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.
21 #include "kvm_cache_regs.h"
23 #include <linux/kvm_host.h>
24 #include <linux/types.h>
25 #include <linux/string.h>
27 #include <linux/highmem.h>
28 #include <linux/module.h>
29 #include <linux/swap.h>
30 #include <linux/hugetlb.h>
31 #include <linux/compiler.h>
34 #include <asm/cmpxchg.h>
39 * When setting this variable to true it enables Two-Dimensional-Paging
40 * where the hardware walks 2 page tables:
41 * 1. the guest-virtual to guest-physical
42 * 2. while doing 1. it walks guest-physical to host-physical
43 * If the hardware supports that we don't need to do shadow paging.
45 bool tdp_enabled = false;
52 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
54 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
59 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
60 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
64 #define pgprintk(x...) do { } while (0)
65 #define rmap_printk(x...) do { } while (0)
69 #if defined(MMU_DEBUG) || defined(AUDIT)
71 module_param(dbg, bool, 0644);
74 static int oos_shadow = 1;
75 module_param(oos_shadow, bool, 0644);
78 #define ASSERT(x) do { } while (0)
82 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
83 __FILE__, __LINE__, #x); \
87 #define PT_FIRST_AVAIL_BITS_SHIFT 9
88 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
90 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
92 #define PT64_LEVEL_BITS 9
94 #define PT64_LEVEL_SHIFT(level) \
95 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
97 #define PT64_LEVEL_MASK(level) \
98 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
100 #define PT64_INDEX(address, level)\
101 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
104 #define PT32_LEVEL_BITS 10
106 #define PT32_LEVEL_SHIFT(level) \
107 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
109 #define PT32_LEVEL_MASK(level) \
110 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
111 #define PT32_LVL_OFFSET_MASK(level) \
112 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
113 * PT32_LEVEL_BITS))) - 1))
115 #define PT32_INDEX(address, level)\
116 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
119 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
120 #define PT64_DIR_BASE_ADDR_MASK \
121 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
122 #define PT64_LVL_ADDR_MASK(level) \
123 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
124 * PT64_LEVEL_BITS))) - 1))
125 #define PT64_LVL_OFFSET_MASK(level) \
126 (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
127 * PT64_LEVEL_BITS))) - 1))
129 #define PT32_BASE_ADDR_MASK PAGE_MASK
130 #define PT32_DIR_BASE_ADDR_MASK \
131 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
132 #define PT32_LVL_ADDR_MASK(level) \
133 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
134 * PT32_LEVEL_BITS))) - 1))
136 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
139 #define PFERR_PRESENT_MASK (1U << 0)
140 #define PFERR_WRITE_MASK (1U << 1)
141 #define PFERR_USER_MASK (1U << 2)
142 #define PFERR_RSVD_MASK (1U << 3)
143 #define PFERR_FETCH_MASK (1U << 4)
145 #define PT_PDPE_LEVEL 3
146 #define PT_DIRECTORY_LEVEL 2
147 #define PT_PAGE_TABLE_LEVEL 1
151 #define ACC_EXEC_MASK 1
152 #define ACC_WRITE_MASK PT_WRITABLE_MASK
153 #define ACC_USER_MASK PT_USER_MASK
154 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
156 #define CREATE_TRACE_POINTS
157 #include "mmutrace.h"
159 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
161 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
163 struct kvm_rmap_desc {
164 u64 *sptes[RMAP_EXT];
165 struct kvm_rmap_desc *more;
168 struct kvm_shadow_walk_iterator {
176 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
177 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
178 shadow_walk_okay(&(_walker)); \
179 shadow_walk_next(&(_walker)))
182 struct kvm_unsync_walk {
183 int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
186 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
188 static struct kmem_cache *pte_chain_cache;
189 static struct kmem_cache *rmap_desc_cache;
190 static struct kmem_cache *mmu_page_header_cache;
192 static u64 __read_mostly shadow_trap_nonpresent_pte;
193 static u64 __read_mostly shadow_notrap_nonpresent_pte;
194 static u64 __read_mostly shadow_base_present_pte;
195 static u64 __read_mostly shadow_nx_mask;
196 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
197 static u64 __read_mostly shadow_user_mask;
198 static u64 __read_mostly shadow_accessed_mask;
199 static u64 __read_mostly shadow_dirty_mask;
201 static inline u64 rsvd_bits(int s, int e)
203 return ((1ULL << (e - s + 1)) - 1) << s;
206 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
208 shadow_trap_nonpresent_pte = trap_pte;
209 shadow_notrap_nonpresent_pte = notrap_pte;
211 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
213 void kvm_mmu_set_base_ptes(u64 base_pte)
215 shadow_base_present_pte = base_pte;
217 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
219 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
220 u64 dirty_mask, u64 nx_mask, u64 x_mask)
222 shadow_user_mask = user_mask;
223 shadow_accessed_mask = accessed_mask;
224 shadow_dirty_mask = dirty_mask;
225 shadow_nx_mask = nx_mask;
226 shadow_x_mask = x_mask;
228 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
230 static int is_write_protection(struct kvm_vcpu *vcpu)
232 return vcpu->arch.cr0 & X86_CR0_WP;
235 static int is_cpuid_PSE36(void)
240 static int is_nx(struct kvm_vcpu *vcpu)
242 return vcpu->arch.shadow_efer & EFER_NX;
245 static int is_shadow_present_pte(u64 pte)
247 return pte != shadow_trap_nonpresent_pte
248 && pte != shadow_notrap_nonpresent_pte;
251 static int is_large_pte(u64 pte)
253 return pte & PT_PAGE_SIZE_MASK;
256 static int is_writeble_pte(unsigned long pte)
258 return pte & PT_WRITABLE_MASK;
261 static int is_dirty_gpte(unsigned long pte)
263 return pte & PT_DIRTY_MASK;
266 static int is_rmap_spte(u64 pte)
268 return is_shadow_present_pte(pte);
271 static int is_last_spte(u64 pte, int level)
273 if (level == PT_PAGE_TABLE_LEVEL)
275 if (is_large_pte(pte))
280 static pfn_t spte_to_pfn(u64 pte)
282 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
285 static gfn_t pse36_gfn_delta(u32 gpte)
287 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
289 return (gpte & PT32_DIR_PSE36_MASK) << shift;
292 static void __set_spte(u64 *sptep, u64 spte)
295 set_64bit((unsigned long *)sptep, spte);
297 set_64bit((unsigned long long *)sptep, spte);
301 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
302 struct kmem_cache *base_cache, int min)
306 if (cache->nobjs >= min)
308 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
309 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
312 cache->objects[cache->nobjs++] = obj;
317 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
320 kfree(mc->objects[--mc->nobjs]);
323 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
328 if (cache->nobjs >= min)
330 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
331 page = alloc_page(GFP_KERNEL);
334 set_page_private(page, 0);
335 cache->objects[cache->nobjs++] = page_address(page);
340 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
343 free_page((unsigned long)mc->objects[--mc->nobjs]);
346 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
350 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
354 r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
358 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
361 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
362 mmu_page_header_cache, 4);
367 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
369 mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
370 mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
371 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
372 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
375 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
381 p = mc->objects[--mc->nobjs];
385 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
387 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
388 sizeof(struct kvm_pte_chain));
391 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
396 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
398 return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
399 sizeof(struct kvm_rmap_desc));
402 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
408 * Return the pointer to the largepage write count for a given
409 * gfn, handling slots that are not large page aligned.
411 static int *slot_largepage_idx(gfn_t gfn,
412 struct kvm_memory_slot *slot,
417 idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
418 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
419 return &slot->lpage_info[level - 2][idx].write_count;
422 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
424 struct kvm_memory_slot *slot;
428 gfn = unalias_gfn(kvm, gfn);
430 slot = gfn_to_memslot_unaliased(kvm, gfn);
431 for (i = PT_DIRECTORY_LEVEL;
432 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
433 write_count = slot_largepage_idx(gfn, slot, i);
438 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
440 struct kvm_memory_slot *slot;
444 gfn = unalias_gfn(kvm, gfn);
445 for (i = PT_DIRECTORY_LEVEL;
446 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
447 slot = gfn_to_memslot_unaliased(kvm, gfn);
448 write_count = slot_largepage_idx(gfn, slot, i);
450 WARN_ON(*write_count < 0);
454 static int has_wrprotected_page(struct kvm *kvm,
458 struct kvm_memory_slot *slot;
461 gfn = unalias_gfn(kvm, gfn);
462 slot = gfn_to_memslot_unaliased(kvm, gfn);
464 largepage_idx = slot_largepage_idx(gfn, slot, level);
465 return *largepage_idx;
471 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
473 unsigned long page_size = PAGE_SIZE;
474 struct vm_area_struct *vma;
478 addr = gfn_to_hva(kvm, gfn);
479 if (kvm_is_error_hva(addr))
480 return PT_PAGE_TABLE_LEVEL;
482 down_read(¤t->mm->mmap_sem);
483 vma = find_vma(current->mm, addr);
487 page_size = vma_kernel_pagesize(vma);
490 up_read(¤t->mm->mmap_sem);
492 for (i = PT_PAGE_TABLE_LEVEL;
493 i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
494 if (page_size >= KVM_HPAGE_SIZE(i))
503 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
505 struct kvm_memory_slot *slot;
507 int level = PT_PAGE_TABLE_LEVEL;
509 slot = gfn_to_memslot(vcpu->kvm, large_gfn);
510 if (slot && slot->dirty_bitmap)
511 return PT_PAGE_TABLE_LEVEL;
513 host_level = host_mapping_level(vcpu->kvm, large_gfn);
515 if (host_level == PT_PAGE_TABLE_LEVEL)
518 for (level = PT_DIRECTORY_LEVEL; level <= host_level; ++level)
519 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
526 * Take gfn and return the reverse mapping to it.
527 * Note: gfn must be unaliased before this function get called
530 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
532 struct kvm_memory_slot *slot;
535 slot = gfn_to_memslot(kvm, gfn);
536 if (likely(level == PT_PAGE_TABLE_LEVEL))
537 return &slot->rmap[gfn - slot->base_gfn];
539 idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
540 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
542 return &slot->lpage_info[level - 2][idx].rmap_pde;
546 * Reverse mapping data structures:
548 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
549 * that points to page_address(page).
551 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
552 * containing more mappings.
554 * Returns the number of rmap entries before the spte was added or zero if
555 * the spte was not added.
558 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
560 struct kvm_mmu_page *sp;
561 struct kvm_rmap_desc *desc;
562 unsigned long *rmapp;
565 if (!is_rmap_spte(*spte))
567 gfn = unalias_gfn(vcpu->kvm, gfn);
568 sp = page_header(__pa(spte));
569 sp->gfns[spte - sp->spt] = gfn;
570 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
572 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
573 *rmapp = (unsigned long)spte;
574 } else if (!(*rmapp & 1)) {
575 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
576 desc = mmu_alloc_rmap_desc(vcpu);
577 desc->sptes[0] = (u64 *)*rmapp;
578 desc->sptes[1] = spte;
579 *rmapp = (unsigned long)desc | 1;
581 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
582 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
583 while (desc->sptes[RMAP_EXT-1] && desc->more) {
587 if (desc->sptes[RMAP_EXT-1]) {
588 desc->more = mmu_alloc_rmap_desc(vcpu);
591 for (i = 0; desc->sptes[i]; ++i)
593 desc->sptes[i] = spte;
598 static void rmap_desc_remove_entry(unsigned long *rmapp,
599 struct kvm_rmap_desc *desc,
601 struct kvm_rmap_desc *prev_desc)
605 for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
607 desc->sptes[i] = desc->sptes[j];
608 desc->sptes[j] = NULL;
611 if (!prev_desc && !desc->more)
612 *rmapp = (unsigned long)desc->sptes[0];
615 prev_desc->more = desc->more;
617 *rmapp = (unsigned long)desc->more | 1;
618 mmu_free_rmap_desc(desc);
621 static void rmap_remove(struct kvm *kvm, u64 *spte)
623 struct kvm_rmap_desc *desc;
624 struct kvm_rmap_desc *prev_desc;
625 struct kvm_mmu_page *sp;
627 unsigned long *rmapp;
630 if (!is_rmap_spte(*spte))
632 sp = page_header(__pa(spte));
633 pfn = spte_to_pfn(*spte);
634 if (*spte & shadow_accessed_mask)
635 kvm_set_pfn_accessed(pfn);
636 if (is_writeble_pte(*spte))
637 kvm_set_pfn_dirty(pfn);
638 rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], sp->role.level);
640 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
642 } else if (!(*rmapp & 1)) {
643 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
644 if ((u64 *)*rmapp != spte) {
645 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
651 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
652 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
655 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
656 if (desc->sptes[i] == spte) {
657 rmap_desc_remove_entry(rmapp,
665 pr_err("rmap_remove: %p %llx many->many\n", spte, *spte);
670 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
672 struct kvm_rmap_desc *desc;
673 struct kvm_rmap_desc *prev_desc;
679 else if (!(*rmapp & 1)) {
681 return (u64 *)*rmapp;
684 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
688 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
689 if (prev_spte == spte)
690 return desc->sptes[i];
691 prev_spte = desc->sptes[i];
698 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
700 unsigned long *rmapp;
702 int i, write_protected = 0;
704 gfn = unalias_gfn(kvm, gfn);
705 rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
707 spte = rmap_next(kvm, rmapp, NULL);
710 BUG_ON(!(*spte & PT_PRESENT_MASK));
711 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
712 if (is_writeble_pte(*spte)) {
713 __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
716 spte = rmap_next(kvm, rmapp, spte);
718 if (write_protected) {
721 spte = rmap_next(kvm, rmapp, NULL);
722 pfn = spte_to_pfn(*spte);
723 kvm_set_pfn_dirty(pfn);
726 /* check for huge page mappings */
727 for (i = PT_DIRECTORY_LEVEL;
728 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
729 rmapp = gfn_to_rmap(kvm, gfn, i);
730 spte = rmap_next(kvm, rmapp, NULL);
733 BUG_ON(!(*spte & PT_PRESENT_MASK));
734 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
735 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
736 if (is_writeble_pte(*spte)) {
737 rmap_remove(kvm, spte);
739 __set_spte(spte, shadow_trap_nonpresent_pte);
743 spte = rmap_next(kvm, rmapp, spte);
747 return write_protected;
750 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
754 int need_tlb_flush = 0;
756 while ((spte = rmap_next(kvm, rmapp, NULL))) {
757 BUG_ON(!(*spte & PT_PRESENT_MASK));
758 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
759 rmap_remove(kvm, spte);
760 __set_spte(spte, shadow_trap_nonpresent_pte);
763 return need_tlb_flush;
766 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
771 pte_t *ptep = (pte_t *)data;
774 WARN_ON(pte_huge(*ptep));
775 new_pfn = pte_pfn(*ptep);
776 spte = rmap_next(kvm, rmapp, NULL);
778 BUG_ON(!is_shadow_present_pte(*spte));
779 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
781 if (pte_write(*ptep)) {
782 rmap_remove(kvm, spte);
783 __set_spte(spte, shadow_trap_nonpresent_pte);
784 spte = rmap_next(kvm, rmapp, NULL);
786 new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
787 new_spte |= (u64)new_pfn << PAGE_SHIFT;
789 new_spte &= ~PT_WRITABLE_MASK;
790 new_spte &= ~SPTE_HOST_WRITEABLE;
791 if (is_writeble_pte(*spte))
792 kvm_set_pfn_dirty(spte_to_pfn(*spte));
793 __set_spte(spte, new_spte);
794 spte = rmap_next(kvm, rmapp, spte);
798 kvm_flush_remote_tlbs(kvm);
803 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
805 int (*handler)(struct kvm *kvm, unsigned long *rmapp,
812 * If mmap_sem isn't taken, we can look the memslots with only
813 * the mmu_lock by skipping over the slots with userspace_addr == 0.
815 for (i = 0; i < kvm->nmemslots; i++) {
816 struct kvm_memory_slot *memslot = &kvm->memslots[i];
817 unsigned long start = memslot->userspace_addr;
820 /* mmu_lock protects userspace_addr */
824 end = start + (memslot->npages << PAGE_SHIFT);
825 if (hva >= start && hva < end) {
826 gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
828 retval |= handler(kvm, &memslot->rmap[gfn_offset],
831 for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
832 int idx = gfn_offset;
833 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
834 retval |= handler(kvm,
835 &memslot->lpage_info[j][idx].rmap_pde,
844 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
846 return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
849 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
851 kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
854 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
860 /* always return old for EPT */
861 if (!shadow_accessed_mask)
864 spte = rmap_next(kvm, rmapp, NULL);
868 BUG_ON(!(_spte & PT_PRESENT_MASK));
869 _young = _spte & PT_ACCESSED_MASK;
872 clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
874 spte = rmap_next(kvm, rmapp, spte);
879 #define RMAP_RECYCLE_THRESHOLD 1000
881 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
883 unsigned long *rmapp;
884 struct kvm_mmu_page *sp;
886 sp = page_header(__pa(spte));
888 gfn = unalias_gfn(vcpu->kvm, gfn);
889 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
891 kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
892 kvm_flush_remote_tlbs(vcpu->kvm);
895 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
897 return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
901 static int is_empty_shadow_page(u64 *spt)
906 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
907 if (is_shadow_present_pte(*pos)) {
908 printk(KERN_ERR "%s: %p %llx\n", __func__,
916 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
918 ASSERT(is_empty_shadow_page(sp->spt));
920 __free_page(virt_to_page(sp->spt));
921 __free_page(virt_to_page(sp->gfns));
923 ++kvm->arch.n_free_mmu_pages;
926 static unsigned kvm_page_table_hashfn(gfn_t gfn)
928 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
931 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
934 struct kvm_mmu_page *sp;
936 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
937 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
938 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
939 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
940 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
941 INIT_LIST_HEAD(&sp->oos_link);
942 bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
944 sp->parent_pte = parent_pte;
945 --vcpu->kvm->arch.n_free_mmu_pages;
949 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
950 struct kvm_mmu_page *sp, u64 *parent_pte)
952 struct kvm_pte_chain *pte_chain;
953 struct hlist_node *node;
958 if (!sp->multimapped) {
959 u64 *old = sp->parent_pte;
962 sp->parent_pte = parent_pte;
966 pte_chain = mmu_alloc_pte_chain(vcpu);
967 INIT_HLIST_HEAD(&sp->parent_ptes);
968 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
969 pte_chain->parent_ptes[0] = old;
971 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
972 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
974 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
975 if (!pte_chain->parent_ptes[i]) {
976 pte_chain->parent_ptes[i] = parent_pte;
980 pte_chain = mmu_alloc_pte_chain(vcpu);
982 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
983 pte_chain->parent_ptes[0] = parent_pte;
986 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
989 struct kvm_pte_chain *pte_chain;
990 struct hlist_node *node;
993 if (!sp->multimapped) {
994 BUG_ON(sp->parent_pte != parent_pte);
995 sp->parent_pte = NULL;
998 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
999 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1000 if (!pte_chain->parent_ptes[i])
1002 if (pte_chain->parent_ptes[i] != parent_pte)
1004 while (i + 1 < NR_PTE_CHAIN_ENTRIES
1005 && pte_chain->parent_ptes[i + 1]) {
1006 pte_chain->parent_ptes[i]
1007 = pte_chain->parent_ptes[i + 1];
1010 pte_chain->parent_ptes[i] = NULL;
1012 hlist_del(&pte_chain->link);
1013 mmu_free_pte_chain(pte_chain);
1014 if (hlist_empty(&sp->parent_ptes)) {
1015 sp->multimapped = 0;
1016 sp->parent_pte = NULL;
1025 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1026 mmu_parent_walk_fn fn)
1028 struct kvm_pte_chain *pte_chain;
1029 struct hlist_node *node;
1030 struct kvm_mmu_page *parent_sp;
1033 if (!sp->multimapped && sp->parent_pte) {
1034 parent_sp = page_header(__pa(sp->parent_pte));
1035 fn(vcpu, parent_sp);
1036 mmu_parent_walk(vcpu, parent_sp, fn);
1039 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1040 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1041 if (!pte_chain->parent_ptes[i])
1043 parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
1044 fn(vcpu, parent_sp);
1045 mmu_parent_walk(vcpu, parent_sp, fn);
1049 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1052 struct kvm_mmu_page *sp = page_header(__pa(spte));
1054 index = spte - sp->spt;
1055 if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1056 sp->unsync_children++;
1057 WARN_ON(!sp->unsync_children);
1060 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1062 struct kvm_pte_chain *pte_chain;
1063 struct hlist_node *node;
1066 if (!sp->parent_pte)
1069 if (!sp->multimapped) {
1070 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1074 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1075 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1076 if (!pte_chain->parent_ptes[i])
1078 kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1082 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1084 kvm_mmu_update_parents_unsync(sp);
1088 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
1089 struct kvm_mmu_page *sp)
1091 mmu_parent_walk(vcpu, sp, unsync_walk_fn);
1092 kvm_mmu_update_parents_unsync(sp);
1095 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1096 struct kvm_mmu_page *sp)
1100 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1101 sp->spt[i] = shadow_trap_nonpresent_pte;
1104 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1105 struct kvm_mmu_page *sp)
1110 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1114 #define KVM_PAGE_ARRAY_NR 16
1116 struct kvm_mmu_pages {
1117 struct mmu_page_and_offset {
1118 struct kvm_mmu_page *sp;
1120 } page[KVM_PAGE_ARRAY_NR];
1124 #define for_each_unsync_children(bitmap, idx) \
1125 for (idx = find_first_bit(bitmap, 512); \
1127 idx = find_next_bit(bitmap, 512, idx+1))
1129 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1135 for (i=0; i < pvec->nr; i++)
1136 if (pvec->page[i].sp == sp)
1139 pvec->page[pvec->nr].sp = sp;
1140 pvec->page[pvec->nr].idx = idx;
1142 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1145 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1146 struct kvm_mmu_pages *pvec)
1148 int i, ret, nr_unsync_leaf = 0;
1150 for_each_unsync_children(sp->unsync_child_bitmap, i) {
1151 u64 ent = sp->spt[i];
1153 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1154 struct kvm_mmu_page *child;
1155 child = page_header(ent & PT64_BASE_ADDR_MASK);
1157 if (child->unsync_children) {
1158 if (mmu_pages_add(pvec, child, i))
1161 ret = __mmu_unsync_walk(child, pvec);
1163 __clear_bit(i, sp->unsync_child_bitmap);
1165 nr_unsync_leaf += ret;
1170 if (child->unsync) {
1172 if (mmu_pages_add(pvec, child, i))
1178 if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1179 sp->unsync_children = 0;
1181 return nr_unsync_leaf;
1184 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1185 struct kvm_mmu_pages *pvec)
1187 if (!sp->unsync_children)
1190 mmu_pages_add(pvec, sp, 0);
1191 return __mmu_unsync_walk(sp, pvec);
1194 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1197 struct hlist_head *bucket;
1198 struct kvm_mmu_page *sp;
1199 struct hlist_node *node;
1201 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1202 index = kvm_page_table_hashfn(gfn);
1203 bucket = &kvm->arch.mmu_page_hash[index];
1204 hlist_for_each_entry(sp, node, bucket, hash_link)
1205 if (sp->gfn == gfn && !sp->role.direct
1206 && !sp->role.invalid) {
1207 pgprintk("%s: found role %x\n",
1208 __func__, sp->role.word);
1214 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1216 WARN_ON(!sp->unsync);
1218 --kvm->stat.mmu_unsync;
1221 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1223 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1225 if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1226 kvm_mmu_zap_page(vcpu->kvm, sp);
1230 trace_kvm_mmu_sync_page(sp);
1231 if (rmap_write_protect(vcpu->kvm, sp->gfn))
1232 kvm_flush_remote_tlbs(vcpu->kvm);
1233 kvm_unlink_unsync_page(vcpu->kvm, sp);
1234 if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1235 kvm_mmu_zap_page(vcpu->kvm, sp);
1239 kvm_mmu_flush_tlb(vcpu);
1243 struct mmu_page_path {
1244 struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1245 unsigned int idx[PT64_ROOT_LEVEL-1];
1248 #define for_each_sp(pvec, sp, parents, i) \
1249 for (i = mmu_pages_next(&pvec, &parents, -1), \
1250 sp = pvec.page[i].sp; \
1251 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1252 i = mmu_pages_next(&pvec, &parents, i))
1254 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1255 struct mmu_page_path *parents,
1260 for (n = i+1; n < pvec->nr; n++) {
1261 struct kvm_mmu_page *sp = pvec->page[n].sp;
1263 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1264 parents->idx[0] = pvec->page[n].idx;
1268 parents->parent[sp->role.level-2] = sp;
1269 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1275 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1277 struct kvm_mmu_page *sp;
1278 unsigned int level = 0;
1281 unsigned int idx = parents->idx[level];
1283 sp = parents->parent[level];
1287 --sp->unsync_children;
1288 WARN_ON((int)sp->unsync_children < 0);
1289 __clear_bit(idx, sp->unsync_child_bitmap);
1291 } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1294 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1295 struct mmu_page_path *parents,
1296 struct kvm_mmu_pages *pvec)
1298 parents->parent[parent->role.level-1] = NULL;
1302 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1303 struct kvm_mmu_page *parent)
1306 struct kvm_mmu_page *sp;
1307 struct mmu_page_path parents;
1308 struct kvm_mmu_pages pages;
1310 kvm_mmu_pages_init(parent, &parents, &pages);
1311 while (mmu_unsync_walk(parent, &pages)) {
1314 for_each_sp(pages, sp, parents, i)
1315 protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1318 kvm_flush_remote_tlbs(vcpu->kvm);
1320 for_each_sp(pages, sp, parents, i) {
1321 kvm_sync_page(vcpu, sp);
1322 mmu_pages_clear_parents(&parents);
1324 cond_resched_lock(&vcpu->kvm->mmu_lock);
1325 kvm_mmu_pages_init(parent, &parents, &pages);
1329 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1337 union kvm_mmu_page_role role;
1340 struct hlist_head *bucket;
1341 struct kvm_mmu_page *sp;
1342 struct hlist_node *node, *tmp;
1344 role = vcpu->arch.mmu.base_role;
1346 role.direct = direct;
1347 role.access = access;
1348 if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1349 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1350 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1351 role.quadrant = quadrant;
1353 index = kvm_page_table_hashfn(gfn);
1354 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1355 hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1356 if (sp->gfn == gfn) {
1358 if (kvm_sync_page(vcpu, sp))
1361 if (sp->role.word != role.word)
1364 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1365 if (sp->unsync_children) {
1366 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1367 kvm_mmu_mark_parents_unsync(vcpu, sp);
1369 trace_kvm_mmu_get_page(sp, false);
1372 ++vcpu->kvm->stat.mmu_cache_miss;
1373 sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1378 hlist_add_head(&sp->hash_link, bucket);
1380 if (rmap_write_protect(vcpu->kvm, gfn))
1381 kvm_flush_remote_tlbs(vcpu->kvm);
1382 account_shadowed(vcpu->kvm, gfn);
1384 if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1385 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1387 nonpaging_prefetch_page(vcpu, sp);
1388 trace_kvm_mmu_get_page(sp, true);
1392 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1393 struct kvm_vcpu *vcpu, u64 addr)
1395 iterator->addr = addr;
1396 iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1397 iterator->level = vcpu->arch.mmu.shadow_root_level;
1398 if (iterator->level == PT32E_ROOT_LEVEL) {
1399 iterator->shadow_addr
1400 = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1401 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1403 if (!iterator->shadow_addr)
1404 iterator->level = 0;
1408 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1410 if (iterator->level < PT_PAGE_TABLE_LEVEL)
1413 if (iterator->level == PT_PAGE_TABLE_LEVEL)
1414 if (is_large_pte(*iterator->sptep))
1417 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1418 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1422 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1424 iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1428 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1429 struct kvm_mmu_page *sp)
1437 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1440 if (is_shadow_present_pte(ent)) {
1441 if (!is_last_spte(ent, sp->role.level)) {
1442 ent &= PT64_BASE_ADDR_MASK;
1443 mmu_page_remove_parent_pte(page_header(ent),
1446 if (is_large_pte(ent))
1448 rmap_remove(kvm, &pt[i]);
1451 pt[i] = shadow_trap_nonpresent_pte;
1455 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1457 mmu_page_remove_parent_pte(sp, parent_pte);
1460 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1463 struct kvm_vcpu *vcpu;
1465 kvm_for_each_vcpu(i, vcpu, kvm)
1466 vcpu->arch.last_pte_updated = NULL;
1469 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1473 while (sp->multimapped || sp->parent_pte) {
1474 if (!sp->multimapped)
1475 parent_pte = sp->parent_pte;
1477 struct kvm_pte_chain *chain;
1479 chain = container_of(sp->parent_ptes.first,
1480 struct kvm_pte_chain, link);
1481 parent_pte = chain->parent_ptes[0];
1483 BUG_ON(!parent_pte);
1484 kvm_mmu_put_page(sp, parent_pte);
1485 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1489 static int mmu_zap_unsync_children(struct kvm *kvm,
1490 struct kvm_mmu_page *parent)
1493 struct mmu_page_path parents;
1494 struct kvm_mmu_pages pages;
1496 if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1499 kvm_mmu_pages_init(parent, &parents, &pages);
1500 while (mmu_unsync_walk(parent, &pages)) {
1501 struct kvm_mmu_page *sp;
1503 for_each_sp(pages, sp, parents, i) {
1504 kvm_mmu_zap_page(kvm, sp);
1505 mmu_pages_clear_parents(&parents);
1508 kvm_mmu_pages_init(parent, &parents, &pages);
1514 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1518 trace_kvm_mmu_zap_page(sp);
1519 ++kvm->stat.mmu_shadow_zapped;
1520 ret = mmu_zap_unsync_children(kvm, sp);
1521 kvm_mmu_page_unlink_children(kvm, sp);
1522 kvm_mmu_unlink_parents(kvm, sp);
1523 kvm_flush_remote_tlbs(kvm);
1524 if (!sp->role.invalid && !sp->role.direct)
1525 unaccount_shadowed(kvm, sp->gfn);
1527 kvm_unlink_unsync_page(kvm, sp);
1528 if (!sp->root_count) {
1529 hlist_del(&sp->hash_link);
1530 kvm_mmu_free_page(kvm, sp);
1532 sp->role.invalid = 1;
1533 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1534 kvm_reload_remote_mmus(kvm);
1536 kvm_mmu_reset_last_pte_updated(kvm);
1541 * Changing the number of mmu pages allocated to the vm
1542 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1544 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1548 used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1549 used_pages = max(0, used_pages);
1552 * If we set the number of mmu pages to be smaller be than the
1553 * number of actived pages , we must to free some mmu pages before we
1557 if (used_pages > kvm_nr_mmu_pages) {
1558 while (used_pages > kvm_nr_mmu_pages) {
1559 struct kvm_mmu_page *page;
1561 page = container_of(kvm->arch.active_mmu_pages.prev,
1562 struct kvm_mmu_page, link);
1563 kvm_mmu_zap_page(kvm, page);
1566 kvm->arch.n_free_mmu_pages = 0;
1569 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1570 - kvm->arch.n_alloc_mmu_pages;
1572 kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1575 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1578 struct hlist_head *bucket;
1579 struct kvm_mmu_page *sp;
1580 struct hlist_node *node, *n;
1583 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1585 index = kvm_page_table_hashfn(gfn);
1586 bucket = &kvm->arch.mmu_page_hash[index];
1587 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1588 if (sp->gfn == gfn && !sp->role.direct) {
1589 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1592 if (kvm_mmu_zap_page(kvm, sp))
1598 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1601 struct hlist_head *bucket;
1602 struct kvm_mmu_page *sp;
1603 struct hlist_node *node, *nn;
1605 index = kvm_page_table_hashfn(gfn);
1606 bucket = &kvm->arch.mmu_page_hash[index];
1607 hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1608 if (sp->gfn == gfn && !sp->role.direct
1609 && !sp->role.invalid) {
1610 pgprintk("%s: zap %lx %x\n",
1611 __func__, gfn, sp->role.word);
1612 kvm_mmu_zap_page(kvm, sp);
1617 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1619 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1620 struct kvm_mmu_page *sp = page_header(__pa(pte));
1622 __set_bit(slot, sp->slot_bitmap);
1625 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1630 if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1633 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1634 if (pt[i] == shadow_notrap_nonpresent_pte)
1635 __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1639 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1643 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1645 if (gpa == UNMAPPED_GVA)
1648 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1654 * The function is based on mtrr_type_lookup() in
1655 * arch/x86/kernel/cpu/mtrr/generic.c
1657 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1662 u8 prev_match, curr_match;
1663 int num_var_ranges = KVM_NR_VAR_MTRR;
1665 if (!mtrr_state->enabled)
1668 /* Make end inclusive end, instead of exclusive */
1671 /* Look in fixed ranges. Just return the type as per start */
1672 if (mtrr_state->have_fixed && (start < 0x100000)) {
1675 if (start < 0x80000) {
1677 idx += (start >> 16);
1678 return mtrr_state->fixed_ranges[idx];
1679 } else if (start < 0xC0000) {
1681 idx += ((start - 0x80000) >> 14);
1682 return mtrr_state->fixed_ranges[idx];
1683 } else if (start < 0x1000000) {
1685 idx += ((start - 0xC0000) >> 12);
1686 return mtrr_state->fixed_ranges[idx];
1691 * Look in variable ranges
1692 * Look of multiple ranges matching this address and pick type
1693 * as per MTRR precedence
1695 if (!(mtrr_state->enabled & 2))
1696 return mtrr_state->def_type;
1699 for (i = 0; i < num_var_ranges; ++i) {
1700 unsigned short start_state, end_state;
1702 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1705 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1706 (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1707 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1708 (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1710 start_state = ((start & mask) == (base & mask));
1711 end_state = ((end & mask) == (base & mask));
1712 if (start_state != end_state)
1715 if ((start & mask) != (base & mask))
1718 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1719 if (prev_match == 0xFF) {
1720 prev_match = curr_match;
1724 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1725 curr_match == MTRR_TYPE_UNCACHABLE)
1726 return MTRR_TYPE_UNCACHABLE;
1728 if ((prev_match == MTRR_TYPE_WRBACK &&
1729 curr_match == MTRR_TYPE_WRTHROUGH) ||
1730 (prev_match == MTRR_TYPE_WRTHROUGH &&
1731 curr_match == MTRR_TYPE_WRBACK)) {
1732 prev_match = MTRR_TYPE_WRTHROUGH;
1733 curr_match = MTRR_TYPE_WRTHROUGH;
1736 if (prev_match != curr_match)
1737 return MTRR_TYPE_UNCACHABLE;
1740 if (prev_match != 0xFF)
1743 return mtrr_state->def_type;
1746 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1750 mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1751 (gfn << PAGE_SHIFT) + PAGE_SIZE);
1752 if (mtrr == 0xfe || mtrr == 0xff)
1753 mtrr = MTRR_TYPE_WRBACK;
1756 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1758 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1761 struct hlist_head *bucket;
1762 struct kvm_mmu_page *s;
1763 struct hlist_node *node, *n;
1765 trace_kvm_mmu_unsync_page(sp);
1766 index = kvm_page_table_hashfn(sp->gfn);
1767 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1768 /* don't unsync if pagetable is shadowed with multiple roles */
1769 hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1770 if (s->gfn != sp->gfn || s->role.direct)
1772 if (s->role.word != sp->role.word)
1775 ++vcpu->kvm->stat.mmu_unsync;
1778 kvm_mmu_mark_parents_unsync(vcpu, sp);
1780 mmu_convert_notrap(sp);
1784 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1787 struct kvm_mmu_page *shadow;
1789 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1791 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1795 if (can_unsync && oos_shadow)
1796 return kvm_unsync_page(vcpu, shadow);
1802 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1803 unsigned pte_access, int user_fault,
1804 int write_fault, int dirty, int level,
1805 gfn_t gfn, pfn_t pfn, bool speculative,
1806 bool can_unsync, bool reset_host_protection)
1812 * We don't set the accessed bit, since we sometimes want to see
1813 * whether the guest actually used the pte (in order to detect
1816 spte = shadow_base_present_pte | shadow_dirty_mask;
1818 spte |= shadow_accessed_mask;
1820 pte_access &= ~ACC_WRITE_MASK;
1821 if (pte_access & ACC_EXEC_MASK)
1822 spte |= shadow_x_mask;
1824 spte |= shadow_nx_mask;
1825 if (pte_access & ACC_USER_MASK)
1826 spte |= shadow_user_mask;
1827 if (level > PT_PAGE_TABLE_LEVEL)
1828 spte |= PT_PAGE_SIZE_MASK;
1830 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1831 kvm_is_mmio_pfn(pfn));
1833 if (reset_host_protection)
1834 spte |= SPTE_HOST_WRITEABLE;
1836 spte |= (u64)pfn << PAGE_SHIFT;
1838 if ((pte_access & ACC_WRITE_MASK)
1839 || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1841 if (level > PT_PAGE_TABLE_LEVEL &&
1842 has_wrprotected_page(vcpu->kvm, gfn, level)) {
1844 spte = shadow_trap_nonpresent_pte;
1848 spte |= PT_WRITABLE_MASK;
1851 * Optimization: for pte sync, if spte was writable the hash
1852 * lookup is unnecessary (and expensive). Write protection
1853 * is responsibility of mmu_get_page / kvm_sync_page.
1854 * Same reasoning can be applied to dirty page accounting.
1856 if (!can_unsync && is_writeble_pte(*sptep))
1859 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1860 pgprintk("%s: found shadow page for %lx, marking ro\n",
1863 pte_access &= ~ACC_WRITE_MASK;
1864 if (is_writeble_pte(spte))
1865 spte &= ~PT_WRITABLE_MASK;
1869 if (pte_access & ACC_WRITE_MASK)
1870 mark_page_dirty(vcpu->kvm, gfn);
1873 __set_spte(sptep, spte);
1877 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1878 unsigned pt_access, unsigned pte_access,
1879 int user_fault, int write_fault, int dirty,
1880 int *ptwrite, int level, gfn_t gfn,
1881 pfn_t pfn, bool speculative,
1882 bool reset_host_protection)
1884 int was_rmapped = 0;
1885 int was_writeble = is_writeble_pte(*sptep);
1888 pgprintk("%s: spte %llx access %x write_fault %d"
1889 " user_fault %d gfn %lx\n",
1890 __func__, *sptep, pt_access,
1891 write_fault, user_fault, gfn);
1893 if (is_rmap_spte(*sptep)) {
1895 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1896 * the parent of the now unreachable PTE.
1898 if (level > PT_PAGE_TABLE_LEVEL &&
1899 !is_large_pte(*sptep)) {
1900 struct kvm_mmu_page *child;
1903 child = page_header(pte & PT64_BASE_ADDR_MASK);
1904 mmu_page_remove_parent_pte(child, sptep);
1905 } else if (pfn != spte_to_pfn(*sptep)) {
1906 pgprintk("hfn old %lx new %lx\n",
1907 spte_to_pfn(*sptep), pfn);
1908 rmap_remove(vcpu->kvm, sptep);
1913 if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1914 dirty, level, gfn, pfn, speculative, true,
1915 reset_host_protection)) {
1918 kvm_x86_ops->tlb_flush(vcpu);
1921 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1922 pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1923 is_large_pte(*sptep)? "2MB" : "4kB",
1924 *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1926 if (!was_rmapped && is_large_pte(*sptep))
1927 ++vcpu->kvm->stat.lpages;
1929 page_header_update_slot(vcpu->kvm, sptep, gfn);
1931 rmap_count = rmap_add(vcpu, sptep, gfn);
1932 kvm_release_pfn_clean(pfn);
1933 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1934 rmap_recycle(vcpu, sptep, gfn);
1937 kvm_release_pfn_dirty(pfn);
1939 kvm_release_pfn_clean(pfn);
1942 vcpu->arch.last_pte_updated = sptep;
1943 vcpu->arch.last_pte_gfn = gfn;
1947 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1951 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1952 int level, gfn_t gfn, pfn_t pfn)
1954 struct kvm_shadow_walk_iterator iterator;
1955 struct kvm_mmu_page *sp;
1959 for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1960 if (iterator.level == level) {
1961 mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1962 0, write, 1, &pt_write,
1963 level, gfn, pfn, false, true);
1964 ++vcpu->stat.pf_fixed;
1968 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1969 pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1970 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1972 1, ACC_ALL, iterator.sptep);
1974 pgprintk("nonpaging_map: ENOMEM\n");
1975 kvm_release_pfn_clean(pfn);
1979 __set_spte(iterator.sptep,
1981 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1982 | shadow_user_mask | shadow_x_mask);
1988 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1993 unsigned long mmu_seq;
1995 level = mapping_level(vcpu, gfn);
1998 * This path builds a PAE pagetable - so we can map 2mb pages at
1999 * maximum. Therefore check if the level is larger than that.
2001 if (level > PT_DIRECTORY_LEVEL)
2002 level = PT_DIRECTORY_LEVEL;
2004 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2006 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2008 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2011 if (is_error_pfn(pfn)) {
2012 kvm_release_pfn_clean(pfn);
2016 spin_lock(&vcpu->kvm->mmu_lock);
2017 if (mmu_notifier_retry(vcpu, mmu_seq))
2019 kvm_mmu_free_some_pages(vcpu);
2020 r = __direct_map(vcpu, v, write, level, gfn, pfn);
2021 spin_unlock(&vcpu->kvm->mmu_lock);
2027 spin_unlock(&vcpu->kvm->mmu_lock);
2028 kvm_release_pfn_clean(pfn);
2033 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2036 struct kvm_mmu_page *sp;
2038 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2040 spin_lock(&vcpu->kvm->mmu_lock);
2041 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2042 hpa_t root = vcpu->arch.mmu.root_hpa;
2044 sp = page_header(root);
2046 if (!sp->root_count && sp->role.invalid)
2047 kvm_mmu_zap_page(vcpu->kvm, sp);
2048 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2049 spin_unlock(&vcpu->kvm->mmu_lock);
2052 for (i = 0; i < 4; ++i) {
2053 hpa_t root = vcpu->arch.mmu.pae_root[i];
2056 root &= PT64_BASE_ADDR_MASK;
2057 sp = page_header(root);
2059 if (!sp->root_count && sp->role.invalid)
2060 kvm_mmu_zap_page(vcpu->kvm, sp);
2062 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2064 spin_unlock(&vcpu->kvm->mmu_lock);
2065 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2068 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2072 if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2073 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2080 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2084 struct kvm_mmu_page *sp;
2088 root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2090 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2091 hpa_t root = vcpu->arch.mmu.root_hpa;
2093 ASSERT(!VALID_PAGE(root));
2096 if (mmu_check_root(vcpu, root_gfn))
2098 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2099 PT64_ROOT_LEVEL, direct,
2101 root = __pa(sp->spt);
2103 vcpu->arch.mmu.root_hpa = root;
2106 direct = !is_paging(vcpu);
2109 for (i = 0; i < 4; ++i) {
2110 hpa_t root = vcpu->arch.mmu.pae_root[i];
2112 ASSERT(!VALID_PAGE(root));
2113 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2114 pdptr = kvm_pdptr_read(vcpu, i);
2115 if (!is_present_gpte(pdptr)) {
2116 vcpu->arch.mmu.pae_root[i] = 0;
2119 root_gfn = pdptr >> PAGE_SHIFT;
2120 } else if (vcpu->arch.mmu.root_level == 0)
2122 if (mmu_check_root(vcpu, root_gfn))
2124 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2125 PT32_ROOT_LEVEL, direct,
2127 root = __pa(sp->spt);
2129 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2131 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2135 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2138 struct kvm_mmu_page *sp;
2140 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2142 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2143 hpa_t root = vcpu->arch.mmu.root_hpa;
2144 sp = page_header(root);
2145 mmu_sync_children(vcpu, sp);
2148 for (i = 0; i < 4; ++i) {
2149 hpa_t root = vcpu->arch.mmu.pae_root[i];
2151 if (root && VALID_PAGE(root)) {
2152 root &= PT64_BASE_ADDR_MASK;
2153 sp = page_header(root);
2154 mmu_sync_children(vcpu, sp);
2159 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2161 spin_lock(&vcpu->kvm->mmu_lock);
2162 mmu_sync_roots(vcpu);
2163 spin_unlock(&vcpu->kvm->mmu_lock);
2166 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2171 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2177 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2178 r = mmu_topup_memory_caches(vcpu);
2183 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2185 gfn = gva >> PAGE_SHIFT;
2187 return nonpaging_map(vcpu, gva & PAGE_MASK,
2188 error_code & PFERR_WRITE_MASK, gfn);
2191 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2197 gfn_t gfn = gpa >> PAGE_SHIFT;
2198 unsigned long mmu_seq;
2201 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2203 r = mmu_topup_memory_caches(vcpu);
2207 level = mapping_level(vcpu, gfn);
2209 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2211 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2213 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2214 if (is_error_pfn(pfn)) {
2215 kvm_release_pfn_clean(pfn);
2218 spin_lock(&vcpu->kvm->mmu_lock);
2219 if (mmu_notifier_retry(vcpu, mmu_seq))
2221 kvm_mmu_free_some_pages(vcpu);
2222 r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2224 spin_unlock(&vcpu->kvm->mmu_lock);
2229 spin_unlock(&vcpu->kvm->mmu_lock);
2230 kvm_release_pfn_clean(pfn);
2234 static void nonpaging_free(struct kvm_vcpu *vcpu)
2236 mmu_free_roots(vcpu);
2239 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2241 struct kvm_mmu *context = &vcpu->arch.mmu;
2243 context->new_cr3 = nonpaging_new_cr3;
2244 context->page_fault = nonpaging_page_fault;
2245 context->gva_to_gpa = nonpaging_gva_to_gpa;
2246 context->free = nonpaging_free;
2247 context->prefetch_page = nonpaging_prefetch_page;
2248 context->sync_page = nonpaging_sync_page;
2249 context->invlpg = nonpaging_invlpg;
2250 context->root_level = 0;
2251 context->shadow_root_level = PT32E_ROOT_LEVEL;
2252 context->root_hpa = INVALID_PAGE;
2256 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2258 ++vcpu->stat.tlb_flush;
2259 kvm_x86_ops->tlb_flush(vcpu);
2262 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2264 pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2265 mmu_free_roots(vcpu);
2268 static void inject_page_fault(struct kvm_vcpu *vcpu,
2272 kvm_inject_page_fault(vcpu, addr, err_code);
2275 static void paging_free(struct kvm_vcpu *vcpu)
2277 nonpaging_free(vcpu);
2280 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2284 bit7 = (gpte >> 7) & 1;
2285 return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2289 #include "paging_tmpl.h"
2293 #include "paging_tmpl.h"
2296 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2298 struct kvm_mmu *context = &vcpu->arch.mmu;
2299 int maxphyaddr = cpuid_maxphyaddr(vcpu);
2300 u64 exb_bit_rsvd = 0;
2303 exb_bit_rsvd = rsvd_bits(63, 63);
2305 case PT32_ROOT_LEVEL:
2306 /* no rsvd bits for 2 level 4K page table entries */
2307 context->rsvd_bits_mask[0][1] = 0;
2308 context->rsvd_bits_mask[0][0] = 0;
2309 if (is_cpuid_PSE36())
2310 /* 36bits PSE 4MB page */
2311 context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2313 /* 32 bits PSE 4MB page */
2314 context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2315 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2317 case PT32E_ROOT_LEVEL:
2318 context->rsvd_bits_mask[0][2] =
2319 rsvd_bits(maxphyaddr, 63) |
2320 rsvd_bits(7, 8) | rsvd_bits(1, 2); /* PDPTE */
2321 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2322 rsvd_bits(maxphyaddr, 62); /* PDE */
2323 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2324 rsvd_bits(maxphyaddr, 62); /* PTE */
2325 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2326 rsvd_bits(maxphyaddr, 62) |
2327 rsvd_bits(13, 20); /* large page */
2328 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2330 case PT64_ROOT_LEVEL:
2331 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2332 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2333 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2334 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2335 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2336 rsvd_bits(maxphyaddr, 51);
2337 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2338 rsvd_bits(maxphyaddr, 51);
2339 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2340 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2341 rsvd_bits(maxphyaddr, 51) |
2343 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2344 rsvd_bits(maxphyaddr, 51) |
2345 rsvd_bits(13, 20); /* large page */
2346 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2351 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2353 struct kvm_mmu *context = &vcpu->arch.mmu;
2355 ASSERT(is_pae(vcpu));
2356 context->new_cr3 = paging_new_cr3;
2357 context->page_fault = paging64_page_fault;
2358 context->gva_to_gpa = paging64_gva_to_gpa;
2359 context->prefetch_page = paging64_prefetch_page;
2360 context->sync_page = paging64_sync_page;
2361 context->invlpg = paging64_invlpg;
2362 context->free = paging_free;
2363 context->root_level = level;
2364 context->shadow_root_level = level;
2365 context->root_hpa = INVALID_PAGE;
2369 static int paging64_init_context(struct kvm_vcpu *vcpu)
2371 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2372 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2375 static int paging32_init_context(struct kvm_vcpu *vcpu)
2377 struct kvm_mmu *context = &vcpu->arch.mmu;
2379 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2380 context->new_cr3 = paging_new_cr3;
2381 context->page_fault = paging32_page_fault;
2382 context->gva_to_gpa = paging32_gva_to_gpa;
2383 context->free = paging_free;
2384 context->prefetch_page = paging32_prefetch_page;
2385 context->sync_page = paging32_sync_page;
2386 context->invlpg = paging32_invlpg;
2387 context->root_level = PT32_ROOT_LEVEL;
2388 context->shadow_root_level = PT32E_ROOT_LEVEL;
2389 context->root_hpa = INVALID_PAGE;
2393 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2395 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2396 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2399 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2401 struct kvm_mmu *context = &vcpu->arch.mmu;
2403 context->new_cr3 = nonpaging_new_cr3;
2404 context->page_fault = tdp_page_fault;
2405 context->free = nonpaging_free;
2406 context->prefetch_page = nonpaging_prefetch_page;
2407 context->sync_page = nonpaging_sync_page;
2408 context->invlpg = nonpaging_invlpg;
2409 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2410 context->root_hpa = INVALID_PAGE;
2412 if (!is_paging(vcpu)) {
2413 context->gva_to_gpa = nonpaging_gva_to_gpa;
2414 context->root_level = 0;
2415 } else if (is_long_mode(vcpu)) {
2416 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2417 context->gva_to_gpa = paging64_gva_to_gpa;
2418 context->root_level = PT64_ROOT_LEVEL;
2419 } else if (is_pae(vcpu)) {
2420 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2421 context->gva_to_gpa = paging64_gva_to_gpa;
2422 context->root_level = PT32E_ROOT_LEVEL;
2424 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2425 context->gva_to_gpa = paging32_gva_to_gpa;
2426 context->root_level = PT32_ROOT_LEVEL;
2432 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2437 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2439 if (!is_paging(vcpu))
2440 r = nonpaging_init_context(vcpu);
2441 else if (is_long_mode(vcpu))
2442 r = paging64_init_context(vcpu);
2443 else if (is_pae(vcpu))
2444 r = paging32E_init_context(vcpu);
2446 r = paging32_init_context(vcpu);
2448 vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2453 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2455 vcpu->arch.update_pte.pfn = bad_pfn;
2458 return init_kvm_tdp_mmu(vcpu);
2460 return init_kvm_softmmu(vcpu);
2463 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2466 if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2467 vcpu->arch.mmu.free(vcpu);
2468 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2472 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2474 destroy_kvm_mmu(vcpu);
2475 return init_kvm_mmu(vcpu);
2477 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2479 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2483 r = mmu_topup_memory_caches(vcpu);
2486 spin_lock(&vcpu->kvm->mmu_lock);
2487 kvm_mmu_free_some_pages(vcpu);
2488 r = mmu_alloc_roots(vcpu);
2489 mmu_sync_roots(vcpu);
2490 spin_unlock(&vcpu->kvm->mmu_lock);
2493 /* set_cr3() should ensure TLB has been flushed */
2494 kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2498 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2500 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2502 mmu_free_roots(vcpu);
2505 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2506 struct kvm_mmu_page *sp,
2510 struct kvm_mmu_page *child;
2513 if (is_shadow_present_pte(pte)) {
2514 if (is_last_spte(pte, sp->role.level))
2515 rmap_remove(vcpu->kvm, spte);
2517 child = page_header(pte & PT64_BASE_ADDR_MASK);
2518 mmu_page_remove_parent_pte(child, spte);
2521 __set_spte(spte, shadow_trap_nonpresent_pte);
2522 if (is_large_pte(pte))
2523 --vcpu->kvm->stat.lpages;
2526 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2527 struct kvm_mmu_page *sp,
2531 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2532 ++vcpu->kvm->stat.mmu_pde_zapped;
2536 ++vcpu->kvm->stat.mmu_pte_updated;
2537 if (sp->role.glevels == PT32_ROOT_LEVEL)
2538 paging32_update_pte(vcpu, sp, spte, new);
2540 paging64_update_pte(vcpu, sp, spte, new);
2543 static bool need_remote_flush(u64 old, u64 new)
2545 if (!is_shadow_present_pte(old))
2547 if (!is_shadow_present_pte(new))
2549 if ((old ^ new) & PT64_BASE_ADDR_MASK)
2551 old ^= PT64_NX_MASK;
2552 new ^= PT64_NX_MASK;
2553 return (old & ~new & PT64_PERM_MASK) != 0;
2556 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2558 if (need_remote_flush(old, new))
2559 kvm_flush_remote_tlbs(vcpu->kvm);
2561 kvm_mmu_flush_tlb(vcpu);
2564 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2566 u64 *spte = vcpu->arch.last_pte_updated;
2568 return !!(spte && (*spte & shadow_accessed_mask));
2571 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2572 const u8 *new, int bytes)
2579 if (bytes != 4 && bytes != 8)
2583 * Assume that the pte write on a page table of the same type
2584 * as the current vcpu paging mode. This is nearly always true
2585 * (might be false while changing modes). Note it is verified later
2589 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2590 if ((bytes == 4) && (gpa % 4 == 0)) {
2591 r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2594 memcpy((void *)&gpte + (gpa % 8), new, 4);
2595 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2596 memcpy((void *)&gpte, new, 8);
2599 if ((bytes == 4) && (gpa % 4 == 0))
2600 memcpy((void *)&gpte, new, 4);
2602 if (!is_present_gpte(gpte))
2604 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2606 vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2608 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2610 if (is_error_pfn(pfn)) {
2611 kvm_release_pfn_clean(pfn);
2614 vcpu->arch.update_pte.gfn = gfn;
2615 vcpu->arch.update_pte.pfn = pfn;
2618 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2620 u64 *spte = vcpu->arch.last_pte_updated;
2623 && vcpu->arch.last_pte_gfn == gfn
2624 && shadow_accessed_mask
2625 && !(*spte & shadow_accessed_mask)
2626 && is_shadow_present_pte(*spte))
2627 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2630 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2631 const u8 *new, int bytes,
2632 bool guest_initiated)
2634 gfn_t gfn = gpa >> PAGE_SHIFT;
2635 struct kvm_mmu_page *sp;
2636 struct hlist_node *node, *n;
2637 struct hlist_head *bucket;
2641 unsigned offset = offset_in_page(gpa);
2643 unsigned page_offset;
2644 unsigned misaligned;
2651 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2652 mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2653 spin_lock(&vcpu->kvm->mmu_lock);
2654 kvm_mmu_access_page(vcpu, gfn);
2655 kvm_mmu_free_some_pages(vcpu);
2656 ++vcpu->kvm->stat.mmu_pte_write;
2657 kvm_mmu_audit(vcpu, "pre pte write");
2658 if (guest_initiated) {
2659 if (gfn == vcpu->arch.last_pt_write_gfn
2660 && !last_updated_pte_accessed(vcpu)) {
2661 ++vcpu->arch.last_pt_write_count;
2662 if (vcpu->arch.last_pt_write_count >= 3)
2665 vcpu->arch.last_pt_write_gfn = gfn;
2666 vcpu->arch.last_pt_write_count = 1;
2667 vcpu->arch.last_pte_updated = NULL;
2670 index = kvm_page_table_hashfn(gfn);
2671 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2672 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2673 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2675 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2676 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2677 misaligned |= bytes < 4;
2678 if (misaligned || flooded) {
2680 * Misaligned accesses are too much trouble to fix
2681 * up; also, they usually indicate a page is not used
2684 * If we're seeing too many writes to a page,
2685 * it may no longer be a page table, or we may be
2686 * forking, in which case it is better to unmap the
2689 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2690 gpa, bytes, sp->role.word);
2691 if (kvm_mmu_zap_page(vcpu->kvm, sp))
2693 ++vcpu->kvm->stat.mmu_flooded;
2696 page_offset = offset;
2697 level = sp->role.level;
2699 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2700 page_offset <<= 1; /* 32->64 */
2702 * A 32-bit pde maps 4MB while the shadow pdes map
2703 * only 2MB. So we need to double the offset again
2704 * and zap two pdes instead of one.
2706 if (level == PT32_ROOT_LEVEL) {
2707 page_offset &= ~7; /* kill rounding error */
2711 quadrant = page_offset >> PAGE_SHIFT;
2712 page_offset &= ~PAGE_MASK;
2713 if (quadrant != sp->role.quadrant)
2716 spte = &sp->spt[page_offset / sizeof(*spte)];
2717 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2719 r = kvm_read_guest_atomic(vcpu->kvm,
2720 gpa & ~(u64)(pte_size - 1),
2722 new = (const void *)&gentry;
2728 mmu_pte_write_zap_pte(vcpu, sp, spte);
2730 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2731 mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2735 kvm_mmu_audit(vcpu, "post pte write");
2736 spin_unlock(&vcpu->kvm->mmu_lock);
2737 if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2738 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2739 vcpu->arch.update_pte.pfn = bad_pfn;
2743 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2751 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2753 spin_lock(&vcpu->kvm->mmu_lock);
2754 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2755 spin_unlock(&vcpu->kvm->mmu_lock);
2758 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2760 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2762 while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2763 !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2764 struct kvm_mmu_page *sp;
2766 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2767 struct kvm_mmu_page, link);
2768 kvm_mmu_zap_page(vcpu->kvm, sp);
2769 ++vcpu->kvm->stat.mmu_recycled;
2773 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2776 enum emulation_result er;
2778 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2787 r = mmu_topup_memory_caches(vcpu);
2791 er = emulate_instruction(vcpu, cr2, error_code, 0);
2796 case EMULATE_DO_MMIO:
2797 ++vcpu->stat.mmio_exits;
2800 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2801 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2802 vcpu->run->internal.ndata = 0;
2810 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2812 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2814 vcpu->arch.mmu.invlpg(vcpu, gva);
2815 kvm_mmu_flush_tlb(vcpu);
2816 ++vcpu->stat.invlpg;
2818 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2820 void kvm_enable_tdp(void)
2824 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2826 void kvm_disable_tdp(void)
2828 tdp_enabled = false;
2830 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2832 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2834 free_page((unsigned long)vcpu->arch.mmu.pae_root);
2837 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2845 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2846 * Therefore we need to allocate shadow page tables in the first
2847 * 4GB of memory, which happens to fit the DMA32 zone.
2849 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2852 vcpu->arch.mmu.pae_root = page_address(page);
2853 for (i = 0; i < 4; ++i)
2854 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2859 free_mmu_pages(vcpu);
2863 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2866 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2868 return alloc_mmu_pages(vcpu);
2871 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2874 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2876 return init_kvm_mmu(vcpu);
2879 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2883 destroy_kvm_mmu(vcpu);
2884 free_mmu_pages(vcpu);
2885 mmu_free_memory_caches(vcpu);
2888 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2890 struct kvm_mmu_page *sp;
2892 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2896 if (!test_bit(slot, sp->slot_bitmap))
2900 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2902 if (pt[i] & PT_WRITABLE_MASK)
2903 pt[i] &= ~PT_WRITABLE_MASK;
2905 kvm_flush_remote_tlbs(kvm);
2908 void kvm_mmu_zap_all(struct kvm *kvm)
2910 struct kvm_mmu_page *sp, *node;
2912 spin_lock(&kvm->mmu_lock);
2913 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2914 if (kvm_mmu_zap_page(kvm, sp))
2915 node = container_of(kvm->arch.active_mmu_pages.next,
2916 struct kvm_mmu_page, link);
2917 spin_unlock(&kvm->mmu_lock);
2919 kvm_flush_remote_tlbs(kvm);
2922 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2924 struct kvm_mmu_page *page;
2926 page = container_of(kvm->arch.active_mmu_pages.prev,
2927 struct kvm_mmu_page, link);
2928 kvm_mmu_zap_page(kvm, page);
2931 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2934 struct kvm *kvm_freed = NULL;
2935 int cache_count = 0;
2937 spin_lock(&kvm_lock);
2939 list_for_each_entry(kvm, &vm_list, vm_list) {
2942 if (!down_read_trylock(&kvm->slots_lock))
2944 spin_lock(&kvm->mmu_lock);
2945 npages = kvm->arch.n_alloc_mmu_pages -
2946 kvm->arch.n_free_mmu_pages;
2947 cache_count += npages;
2948 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2949 kvm_mmu_remove_one_alloc_mmu_page(kvm);
2955 spin_unlock(&kvm->mmu_lock);
2956 up_read(&kvm->slots_lock);
2959 list_move_tail(&kvm_freed->vm_list, &vm_list);
2961 spin_unlock(&kvm_lock);
2966 static struct shrinker mmu_shrinker = {
2967 .shrink = mmu_shrink,
2968 .seeks = DEFAULT_SEEKS * 10,
2971 static void mmu_destroy_caches(void)
2973 if (pte_chain_cache)
2974 kmem_cache_destroy(pte_chain_cache);
2975 if (rmap_desc_cache)
2976 kmem_cache_destroy(rmap_desc_cache);
2977 if (mmu_page_header_cache)
2978 kmem_cache_destroy(mmu_page_header_cache);
2981 void kvm_mmu_module_exit(void)
2983 mmu_destroy_caches();
2984 unregister_shrinker(&mmu_shrinker);
2987 int kvm_mmu_module_init(void)
2989 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2990 sizeof(struct kvm_pte_chain),
2992 if (!pte_chain_cache)
2994 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2995 sizeof(struct kvm_rmap_desc),
2997 if (!rmap_desc_cache)
3000 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
3001 sizeof(struct kvm_mmu_page),
3003 if (!mmu_page_header_cache)
3006 register_shrinker(&mmu_shrinker);
3011 mmu_destroy_caches();
3016 * Caculate mmu pages needed for kvm.
3018 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3021 unsigned int nr_mmu_pages;
3022 unsigned int nr_pages = 0;
3024 for (i = 0; i < kvm->nmemslots; i++)
3025 nr_pages += kvm->memslots[i].npages;
3027 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3028 nr_mmu_pages = max(nr_mmu_pages,
3029 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3031 return nr_mmu_pages;
3034 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3037 if (len > buffer->len)
3042 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3047 ret = pv_mmu_peek_buffer(buffer, len);
3052 buffer->processed += len;
3056 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3057 gpa_t addr, gpa_t value)
3062 if (!is_long_mode(vcpu) && !is_pae(vcpu))
3065 r = mmu_topup_memory_caches(vcpu);
3069 if (!emulator_write_phys(vcpu, addr, &value, bytes))
3075 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3077 kvm_set_cr3(vcpu, vcpu->arch.cr3);
3081 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3083 spin_lock(&vcpu->kvm->mmu_lock);
3084 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3085 spin_unlock(&vcpu->kvm->mmu_lock);
3089 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3090 struct kvm_pv_mmu_op_buffer *buffer)
3092 struct kvm_mmu_op_header *header;
3094 header = pv_mmu_peek_buffer(buffer, sizeof *header);
3097 switch (header->op) {
3098 case KVM_MMU_OP_WRITE_PTE: {
3099 struct kvm_mmu_op_write_pte *wpte;
3101 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3104 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3107 case KVM_MMU_OP_FLUSH_TLB: {
3108 struct kvm_mmu_op_flush_tlb *ftlb;
3110 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3113 return kvm_pv_mmu_flush_tlb(vcpu);
3115 case KVM_MMU_OP_RELEASE_PT: {
3116 struct kvm_mmu_op_release_pt *rpt;
3118 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3121 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3127 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3128 gpa_t addr, unsigned long *ret)
3131 struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3133 buffer->ptr = buffer->buf;
3134 buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3135 buffer->processed = 0;
3137 r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3141 while (buffer->len) {
3142 r = kvm_pv_mmu_op_one(vcpu, buffer);
3151 *ret = buffer->processed;
3155 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3157 struct kvm_shadow_walk_iterator iterator;
3160 spin_lock(&vcpu->kvm->mmu_lock);
3161 for_each_shadow_entry(vcpu, addr, iterator) {
3162 sptes[iterator.level-1] = *iterator.sptep;
3164 if (!is_shadow_present_pte(*iterator.sptep))
3167 spin_unlock(&vcpu->kvm->mmu_lock);
3171 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3175 static const char *audit_msg;
3177 static gva_t canonicalize(gva_t gva)
3179 #ifdef CONFIG_X86_64
3180 gva = (long long)(gva << 16) >> 16;
3186 typedef void (*inspect_spte_fn) (struct kvm *kvm, struct kvm_mmu_page *sp,
3189 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3194 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3195 u64 ent = sp->spt[i];
3197 if (is_shadow_present_pte(ent)) {
3198 if (!is_last_spte(ent, sp->role.level)) {
3199 struct kvm_mmu_page *child;
3200 child = page_header(ent & PT64_BASE_ADDR_MASK);
3201 __mmu_spte_walk(kvm, child, fn);
3203 fn(kvm, sp, &sp->spt[i]);
3208 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3211 struct kvm_mmu_page *sp;
3213 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3215 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3216 hpa_t root = vcpu->arch.mmu.root_hpa;
3217 sp = page_header(root);
3218 __mmu_spte_walk(vcpu->kvm, sp, fn);
3221 for (i = 0; i < 4; ++i) {
3222 hpa_t root = vcpu->arch.mmu.pae_root[i];
3224 if (root && VALID_PAGE(root)) {
3225 root &= PT64_BASE_ADDR_MASK;
3226 sp = page_header(root);
3227 __mmu_spte_walk(vcpu->kvm, sp, fn);
3233 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3234 gva_t va, int level)
3236 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3238 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3240 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3243 if (ent == shadow_trap_nonpresent_pte)
3246 va = canonicalize(va);
3247 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3248 audit_mappings_page(vcpu, ent, va, level - 1);
3250 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3251 gfn_t gfn = gpa >> PAGE_SHIFT;
3252 pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3253 hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3255 if (is_error_pfn(pfn)) {
3256 kvm_release_pfn_clean(pfn);
3260 if (is_shadow_present_pte(ent)
3261 && (ent & PT64_BASE_ADDR_MASK) != hpa)
3262 printk(KERN_ERR "xx audit error: (%s) levels %d"
3263 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3264 audit_msg, vcpu->arch.mmu.root_level,
3266 is_shadow_present_pte(ent));
3267 else if (ent == shadow_notrap_nonpresent_pte
3268 && !is_error_hpa(hpa))
3269 printk(KERN_ERR "audit: (%s) notrap shadow,"
3270 " valid guest gva %lx\n", audit_msg, va);
3271 kvm_release_pfn_clean(pfn);
3277 static void audit_mappings(struct kvm_vcpu *vcpu)
3281 if (vcpu->arch.mmu.root_level == 4)
3282 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3284 for (i = 0; i < 4; ++i)
3285 if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3286 audit_mappings_page(vcpu,
3287 vcpu->arch.mmu.pae_root[i],
3292 static int count_rmaps(struct kvm_vcpu *vcpu)
3297 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3298 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3299 struct kvm_rmap_desc *d;
3301 for (j = 0; j < m->npages; ++j) {
3302 unsigned long *rmapp = &m->rmap[j];
3306 if (!(*rmapp & 1)) {
3310 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3312 for (k = 0; k < RMAP_EXT; ++k)
3324 void inspect_spte_has_rmap(struct kvm *kvm, struct kvm_mmu_page *sp, u64 *sptep)
3326 unsigned long *rmapp;
3327 struct kvm_mmu_page *rev_sp;
3330 if (*sptep & PT_WRITABLE_MASK) {
3331 rev_sp = page_header(__pa(sptep));
3332 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3334 if (!gfn_to_memslot(kvm, gfn)) {
3335 if (!printk_ratelimit())
3337 printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3339 printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3340 audit_msg, sptep - rev_sp->spt,
3346 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3347 is_large_pte(*sptep));
3349 if (!printk_ratelimit())
3351 printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3359 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3361 mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3364 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3366 struct kvm_mmu_page *sp;
3369 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3372 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3375 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3378 if (!(ent & PT_PRESENT_MASK))
3380 if (!(ent & PT_WRITABLE_MASK))
3382 inspect_spte_has_rmap(vcpu->kvm, sp, &pt[i]);
3388 static void audit_rmap(struct kvm_vcpu *vcpu)
3390 check_writable_mappings_rmap(vcpu);
3394 static void audit_write_protection(struct kvm_vcpu *vcpu)
3396 struct kvm_mmu_page *sp;
3397 struct kvm_memory_slot *slot;
3398 unsigned long *rmapp;
3402 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3403 if (sp->role.direct)
3408 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3409 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3410 rmapp = &slot->rmap[gfn - slot->base_gfn];
3412 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3414 if (*spte & PT_WRITABLE_MASK)
3415 printk(KERN_ERR "%s: (%s) shadow page has "
3416 "writable mappings: gfn %lx role %x\n",
3417 __func__, audit_msg, sp->gfn,
3419 spte = rmap_next(vcpu->kvm, rmapp, spte);
3424 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3431 audit_write_protection(vcpu);
3432 if (strcmp("pre pte write", audit_msg) != 0)
3433 audit_mappings(vcpu);
3434 audit_writable_sptes_have_rmaps(vcpu);