
1639 struct kvm_pte_chain *chain;
1641 chain = container_of(sp->parent_ptes.first,
1642 struct kvm_pte_chain, link);
1643 parent_pte = chain->parent_ptes[0];
1645 BUG_ON(!parent_pte);
1646 kvm_mmu_put_page(sp, parent_pte);
1647 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1651 static int mmu_zap_unsync_children(struct kvm *kvm,
1652 struct kvm_mmu_page *parent,
1653 struct list_head *invalid_list)
1656 struct mmu_page_path parents;
1657 struct kvm_mmu_pages pages;
1659 if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1662 kvm_mmu_pages_init(parent, &parents, &pages);
1663 while (mmu_unsync_walk(parent, &pages)) {
1664 struct kvm_mmu_page *sp;
1666 for_each_sp(pages, sp, parents, i) {
1667 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
1668 mmu_pages_clear_parents(&parents);
1671 kvm_mmu_pages_init(parent, &parents, &pages);
1677 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1678 struct list_head *invalid_list)
1682 trace_kvm_mmu_prepare_zap_page(sp);
1683 ++kvm->stat.mmu_shadow_zapped;
1684 ret = mmu_zap_unsync_children(kvm, sp, invalid_list);
1685 kvm_mmu_page_unlink_children(kvm, sp);
1686 kvm_mmu_unlink_parents(kvm, sp);
1687 if (!sp->role.invalid && !sp->role.direct)
1688 unaccount_shadowed(kvm, sp->gfn);
1690 kvm_unlink_unsync_page(kvm, sp);
1691 if (!sp->root_count) {
1694 list_move(&sp->link, invalid_list);
1696 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1697 kvm_reload_remote_mmus(kvm);
1700 sp->role.invalid = 1;
1701 kvm_mmu_reset_last_pte_updated(kvm);
1705 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1706 struct list_head *invalid_list)
1708 struct kvm_mmu_page *sp;
1710 if (list_empty(invalid_list))
1713 kvm_flush_remote_tlbs(kvm);
1716 sp = list_first_entry(invalid_list, struct kvm_mmu_page, link);
1717 WARN_ON(!sp->role.invalid || sp->root_count);
1718 kvm_mmu_free_page(kvm, sp);
1719 } while (!list_empty(invalid_list));
1724 * Changing the number of mmu pages allocated to the vm
1725 * Note: if goal_nr_mmu_pages is too small, you will get dead lock
1727 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int goal_nr_mmu_pages)
1729 LIST_HEAD(invalid_list);
1731 * If we set the number of mmu pages to be smaller be than the
1732 * number of actived pages , we must to free some mmu pages before we
1736 if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) {
1737 while (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages &&
1738 !list_empty(&kvm->arch.active_mmu_pages)) {
1739 struct kvm_mmu_page *page;
1741 page = container_of(kvm->arch.active_mmu_pages.prev,
1742 struct kvm_mmu_page, link);
1743 kvm_mmu_prepare_zap_page(kvm, page, &invalid_list);
1744 kvm_mmu_commit_zap_page(kvm, &invalid_list);
1746 goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages;
1749 kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages;
1752 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1754 struct kvm_mmu_page *sp;
1755 struct hlist_node *node;
1756 LIST_HEAD(invalid_list);
1759 pgprintk("%s: looking for gfn %llx\n", __func__, gfn);
1762 for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
1763 pgprintk("%s: gfn %llx role %x\n", __func__, gfn,
1766 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
1768 kvm_mmu_commit_zap_page(kvm, &invalid_list);
1772 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1774 struct kvm_mmu_page *sp;
1775 struct hlist_node *node;
1776 LIST_HEAD(invalid_list);
1778 for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
1779 pgprintk("%s: zap %llx %x\n",
1780 __func__, gfn, sp->role.word);
1781 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
1783 kvm_mmu_commit_zap_page(kvm, &invalid_list);
1786 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1788 int slot = memslot_id(kvm, gfn);
1789 struct kvm_mmu_page *sp = page_header(__pa(pte));
1791 __set_bit(slot, sp->slot_bitmap);
1794 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1799 if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1802 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1803 if (pt[i] == shadow_notrap_nonpresent_pte)
1804 __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1809 * The function is based on mtrr_type_lookup() in
1810 * arch/x86/kernel/cpu/mtrr/generic.c
1812 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1817 u8 prev_match, curr_match;
1818 int num_var_ranges = KVM_NR_VAR_MTRR;
1820 if (!mtrr_state->enabled)
1823 /* Make end inclusive end, instead of exclusive */
1826 /* Look in fixed ranges. Just return the type as per start */
1827 if (mtrr_state->have_fixed && (start < 0x100000)) {
1830 if (start < 0x80000) {
1832 idx += (start >> 16);
1833 return mtrr_state->fixed_ranges[idx];
1834 } else if (start < 0xC0000) {
1836 idx += ((start - 0x80000) >> 14);
1837 return mtrr_state->fixed_ranges[idx];
1838 } else if (start < 0x1000000) {
1840 idx += ((start - 0xC0000) >> 12);
1841 return mtrr_state->fixed_ranges[idx];
1846 * Look in variable ranges
1847 * Look of multiple ranges matching this address and pick type
1848 * as per MTRR precedence
1850 if (!(mtrr_state->enabled & 2))
1851 return mtrr_state->def_type;
1854 for (i = 0; i < num_var_ranges; ++i) {
1855 unsigned short start_state, end_state;
1857 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1860 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1861 (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1862 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1863 (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1865 start_state = ((start & mask) == (base & mask));
1866 end_state = ((end & mask) == (base & mask));
1867 if (start_state != end_state)
1870 if ((start & mask) != (base & mask))
1873 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1874 if (prev_match == 0xFF) {
1875 prev_match = curr_match;
1879 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1880 curr_match == MTRR_TYPE_UNCACHABLE)
1881 return MTRR_TYPE_UNCACHABLE;
1883 if ((prev_match == MTRR_TYPE_WRBACK &&
1884 curr_match == MTRR_TYPE_WRTHROUGH) ||
1885 (prev_match == MTRR_TYPE_WRTHROUGH &&
1886 curr_match == MTRR_TYPE_WRBACK)) {
1887 prev_match = MTRR_TYPE_WRTHROUGH;
1888 curr_match = MTRR_TYPE_WRTHROUGH;
1891 if (prev_match != curr_match)
1892 return MTRR_TYPE_UNCACHABLE;
1895 if (prev_match != 0xFF)
1898 return mtrr_state->def_type;
1901 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1905 mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1906 (gfn << PAGE_SHIFT) + PAGE_SIZE);
1907 if (mtrr == 0xfe || mtrr == 0xff)
1908 mtrr = MTRR_TYPE_WRBACK;
1911 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1913 static void __kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1915 trace_kvm_mmu_unsync_page(sp);
1916 ++vcpu->kvm->stat.mmu_unsync;
1919 kvm_mmu_mark_parents_unsync(sp);
1920 mmu_convert_notrap(sp);
1923 static void kvm_unsync_pages(struct kvm_vcpu *vcpu, gfn_t gfn)
1925 struct kvm_mmu_page *s;
1926 struct hlist_node *node;
1928 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1931 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1932 __kvm_unsync_page(vcpu, s);
1936 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1939 struct kvm_mmu_page *s;
1940 struct hlist_node *node;
1941 bool need_unsync = false;
1943 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1947 if (s->role.level != PT_PAGE_TABLE_LEVEL)
1950 if (!need_unsync && !s->unsync) {
1957 kvm_unsync_pages(vcpu, gfn);
1961 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1962 unsigned pte_access, int user_fault,
1963 int write_fault, int dirty, int level,
1964 gfn_t gfn, pfn_t pfn, bool speculative,
1965 bool can_unsync, bool reset_host_protection)
1971 * We don't set the accessed bit, since we sometimes want to see
1972 * whether the guest actually used the pte (in order to detect
1975 spte = shadow_base_present_pte;
1977 spte |= shadow_accessed_mask;
1979 pte_access &= ~ACC_WRITE_MASK;
1980 if (pte_access & ACC_EXEC_MASK)
1981 spte |= shadow_x_mask;
1983 spte |= shadow_nx_mask;
1984 if (pte_access & ACC_USER_MASK)
1985 spte |= shadow_user_mask;
1986 if (level > PT_PAGE_TABLE_LEVEL)
1987 spte |= PT_PAGE_SIZE_MASK;
1989 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1990 kvm_is_mmio_pfn(pfn));
1992 if (reset_host_protection)
1993 spte |= SPTE_HOST_WRITEABLE;
1995 spte |= (u64)pfn << PAGE_SHIFT;
1997 if ((pte_access & ACC_WRITE_MASK)
1998 || (!vcpu->arch.mmu.direct_map && write_fault
1999 && !is_write_protection(vcpu) && !user_fault)) {
2001 if (level > PT_PAGE_TABLE_LEVEL &&
2002 has_wrprotected_page(vcpu->kvm, gfn, level)) {
2004 drop_spte(vcpu->kvm, sptep, shadow_trap_nonpresent_pte);
2008 spte |= PT_WRITABLE_MASK;
2010 if (!vcpu->arch.mmu.direct_map
2011 && !(pte_access & ACC_WRITE_MASK))
2012 spte &= ~PT_USER_MASK;
2015 * Optimization: for pte sync, if spte was writable the hash
2016 * lookup is unnecessary (and expensive). Write protection
2017 * is responsibility of mmu_get_page / kvm_sync_page.
2018 * Same reasoning can be applied to dirty page accounting.
2020 if (!can_unsync && is_writable_pte(*sptep))
2023 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
2024 pgprintk("%s: found shadow page for %llx, marking ro\n",
2027 pte_access &= ~ACC_WRITE_MASK;
2028 if (is_writable_pte(spte))
2029 spte &= ~PT_WRITABLE_MASK;
2033 if (pte_access & ACC_WRITE_MASK)
2034 mark_page_dirty(vcpu->kvm, gfn);
2037 update_spte(sptep, spte);
2042 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2043 unsigned pt_access, unsigned pte_access,
2044 int user_fault, int write_fault, int dirty,
2045 int *ptwrite, int level, gfn_t gfn,
2046 pfn_t pfn, bool speculative,
2047 bool reset_host_protection)
2049 int was_rmapped = 0;
2052 pgprintk("%s: spte %llx access %x write_fault %d"
2053 " user_fault %d gfn %llx\n",
2054 __func__, *sptep, pt_access,
2055 write_fault, user_fault, gfn);
2057 if (is_rmap_spte(*sptep)) {
2059 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
2060 * the parent of the now unreachable PTE.
2062 if (level > PT_PAGE_TABLE_LEVEL &&
2063 !is_large_pte(*sptep)) {
2064 struct kvm_mmu_page *child;
2067 child = page_header(pte & PT64_BASE_ADDR_MASK);
2068 mmu_page_remove_parent_pte(child, sptep);
2069 __set_spte(sptep, shadow_trap_nonpresent_pte);
2070 kvm_flush_remote_tlbs(vcpu->kvm);
2071 } else if (pfn != spte_to_pfn(*sptep)) {
2072 pgprintk("hfn old %llx new %llx\n",
2073 spte_to_pfn(*sptep), pfn);
2074 drop_spte(vcpu->kvm, sptep, shadow_trap_nonpresent_pte);
2075 kvm_flush_remote_tlbs(vcpu->kvm);
2080 if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
2081 dirty, level, gfn, pfn, speculative, true,
2082 reset_host_protection)) {
2085 kvm_mmu_flush_tlb(vcpu);
2088 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
2089 pgprintk("instantiating %s PTE (%s) at %llx (%llx) addr %p\n",
2090 is_large_pte(*sptep)? "2MB" : "4kB",
2091 *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
2093 if (!was_rmapped && is_large_pte(*sptep))
2094 ++vcpu->kvm->stat.lpages;
2096 page_header_update_slot(vcpu->kvm, sptep, gfn);
2098 rmap_count = rmap_add(vcpu, sptep, gfn);
2099 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
2100 rmap_recycle(vcpu, sptep, gfn);
2102 kvm_release_pfn_clean(pfn);
2104 vcpu->arch.last_pte_updated = sptep;
2105 vcpu->arch.last_pte_gfn = gfn;
2109 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
2113 static struct kvm_memory_slot *
2114 pte_prefetch_gfn_to_memslot(struct kvm_vcpu *vcpu, gfn_t gfn, bool no_dirty_log)
2116 struct kvm_memory_slot *slot;
2118 slot = gfn_to_memslot(vcpu->kvm, gfn);
2119 if (!slot || slot->flags & KVM_MEMSLOT_INVALID ||
2120 (no_dirty_log && slot->dirty_bitmap))
2126 static pfn_t pte_prefetch_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn,
2129 struct kvm_memory_slot *slot;
2132 slot = pte_prefetch_gfn_to_memslot(vcpu, gfn, no_dirty_log);
2135 return page_to_pfn(bad_page);
2138 hva = gfn_to_hva_memslot(slot, gfn);
2140 return hva_to_pfn_atomic(vcpu->kvm, hva);
2143 static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu,
2144 struct kvm_mmu_page *sp,
2145 u64 *start, u64 *end)
2147 struct page *pages[PTE_PREFETCH_NUM];
2148 unsigned access = sp->role.access;
2152 gfn = kvm_mmu_page_get_gfn(sp, start - sp->spt);
2153 if (!pte_prefetch_gfn_to_memslot(vcpu, gfn, access & ACC_WRITE_MASK))
2156 ret = gfn_to_page_many_atomic(vcpu->kvm, gfn, pages, end - start);
2160 for (i = 0; i < ret; i++, gfn++, start++)
2161 mmu_set_spte(vcpu, start, ACC_ALL,
2162 access, 0, 0, 1, NULL,
2163 sp->role.level, gfn,
2164 page_to_pfn(pages[i]), true, true);
2169 static void __direct_pte_prefetch(struct kvm_vcpu *vcpu,
2170 struct kvm_mmu_page *sp, u64 *sptep)
2172 u64 *spte, *start = NULL;
2175 WARN_ON(!sp->role.direct);
2177 i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
2180 for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
2181 if (*spte != shadow_trap_nonpresent_pte || spte == sptep) {
2184 if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0)
2192 static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
2194 struct kvm_mmu_page *sp;
2197 * Since it's no accessed bit on EPT, it's no way to
2198 * distinguish between actually accessed translations
2199 * and prefetched, so disable pte prefetch if EPT is
2202 if (!shadow_accessed_mask)
2205 sp = page_header(__pa(sptep));
2206 if (sp->role.level > PT_PAGE_TABLE_LEVEL)
2209 __direct_pte_prefetch(vcpu, sp, sptep);
2212 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
2213 int level, gfn_t gfn, pfn_t pfn)
2215 struct kvm_shadow_walk_iterator iterator;
2216 struct kvm_mmu_page *sp;
2220 for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
2221 if (iterator.level == level) {
2222 mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
2223 0, write, 1, &pt_write,
2224 level, gfn, pfn, false, true);
2225 direct_pte_prefetch(vcpu, iterator.sptep);
2226 ++vcpu->stat.pf_fixed;
2230 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
2231 u64 base_addr = iterator.addr;
2233 base_addr &= PT64_LVL_ADDR_MASK(iterator.level);
2234 pseudo_gfn = base_addr >> PAGE_SHIFT;
2235 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
2237 1, ACC_ALL, iterator.sptep);
2239 pgprintk("nonpaging_map: ENOMEM\n");
2240 kvm_release_pfn_clean(pfn);
2244 __set_spte(iterator.sptep,
2246 | PT_PRESENT_MASK | PT_WRITABLE_MASK
2247 | shadow_user_mask | shadow_x_mask
2248 | shadow_accessed_mask);
2254 static void kvm_send_hwpoison_signal(unsigned long address, struct task_struct *tsk)
2258 info.si_signo = SIGBUS;
2260 info.si_code = BUS_MCEERR_AR;
2261 info.si_addr = (void __user *)address;
2262 info.si_addr_lsb = PAGE_SHIFT;
2264 send_sig_info(SIGBUS, &info, tsk);
2267 static int kvm_handle_bad_page(struct kvm *kvm, gfn_t gfn, pfn_t pfn)
2269 kvm_release_pfn_clean(pfn);
2270 if (is_hwpoison_pfn(pfn)) {
2271 kvm_send_hwpoison_signal(gfn_to_hva(kvm, gfn), current);
2273 } else if (is_fault_pfn(pfn))
2279 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
2284 unsigned long mmu_seq;
2286 level = mapping_level(vcpu, gfn);
2289 * This path builds a PAE pagetable - so we can map 2mb pages at
2290 * maximum. Therefore check if the level is larger than that.
2292 if (level > PT_DIRECTORY_LEVEL)
2293 level = PT_DIRECTORY_LEVEL;
2295 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2297 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2299 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2302 if (is_error_pfn(pfn))
2303 return kvm_handle_bad_page(vcpu->kvm, gfn, pfn);
2305 spin_lock(&vcpu->kvm->mmu_lock);
2306 if (mmu_notifier_retry(vcpu, mmu_seq))
2308 kvm_mmu_free_some_pages(vcpu);
2309 r = __direct_map(vcpu, v, write, level, gfn, pfn);
2310 spin_unlock(&vcpu->kvm->mmu_lock);
2316 spin_unlock(&vcpu->kvm->mmu_lock);
2317 kvm_release_pfn_clean(pfn);
2322 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2325 struct kvm_mmu_page *sp;
2326 LIST_HEAD(invalid_list);
2328 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2330 spin_lock(&vcpu->kvm->mmu_lock);
2331 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL &&
2332 (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL ||
2333 vcpu->arch.mmu.direct_map)) {
2334 hpa_t root = vcpu->arch.mmu.root_hpa;
2336 sp = page_header(root);
2338 if (!sp->root_count && sp->role.invalid) {
2339 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
2340 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2342 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2343 spin_unlock(&vcpu->kvm->mmu_lock);
2346 for (i = 0; i < 4; ++i) {
2347 hpa_t root = vcpu->arch.mmu.pae_root[i];
2350 root &= PT64_BASE_ADDR_MASK;
2351 sp = page_header(root);
2353 if (!sp->root_count && sp->role.invalid)
2354 kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2357 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2359 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2360 spin_unlock(&vcpu->kvm->mmu_lock);
2361 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2364 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2368 if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2369 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
2376 static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
2378 struct kvm_mmu_page *sp;
2381 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2382 spin_lock(&vcpu->kvm->mmu_lock);
2383 kvm_mmu_free_some_pages(vcpu);
2384 sp = kvm_mmu_get_page(vcpu, 0, 0, PT64_ROOT_LEVEL,
2387 spin_unlock(&vcpu->kvm->mmu_lock);
2388 vcpu->arch.mmu.root_hpa = __pa(sp->spt);
2389 } else if (vcpu->arch.mmu.shadow_root_level == PT32E_ROOT_LEVEL) {
2390 for (i = 0; i < 4; ++i) {
2391 hpa_t root = vcpu->arch.mmu.pae_root[i];
2393 ASSERT(!VALID_PAGE(root));
2394 spin_lock(&vcpu->kvm->mmu_lock);
2395 kvm_mmu_free_some_pages(vcpu);
2396 sp = kvm_mmu_get_page(vcpu, i << 30, i << 30,
2397 PT32_ROOT_LEVEL, 1, ACC_ALL,
2399 root = __pa(sp->spt);
2401 spin_unlock(&vcpu->kvm->mmu_lock);
2402 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2404 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2411 static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
2413 struct kvm_mmu_page *sp;
2418 root_gfn = vcpu->arch.mmu.get_cr3(vcpu) >> PAGE_SHIFT;
2420 if (mmu_check_root(vcpu, root_gfn))
2424 * Do we shadow a long mode page table? If so we need to
2425 * write-protect the guests page table root.
2427 if (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL) {
2428 hpa_t root = vcpu->arch.mmu.root_hpa;
2430 ASSERT(!VALID_PAGE(root));
2432 spin_lock(&vcpu->kvm->mmu_lock);
2433 kvm_mmu_free_some_pages(vcpu);
2434 sp = kvm_mmu_get_page(vcpu, root_gfn, 0, PT64_ROOT_LEVEL,
2436 root = __pa(sp->spt);
2438 spin_unlock(&vcpu->kvm->mmu_lock);
2439 vcpu->arch.mmu.root_hpa = root;
2444 * We shadow a 32 bit page table. This may be a legacy 2-level
2445 * or a PAE 3-level page table. In either case we need to be aware that
2446 * the shadow page table may be a PAE or a long mode page table.
2448 pm_mask = PT_PRESENT_MASK;
2449 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL)
2450 pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK;
2452 for (i = 0; i < 4; ++i) {
2453 hpa_t root = vcpu->arch.mmu.pae_root[i];
2455 ASSERT(!VALID_PAGE(root));
2456 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2457 pdptr = kvm_pdptr_read_mmu(vcpu, &vcpu->arch.mmu, i);
2458 if (!is_present_gpte(pdptr)) {
2459 vcpu->arch.mmu.pae_root[i] = 0;
2462 root_gfn = pdptr >> PAGE_SHIFT;
2463 if (mmu_check_root(vcpu, root_gfn))
2466 spin_lock(&vcpu->kvm->mmu_lock);
2467 kvm_mmu_free_some_pages(vcpu);
2468 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2471 root = __pa(sp->spt);
2473 spin_unlock(&vcpu->kvm->mmu_lock);
2475 vcpu->arch.mmu.pae_root[i] = root | pm_mask;
2477 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2480 * If we shadow a 32 bit page table with a long mode page
2481 * table we enter this path.
2483 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2484 if (vcpu->arch.mmu.lm_root == NULL) {
2486 * The additional page necessary for this is only
2487 * allocated on demand.
2492 lm_root = (void*)get_zeroed_page(GFP_KERNEL);
2493 if (lm_root == NULL)
2496 lm_root[0] = __pa(vcpu->arch.mmu.pae_root) | pm_mask;
2498 vcpu->arch.mmu.lm_root = lm_root;
2501 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.lm_root);
2507 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2509 if (vcpu->arch.mmu.direct_map)
2510 return mmu_alloc_direct_roots(vcpu);
2512 return mmu_alloc_shadow_roots(vcpu);
2515 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2518 struct kvm_mmu_page *sp;
2520 if (vcpu->arch.mmu.direct_map)
2523 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2526 trace_kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
2527 if (vcpu->arch.mmu.root_level == PT64_ROOT_LEVEL) {
2528 hpa_t root = vcpu->arch.mmu.root_hpa;
2529 sp = page_header(root);
2530 mmu_sync_children(vcpu, sp);
2533 for (i = 0; i < 4; ++i) {
2534 hpa_t root = vcpu->arch.mmu.pae_root[i];
2536 if (root && VALID_PAGE(root)) {
2537 root &= PT64_BASE_ADDR_MASK;
2538 sp = page_header(root);
2539 mmu_sync_children(vcpu, sp);
2542 trace_kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
2545 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2547 spin_lock(&vcpu->kvm->mmu_lock);
2548 mmu_sync_roots(vcpu);
2549 spin_unlock(&vcpu->kvm->mmu_lock);
2552 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2553 u32 access, u32 *error)
2560 static gpa_t nonpaging_gva_to_gpa_nested(struct kvm_vcpu *vcpu, gva_t vaddr,
2561 u32 access, u32 *error)
2565 return vcpu->arch.nested_mmu.translate_gpa(vcpu, vaddr, access);
2568 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2574 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2575 r = mmu_topup_memory_caches(vcpu);
2580 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2582 gfn = gva >> PAGE_SHIFT;
2584 return nonpaging_map(vcpu, gva & PAGE_MASK,
2585 error_code & PFERR_WRITE_MASK, gfn);
2588 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2594 gfn_t gfn = gpa >> PAGE_SHIFT;
2595 unsigned long mmu_seq;
2598 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2600 r = mmu_topup_memory_caches(vcpu);
2604 level = mapping_level(vcpu, gfn);
2606 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2608 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2610 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2611 if (is_error_pfn(pfn))
2612 return kvm_handle_bad_page(vcpu->kvm, gfn, pfn);
2613 spin_lock(&vcpu->kvm->mmu_lock);
2614 if (mmu_notifier_retry(vcpu, mmu_seq))
2616 kvm_mmu_free_some_pages(vcpu);
2617 r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2619 spin_unlock(&vcpu->kvm->mmu_lock);
2624 spin_unlock(&vcpu->kvm->mmu_lock);
2625 kvm_release_pfn_clean(pfn);
2629 static void nonpaging_free(struct kvm_vcpu *vcpu)
2631 mmu_free_roots(vcpu);
2634 static int nonpaging_init_context(struct kvm_vcpu *vcpu,
2635 struct kvm_mmu *context)
2637 context->new_cr3 = nonpaging_new_cr3;
2638 context->page_fault = nonpaging_page_fault;
2639 context->gva_to_gpa = nonpaging_gva_to_gpa;
2640 context->free = nonpaging_free;
2641 context->prefetch_page = nonpaging_prefetch_page;
2642 context->sync_page = nonpaging_sync_page;
2643 context->invlpg = nonpaging_invlpg;
2644 context->root_level = 0;
2645 context->shadow_root_level = PT32E_ROOT_LEVEL;
2646 context->root_hpa = INVALID_PAGE;
2647 context->direct_map = true;
2648 context->nx = false;
2652 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2654 ++vcpu->stat.tlb_flush;
2655 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
2658 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2660 pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2661 mmu_free_roots(vcpu);
2664 static unsigned long get_cr3(struct kvm_vcpu *vcpu)
2666 return vcpu->arch.cr3;
2669 static void inject_page_fault(struct kvm_vcpu *vcpu)
2671 vcpu->arch.mmu.inject_page_fault(vcpu);
2674 static void paging_free(struct kvm_vcpu *vcpu)
2676 nonpaging_free(vcpu);
2679 static bool is_rsvd_bits_set(struct kvm_mmu *mmu, u64 gpte, int level)
2683 bit7 = (gpte >> 7) & 1;
2684 return (gpte & mmu->rsvd_bits_mask[bit7][level-1]) != 0;
2688 #include "paging_tmpl.h"
2692 #include "paging_tmpl.h"
2695 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
2696 struct kvm_mmu *context,
2699 int maxphyaddr = cpuid_maxphyaddr(vcpu);
2700 u64 exb_bit_rsvd = 0;
2703 exb_bit_rsvd = rsvd_bits(63, 63);
2705 case PT32_ROOT_LEVEL:
2706 /* no rsvd bits for 2 level 4K page table entries */
2707 context->rsvd_bits_mask[0][1] = 0;
2708 context->rsvd_bits_mask[0][0] = 0;
2709 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2711 if (!is_pse(vcpu)) {
2712 context->rsvd_bits_mask[1][1] = 0;
2716 if (is_cpuid_PSE36())
2717 /* 36bits PSE 4MB page */
2718 context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2720 /* 32 bits PSE 4MB page */
2721 context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2723 case PT32E_ROOT_LEVEL:
2724 context->rsvd_bits_mask[0][2] =
2725 rsvd_bits(maxphyaddr, 63) |
2726 rsvd_bits(7, 8) | rsvd_bits(1, 2); /* PDPTE */
2727 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2728 rsvd_bits(maxphyaddr, 62); /* PDE */
2729 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2730 rsvd_bits(maxphyaddr, 62); /* PTE */
2731 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2732 rsvd_bits(maxphyaddr, 62) |
2733 rsvd_bits(13, 20); /* large page */
2734 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2736 case PT64_ROOT_LEVEL:
2737 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2738 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2739 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2740 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2741 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2742 rsvd_bits(maxphyaddr, 51);
2743 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2744 rsvd_bits(maxphyaddr, 51);
2745 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2746 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2747 rsvd_bits(maxphyaddr, 51) |
2749 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2750 rsvd_bits(maxphyaddr, 51) |
2751 rsvd_bits(13, 20); /* large page */
2752 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2757 static int paging64_init_context_common(struct kvm_vcpu *vcpu,
2758 struct kvm_mmu *context,
2761 context->nx = is_nx(vcpu);
2763 reset_rsvds_bits_mask(vcpu, context, level);
2765 ASSERT(is_pae(vcpu));
2766 context->new_cr3 = paging_new_cr3;
2767 context->page_fault = paging64_page_fault;
2768 context->gva_to_gpa = paging64_gva_to_gpa;
2769 context->prefetch_page = paging64_prefetch_page;
2770 context->sync_page = paging64_sync_page;
2771 context->invlpg = paging64_invlpg;
2772 context->free = paging_free;
2773 context->root_level = level;
2774 context->shadow_root_level = level;
2775 context->root_hpa = INVALID_PAGE;
2776 context->direct_map = false;
2780 static int paging64_init_context(struct kvm_vcpu *vcpu,
2781 struct kvm_mmu *context)
2783 return paging64_init_context_common(vcpu, context, PT64_ROOT_LEVEL);
2786 static int paging32_init_context(struct kvm_vcpu *vcpu,
2787 struct kvm_mmu *context)
2789 context->nx = false;
2791 reset_rsvds_bits_mask(vcpu, context, PT32_ROOT_LEVEL);
2793 context->new_cr3 = paging_new_cr3;
2794 context->page_fault = paging32_page_fault;
2795 context->gva_to_gpa = paging32_gva_to_gpa;
2796 context->free = paging_free;
2797 context->prefetch_page = paging32_prefetch_page;
2798 context->sync_page = paging32_sync_page;
2799 context->invlpg = paging32_invlpg;
2800 context->root_level = PT32_ROOT_LEVEL;
2801 context->shadow_root_level = PT32E_ROOT_LEVEL;
2802 context->root_hpa = INVALID_PAGE;
2803 context->direct_map = false;
2807 static int paging32E_init_context(struct kvm_vcpu *vcpu,
2808 struct kvm_mmu *context)
2810 return paging64_init_context_common(vcpu, context, PT32E_ROOT_LEVEL);
2813 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2815 struct kvm_mmu *context = vcpu->arch.walk_mmu;
2817 context->new_cr3 = nonpaging_new_cr3;
2818 context->page_fault = tdp_page_fault;
2819 context->free = nonpaging_free;
2820 context->prefetch_page = nonpaging_prefetch_page;
2821 context->sync_page = nonpaging_sync_page;
2822 context->invlpg = nonpaging_invlpg;
2823 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2824 context->root_hpa = INVALID_PAGE;
2825 context->direct_map = true;
2826 context->set_cr3 = kvm_x86_ops->set_tdp_cr3;
2827 context->get_cr3 = get_cr3;
2828 context->inject_page_fault = kvm_inject_page_fault;
2829 context->nx = is_nx(vcpu);
2831 if (!is_paging(vcpu)) {
2832 context->nx = false;
2833 context->gva_to_gpa = nonpaging_gva_to_gpa;
2834 context->root_level = 0;
2835 } else if (is_long_mode(vcpu)) {
2836 context->nx = is_nx(vcpu);
2837 reset_rsvds_bits_mask(vcpu, context, PT64_ROOT_LEVEL);
2838 context->gva_to_gpa = paging64_gva_to_gpa;
2839 context->root_level = PT64_ROOT_LEVEL;
2840 } else if (is_pae(vcpu)) {
2841 context->nx = is_nx(vcpu);
2842 reset_rsvds_bits_mask(vcpu, context, PT32E_ROOT_LEVEL);
2843 context->gva_to_gpa = paging64_gva_to_gpa;
2844 context->root_level = PT32E_ROOT_LEVEL;
2846 context->nx = false;
2847 reset_rsvds_bits_mask(vcpu, context, PT32_ROOT_LEVEL);
2848 context->gva_to_gpa = paging32_gva_to_gpa;
2849 context->root_level = PT32_ROOT_LEVEL;
2855 int kvm_init_shadow_mmu(struct kvm_vcpu *vcpu, struct kvm_mmu *context)
2859 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2861 if (!is_paging(vcpu))
2862 r = nonpaging_init_context(vcpu, context);
2863 else if (is_long_mode(vcpu))
2864 r = paging64_init_context(vcpu, context);
2865 else if (is_pae(vcpu))
2866 r = paging32E_init_context(vcpu, context);
2868 r = paging32_init_context(vcpu, context);
2870 vcpu->arch.mmu.base_role.cr4_pae = !!is_pae(vcpu);
2871 vcpu->arch.mmu.base_role.cr0_wp = is_write_protection(vcpu);
2875 EXPORT_SYMBOL_GPL(kvm_init_shadow_mmu);
2877 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2879 int r = kvm_init_shadow_mmu(vcpu, vcpu->arch.walk_mmu);
2881 vcpu->arch.walk_mmu->set_cr3 = kvm_x86_ops->set_cr3;
2882 vcpu->arch.walk_mmu->get_cr3 = get_cr3;
2883 vcpu->arch.walk_mmu->inject_page_fault = kvm_inject_page_fault;
2888 static int init_kvm_nested_mmu(struct kvm_vcpu *vcpu)
2890 struct kvm_mmu *g_context = &vcpu->arch.nested_mmu;
2892 g_context->get_cr3 = get_cr3;
2893 g_context->inject_page_fault = kvm_inject_page_fault;
2896 * Note that arch.mmu.gva_to_gpa translates l2_gva to l1_gpa. The
2897 * translation of l2_gpa to l1_gpa addresses is done using the
2898 * arch.nested_mmu.gva_to_gpa function. Basically the gva_to_gpa
2899 * functions between mmu and nested_mmu are swapped.
2901 if (!is_paging(vcpu)) {
2902 g_context->nx = false;
2903 g_context->root_level = 0;
2904 g_context->gva_to_gpa = nonpaging_gva_to_gpa_nested;
2905 } else if (is_long_mode(vcpu)) {
2906 g_context->nx = is_nx(vcpu);
2907 reset_rsvds_bits_mask(vcpu, g_context, PT64_ROOT_LEVEL);
2908 g_context->root_level = PT64_ROOT_LEVEL;
2909 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
2910 } else if (is_pae(vcpu)) {
2911 g_context->nx = is_nx(vcpu);
2912 reset_rsvds_bits_mask(vcpu, g_context, PT32E_ROOT_LEVEL);
2913 g_context->root_level = PT32E_ROOT_LEVEL;
2914 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
2916 g_context->nx = false;
2917 reset_rsvds_bits_mask(vcpu, g_context, PT32_ROOT_LEVEL);
2918 g_context->root_level = PT32_ROOT_LEVEL;
2919 g_context->gva_to_gpa = paging32_gva_to_gpa_nested;
2925 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2927 vcpu->arch.update_pte.pfn = bad_pfn;
2929 if (mmu_is_nested(vcpu))
2930 return init_kvm_nested_mmu(vcpu);
2931 else if (tdp_enabled)
2932 return init_kvm_tdp_mmu(vcpu);
2934 return init_kvm_softmmu(vcpu);
2937 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2940 if (VALID_PAGE(vcpu->arch.mmu.root_hpa))
2941 /* mmu.free() should set root_hpa = INVALID_PAGE */
2942 vcpu->arch.mmu.free(vcpu);
2945 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2947 destroy_kvm_mmu(vcpu);
2948 return init_kvm_mmu(vcpu);
2950 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2952 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2956 r = mmu_topup_memory_caches(vcpu);
2959 r = mmu_alloc_roots(vcpu);
2960 spin_lock(&vcpu->kvm->mmu_lock);
2961 mmu_sync_roots(vcpu);
2962 spin_unlock(&vcpu->kvm->mmu_lock);
2965 /* set_cr3() should ensure TLB has been flushed */
2966 vcpu->arch.mmu.set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2970 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2972 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2974 mmu_free_roots(vcpu);
2976 EXPORT_SYMBOL_GPL(kvm_mmu_unload);
2978 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2979 struct kvm_mmu_page *sp,
2983 struct kvm_mmu_page *child;
2986 if (is_shadow_present_pte(pte)) {
2987 if (is_last_spte(pte, sp->role.level))
2988 drop_spte(vcpu->kvm, spte, shadow_trap_nonpresent_pte);
2990 child = page_header(pte & PT64_BASE_ADDR_MASK);
2991 mmu_page_remove_parent_pte(child, spte);
2994 __set_spte(spte, shadow_trap_nonpresent_pte);
2995 if (is_large_pte(pte))
2996 --vcpu->kvm->stat.lpages;
2999 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
3000 struct kvm_mmu_page *sp,
3004 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
3005 ++vcpu->kvm->stat.mmu_pde_zapped;
3009 if (is_rsvd_bits_set(&vcpu->arch.mmu, *(u64 *)new, PT_PAGE_TABLE_LEVEL))
3012 ++vcpu->kvm->stat.mmu_pte_updated;
3013 if (!sp->role.cr4_pae)
3014 paging32_update_pte(vcpu, sp, spte, new);
3016 paging64_update_pte(vcpu, sp, spte, new);
3019 static bool need_remote_flush(u64 old, u64 new)
3021 if (!is_shadow_present_pte(old))
3023 if (!is_shadow_present_pte(new))
3025 if ((old ^ new) & PT64_BASE_ADDR_MASK)
3027 old ^= PT64_NX_MASK;
3028 new ^= PT64_NX_MASK;
3029 return (old & ~new & PT64_PERM_MASK) != 0;
3032 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, bool zap_page,
3033 bool remote_flush, bool local_flush)
3039 kvm_flush_remote_tlbs(vcpu->kvm);
3040 else if (local_flush)
3041 kvm_mmu_flush_tlb(vcpu);
3044 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
3046 u64 *spte = vcpu->arch.last_pte_updated;
3048 return !!(spte && (*spte & shadow_accessed_mask));
3051 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
3057 if (!is_present_gpte(gpte))
3059 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
3061 vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
3063 pfn = gfn_to_pfn(vcpu->kvm, gfn);
3065 if (is_error_pfn(pfn)) {
3066 kvm_release_pfn_clean(pfn);
3069 vcpu->arch.update_pte.gfn = gfn;
3070 vcpu->arch.update_pte.pfn = pfn;
3073 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
3075 u64 *spte = vcpu->arch.last_pte_updated;
3078 && vcpu->arch.last_pte_gfn == gfn
3079 && shadow_accessed_mask
3080 && !(*spte & shadow_accessed_mask)
3081 && is_shadow_present_pte(*spte))
3082 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
3085 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
3086 const u8 *new, int bytes,
3087 bool guest_initiated)
3089 gfn_t gfn = gpa >> PAGE_SHIFT;
3090 union kvm_mmu_page_role mask = { .word = 0 };
3091 struct kvm_mmu_page *sp;
3092 struct hlist_node *node;
3093 LIST_HEAD(invalid_list);
3096 unsigned offset = offset_in_page(gpa);
3098 unsigned page_offset;
3099 unsigned misaligned;
3106 bool remote_flush, local_flush, zap_page;
3108 zap_page = remote_flush = local_flush = false;
3110 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
3112 invlpg_counter = atomic_read(&vcpu->kvm->arch.invlpg_counter);
3115 * Assume that the pte write on a page table of the same type
3116 * as the current vcpu paging mode. This is nearly always true
3117 * (might be false while changing modes). Note it is verified later
3120 if ((is_pae(vcpu) && bytes == 4) || !new) {
3121 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
3126 r = kvm_read_guest(vcpu->kvm, gpa, &gentry, min(bytes, 8));
3129 new = (const u8 *)&gentry;
3134 gentry = *(const u32 *)new;
3137 gentry = *(const u64 *)new;
3144 mmu_guess_page_from_pte_write(vcpu, gpa, gentry);
3145 spin_lock(&vcpu->kvm->mmu_lock);
3146 if (atomic_read(&vcpu->kvm->arch.invlpg_counter) != invlpg_counter)
3148 kvm_mmu_access_page(vcpu, gfn);
3149 kvm_mmu_free_some_pages(vcpu);
3150 ++vcpu->kvm->stat.mmu_pte_write;
3151 trace_kvm_mmu_audit(vcpu, AUDIT_PRE_PTE_WRITE);
3152 if (guest_initiated) {
3153 if (gfn == vcpu->arch.last_pt_write_gfn
3154 && !last_updated_pte_accessed(vcpu)) {
3155 ++vcpu->arch.last_pt_write_count;
3156 if (vcpu->arch.last_pt_write_count >= 3)
3159 vcpu->arch.last_pt_write_gfn = gfn;
3160 vcpu->arch.last_pt_write_count = 1;
3161 vcpu->arch.last_pte_updated = NULL;
3165 mask.cr0_wp = mask.cr4_pae = mask.nxe = 1;
3166 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn, node) {
3167 pte_size = sp->role.cr4_pae ? 8 : 4;
3168 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
3169 misaligned |= bytes < 4;
3170 if (misaligned || flooded) {
3172 * Misaligned accesses are too much trouble to fix
3173 * up; also, they usually indicate a page is not used
3176 * If we're seeing too many writes to a page,
3177 * it may no longer be a page table, or we may be
3178 * forking, in which case it is better to unmap the
3181 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
3182 gpa, bytes, sp->role.word);
3183 zap_page |= !!kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
3185 ++vcpu->kvm->stat.mmu_flooded;
3188 page_offset = offset;
3189 level = sp->role.level;
3191 if (!sp->role.cr4_pae) {
3192 page_offset <<= 1; /* 32->64 */
3194 * A 32-bit pde maps 4MB while the shadow pdes map
3195 * only 2MB. So we need to double the offset again
3196 * and zap two pdes instead of one.
3198 if (level == PT32_ROOT_LEVEL) {
3199 page_offset &= ~7; /* kill rounding error */
3203 quadrant = page_offset >> PAGE_SHIFT;
3204 page_offset &= ~PAGE_MASK;
3205 if (quadrant != sp->role.quadrant)
3209 spte = &sp->spt[page_offset / sizeof(*spte)];
3212 mmu_pte_write_zap_pte(vcpu, sp, spte);
3214 !((sp->role.word ^ vcpu->arch.mmu.base_role.word)
3216 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
3217 if (!remote_flush && need_remote_flush(entry, *spte))
3218 remote_flush = true;
3222 mmu_pte_write_flush_tlb(vcpu, zap_page, remote_flush, local_flush);
3223 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
3224 trace_kvm_mmu_audit(vcpu, AUDIT_POST_PTE_WRITE);
3225 spin_unlock(&vcpu->kvm->mmu_lock);
3226 if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
3227 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
3228 vcpu->arch.update_pte.pfn = bad_pfn;
3232 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
3237 if (vcpu->arch.mmu.direct_map)
3240 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
3242 spin_lock(&vcpu->kvm->mmu_lock);
3243 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
3244 spin_unlock(&vcpu->kvm->mmu_lock);
3247 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
3249 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
3251 LIST_HEAD(invalid_list);
3253 while (kvm_mmu_available_pages(vcpu->kvm) < KVM_REFILL_PAGES &&
3254 !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
3255 struct kvm_mmu_page *sp;
3257 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
3258 struct kvm_mmu_page, link);
3259 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
3260 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
3261 ++vcpu->kvm->stat.mmu_recycled;
3265 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
3268 enum emulation_result er;
3270 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
3279 r = mmu_topup_memory_caches(vcpu);
3283 er = emulate_instruction(vcpu, cr2, error_code, 0);
3288 case EMULATE_DO_MMIO:
3289 ++vcpu->stat.mmio_exits;
3299 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
3301 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
3303 vcpu->arch.mmu.invlpg(vcpu, gva);
3304 kvm_mmu_flush_tlb(vcpu);
3305 ++vcpu->stat.invlpg;
3307 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
3309 void kvm_enable_tdp(void)
3313 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
3315 void kvm_disable_tdp(void)
3317 tdp_enabled = false;
3319 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
3321 static void free_mmu_pages(struct kvm_vcpu *vcpu)
3323 free_page((unsigned long)vcpu->arch.mmu.pae_root);
3324 if (vcpu->arch.mmu.lm_root != NULL)
3325 free_page((unsigned long)vcpu->arch.mmu.lm_root);
3328 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
3336 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
3337 * Therefore we need to allocate shadow page tables in the first
3338 * 4GB of memory, which happens to fit the DMA32 zone.
3340 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
3344 vcpu->arch.mmu.pae_root = page_address(page);
3345 for (i = 0; i < 4; ++i)
3346 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
3351 int kvm_mmu_create(struct kvm_vcpu *vcpu)
3354 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3356 return alloc_mmu_pages(vcpu);
3359 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
3362 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3364 return init_kvm_mmu(vcpu);
3367 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
3369 struct kvm_mmu_page *sp;
3371 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
3375 if (!test_bit(slot, sp->slot_bitmap))
3379 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
3381 if (is_writable_pte(pt[i]))
3382 pt[i] &= ~PT_WRITABLE_MASK;
3384 kvm_flush_remote_tlbs(kvm);
3387 void kvm_mmu_zap_all(struct kvm *kvm)
3389 struct kvm_mmu_page *sp, *node;
3390 LIST_HEAD(invalid_list);
3392 spin_lock(&kvm->mmu_lock);
3394 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
3395 if (kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list))
3398 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3399 spin_unlock(&kvm->mmu_lock);
3402 static int kvm_mmu_remove_some_alloc_mmu_pages(struct kvm *kvm,
3403 struct list_head *invalid_list)
3405 struct kvm_mmu_page *page;
3407 page = container_of(kvm->arch.active_mmu_pages.prev,
3408 struct kvm_mmu_page, link);
3409 return kvm_mmu_prepare_zap_page(kvm, page, invalid_list);
3412 static int mmu_shrink(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
3415 struct kvm *kvm_freed = NULL;
3417 if (nr_to_scan == 0)
3420 spin_lock(&kvm_lock);
3422 list_for_each_entry(kvm, &vm_list, vm_list) {
3423 int idx, freed_pages;
3424 LIST_HEAD(invalid_list);
3426 idx = srcu_read_lock(&kvm->srcu);
3427 spin_lock(&kvm->mmu_lock);
3428 if (!kvm_freed && nr_to_scan > 0 &&
3429 kvm->arch.n_used_mmu_pages > 0) {
3430 freed_pages = kvm_mmu_remove_some_alloc_mmu_pages(kvm,
3436 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3437 spin_unlock(&kvm->mmu_lock);
3438 srcu_read_unlock(&kvm->srcu, idx);
3441 list_move_tail(&kvm_freed->vm_list, &vm_list);
3443 spin_unlock(&kvm_lock);
3446 return percpu_counter_read_positive(&kvm_total_used_mmu_pages);
3449 static struct shrinker mmu_shrinker = {
3450 .shrink = mmu_shrink,
3451 .seeks = DEFAULT_SEEKS * 10,
3454 static void mmu_destroy_caches(void)
3456 if (pte_chain_cache)
3457 kmem_cache_destroy(pte_chain_cache);
3458 if (rmap_desc_cache)
3459 kmem_cache_destroy(rmap_desc_cache);
3460 if (mmu_page_header_cache)
3461 kmem_cache_destroy(mmu_page_header_cache);
3464 void kvm_mmu_module_exit(void)
3466 mmu_destroy_caches();
3467 percpu_counter_destroy(&kvm_total_used_mmu_pages);
3468 unregister_shrinker(&mmu_shrinker);
3471 int kvm_mmu_module_init(void)
3473 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
3474 sizeof(struct kvm_pte_chain),
3476 if (!pte_chain_cache)
3478 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
3479 sizeof(struct kvm_rmap_desc),
3481 if (!rmap_desc_cache)
3484 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
3485 sizeof(struct kvm_mmu_page),
3487 if (!mmu_page_header_cache)
3490 if (percpu_counter_init(&kvm_total_used_mmu_pages, 0))
3493 register_shrinker(&mmu_shrinker);
3498 mmu_destroy_caches();
3503 * Caculate mmu pages needed for kvm.
3505 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3508 unsigned int nr_mmu_pages;
3509 unsigned int nr_pages = 0;
3510 struct kvm_memslots *slots;
3512 slots = kvm_memslots(kvm);
3514 for (i = 0; i < slots->nmemslots; i++)
3515 nr_pages += slots->memslots[i].npages;
3517 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3518 nr_mmu_pages = max(nr_mmu_pages,
3519 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3521 return nr_mmu_pages;
3524 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3527 if (len > buffer->len)
3532 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3537 ret = pv_mmu_peek_buffer(buffer, len);
3542 buffer->processed += len;
3546 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3547 gpa_t addr, gpa_t value)
3552 if (!is_long_mode(vcpu) && !is_pae(vcpu))
3555 r = mmu_topup_memory_caches(vcpu);
3559 if (!emulator_write_phys(vcpu, addr, &value, bytes))
3565 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3567 (void)kvm_set_cr3(vcpu, vcpu->arch.cr3);
3571 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3573 spin_lock(&vcpu->kvm->mmu_lock);
3574 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3575 spin_unlock(&vcpu->kvm->mmu_lock);
3579 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3580 struct kvm_pv_mmu_op_buffer *buffer)
3582 struct kvm_mmu_op_header *header;
3584 header = pv_mmu_peek_buffer(buffer, sizeof *header);
3587 switch (header->op) {
3588 case KVM_MMU_OP_WRITE_PTE: {
3589 struct kvm_mmu_op_write_pte *wpte;
3591 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3594 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3597 case KVM_MMU_OP_FLUSH_TLB: {
3598 struct kvm_mmu_op_flush_tlb *ftlb;
3600 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3603 return kvm_pv_mmu_flush_tlb(vcpu);
3605 case KVM_MMU_OP_RELEASE_PT: {
3606 struct kvm_mmu_op_release_pt *rpt;
3608 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3611 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3617 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3618 gpa_t addr, unsigned long *ret)
3621 struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3623 buffer->ptr = buffer->buf;
3624 buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3625 buffer->processed = 0;
3627 r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3631 while (buffer->len) {
3632 r = kvm_pv_mmu_op_one(vcpu, buffer);
3641 *ret = buffer->processed;
3645 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3647 struct kvm_shadow_walk_iterator iterator;
3650 spin_lock(&vcpu->kvm->mmu_lock);
3651 for_each_shadow_entry(vcpu, addr, iterator) {
3652 sptes[iterator.level-1] = *iterator.sptep;
3654 if (!is_shadow_present_pte(*iterator.sptep))
3657 spin_unlock(&vcpu->kvm->mmu_lock);
3661 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3663 #ifdef CONFIG_KVM_MMU_AUDIT
3664 #include "mmu_audit.c"
3666 static void mmu_audit_disable(void) { }
3669 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
3673 destroy_kvm_mmu(vcpu);
3674 free_mmu_pages(vcpu);
3675 mmu_free_memory_caches(vcpu);
3676 mmu_audit_disable();