2 * Kernel-based Virtual Machine driver for Linux
4 * This module enables machines with Intel VT-x extensions to run virtual
5 * machines without emulation or binary translation.
9 * Copyright (C) 2006 Qumranet, Inc.
12 * Yaniv Kamay <yaniv@qumranet.com>
13 * Avi Kivity <avi@qumranet.com>
15 * This work is licensed under the terms of the GNU GPL, version 2. See
16 * the COPYING file in the top-level directory.
22 #include "kvm_cache_regs.h"
24 #include <linux/kvm_host.h>
25 #include <linux/types.h>
26 #include <linux/string.h>
28 #include <linux/highmem.h>
29 #include <linux/module.h>
30 #include <linux/swap.h>
31 #include <linux/hugetlb.h>
32 #include <linux/compiler.h>
33 #include <linux/srcu.h>
34 #include <linux/slab.h>
37 #include <asm/cmpxchg.h>
42 * When setting this variable to true it enables Two-Dimensional-Paging
43 * where the hardware walks 2 page tables:
44 * 1. the guest-virtual to guest-physical
45 * 2. while doing 1. it walks guest-physical to host-physical
46 * If the hardware supports that we don't need to do shadow paging.
48 bool tdp_enabled = false;
55 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
57 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
62 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
63 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
67 #define pgprintk(x...) do { } while (0)
68 #define rmap_printk(x...) do { } while (0)
72 #if defined(MMU_DEBUG) || defined(AUDIT)
74 module_param(dbg, bool, 0644);
77 static int oos_shadow = 1;
78 module_param(oos_shadow, bool, 0644);
81 #define ASSERT(x) do { } while (0)
85 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
86 __FILE__, __LINE__, #x); \
90 #define PT_FIRST_AVAIL_BITS_SHIFT 9
91 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
93 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
95 #define PT64_LEVEL_BITS 9
97 #define PT64_LEVEL_SHIFT(level) \
98 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
100 #define PT64_LEVEL_MASK(level) \
101 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
103 #define PT64_INDEX(address, level)\
104 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
107 #define PT32_LEVEL_BITS 10
109 #define PT32_LEVEL_SHIFT(level) \
110 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
112 #define PT32_LEVEL_MASK(level) \
113 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
114 #define PT32_LVL_OFFSET_MASK(level) \
115 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
116 * PT32_LEVEL_BITS))) - 1))
118 #define PT32_INDEX(address, level)\
119 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
122 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
123 #define PT64_DIR_BASE_ADDR_MASK \
124 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
125 #define PT64_LVL_ADDR_MASK(level) \
126 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
127 * PT64_LEVEL_BITS))) - 1))
128 #define PT64_LVL_OFFSET_MASK(level) \
129 (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
130 * PT64_LEVEL_BITS))) - 1))
132 #define PT32_BASE_ADDR_MASK PAGE_MASK
133 #define PT32_DIR_BASE_ADDR_MASK \
134 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
135 #define PT32_LVL_ADDR_MASK(level) \
136 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
137 * PT32_LEVEL_BITS))) - 1))
139 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
144 #define ACC_EXEC_MASK 1
145 #define ACC_WRITE_MASK PT_WRITABLE_MASK
146 #define ACC_USER_MASK PT_USER_MASK
147 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
149 #include <trace/events/kvm.h>
151 #undef TRACE_INCLUDE_FILE
152 #define CREATE_TRACE_POINTS
153 #include "mmutrace.h"
155 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
157 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
159 struct kvm_rmap_desc {
160 u64 *sptes[RMAP_EXT];
161 struct kvm_rmap_desc *more;
164 struct kvm_shadow_walk_iterator {
172 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
173 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
174 shadow_walk_okay(&(_walker)); \
175 shadow_walk_next(&(_walker)))
178 struct kvm_unsync_walk {
179 int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
182 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
184 static struct kmem_cache *pte_chain_cache;
185 static struct kmem_cache *rmap_desc_cache;
186 static struct kmem_cache *mmu_page_header_cache;
188 static u64 __read_mostly shadow_trap_nonpresent_pte;
189 static u64 __read_mostly shadow_notrap_nonpresent_pte;
190 static u64 __read_mostly shadow_base_present_pte;
191 static u64 __read_mostly shadow_nx_mask;
192 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
193 static u64 __read_mostly shadow_user_mask;
194 static u64 __read_mostly shadow_accessed_mask;
195 static u64 __read_mostly shadow_dirty_mask;
197 static inline u64 rsvd_bits(int s, int e)
199 return ((1ULL << (e - s + 1)) - 1) << s;
202 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
204 shadow_trap_nonpresent_pte = trap_pte;
205 shadow_notrap_nonpresent_pte = notrap_pte;
207 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
209 void kvm_mmu_set_base_ptes(u64 base_pte)
211 shadow_base_present_pte = base_pte;
213 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
215 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
216 u64 dirty_mask, u64 nx_mask, u64 x_mask)
218 shadow_user_mask = user_mask;
219 shadow_accessed_mask = accessed_mask;
220 shadow_dirty_mask = dirty_mask;
221 shadow_nx_mask = nx_mask;
222 shadow_x_mask = x_mask;
224 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
226 static int is_write_protection(struct kvm_vcpu *vcpu)
228 return kvm_read_cr0_bits(vcpu, X86_CR0_WP);
231 static int is_cpuid_PSE36(void)
236 static int is_nx(struct kvm_vcpu *vcpu)
238 return vcpu->arch.efer & EFER_NX;
241 static int is_shadow_present_pte(u64 pte)
243 return pte != shadow_trap_nonpresent_pte
244 && pte != shadow_notrap_nonpresent_pte;
247 static int is_large_pte(u64 pte)
249 return pte & PT_PAGE_SIZE_MASK;
252 static int is_writable_pte(unsigned long pte)
254 return pte & PT_WRITABLE_MASK;
257 static int is_dirty_gpte(unsigned long pte)
259 return pte & PT_DIRTY_MASK;
262 static int is_rmap_spte(u64 pte)
264 return is_shadow_present_pte(pte);
267 static int is_last_spte(u64 pte, int level)
269 if (level == PT_PAGE_TABLE_LEVEL)
271 if (is_large_pte(pte))
276 static pfn_t spte_to_pfn(u64 pte)
278 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
281 static gfn_t pse36_gfn_delta(u32 gpte)
283 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
285 return (gpte & PT32_DIR_PSE36_MASK) << shift;
288 static void __set_spte(u64 *sptep, u64 spte)
291 set_64bit((unsigned long *)sptep, spte);
293 set_64bit((unsigned long long *)sptep, spte);
297 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
298 struct kmem_cache *base_cache, int min)
302 if (cache->nobjs >= min)
304 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
305 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
308 cache->objects[cache->nobjs++] = obj;
313 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
316 kfree(mc->objects[--mc->nobjs]);
319 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
324 if (cache->nobjs >= min)
326 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
327 page = alloc_page(GFP_KERNEL);
330 set_page_private(page, 0);
331 cache->objects[cache->nobjs++] = page_address(page);
336 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
339 free_page((unsigned long)mc->objects[--mc->nobjs]);
342 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
346 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
350 r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
354 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
357 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
358 mmu_page_header_cache, 4);
363 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
365 mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
366 mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
367 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
368 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
371 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
377 p = mc->objects[--mc->nobjs];
381 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
383 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
384 sizeof(struct kvm_pte_chain));
387 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
392 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
394 return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
395 sizeof(struct kvm_rmap_desc));
398 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
404 * Return the pointer to the largepage write count for a given
405 * gfn, handling slots that are not large page aligned.
407 static int *slot_largepage_idx(gfn_t gfn,
408 struct kvm_memory_slot *slot,
413 idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
414 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
415 return &slot->lpage_info[level - 2][idx].write_count;
418 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
420 struct kvm_memory_slot *slot;
424 gfn = unalias_gfn(kvm, gfn);
426 slot = gfn_to_memslot_unaliased(kvm, gfn);
427 for (i = PT_DIRECTORY_LEVEL;
428 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
429 write_count = slot_largepage_idx(gfn, slot, i);
434 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
436 struct kvm_memory_slot *slot;
440 gfn = unalias_gfn(kvm, gfn);
441 for (i = PT_DIRECTORY_LEVEL;
442 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
443 slot = gfn_to_memslot_unaliased(kvm, gfn);
444 write_count = slot_largepage_idx(gfn, slot, i);
446 WARN_ON(*write_count < 0);
450 static int has_wrprotected_page(struct kvm *kvm,
454 struct kvm_memory_slot *slot;
457 gfn = unalias_gfn(kvm, gfn);
458 slot = gfn_to_memslot_unaliased(kvm, gfn);
460 largepage_idx = slot_largepage_idx(gfn, slot, level);
461 return *largepage_idx;
467 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
469 unsigned long page_size;
472 page_size = kvm_host_page_size(kvm, gfn);
474 for (i = PT_PAGE_TABLE_LEVEL;
475 i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
476 if (page_size >= KVM_HPAGE_SIZE(i))
485 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
487 struct kvm_memory_slot *slot;
488 int host_level, level, max_level;
490 slot = gfn_to_memslot(vcpu->kvm, large_gfn);
491 if (slot && slot->dirty_bitmap)
492 return PT_PAGE_TABLE_LEVEL;
494 host_level = host_mapping_level(vcpu->kvm, large_gfn);
496 if (host_level == PT_PAGE_TABLE_LEVEL)
499 max_level = kvm_x86_ops->get_lpage_level() < host_level ?
500 kvm_x86_ops->get_lpage_level() : host_level;
502 for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
503 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
510 * Take gfn and return the reverse mapping to it.
511 * Note: gfn must be unaliased before this function get called
514 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
516 struct kvm_memory_slot *slot;
519 slot = gfn_to_memslot(kvm, gfn);
520 if (likely(level == PT_PAGE_TABLE_LEVEL))
521 return &slot->rmap[gfn - slot->base_gfn];
523 idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
524 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
526 return &slot->lpage_info[level - 2][idx].rmap_pde;
530 * Reverse mapping data structures:
532 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
533 * that points to page_address(page).
535 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
536 * containing more mappings.
538 * Returns the number of rmap entries before the spte was added or zero if
539 * the spte was not added.
542 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
544 struct kvm_mmu_page *sp;
545 struct kvm_rmap_desc *desc;
546 unsigned long *rmapp;
549 if (!is_rmap_spte(*spte))
551 gfn = unalias_gfn(vcpu->kvm, gfn);
552 sp = page_header(__pa(spte));
553 sp->gfns[spte - sp->spt] = gfn;
554 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
556 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
557 *rmapp = (unsigned long)spte;
558 } else if (!(*rmapp & 1)) {
559 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
560 desc = mmu_alloc_rmap_desc(vcpu);
561 desc->sptes[0] = (u64 *)*rmapp;
562 desc->sptes[1] = spte;
563 *rmapp = (unsigned long)desc | 1;
565 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
566 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
567 while (desc->sptes[RMAP_EXT-1] && desc->more) {
571 if (desc->sptes[RMAP_EXT-1]) {
572 desc->more = mmu_alloc_rmap_desc(vcpu);
575 for (i = 0; desc->sptes[i]; ++i)
577 desc->sptes[i] = spte;
582 static void rmap_desc_remove_entry(unsigned long *rmapp,
583 struct kvm_rmap_desc *desc,
585 struct kvm_rmap_desc *prev_desc)
589 for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
591 desc->sptes[i] = desc->sptes[j];
592 desc->sptes[j] = NULL;
595 if (!prev_desc && !desc->more)
596 *rmapp = (unsigned long)desc->sptes[0];
599 prev_desc->more = desc->more;
601 *rmapp = (unsigned long)desc->more | 1;
602 mmu_free_rmap_desc(desc);
605 static void rmap_remove(struct kvm *kvm, u64 *spte)
607 struct kvm_rmap_desc *desc;
608 struct kvm_rmap_desc *prev_desc;
609 struct kvm_mmu_page *sp;
611 unsigned long *rmapp;
614 if (!is_rmap_spte(*spte))
616 sp = page_header(__pa(spte));
617 pfn = spte_to_pfn(*spte);
618 if (*spte & shadow_accessed_mask)
619 kvm_set_pfn_accessed(pfn);
620 if (is_writable_pte(*spte))
621 kvm_set_pfn_dirty(pfn);
622 rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], sp->role.level);
624 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
626 } else if (!(*rmapp & 1)) {
627 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
628 if ((u64 *)*rmapp != spte) {
629 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
635 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
636 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
639 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
640 if (desc->sptes[i] == spte) {
641 rmap_desc_remove_entry(rmapp,
649 pr_err("rmap_remove: %p %llx many->many\n", spte, *spte);
654 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
656 struct kvm_rmap_desc *desc;
657 struct kvm_rmap_desc *prev_desc;
663 else if (!(*rmapp & 1)) {
665 return (u64 *)*rmapp;
668 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
672 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
673 if (prev_spte == spte)
674 return desc->sptes[i];
675 prev_spte = desc->sptes[i];
682 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
684 unsigned long *rmapp;
686 int i, write_protected = 0;
688 gfn = unalias_gfn(kvm, gfn);
689 rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
691 spte = rmap_next(kvm, rmapp, NULL);
694 BUG_ON(!(*spte & PT_PRESENT_MASK));
695 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
696 if (is_writable_pte(*spte)) {
697 __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
700 spte = rmap_next(kvm, rmapp, spte);
702 if (write_protected) {
705 spte = rmap_next(kvm, rmapp, NULL);
706 pfn = spte_to_pfn(*spte);
707 kvm_set_pfn_dirty(pfn);
710 /* check for huge page mappings */
711 for (i = PT_DIRECTORY_LEVEL;
712 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
713 rmapp = gfn_to_rmap(kvm, gfn, i);
714 spte = rmap_next(kvm, rmapp, NULL);
717 BUG_ON(!(*spte & PT_PRESENT_MASK));
718 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
719 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
720 if (is_writable_pte(*spte)) {
721 rmap_remove(kvm, spte);
723 __set_spte(spte, shadow_trap_nonpresent_pte);
727 spte = rmap_next(kvm, rmapp, spte);
731 return write_protected;
734 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
738 int need_tlb_flush = 0;
740 while ((spte = rmap_next(kvm, rmapp, NULL))) {
741 BUG_ON(!(*spte & PT_PRESENT_MASK));
742 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
743 rmap_remove(kvm, spte);
744 __set_spte(spte, shadow_trap_nonpresent_pte);
747 return need_tlb_flush;
750 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
755 pte_t *ptep = (pte_t *)data;
758 WARN_ON(pte_huge(*ptep));
759 new_pfn = pte_pfn(*ptep);
760 spte = rmap_next(kvm, rmapp, NULL);
762 BUG_ON(!is_shadow_present_pte(*spte));
763 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
765 if (pte_write(*ptep)) {
766 rmap_remove(kvm, spte);
767 __set_spte(spte, shadow_trap_nonpresent_pte);
768 spte = rmap_next(kvm, rmapp, NULL);
770 new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
771 new_spte |= (u64)new_pfn << PAGE_SHIFT;
773 new_spte &= ~PT_WRITABLE_MASK;
774 new_spte &= ~SPTE_HOST_WRITEABLE;
775 if (is_writable_pte(*spte))
776 kvm_set_pfn_dirty(spte_to_pfn(*spte));
777 __set_spte(spte, new_spte);
778 spte = rmap_next(kvm, rmapp, spte);
782 kvm_flush_remote_tlbs(kvm);
787 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
789 int (*handler)(struct kvm *kvm, unsigned long *rmapp,
795 struct kvm_memslots *slots;
797 slots = rcu_dereference(kvm->memslots);
799 for (i = 0; i < slots->nmemslots; i++) {
800 struct kvm_memory_slot *memslot = &slots->memslots[i];
801 unsigned long start = memslot->userspace_addr;
804 end = start + (memslot->npages << PAGE_SHIFT);
805 if (hva >= start && hva < end) {
806 gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
808 ret = handler(kvm, &memslot->rmap[gfn_offset], data);
810 for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
811 int idx = gfn_offset;
812 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
814 &memslot->lpage_info[j][idx].rmap_pde,
817 trace_kvm_age_page(hva, memslot, ret);
825 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
827 return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
830 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
832 kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
835 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
842 * Emulate the accessed bit for EPT, by checking if this page has
843 * an EPT mapping, and clearing it if it does. On the next access,
844 * a new EPT mapping will be established.
845 * This has some overhead, but not as much as the cost of swapping
846 * out actively used pages or breaking up actively used hugepages.
848 if (!shadow_accessed_mask)
849 return kvm_unmap_rmapp(kvm, rmapp, data);
851 spte = rmap_next(kvm, rmapp, NULL);
855 BUG_ON(!(_spte & PT_PRESENT_MASK));
856 _young = _spte & PT_ACCESSED_MASK;
859 clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
861 spte = rmap_next(kvm, rmapp, spte);
866 #define RMAP_RECYCLE_THRESHOLD 1000
868 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
870 unsigned long *rmapp;
871 struct kvm_mmu_page *sp;
873 sp = page_header(__pa(spte));
875 gfn = unalias_gfn(vcpu->kvm, gfn);
876 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
878 kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
879 kvm_flush_remote_tlbs(vcpu->kvm);
882 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
884 return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
888 static int is_empty_shadow_page(u64 *spt)
893 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
894 if (is_shadow_present_pte(*pos)) {
895 printk(KERN_ERR "%s: %p %llx\n", __func__,
903 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
905 ASSERT(is_empty_shadow_page(sp->spt));
907 __free_page(virt_to_page(sp->spt));
908 __free_page(virt_to_page(sp->gfns));
910 ++kvm->arch.n_free_mmu_pages;
913 static unsigned kvm_page_table_hashfn(gfn_t gfn)
915 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
918 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
921 struct kvm_mmu_page *sp;
923 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
924 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
925 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
926 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
927 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
928 INIT_LIST_HEAD(&sp->oos_link);
929 bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
931 sp->parent_pte = parent_pte;
932 --vcpu->kvm->arch.n_free_mmu_pages;
936 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
937 struct kvm_mmu_page *sp, u64 *parent_pte)
939 struct kvm_pte_chain *pte_chain;
940 struct hlist_node *node;
945 if (!sp->multimapped) {
946 u64 *old = sp->parent_pte;
949 sp->parent_pte = parent_pte;
953 pte_chain = mmu_alloc_pte_chain(vcpu);
954 INIT_HLIST_HEAD(&sp->parent_ptes);
955 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
956 pte_chain->parent_ptes[0] = old;
958 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
959 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
961 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
962 if (!pte_chain->parent_ptes[i]) {
963 pte_chain->parent_ptes[i] = parent_pte;
967 pte_chain = mmu_alloc_pte_chain(vcpu);
969 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
970 pte_chain->parent_ptes[0] = parent_pte;
973 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
976 struct kvm_pte_chain *pte_chain;
977 struct hlist_node *node;
980 if (!sp->multimapped) {
981 BUG_ON(sp->parent_pte != parent_pte);
982 sp->parent_pte = NULL;
985 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
986 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
987 if (!pte_chain->parent_ptes[i])
989 if (pte_chain->parent_ptes[i] != parent_pte)
991 while (i + 1 < NR_PTE_CHAIN_ENTRIES
992 && pte_chain->parent_ptes[i + 1]) {
993 pte_chain->parent_ptes[i]
994 = pte_chain->parent_ptes[i + 1];
997 pte_chain->parent_ptes[i] = NULL;
999 hlist_del(&pte_chain->link);
1000 mmu_free_pte_chain(pte_chain);
1001 if (hlist_empty(&sp->parent_ptes)) {
1002 sp->multimapped = 0;
1003 sp->parent_pte = NULL;
1012 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1013 mmu_parent_walk_fn fn)
1015 struct kvm_pte_chain *pte_chain;
1016 struct hlist_node *node;
1017 struct kvm_mmu_page *parent_sp;
1020 if (!sp->multimapped && sp->parent_pte) {
1021 parent_sp = page_header(__pa(sp->parent_pte));
1022 fn(vcpu, parent_sp);
1023 mmu_parent_walk(vcpu, parent_sp, fn);
1026 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1027 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1028 if (!pte_chain->parent_ptes[i])
1030 parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
1031 fn(vcpu, parent_sp);
1032 mmu_parent_walk(vcpu, parent_sp, fn);
1036 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1039 struct kvm_mmu_page *sp = page_header(__pa(spte));
1041 index = spte - sp->spt;
1042 if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1043 sp->unsync_children++;
1044 WARN_ON(!sp->unsync_children);
1047 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1049 struct kvm_pte_chain *pte_chain;
1050 struct hlist_node *node;
1053 if (!sp->parent_pte)
1056 if (!sp->multimapped) {
1057 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1061 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1062 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1063 if (!pte_chain->parent_ptes[i])
1065 kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1069 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1071 kvm_mmu_update_parents_unsync(sp);
1075 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
1076 struct kvm_mmu_page *sp)
1078 mmu_parent_walk(vcpu, sp, unsync_walk_fn);
1079 kvm_mmu_update_parents_unsync(sp);
1082 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1083 struct kvm_mmu_page *sp)
1087 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1088 sp->spt[i] = shadow_trap_nonpresent_pte;
1091 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1092 struct kvm_mmu_page *sp)
1097 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1101 #define KVM_PAGE_ARRAY_NR 16
1103 struct kvm_mmu_pages {
1104 struct mmu_page_and_offset {
1105 struct kvm_mmu_page *sp;
1107 } page[KVM_PAGE_ARRAY_NR];
1111 #define for_each_unsync_children(bitmap, idx) \
1112 for (idx = find_first_bit(bitmap, 512); \
1114 idx = find_next_bit(bitmap, 512, idx+1))
1116 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1122 for (i=0; i < pvec->nr; i++)
1123 if (pvec->page[i].sp == sp)
1126 pvec->page[pvec->nr].sp = sp;
1127 pvec->page[pvec->nr].idx = idx;
1129 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1132 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1133 struct kvm_mmu_pages *pvec)
1135 int i, ret, nr_unsync_leaf = 0;
1137 for_each_unsync_children(sp->unsync_child_bitmap, i) {
1138 u64 ent = sp->spt[i];
1140 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1141 struct kvm_mmu_page *child;
1142 child = page_header(ent & PT64_BASE_ADDR_MASK);
1144 if (child->unsync_children) {
1145 if (mmu_pages_add(pvec, child, i))
1148 ret = __mmu_unsync_walk(child, pvec);
1150 __clear_bit(i, sp->unsync_child_bitmap);
1152 nr_unsync_leaf += ret;
1157 if (child->unsync) {
1159 if (mmu_pages_add(pvec, child, i))
1165 if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1166 sp->unsync_children = 0;
1168 return nr_unsync_leaf;
1171 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1172 struct kvm_mmu_pages *pvec)
1174 if (!sp->unsync_children)
1177 mmu_pages_add(pvec, sp, 0);
1178 return __mmu_unsync_walk(sp, pvec);
1181 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1184 struct hlist_head *bucket;
1185 struct kvm_mmu_page *sp;
1186 struct hlist_node *node;
1188 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1189 index = kvm_page_table_hashfn(gfn);
1190 bucket = &kvm->arch.mmu_page_hash[index];
1191 hlist_for_each_entry(sp, node, bucket, hash_link)
1192 if (sp->gfn == gfn && !sp->role.direct
1193 && !sp->role.invalid) {
1194 pgprintk("%s: found role %x\n",
1195 __func__, sp->role.word);
1201 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1203 WARN_ON(!sp->unsync);
1205 --kvm->stat.mmu_unsync;
1208 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1210 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1212 if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1213 kvm_mmu_zap_page(vcpu->kvm, sp);
1217 trace_kvm_mmu_sync_page(sp);
1218 if (rmap_write_protect(vcpu->kvm, sp->gfn))
1219 kvm_flush_remote_tlbs(vcpu->kvm);
1220 kvm_unlink_unsync_page(vcpu->kvm, sp);
1221 if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1222 kvm_mmu_zap_page(vcpu->kvm, sp);
1226 kvm_mmu_flush_tlb(vcpu);
1230 struct mmu_page_path {
1231 struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1232 unsigned int idx[PT64_ROOT_LEVEL-1];
1235 #define for_each_sp(pvec, sp, parents, i) \
1236 for (i = mmu_pages_next(&pvec, &parents, -1), \
1237 sp = pvec.page[i].sp; \
1238 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1239 i = mmu_pages_next(&pvec, &parents, i))
1241 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1242 struct mmu_page_path *parents,
1247 for (n = i+1; n < pvec->nr; n++) {
1248 struct kvm_mmu_page *sp = pvec->page[n].sp;
1250 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1251 parents->idx[0] = pvec->page[n].idx;
1255 parents->parent[sp->role.level-2] = sp;
1256 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1262 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1264 struct kvm_mmu_page *sp;
1265 unsigned int level = 0;
1268 unsigned int idx = parents->idx[level];
1270 sp = parents->parent[level];
1274 --sp->unsync_children;
1275 WARN_ON((int)sp->unsync_children < 0);
1276 __clear_bit(idx, sp->unsync_child_bitmap);
1278 } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1281 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1282 struct mmu_page_path *parents,
1283 struct kvm_mmu_pages *pvec)
1285 parents->parent[parent->role.level-1] = NULL;
1289 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1290 struct kvm_mmu_page *parent)
1293 struct kvm_mmu_page *sp;
1294 struct mmu_page_path parents;
1295 struct kvm_mmu_pages pages;
1297 kvm_mmu_pages_init(parent, &parents, &pages);
1298 while (mmu_unsync_walk(parent, &pages)) {
1301 for_each_sp(pages, sp, parents, i)
1302 protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1305 kvm_flush_remote_tlbs(vcpu->kvm);
1307 for_each_sp(pages, sp, parents, i) {
1308 kvm_sync_page(vcpu, sp);
1309 mmu_pages_clear_parents(&parents);
1311 cond_resched_lock(&vcpu->kvm->mmu_lock);
1312 kvm_mmu_pages_init(parent, &parents, &pages);
1316 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1324 union kvm_mmu_page_role role;
1327 struct hlist_head *bucket;
1328 struct kvm_mmu_page *sp;
1329 struct hlist_node *node, *tmp;
1331 role = vcpu->arch.mmu.base_role;
1333 role.direct = direct;
1334 role.access = access;
1335 if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1336 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1337 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1338 role.quadrant = quadrant;
1340 index = kvm_page_table_hashfn(gfn);
1341 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1342 hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1343 if (sp->gfn == gfn) {
1345 if (kvm_sync_page(vcpu, sp))
1348 if (sp->role.word != role.word)
1351 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1352 if (sp->unsync_children) {
1353 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1354 kvm_mmu_mark_parents_unsync(vcpu, sp);
1356 trace_kvm_mmu_get_page(sp, false);
1359 ++vcpu->kvm->stat.mmu_cache_miss;
1360 sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1365 hlist_add_head(&sp->hash_link, bucket);
1367 if (rmap_write_protect(vcpu->kvm, gfn))
1368 kvm_flush_remote_tlbs(vcpu->kvm);
1369 account_shadowed(vcpu->kvm, gfn);
1371 if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1372 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1374 nonpaging_prefetch_page(vcpu, sp);
1375 trace_kvm_mmu_get_page(sp, true);
1379 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1380 struct kvm_vcpu *vcpu, u64 addr)
1382 iterator->addr = addr;
1383 iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1384 iterator->level = vcpu->arch.mmu.shadow_root_level;
1385 if (iterator->level == PT32E_ROOT_LEVEL) {
1386 iterator->shadow_addr
1387 = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1388 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1390 if (!iterator->shadow_addr)
1391 iterator->level = 0;
1395 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1397 if (iterator->level < PT_PAGE_TABLE_LEVEL)
1400 if (iterator->level == PT_PAGE_TABLE_LEVEL)
1401 if (is_large_pte(*iterator->sptep))
1404 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1405 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1409 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1411 iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1415 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1416 struct kvm_mmu_page *sp)
1424 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1427 if (is_shadow_present_pte(ent)) {
1428 if (!is_last_spte(ent, sp->role.level)) {
1429 ent &= PT64_BASE_ADDR_MASK;
1430 mmu_page_remove_parent_pte(page_header(ent),
1433 if (is_large_pte(ent))
1435 rmap_remove(kvm, &pt[i]);
1438 pt[i] = shadow_trap_nonpresent_pte;
1442 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1444 mmu_page_remove_parent_pte(sp, parent_pte);
1447 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1450 struct kvm_vcpu *vcpu;
1452 kvm_for_each_vcpu(i, vcpu, kvm)
1453 vcpu->arch.last_pte_updated = NULL;
1456 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1460 while (sp->multimapped || sp->parent_pte) {
1461 if (!sp->multimapped)
1462 parent_pte = sp->parent_pte;
1464 struct kvm_pte_chain *chain;
1466 chain = container_of(sp->parent_ptes.first,
1467 struct kvm_pte_chain, link);
1468 parent_pte = chain->parent_ptes[0];
1470 BUG_ON(!parent_pte);
1471 kvm_mmu_put_page(sp, parent_pte);
1472 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1476 static int mmu_zap_unsync_children(struct kvm *kvm,
1477 struct kvm_mmu_page *parent)
1480 struct mmu_page_path parents;
1481 struct kvm_mmu_pages pages;
1483 if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1486 kvm_mmu_pages_init(parent, &parents, &pages);
1487 while (mmu_unsync_walk(parent, &pages)) {
1488 struct kvm_mmu_page *sp;
1490 for_each_sp(pages, sp, parents, i) {
1491 kvm_mmu_zap_page(kvm, sp);
1492 mmu_pages_clear_parents(&parents);
1495 kvm_mmu_pages_init(parent, &parents, &pages);
1501 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1505 trace_kvm_mmu_zap_page(sp);
1506 ++kvm->stat.mmu_shadow_zapped;
1507 ret = mmu_zap_unsync_children(kvm, sp);
1508 kvm_mmu_page_unlink_children(kvm, sp);
1509 kvm_mmu_unlink_parents(kvm, sp);
1510 kvm_flush_remote_tlbs(kvm);
1511 if (!sp->role.invalid && !sp->role.direct)
1512 unaccount_shadowed(kvm, sp->gfn);
1514 kvm_unlink_unsync_page(kvm, sp);
1515 if (!sp->root_count) {
1516 hlist_del(&sp->hash_link);
1517 kvm_mmu_free_page(kvm, sp);
1519 sp->role.invalid = 1;
1520 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1521 kvm_reload_remote_mmus(kvm);
1523 kvm_mmu_reset_last_pte_updated(kvm);
1528 * Changing the number of mmu pages allocated to the vm
1529 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1531 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1535 used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1536 used_pages = max(0, used_pages);
1539 * If we set the number of mmu pages to be smaller be than the
1540 * number of actived pages , we must to free some mmu pages before we
1544 if (used_pages > kvm_nr_mmu_pages) {
1545 while (used_pages > kvm_nr_mmu_pages &&
1546 !list_empty(&kvm->arch.active_mmu_pages)) {
1547 struct kvm_mmu_page *page;
1549 page = container_of(kvm->arch.active_mmu_pages.prev,
1550 struct kvm_mmu_page, link);
1551 used_pages -= kvm_mmu_zap_page(kvm, page);
1554 kvm_nr_mmu_pages = used_pages;
1555 kvm->arch.n_free_mmu_pages = 0;
1558 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1559 - kvm->arch.n_alloc_mmu_pages;
1561 kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1564 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1567 struct hlist_head *bucket;
1568 struct kvm_mmu_page *sp;
1569 struct hlist_node *node, *n;
1572 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1574 index = kvm_page_table_hashfn(gfn);
1575 bucket = &kvm->arch.mmu_page_hash[index];
1576 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1577 if (sp->gfn == gfn && !sp->role.direct) {
1578 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1581 if (kvm_mmu_zap_page(kvm, sp))
1587 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1590 struct hlist_head *bucket;
1591 struct kvm_mmu_page *sp;
1592 struct hlist_node *node, *nn;
1594 index = kvm_page_table_hashfn(gfn);
1595 bucket = &kvm->arch.mmu_page_hash[index];
1596 hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1597 if (sp->gfn == gfn && !sp->role.direct
1598 && !sp->role.invalid) {
1599 pgprintk("%s: zap %lx %x\n",
1600 __func__, gfn, sp->role.word);
1601 if (kvm_mmu_zap_page(kvm, sp))
1607 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1609 int slot = memslot_id(kvm, gfn);
1610 struct kvm_mmu_page *sp = page_header(__pa(pte));
1612 __set_bit(slot, sp->slot_bitmap);
1615 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1620 if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1623 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1624 if (pt[i] == shadow_notrap_nonpresent_pte)
1625 __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1629 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1633 gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
1635 if (gpa == UNMAPPED_GVA)
1638 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1644 * The function is based on mtrr_type_lookup() in
1645 * arch/x86/kernel/cpu/mtrr/generic.c
1647 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1652 u8 prev_match, curr_match;
1653 int num_var_ranges = KVM_NR_VAR_MTRR;
1655 if (!mtrr_state->enabled)
1658 /* Make end inclusive end, instead of exclusive */
1661 /* Look in fixed ranges. Just return the type as per start */
1662 if (mtrr_state->have_fixed && (start < 0x100000)) {
1665 if (start < 0x80000) {
1667 idx += (start >> 16);
1668 return mtrr_state->fixed_ranges[idx];
1669 } else if (start < 0xC0000) {
1671 idx += ((start - 0x80000) >> 14);
1672 return mtrr_state->fixed_ranges[idx];
1673 } else if (start < 0x1000000) {
1675 idx += ((start - 0xC0000) >> 12);
1676 return mtrr_state->fixed_ranges[idx];
1681 * Look in variable ranges
1682 * Look of multiple ranges matching this address and pick type
1683 * as per MTRR precedence
1685 if (!(mtrr_state->enabled & 2))
1686 return mtrr_state->def_type;
1689 for (i = 0; i < num_var_ranges; ++i) {
1690 unsigned short start_state, end_state;
1692 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1695 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1696 (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1697 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1698 (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1700 start_state = ((start & mask) == (base & mask));
1701 end_state = ((end & mask) == (base & mask));
1702 if (start_state != end_state)
1705 if ((start & mask) != (base & mask))
1708 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1709 if (prev_match == 0xFF) {
1710 prev_match = curr_match;
1714 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1715 curr_match == MTRR_TYPE_UNCACHABLE)
1716 return MTRR_TYPE_UNCACHABLE;
1718 if ((prev_match == MTRR_TYPE_WRBACK &&
1719 curr_match == MTRR_TYPE_WRTHROUGH) ||
1720 (prev_match == MTRR_TYPE_WRTHROUGH &&
1721 curr_match == MTRR_TYPE_WRBACK)) {
1722 prev_match = MTRR_TYPE_WRTHROUGH;
1723 curr_match = MTRR_TYPE_WRTHROUGH;
1726 if (prev_match != curr_match)
1727 return MTRR_TYPE_UNCACHABLE;
1730 if (prev_match != 0xFF)
1733 return mtrr_state->def_type;
1736 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1740 mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1741 (gfn << PAGE_SHIFT) + PAGE_SIZE);
1742 if (mtrr == 0xfe || mtrr == 0xff)
1743 mtrr = MTRR_TYPE_WRBACK;
1746 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1748 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1751 struct hlist_head *bucket;
1752 struct kvm_mmu_page *s;
1753 struct hlist_node *node, *n;
1755 trace_kvm_mmu_unsync_page(sp);
1756 index = kvm_page_table_hashfn(sp->gfn);
1757 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1758 /* don't unsync if pagetable is shadowed with multiple roles */
1759 hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1760 if (s->gfn != sp->gfn || s->role.direct)
1762 if (s->role.word != sp->role.word)
1765 ++vcpu->kvm->stat.mmu_unsync;
1768 kvm_mmu_mark_parents_unsync(vcpu, sp);
1770 mmu_convert_notrap(sp);
1774 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1777 struct kvm_mmu_page *shadow;
1779 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1781 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1785 if (can_unsync && oos_shadow)
1786 return kvm_unsync_page(vcpu, shadow);
1792 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1793 unsigned pte_access, int user_fault,
1794 int write_fault, int dirty, int level,
1795 gfn_t gfn, pfn_t pfn, bool speculative,
1796 bool can_unsync, bool reset_host_protection)
1802 * We don't set the accessed bit, since we sometimes want to see
1803 * whether the guest actually used the pte (in order to detect
1806 spte = shadow_base_present_pte | shadow_dirty_mask;
1808 spte |= shadow_accessed_mask;
1810 pte_access &= ~ACC_WRITE_MASK;
1811 if (pte_access & ACC_EXEC_MASK)
1812 spte |= shadow_x_mask;
1814 spte |= shadow_nx_mask;
1815 if (pte_access & ACC_USER_MASK)
1816 spte |= shadow_user_mask;
1817 if (level > PT_PAGE_TABLE_LEVEL)
1818 spte |= PT_PAGE_SIZE_MASK;
1820 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1821 kvm_is_mmio_pfn(pfn));
1823 if (reset_host_protection)
1824 spte |= SPTE_HOST_WRITEABLE;
1826 spte |= (u64)pfn << PAGE_SHIFT;
1828 if ((pte_access & ACC_WRITE_MASK)
1829 || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1831 if (level > PT_PAGE_TABLE_LEVEL &&
1832 has_wrprotected_page(vcpu->kvm, gfn, level)) {
1834 spte = shadow_trap_nonpresent_pte;
1838 spte |= PT_WRITABLE_MASK;
1841 * Optimization: for pte sync, if spte was writable the hash
1842 * lookup is unnecessary (and expensive). Write protection
1843 * is responsibility of mmu_get_page / kvm_sync_page.
1844 * Same reasoning can be applied to dirty page accounting.
1846 if (!can_unsync && is_writable_pte(*sptep))
1849 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1850 pgprintk("%s: found shadow page for %lx, marking ro\n",
1853 pte_access &= ~ACC_WRITE_MASK;
1854 if (is_writable_pte(spte))
1855 spte &= ~PT_WRITABLE_MASK;
1859 if (pte_access & ACC_WRITE_MASK)
1860 mark_page_dirty(vcpu->kvm, gfn);
1863 __set_spte(sptep, spte);
1867 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1868 unsigned pt_access, unsigned pte_access,
1869 int user_fault, int write_fault, int dirty,
1870 int *ptwrite, int level, gfn_t gfn,
1871 pfn_t pfn, bool speculative,
1872 bool reset_host_protection)
1874 int was_rmapped = 0;
1875 int was_writable = is_writable_pte(*sptep);
1878 pgprintk("%s: spte %llx access %x write_fault %d"
1879 " user_fault %d gfn %lx\n",
1880 __func__, *sptep, pt_access,
1881 write_fault, user_fault, gfn);
1883 if (is_rmap_spte(*sptep)) {
1885 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1886 * the parent of the now unreachable PTE.
1888 if (level > PT_PAGE_TABLE_LEVEL &&
1889 !is_large_pte(*sptep)) {
1890 struct kvm_mmu_page *child;
1893 child = page_header(pte & PT64_BASE_ADDR_MASK);
1894 mmu_page_remove_parent_pte(child, sptep);
1895 } else if (pfn != spte_to_pfn(*sptep)) {
1896 pgprintk("hfn old %lx new %lx\n",
1897 spte_to_pfn(*sptep), pfn);
1898 rmap_remove(vcpu->kvm, sptep);
1903 if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1904 dirty, level, gfn, pfn, speculative, true,
1905 reset_host_protection)) {
1908 kvm_x86_ops->tlb_flush(vcpu);
1911 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1912 pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1913 is_large_pte(*sptep)? "2MB" : "4kB",
1914 *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1916 if (!was_rmapped && is_large_pte(*sptep))
1917 ++vcpu->kvm->stat.lpages;
1919 page_header_update_slot(vcpu->kvm, sptep, gfn);
1921 rmap_count = rmap_add(vcpu, sptep, gfn);
1922 kvm_release_pfn_clean(pfn);
1923 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1924 rmap_recycle(vcpu, sptep, gfn);
1927 kvm_release_pfn_dirty(pfn);
1929 kvm_release_pfn_clean(pfn);
1932 vcpu->arch.last_pte_updated = sptep;
1933 vcpu->arch.last_pte_gfn = gfn;
1937 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1941 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1942 int level, gfn_t gfn, pfn_t pfn)
1944 struct kvm_shadow_walk_iterator iterator;
1945 struct kvm_mmu_page *sp;
1949 for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1950 if (iterator.level == level) {
1951 mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1952 0, write, 1, &pt_write,
1953 level, gfn, pfn, false, true);
1954 ++vcpu->stat.pf_fixed;
1958 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1959 pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1960 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1962 1, ACC_ALL, iterator.sptep);
1964 pgprintk("nonpaging_map: ENOMEM\n");
1965 kvm_release_pfn_clean(pfn);
1969 __set_spte(iterator.sptep,
1971 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1972 | shadow_user_mask | shadow_x_mask);
1978 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1983 unsigned long mmu_seq;
1985 level = mapping_level(vcpu, gfn);
1988 * This path builds a PAE pagetable - so we can map 2mb pages at
1989 * maximum. Therefore check if the level is larger than that.
1991 if (level > PT_DIRECTORY_LEVEL)
1992 level = PT_DIRECTORY_LEVEL;
1994 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
1996 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1998 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2001 if (is_error_pfn(pfn)) {
2002 kvm_release_pfn_clean(pfn);
2006 spin_lock(&vcpu->kvm->mmu_lock);
2007 if (mmu_notifier_retry(vcpu, mmu_seq))
2009 kvm_mmu_free_some_pages(vcpu);
2010 r = __direct_map(vcpu, v, write, level, gfn, pfn);
2011 spin_unlock(&vcpu->kvm->mmu_lock);
2017 spin_unlock(&vcpu->kvm->mmu_lock);
2018 kvm_release_pfn_clean(pfn);
2023 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2026 struct kvm_mmu_page *sp;
2028 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2030 spin_lock(&vcpu->kvm->mmu_lock);
2031 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2032 hpa_t root = vcpu->arch.mmu.root_hpa;
2034 sp = page_header(root);
2036 if (!sp->root_count && sp->role.invalid)
2037 kvm_mmu_zap_page(vcpu->kvm, sp);
2038 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2039 spin_unlock(&vcpu->kvm->mmu_lock);
2042 for (i = 0; i < 4; ++i) {
2043 hpa_t root = vcpu->arch.mmu.pae_root[i];
2046 root &= PT64_BASE_ADDR_MASK;
2047 sp = page_header(root);
2049 if (!sp->root_count && sp->role.invalid)
2050 kvm_mmu_zap_page(vcpu->kvm, sp);
2052 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2054 spin_unlock(&vcpu->kvm->mmu_lock);
2055 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2058 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2062 if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2063 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2070 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2074 struct kvm_mmu_page *sp;
2078 root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2080 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2081 hpa_t root = vcpu->arch.mmu.root_hpa;
2083 ASSERT(!VALID_PAGE(root));
2086 if (mmu_check_root(vcpu, root_gfn))
2088 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2089 PT64_ROOT_LEVEL, direct,
2091 root = __pa(sp->spt);
2093 vcpu->arch.mmu.root_hpa = root;
2096 direct = !is_paging(vcpu);
2099 for (i = 0; i < 4; ++i) {
2100 hpa_t root = vcpu->arch.mmu.pae_root[i];
2102 ASSERT(!VALID_PAGE(root));
2103 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2104 pdptr = kvm_pdptr_read(vcpu, i);
2105 if (!is_present_gpte(pdptr)) {
2106 vcpu->arch.mmu.pae_root[i] = 0;
2109 root_gfn = pdptr >> PAGE_SHIFT;
2110 } else if (vcpu->arch.mmu.root_level == 0)
2112 if (mmu_check_root(vcpu, root_gfn))
2114 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2115 PT32_ROOT_LEVEL, direct,
2117 root = __pa(sp->spt);
2119 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2121 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2125 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2128 struct kvm_mmu_page *sp;
2130 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2132 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2133 hpa_t root = vcpu->arch.mmu.root_hpa;
2134 sp = page_header(root);
2135 mmu_sync_children(vcpu, sp);
2138 for (i = 0; i < 4; ++i) {
2139 hpa_t root = vcpu->arch.mmu.pae_root[i];
2141 if (root && VALID_PAGE(root)) {
2142 root &= PT64_BASE_ADDR_MASK;
2143 sp = page_header(root);
2144 mmu_sync_children(vcpu, sp);
2149 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2151 spin_lock(&vcpu->kvm->mmu_lock);
2152 mmu_sync_roots(vcpu);
2153 spin_unlock(&vcpu->kvm->mmu_lock);
2156 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2157 u32 access, u32 *error)
2164 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2170 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2171 r = mmu_topup_memory_caches(vcpu);
2176 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2178 gfn = gva >> PAGE_SHIFT;
2180 return nonpaging_map(vcpu, gva & PAGE_MASK,
2181 error_code & PFERR_WRITE_MASK, gfn);
2184 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2190 gfn_t gfn = gpa >> PAGE_SHIFT;
2191 unsigned long mmu_seq;
2194 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2196 r = mmu_topup_memory_caches(vcpu);
2200 level = mapping_level(vcpu, gfn);
2202 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2204 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2206 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2207 if (is_error_pfn(pfn)) {
2208 kvm_release_pfn_clean(pfn);
2211 spin_lock(&vcpu->kvm->mmu_lock);
2212 if (mmu_notifier_retry(vcpu, mmu_seq))
2214 kvm_mmu_free_some_pages(vcpu);
2215 r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2217 spin_unlock(&vcpu->kvm->mmu_lock);
2222 spin_unlock(&vcpu->kvm->mmu_lock);
2223 kvm_release_pfn_clean(pfn);
2227 static void nonpaging_free(struct kvm_vcpu *vcpu)
2229 mmu_free_roots(vcpu);
2232 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2234 struct kvm_mmu *context = &vcpu->arch.mmu;
2236 context->new_cr3 = nonpaging_new_cr3;
2237 context->page_fault = nonpaging_page_fault;
2238 context->gva_to_gpa = nonpaging_gva_to_gpa;
2239 context->free = nonpaging_free;
2240 context->prefetch_page = nonpaging_prefetch_page;
2241 context->sync_page = nonpaging_sync_page;
2242 context->invlpg = nonpaging_invlpg;
2243 context->root_level = 0;
2244 context->shadow_root_level = PT32E_ROOT_LEVEL;
2245 context->root_hpa = INVALID_PAGE;
2249 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2251 ++vcpu->stat.tlb_flush;
2252 kvm_x86_ops->tlb_flush(vcpu);
2255 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2257 pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2258 mmu_free_roots(vcpu);
2261 static void inject_page_fault(struct kvm_vcpu *vcpu,
2265 kvm_inject_page_fault(vcpu, addr, err_code);
2268 static void paging_free(struct kvm_vcpu *vcpu)
2270 nonpaging_free(vcpu);
2273 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2277 bit7 = (gpte >> 7) & 1;
2278 return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2282 #include "paging_tmpl.h"
2286 #include "paging_tmpl.h"
2289 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2291 struct kvm_mmu *context = &vcpu->arch.mmu;
2292 int maxphyaddr = cpuid_maxphyaddr(vcpu);
2293 u64 exb_bit_rsvd = 0;
2296 exb_bit_rsvd = rsvd_bits(63, 63);
2298 case PT32_ROOT_LEVEL:
2299 /* no rsvd bits for 2 level 4K page table entries */
2300 context->rsvd_bits_mask[0][1] = 0;
2301 context->rsvd_bits_mask[0][0] = 0;
2302 if (is_cpuid_PSE36())
2303 /* 36bits PSE 4MB page */
2304 context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2306 /* 32 bits PSE 4MB page */
2307 context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2308 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2310 case PT32E_ROOT_LEVEL:
2311 context->rsvd_bits_mask[0][2] =
2312 rsvd_bits(maxphyaddr, 63) |
2313 rsvd_bits(7, 8) | rsvd_bits(1, 2); /* PDPTE */
2314 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2315 rsvd_bits(maxphyaddr, 62); /* PDE */
2316 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2317 rsvd_bits(maxphyaddr, 62); /* PTE */
2318 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2319 rsvd_bits(maxphyaddr, 62) |
2320 rsvd_bits(13, 20); /* large page */
2321 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2323 case PT64_ROOT_LEVEL:
2324 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2325 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2326 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2327 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2328 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2329 rsvd_bits(maxphyaddr, 51);
2330 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2331 rsvd_bits(maxphyaddr, 51);
2332 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2333 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2334 rsvd_bits(maxphyaddr, 51) |
2336 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2337 rsvd_bits(maxphyaddr, 51) |
2338 rsvd_bits(13, 20); /* large page */
2339 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2344 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2346 struct kvm_mmu *context = &vcpu->arch.mmu;
2348 ASSERT(is_pae(vcpu));
2349 context->new_cr3 = paging_new_cr3;
2350 context->page_fault = paging64_page_fault;
2351 context->gva_to_gpa = paging64_gva_to_gpa;
2352 context->prefetch_page = paging64_prefetch_page;
2353 context->sync_page = paging64_sync_page;
2354 context->invlpg = paging64_invlpg;
2355 context->free = paging_free;
2356 context->root_level = level;
2357 context->shadow_root_level = level;
2358 context->root_hpa = INVALID_PAGE;
2362 static int paging64_init_context(struct kvm_vcpu *vcpu)
2364 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2365 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2368 static int paging32_init_context(struct kvm_vcpu *vcpu)
2370 struct kvm_mmu *context = &vcpu->arch.mmu;
2372 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2373 context->new_cr3 = paging_new_cr3;
2374 context->page_fault = paging32_page_fault;
2375 context->gva_to_gpa = paging32_gva_to_gpa;
2376 context->free = paging_free;
2377 context->prefetch_page = paging32_prefetch_page;
2378 context->sync_page = paging32_sync_page;
2379 context->invlpg = paging32_invlpg;
2380 context->root_level = PT32_ROOT_LEVEL;
2381 context->shadow_root_level = PT32E_ROOT_LEVEL;
2382 context->root_hpa = INVALID_PAGE;
2386 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2388 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2389 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2392 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2394 struct kvm_mmu *context = &vcpu->arch.mmu;
2396 context->new_cr3 = nonpaging_new_cr3;
2397 context->page_fault = tdp_page_fault;
2398 context->free = nonpaging_free;
2399 context->prefetch_page = nonpaging_prefetch_page;
2400 context->sync_page = nonpaging_sync_page;
2401 context->invlpg = nonpaging_invlpg;
2402 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2403 context->root_hpa = INVALID_PAGE;
2405 if (!is_paging(vcpu)) {
2406 context->gva_to_gpa = nonpaging_gva_to_gpa;
2407 context->root_level = 0;
2408 } else if (is_long_mode(vcpu)) {
2409 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2410 context->gva_to_gpa = paging64_gva_to_gpa;
2411 context->root_level = PT64_ROOT_LEVEL;
2412 } else if (is_pae(vcpu)) {
2413 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2414 context->gva_to_gpa = paging64_gva_to_gpa;
2415 context->root_level = PT32E_ROOT_LEVEL;
2417 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2418 context->gva_to_gpa = paging32_gva_to_gpa;
2419 context->root_level = PT32_ROOT_LEVEL;
2425 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2430 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2432 if (!is_paging(vcpu))
2433 r = nonpaging_init_context(vcpu);
2434 else if (is_long_mode(vcpu))
2435 r = paging64_init_context(vcpu);
2436 else if (is_pae(vcpu))
2437 r = paging32E_init_context(vcpu);
2439 r = paging32_init_context(vcpu);
2441 vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2446 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2448 vcpu->arch.update_pte.pfn = bad_pfn;
2451 return init_kvm_tdp_mmu(vcpu);
2453 return init_kvm_softmmu(vcpu);
2456 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2459 if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2460 vcpu->arch.mmu.free(vcpu);
2461 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2465 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2467 destroy_kvm_mmu(vcpu);
2468 return init_kvm_mmu(vcpu);
2470 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2472 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2476 r = mmu_topup_memory_caches(vcpu);
2479 spin_lock(&vcpu->kvm->mmu_lock);
2480 kvm_mmu_free_some_pages(vcpu);
2481 r = mmu_alloc_roots(vcpu);
2482 mmu_sync_roots(vcpu);
2483 spin_unlock(&vcpu->kvm->mmu_lock);
2486 /* set_cr3() should ensure TLB has been flushed */
2487 kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2491 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2493 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2495 mmu_free_roots(vcpu);
2498 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2499 struct kvm_mmu_page *sp,
2503 struct kvm_mmu_page *child;
2506 if (is_shadow_present_pte(pte)) {
2507 if (is_last_spte(pte, sp->role.level))
2508 rmap_remove(vcpu->kvm, spte);
2510 child = page_header(pte & PT64_BASE_ADDR_MASK);
2511 mmu_page_remove_parent_pte(child, spte);
2514 __set_spte(spte, shadow_trap_nonpresent_pte);
2515 if (is_large_pte(pte))
2516 --vcpu->kvm->stat.lpages;
2519 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2520 struct kvm_mmu_page *sp,
2524 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2525 ++vcpu->kvm->stat.mmu_pde_zapped;
2529 ++vcpu->kvm->stat.mmu_pte_updated;
2530 if (sp->role.glevels == PT32_ROOT_LEVEL)
2531 paging32_update_pte(vcpu, sp, spte, new);
2533 paging64_update_pte(vcpu, sp, spte, new);
2536 static bool need_remote_flush(u64 old, u64 new)
2538 if (!is_shadow_present_pte(old))
2540 if (!is_shadow_present_pte(new))
2542 if ((old ^ new) & PT64_BASE_ADDR_MASK)
2544 old ^= PT64_NX_MASK;
2545 new ^= PT64_NX_MASK;
2546 return (old & ~new & PT64_PERM_MASK) != 0;
2549 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2551 if (need_remote_flush(old, new))
2552 kvm_flush_remote_tlbs(vcpu->kvm);
2554 kvm_mmu_flush_tlb(vcpu);
2557 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2559 u64 *spte = vcpu->arch.last_pte_updated;
2561 return !!(spte && (*spte & shadow_accessed_mask));
2564 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2565 const u8 *new, int bytes)
2572 if (bytes != 4 && bytes != 8)
2576 * Assume that the pte write on a page table of the same type
2577 * as the current vcpu paging mode. This is nearly always true
2578 * (might be false while changing modes). Note it is verified later
2582 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2583 if ((bytes == 4) && (gpa % 4 == 0)) {
2584 r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2587 memcpy((void *)&gpte + (gpa % 8), new, 4);
2588 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2589 memcpy((void *)&gpte, new, 8);
2592 if ((bytes == 4) && (gpa % 4 == 0))
2593 memcpy((void *)&gpte, new, 4);
2595 if (!is_present_gpte(gpte))
2597 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2599 vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2601 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2603 if (is_error_pfn(pfn)) {
2604 kvm_release_pfn_clean(pfn);
2607 vcpu->arch.update_pte.gfn = gfn;
2608 vcpu->arch.update_pte.pfn = pfn;
2611 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2613 u64 *spte = vcpu->arch.last_pte_updated;
2616 && vcpu->arch.last_pte_gfn == gfn
2617 && shadow_accessed_mask
2618 && !(*spte & shadow_accessed_mask)
2619 && is_shadow_present_pte(*spte))
2620 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2623 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2624 const u8 *new, int bytes,
2625 bool guest_initiated)
2627 gfn_t gfn = gpa >> PAGE_SHIFT;
2628 struct kvm_mmu_page *sp;
2629 struct hlist_node *node, *n;
2630 struct hlist_head *bucket;
2634 unsigned offset = offset_in_page(gpa);
2636 unsigned page_offset;
2637 unsigned misaligned;
2644 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2645 mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2646 spin_lock(&vcpu->kvm->mmu_lock);
2647 kvm_mmu_access_page(vcpu, gfn);
2648 kvm_mmu_free_some_pages(vcpu);
2649 ++vcpu->kvm->stat.mmu_pte_write;
2650 kvm_mmu_audit(vcpu, "pre pte write");
2651 if (guest_initiated) {
2652 if (gfn == vcpu->arch.last_pt_write_gfn
2653 && !last_updated_pte_accessed(vcpu)) {
2654 ++vcpu->arch.last_pt_write_count;
2655 if (vcpu->arch.last_pt_write_count >= 3)
2658 vcpu->arch.last_pt_write_gfn = gfn;
2659 vcpu->arch.last_pt_write_count = 1;
2660 vcpu->arch.last_pte_updated = NULL;
2663 index = kvm_page_table_hashfn(gfn);
2664 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2665 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2666 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2668 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2669 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2670 misaligned |= bytes < 4;
2671 if (misaligned || flooded) {
2673 * Misaligned accesses are too much trouble to fix
2674 * up; also, they usually indicate a page is not used
2677 * If we're seeing too many writes to a page,
2678 * it may no longer be a page table, or we may be
2679 * forking, in which case it is better to unmap the
2682 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2683 gpa, bytes, sp->role.word);
2684 if (kvm_mmu_zap_page(vcpu->kvm, sp))
2686 ++vcpu->kvm->stat.mmu_flooded;
2689 page_offset = offset;
2690 level = sp->role.level;
2692 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2693 page_offset <<= 1; /* 32->64 */
2695 * A 32-bit pde maps 4MB while the shadow pdes map
2696 * only 2MB. So we need to double the offset again
2697 * and zap two pdes instead of one.
2699 if (level == PT32_ROOT_LEVEL) {
2700 page_offset &= ~7; /* kill rounding error */
2704 quadrant = page_offset >> PAGE_SHIFT;
2705 page_offset &= ~PAGE_MASK;
2706 if (quadrant != sp->role.quadrant)
2709 spte = &sp->spt[page_offset / sizeof(*spte)];
2710 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2712 r = kvm_read_guest_atomic(vcpu->kvm,
2713 gpa & ~(u64)(pte_size - 1),
2715 new = (const void *)&gentry;
2721 mmu_pte_write_zap_pte(vcpu, sp, spte);
2723 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2724 mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2728 kvm_mmu_audit(vcpu, "post pte write");
2729 spin_unlock(&vcpu->kvm->mmu_lock);
2730 if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2731 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2732 vcpu->arch.update_pte.pfn = bad_pfn;
2736 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2744 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2746 spin_lock(&vcpu->kvm->mmu_lock);
2747 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2748 spin_unlock(&vcpu->kvm->mmu_lock);
2751 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2753 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2755 while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2756 !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2757 struct kvm_mmu_page *sp;
2759 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2760 struct kvm_mmu_page, link);
2761 kvm_mmu_zap_page(vcpu->kvm, sp);
2762 ++vcpu->kvm->stat.mmu_recycled;
2766 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2769 enum emulation_result er;
2771 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2780 r = mmu_topup_memory_caches(vcpu);
2784 er = emulate_instruction(vcpu, cr2, error_code, 0);
2789 case EMULATE_DO_MMIO:
2790 ++vcpu->stat.mmio_exits;
2793 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2794 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2795 vcpu->run->internal.ndata = 0;
2803 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2805 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2807 vcpu->arch.mmu.invlpg(vcpu, gva);
2808 kvm_mmu_flush_tlb(vcpu);
2809 ++vcpu->stat.invlpg;
2811 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2813 void kvm_enable_tdp(void)
2817 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2819 void kvm_disable_tdp(void)
2821 tdp_enabled = false;
2823 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2825 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2827 free_page((unsigned long)vcpu->arch.mmu.pae_root);
2830 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2838 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2839 * Therefore we need to allocate shadow page tables in the first
2840 * 4GB of memory, which happens to fit the DMA32 zone.
2842 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2846 vcpu->arch.mmu.pae_root = page_address(page);
2847 for (i = 0; i < 4; ++i)
2848 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2853 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2856 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2858 return alloc_mmu_pages(vcpu);
2861 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2864 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2866 return init_kvm_mmu(vcpu);
2869 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2873 destroy_kvm_mmu(vcpu);
2874 free_mmu_pages(vcpu);
2875 mmu_free_memory_caches(vcpu);
2878 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2880 struct kvm_mmu_page *sp;
2882 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2886 if (!test_bit(slot, sp->slot_bitmap))
2890 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2892 if (pt[i] & PT_WRITABLE_MASK)
2893 pt[i] &= ~PT_WRITABLE_MASK;
2895 kvm_flush_remote_tlbs(kvm);
2898 void kvm_mmu_zap_all(struct kvm *kvm)
2900 struct kvm_mmu_page *sp, *node;
2902 spin_lock(&kvm->mmu_lock);
2903 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2904 if (kvm_mmu_zap_page(kvm, sp))
2905 node = container_of(kvm->arch.active_mmu_pages.next,
2906 struct kvm_mmu_page, link);
2907 spin_unlock(&kvm->mmu_lock);
2909 kvm_flush_remote_tlbs(kvm);
2912 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2914 struct kvm_mmu_page *page;
2916 page = container_of(kvm->arch.active_mmu_pages.prev,
2917 struct kvm_mmu_page, link);
2918 kvm_mmu_zap_page(kvm, page);
2921 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2924 struct kvm *kvm_freed = NULL;
2925 int cache_count = 0;
2927 spin_lock(&kvm_lock);
2929 list_for_each_entry(kvm, &vm_list, vm_list) {
2932 idx = srcu_read_lock(&kvm->srcu);
2933 spin_lock(&kvm->mmu_lock);
2934 npages = kvm->arch.n_alloc_mmu_pages -
2935 kvm->arch.n_free_mmu_pages;
2936 cache_count += npages;
2937 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2938 kvm_mmu_remove_one_alloc_mmu_page(kvm);
2944 spin_unlock(&kvm->mmu_lock);
2945 srcu_read_unlock(&kvm->srcu, idx);
2948 list_move_tail(&kvm_freed->vm_list, &vm_list);
2950 spin_unlock(&kvm_lock);
2955 static struct shrinker mmu_shrinker = {
2956 .shrink = mmu_shrink,
2957 .seeks = DEFAULT_SEEKS * 10,
2960 static void mmu_destroy_caches(void)
2962 if (pte_chain_cache)
2963 kmem_cache_destroy(pte_chain_cache);
2964 if (rmap_desc_cache)
2965 kmem_cache_destroy(rmap_desc_cache);
2966 if (mmu_page_header_cache)
2967 kmem_cache_destroy(mmu_page_header_cache);
2970 void kvm_mmu_module_exit(void)
2972 mmu_destroy_caches();
2973 unregister_shrinker(&mmu_shrinker);
2976 int kvm_mmu_module_init(void)
2978 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2979 sizeof(struct kvm_pte_chain),
2981 if (!pte_chain_cache)
2983 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2984 sizeof(struct kvm_rmap_desc),
2986 if (!rmap_desc_cache)
2989 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2990 sizeof(struct kvm_mmu_page),
2992 if (!mmu_page_header_cache)
2995 register_shrinker(&mmu_shrinker);
3000 mmu_destroy_caches();
3005 * Caculate mmu pages needed for kvm.
3007 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3010 unsigned int nr_mmu_pages;
3011 unsigned int nr_pages = 0;
3012 struct kvm_memslots *slots;
3014 slots = rcu_dereference(kvm->memslots);
3015 for (i = 0; i < slots->nmemslots; i++)
3016 nr_pages += slots->memslots[i].npages;
3018 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3019 nr_mmu_pages = max(nr_mmu_pages,
3020 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3022 return nr_mmu_pages;
3025 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3028 if (len > buffer->len)
3033 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3038 ret = pv_mmu_peek_buffer(buffer, len);
3043 buffer->processed += len;
3047 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3048 gpa_t addr, gpa_t value)
3053 if (!is_long_mode(vcpu) && !is_pae(vcpu))
3056 r = mmu_topup_memory_caches(vcpu);
3060 if (!emulator_write_phys(vcpu, addr, &value, bytes))
3066 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3068 kvm_set_cr3(vcpu, vcpu->arch.cr3);
3072 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3074 spin_lock(&vcpu->kvm->mmu_lock);
3075 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3076 spin_unlock(&vcpu->kvm->mmu_lock);
3080 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3081 struct kvm_pv_mmu_op_buffer *buffer)
3083 struct kvm_mmu_op_header *header;
3085 header = pv_mmu_peek_buffer(buffer, sizeof *header);
3088 switch (header->op) {
3089 case KVM_MMU_OP_WRITE_PTE: {
3090 struct kvm_mmu_op_write_pte *wpte;
3092 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3095 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3098 case KVM_MMU_OP_FLUSH_TLB: {
3099 struct kvm_mmu_op_flush_tlb *ftlb;
3101 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3104 return kvm_pv_mmu_flush_tlb(vcpu);
3106 case KVM_MMU_OP_RELEASE_PT: {
3107 struct kvm_mmu_op_release_pt *rpt;
3109 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3112 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3118 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3119 gpa_t addr, unsigned long *ret)
3122 struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3124 buffer->ptr = buffer->buf;
3125 buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3126 buffer->processed = 0;
3128 r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3132 while (buffer->len) {
3133 r = kvm_pv_mmu_op_one(vcpu, buffer);
3142 *ret = buffer->processed;
3146 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3148 struct kvm_shadow_walk_iterator iterator;
3151 spin_lock(&vcpu->kvm->mmu_lock);
3152 for_each_shadow_entry(vcpu, addr, iterator) {
3153 sptes[iterator.level-1] = *iterator.sptep;
3155 if (!is_shadow_present_pte(*iterator.sptep))
3158 spin_unlock(&vcpu->kvm->mmu_lock);
3162 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3166 static const char *audit_msg;
3168 static gva_t canonicalize(gva_t gva)
3170 #ifdef CONFIG_X86_64
3171 gva = (long long)(gva << 16) >> 16;
3177 typedef void (*inspect_spte_fn) (struct kvm *kvm, struct kvm_mmu_page *sp,
3180 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3185 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3186 u64 ent = sp->spt[i];
3188 if (is_shadow_present_pte(ent)) {
3189 if (!is_last_spte(ent, sp->role.level)) {
3190 struct kvm_mmu_page *child;
3191 child = page_header(ent & PT64_BASE_ADDR_MASK);
3192 __mmu_spte_walk(kvm, child, fn);
3194 fn(kvm, sp, &sp->spt[i]);
3199 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3202 struct kvm_mmu_page *sp;
3204 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3206 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3207 hpa_t root = vcpu->arch.mmu.root_hpa;
3208 sp = page_header(root);
3209 __mmu_spte_walk(vcpu->kvm, sp, fn);
3212 for (i = 0; i < 4; ++i) {
3213 hpa_t root = vcpu->arch.mmu.pae_root[i];
3215 if (root && VALID_PAGE(root)) {
3216 root &= PT64_BASE_ADDR_MASK;
3217 sp = page_header(root);
3218 __mmu_spte_walk(vcpu->kvm, sp, fn);
3224 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3225 gva_t va, int level)
3227 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3229 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3231 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3234 if (ent == shadow_trap_nonpresent_pte)
3237 va = canonicalize(va);
3238 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3239 audit_mappings_page(vcpu, ent, va, level - 1);
3241 gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, va, NULL);
3242 gfn_t gfn = gpa >> PAGE_SHIFT;
3243 pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3244 hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3246 if (is_error_pfn(pfn)) {
3247 kvm_release_pfn_clean(pfn);
3251 if (is_shadow_present_pte(ent)
3252 && (ent & PT64_BASE_ADDR_MASK) != hpa)
3253 printk(KERN_ERR "xx audit error: (%s) levels %d"
3254 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3255 audit_msg, vcpu->arch.mmu.root_level,
3257 is_shadow_present_pte(ent));
3258 else if (ent == shadow_notrap_nonpresent_pte
3259 && !is_error_hpa(hpa))
3260 printk(KERN_ERR "audit: (%s) notrap shadow,"
3261 " valid guest gva %lx\n", audit_msg, va);
3262 kvm_release_pfn_clean(pfn);
3268 static void audit_mappings(struct kvm_vcpu *vcpu)
3272 if (vcpu->arch.mmu.root_level == 4)
3273 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3275 for (i = 0; i < 4; ++i)
3276 if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3277 audit_mappings_page(vcpu,
3278 vcpu->arch.mmu.pae_root[i],
3283 static int count_rmaps(struct kvm_vcpu *vcpu)
3288 idx = srcu_read_lock(&kvm->srcu);
3289 slots = rcu_dereference(kvm->memslots);
3290 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3291 struct kvm_memory_slot *m = &slots->memslots[i];
3292 struct kvm_rmap_desc *d;
3294 for (j = 0; j < m->npages; ++j) {
3295 unsigned long *rmapp = &m->rmap[j];
3299 if (!(*rmapp & 1)) {
3303 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3305 for (k = 0; k < RMAP_EXT; ++k)
3314 srcu_read_unlock(&kvm->srcu, idx);
3318 void inspect_spte_has_rmap(struct kvm *kvm, struct kvm_mmu_page *sp, u64 *sptep)
3320 unsigned long *rmapp;
3321 struct kvm_mmu_page *rev_sp;
3324 if (*sptep & PT_WRITABLE_MASK) {
3325 rev_sp = page_header(__pa(sptep));
3326 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3328 if (!gfn_to_memslot(kvm, gfn)) {
3329 if (!printk_ratelimit())
3331 printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3333 printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3334 audit_msg, sptep - rev_sp->spt,
3340 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3341 is_large_pte(*sptep));
3343 if (!printk_ratelimit())
3345 printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3353 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3355 mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3358 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3360 struct kvm_mmu_page *sp;
3363 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3366 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3369 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3372 if (!(ent & PT_PRESENT_MASK))
3374 if (!(ent & PT_WRITABLE_MASK))
3376 inspect_spte_has_rmap(vcpu->kvm, sp, &pt[i]);
3382 static void audit_rmap(struct kvm_vcpu *vcpu)
3384 check_writable_mappings_rmap(vcpu);
3388 static void audit_write_protection(struct kvm_vcpu *vcpu)
3390 struct kvm_mmu_page *sp;
3391 struct kvm_memory_slot *slot;
3392 unsigned long *rmapp;
3396 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3397 if (sp->role.direct)
3402 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3403 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3404 rmapp = &slot->rmap[gfn - slot->base_gfn];
3406 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3408 if (*spte & PT_WRITABLE_MASK)
3409 printk(KERN_ERR "%s: (%s) shadow page has "
3410 "writable mappings: gfn %lx role %x\n",
3411 __func__, audit_msg, sp->gfn,
3413 spte = rmap_next(vcpu->kvm, rmapp, spte);
3418 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3425 audit_write_protection(vcpu);
3426 if (strcmp("pre pte write", audit_msg) != 0)
3427 audit_mappings(vcpu);
3428 audit_writable_sptes_have_rmaps(vcpu);