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>
36 #include <asm/cmpxchg.h>
41 * When setting this variable to true it enables Two-Dimensional-Paging
42 * where the hardware walks 2 page tables:
43 * 1. the guest-virtual to guest-physical
44 * 2. while doing 1. it walks guest-physical to host-physical
45 * If the hardware supports that we don't need to do shadow paging.
47 bool tdp_enabled = false;
54 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
56 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
61 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
62 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
66 #define pgprintk(x...) do { } while (0)
67 #define rmap_printk(x...) do { } while (0)
71 #if defined(MMU_DEBUG) || defined(AUDIT)
73 module_param(dbg, bool, 0644);
76 static int oos_shadow = 1;
77 module_param(oos_shadow, bool, 0644);
80 #define ASSERT(x) do { } while (0)
84 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
85 __FILE__, __LINE__, #x); \
89 #define PT_FIRST_AVAIL_BITS_SHIFT 9
90 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
92 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
94 #define PT64_LEVEL_BITS 9
96 #define PT64_LEVEL_SHIFT(level) \
97 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
99 #define PT64_LEVEL_MASK(level) \
100 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
102 #define PT64_INDEX(address, level)\
103 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
106 #define PT32_LEVEL_BITS 10
108 #define PT32_LEVEL_SHIFT(level) \
109 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
111 #define PT32_LEVEL_MASK(level) \
112 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
113 #define PT32_LVL_OFFSET_MASK(level) \
114 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
115 * PT32_LEVEL_BITS))) - 1))
117 #define PT32_INDEX(address, level)\
118 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
121 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
122 #define PT64_DIR_BASE_ADDR_MASK \
123 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
124 #define PT64_LVL_ADDR_MASK(level) \
125 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
126 * PT64_LEVEL_BITS))) - 1))
127 #define PT64_LVL_OFFSET_MASK(level) \
128 (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
129 * PT64_LEVEL_BITS))) - 1))
131 #define PT32_BASE_ADDR_MASK PAGE_MASK
132 #define PT32_DIR_BASE_ADDR_MASK \
133 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
134 #define PT32_LVL_ADDR_MASK(level) \
135 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
136 * PT32_LEVEL_BITS))) - 1))
138 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
143 #define ACC_EXEC_MASK 1
144 #define ACC_WRITE_MASK PT_WRITABLE_MASK
145 #define ACC_USER_MASK PT_USER_MASK
146 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
148 #include <trace/events/kvm.h>
150 #undef TRACE_INCLUDE_FILE
151 #define CREATE_TRACE_POINTS
152 #include "mmutrace.h"
154 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
156 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
158 struct kvm_rmap_desc {
159 u64 *sptes[RMAP_EXT];
160 struct kvm_rmap_desc *more;
163 struct kvm_shadow_walk_iterator {
171 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
172 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
173 shadow_walk_okay(&(_walker)); \
174 shadow_walk_next(&(_walker)))
177 struct kvm_unsync_walk {
178 int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
181 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
183 static struct kmem_cache *pte_chain_cache;
184 static struct kmem_cache *rmap_desc_cache;
185 static struct kmem_cache *mmu_page_header_cache;
187 static u64 __read_mostly shadow_trap_nonpresent_pte;
188 static u64 __read_mostly shadow_notrap_nonpresent_pte;
189 static u64 __read_mostly shadow_base_present_pte;
190 static u64 __read_mostly shadow_nx_mask;
191 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
192 static u64 __read_mostly shadow_user_mask;
193 static u64 __read_mostly shadow_accessed_mask;
194 static u64 __read_mostly shadow_dirty_mask;
196 static inline u64 rsvd_bits(int s, int e)
198 return ((1ULL << (e - s + 1)) - 1) << s;
201 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
203 shadow_trap_nonpresent_pte = trap_pte;
204 shadow_notrap_nonpresent_pte = notrap_pte;
206 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
208 void kvm_mmu_set_base_ptes(u64 base_pte)
210 shadow_base_present_pte = base_pte;
212 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
214 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
215 u64 dirty_mask, u64 nx_mask, u64 x_mask)
217 shadow_user_mask = user_mask;
218 shadow_accessed_mask = accessed_mask;
219 shadow_dirty_mask = dirty_mask;
220 shadow_nx_mask = nx_mask;
221 shadow_x_mask = x_mask;
223 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
225 static int is_write_protection(struct kvm_vcpu *vcpu)
227 return kvm_read_cr0_bits(vcpu, X86_CR0_WP);
230 static int is_cpuid_PSE36(void)
235 static int is_nx(struct kvm_vcpu *vcpu)
237 return vcpu->arch.efer & EFER_NX;
240 static int is_shadow_present_pte(u64 pte)
242 return pte != shadow_trap_nonpresent_pte
243 && pte != shadow_notrap_nonpresent_pte;
246 static int is_large_pte(u64 pte)
248 return pte & PT_PAGE_SIZE_MASK;
251 static int is_writable_pte(unsigned long pte)
253 return pte & PT_WRITABLE_MASK;
256 static int is_dirty_gpte(unsigned long pte)
258 return pte & PT_DIRTY_MASK;
261 static int is_rmap_spte(u64 pte)
263 return is_shadow_present_pte(pte);
266 static int is_last_spte(u64 pte, int level)
268 if (level == PT_PAGE_TABLE_LEVEL)
270 if (is_large_pte(pte))
275 static pfn_t spte_to_pfn(u64 pte)
277 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
280 static gfn_t pse36_gfn_delta(u32 gpte)
282 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
284 return (gpte & PT32_DIR_PSE36_MASK) << shift;
287 static void __set_spte(u64 *sptep, u64 spte)
290 set_64bit((unsigned long *)sptep, spte);
292 set_64bit((unsigned long long *)sptep, spte);
296 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
297 struct kmem_cache *base_cache, int min)
301 if (cache->nobjs >= min)
303 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
304 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
307 cache->objects[cache->nobjs++] = obj;
312 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
315 kfree(mc->objects[--mc->nobjs]);
318 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
323 if (cache->nobjs >= min)
325 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
326 page = alloc_page(GFP_KERNEL);
329 set_page_private(page, 0);
330 cache->objects[cache->nobjs++] = page_address(page);
335 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
338 free_page((unsigned long)mc->objects[--mc->nobjs]);
341 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
345 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
349 r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
353 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
356 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
357 mmu_page_header_cache, 4);
362 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
364 mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
365 mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
366 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
367 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
370 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
376 p = mc->objects[--mc->nobjs];
380 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
382 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
383 sizeof(struct kvm_pte_chain));
386 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
391 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
393 return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
394 sizeof(struct kvm_rmap_desc));
397 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
403 * Return the pointer to the largepage write count for a given
404 * gfn, handling slots that are not large page aligned.
406 static int *slot_largepage_idx(gfn_t gfn,
407 struct kvm_memory_slot *slot,
412 idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
413 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
414 return &slot->lpage_info[level - 2][idx].write_count;
417 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
419 struct kvm_memory_slot *slot;
423 gfn = unalias_gfn(kvm, gfn);
425 slot = gfn_to_memslot_unaliased(kvm, gfn);
426 for (i = PT_DIRECTORY_LEVEL;
427 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
428 write_count = slot_largepage_idx(gfn, slot, i);
433 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
435 struct kvm_memory_slot *slot;
439 gfn = unalias_gfn(kvm, gfn);
440 for (i = PT_DIRECTORY_LEVEL;
441 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
442 slot = gfn_to_memslot_unaliased(kvm, gfn);
443 write_count = slot_largepage_idx(gfn, slot, i);
445 WARN_ON(*write_count < 0);
449 static int has_wrprotected_page(struct kvm *kvm,
453 struct kvm_memory_slot *slot;
456 gfn = unalias_gfn(kvm, gfn);
457 slot = gfn_to_memslot_unaliased(kvm, gfn);
459 largepage_idx = slot_largepage_idx(gfn, slot, level);
460 return *largepage_idx;
466 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
468 unsigned long page_size;
471 page_size = kvm_host_page_size(kvm, gfn);
473 for (i = PT_PAGE_TABLE_LEVEL;
474 i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
475 if (page_size >= KVM_HPAGE_SIZE(i))
484 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
486 struct kvm_memory_slot *slot;
487 int host_level, level, max_level;
489 slot = gfn_to_memslot(vcpu->kvm, large_gfn);
490 if (slot && slot->dirty_bitmap)
491 return PT_PAGE_TABLE_LEVEL;
493 host_level = host_mapping_level(vcpu->kvm, large_gfn);
495 if (host_level == PT_PAGE_TABLE_LEVEL)
498 max_level = kvm_x86_ops->get_lpage_level() < host_level ?
499 kvm_x86_ops->get_lpage_level() : host_level;
501 for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
502 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
509 * Take gfn and return the reverse mapping to it.
510 * Note: gfn must be unaliased before this function get called
513 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
515 struct kvm_memory_slot *slot;
518 slot = gfn_to_memslot(kvm, gfn);
519 if (likely(level == PT_PAGE_TABLE_LEVEL))
520 return &slot->rmap[gfn - slot->base_gfn];
522 idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
523 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
525 return &slot->lpage_info[level - 2][idx].rmap_pde;
529 * Reverse mapping data structures:
531 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
532 * that points to page_address(page).
534 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
535 * containing more mappings.
537 * Returns the number of rmap entries before the spte was added or zero if
538 * the spte was not added.
541 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
543 struct kvm_mmu_page *sp;
544 struct kvm_rmap_desc *desc;
545 unsigned long *rmapp;
548 if (!is_rmap_spte(*spte))
550 gfn = unalias_gfn(vcpu->kvm, gfn);
551 sp = page_header(__pa(spte));
552 sp->gfns[spte - sp->spt] = gfn;
553 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
555 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
556 *rmapp = (unsigned long)spte;
557 } else if (!(*rmapp & 1)) {
558 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
559 desc = mmu_alloc_rmap_desc(vcpu);
560 desc->sptes[0] = (u64 *)*rmapp;
561 desc->sptes[1] = spte;
562 *rmapp = (unsigned long)desc | 1;
564 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
565 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
566 while (desc->sptes[RMAP_EXT-1] && desc->more) {
570 if (desc->sptes[RMAP_EXT-1]) {
571 desc->more = mmu_alloc_rmap_desc(vcpu);
574 for (i = 0; desc->sptes[i]; ++i)
576 desc->sptes[i] = spte;
581 static void rmap_desc_remove_entry(unsigned long *rmapp,
582 struct kvm_rmap_desc *desc,
584 struct kvm_rmap_desc *prev_desc)
588 for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
590 desc->sptes[i] = desc->sptes[j];
591 desc->sptes[j] = NULL;
594 if (!prev_desc && !desc->more)
595 *rmapp = (unsigned long)desc->sptes[0];
598 prev_desc->more = desc->more;
600 *rmapp = (unsigned long)desc->more | 1;
601 mmu_free_rmap_desc(desc);
604 static void rmap_remove(struct kvm *kvm, u64 *spte)
606 struct kvm_rmap_desc *desc;
607 struct kvm_rmap_desc *prev_desc;
608 struct kvm_mmu_page *sp;
610 unsigned long *rmapp;
613 if (!is_rmap_spte(*spte))
615 sp = page_header(__pa(spte));
616 pfn = spte_to_pfn(*spte);
617 if (*spte & shadow_accessed_mask)
618 kvm_set_pfn_accessed(pfn);
619 if (is_writable_pte(*spte))
620 kvm_set_pfn_dirty(pfn);
621 rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], sp->role.level);
623 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
625 } else if (!(*rmapp & 1)) {
626 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
627 if ((u64 *)*rmapp != spte) {
628 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
634 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
635 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
638 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
639 if (desc->sptes[i] == spte) {
640 rmap_desc_remove_entry(rmapp,
648 pr_err("rmap_remove: %p %llx many->many\n", spte, *spte);
653 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
655 struct kvm_rmap_desc *desc;
656 struct kvm_rmap_desc *prev_desc;
662 else if (!(*rmapp & 1)) {
664 return (u64 *)*rmapp;
667 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
671 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
672 if (prev_spte == spte)
673 return desc->sptes[i];
674 prev_spte = desc->sptes[i];
681 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
683 unsigned long *rmapp;
685 int i, write_protected = 0;
687 gfn = unalias_gfn(kvm, gfn);
688 rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
690 spte = rmap_next(kvm, rmapp, NULL);
693 BUG_ON(!(*spte & PT_PRESENT_MASK));
694 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
695 if (is_writable_pte(*spte)) {
696 __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
699 spte = rmap_next(kvm, rmapp, spte);
701 if (write_protected) {
704 spte = rmap_next(kvm, rmapp, NULL);
705 pfn = spte_to_pfn(*spte);
706 kvm_set_pfn_dirty(pfn);
709 /* check for huge page mappings */
710 for (i = PT_DIRECTORY_LEVEL;
711 i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
712 rmapp = gfn_to_rmap(kvm, gfn, i);
713 spte = rmap_next(kvm, rmapp, NULL);
716 BUG_ON(!(*spte & PT_PRESENT_MASK));
717 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
718 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
719 if (is_writable_pte(*spte)) {
720 rmap_remove(kvm, spte);
722 __set_spte(spte, shadow_trap_nonpresent_pte);
726 spte = rmap_next(kvm, rmapp, spte);
730 return write_protected;
733 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
737 int need_tlb_flush = 0;
739 while ((spte = rmap_next(kvm, rmapp, NULL))) {
740 BUG_ON(!(*spte & PT_PRESENT_MASK));
741 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
742 rmap_remove(kvm, spte);
743 __set_spte(spte, shadow_trap_nonpresent_pte);
746 return need_tlb_flush;
749 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
754 pte_t *ptep = (pte_t *)data;
757 WARN_ON(pte_huge(*ptep));
758 new_pfn = pte_pfn(*ptep);
759 spte = rmap_next(kvm, rmapp, NULL);
761 BUG_ON(!is_shadow_present_pte(*spte));
762 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
764 if (pte_write(*ptep)) {
765 rmap_remove(kvm, spte);
766 __set_spte(spte, shadow_trap_nonpresent_pte);
767 spte = rmap_next(kvm, rmapp, NULL);
769 new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
770 new_spte |= (u64)new_pfn << PAGE_SHIFT;
772 new_spte &= ~PT_WRITABLE_MASK;
773 new_spte &= ~SPTE_HOST_WRITEABLE;
774 if (is_writable_pte(*spte))
775 kvm_set_pfn_dirty(spte_to_pfn(*spte));
776 __set_spte(spte, new_spte);
777 spte = rmap_next(kvm, rmapp, spte);
781 kvm_flush_remote_tlbs(kvm);
786 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
788 int (*handler)(struct kvm *kvm, unsigned long *rmapp,
794 struct kvm_memslots *slots;
796 slots = rcu_dereference(kvm->memslots);
798 for (i = 0; i < slots->nmemslots; i++) {
799 struct kvm_memory_slot *memslot = &slots->memslots[i];
800 unsigned long start = memslot->userspace_addr;
803 end = start + (memslot->npages << PAGE_SHIFT);
804 if (hva >= start && hva < end) {
805 gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
807 ret = handler(kvm, &memslot->rmap[gfn_offset], data);
809 for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
810 int idx = gfn_offset;
811 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
813 &memslot->lpage_info[j][idx].rmap_pde,
816 trace_kvm_age_page(hva, memslot, ret);
824 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
826 return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
829 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
831 kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
834 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
841 * Emulate the accessed bit for EPT, by checking if this page has
842 * an EPT mapping, and clearing it if it does. On the next access,
843 * a new EPT mapping will be established.
844 * This has some overhead, but not as much as the cost of swapping
845 * out actively used pages or breaking up actively used hugepages.
847 if (!shadow_accessed_mask)
848 return kvm_unmap_rmapp(kvm, rmapp, data);
850 spte = rmap_next(kvm, rmapp, NULL);
854 BUG_ON(!(_spte & PT_PRESENT_MASK));
855 _young = _spte & PT_ACCESSED_MASK;
858 clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
860 spte = rmap_next(kvm, rmapp, spte);
865 #define RMAP_RECYCLE_THRESHOLD 1000
867 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
869 unsigned long *rmapp;
870 struct kvm_mmu_page *sp;
872 sp = page_header(__pa(spte));
874 gfn = unalias_gfn(vcpu->kvm, gfn);
875 rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
877 kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
878 kvm_flush_remote_tlbs(vcpu->kvm);
881 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
883 return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
887 static int is_empty_shadow_page(u64 *spt)
892 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
893 if (is_shadow_present_pte(*pos)) {
894 printk(KERN_ERR "%s: %p %llx\n", __func__,
902 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
904 ASSERT(is_empty_shadow_page(sp->spt));
906 __free_page(virt_to_page(sp->spt));
907 __free_page(virt_to_page(sp->gfns));
909 ++kvm->arch.n_free_mmu_pages;
912 static unsigned kvm_page_table_hashfn(gfn_t gfn)
914 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
917 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
920 struct kvm_mmu_page *sp;
922 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
923 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
924 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
925 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
926 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
927 INIT_LIST_HEAD(&sp->oos_link);
928 bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
930 sp->parent_pte = parent_pte;
931 --vcpu->kvm->arch.n_free_mmu_pages;
935 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
936 struct kvm_mmu_page *sp, u64 *parent_pte)
938 struct kvm_pte_chain *pte_chain;
939 struct hlist_node *node;
944 if (!sp->multimapped) {
945 u64 *old = sp->parent_pte;
948 sp->parent_pte = parent_pte;
952 pte_chain = mmu_alloc_pte_chain(vcpu);
953 INIT_HLIST_HEAD(&sp->parent_ptes);
954 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
955 pte_chain->parent_ptes[0] = old;
957 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
958 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
960 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
961 if (!pte_chain->parent_ptes[i]) {
962 pte_chain->parent_ptes[i] = parent_pte;
966 pte_chain = mmu_alloc_pte_chain(vcpu);
968 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
969 pte_chain->parent_ptes[0] = parent_pte;
972 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
975 struct kvm_pte_chain *pte_chain;
976 struct hlist_node *node;
979 if (!sp->multimapped) {
980 BUG_ON(sp->parent_pte != parent_pte);
981 sp->parent_pte = NULL;
984 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
985 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
986 if (!pte_chain->parent_ptes[i])
988 if (pte_chain->parent_ptes[i] != parent_pte)
990 while (i + 1 < NR_PTE_CHAIN_ENTRIES
991 && pte_chain->parent_ptes[i + 1]) {
992 pte_chain->parent_ptes[i]
993 = pte_chain->parent_ptes[i + 1];
996 pte_chain->parent_ptes[i] = NULL;
998 hlist_del(&pte_chain->link);
999 mmu_free_pte_chain(pte_chain);
1000 if (hlist_empty(&sp->parent_ptes)) {
1001 sp->multimapped = 0;
1002 sp->parent_pte = NULL;
1011 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1012 mmu_parent_walk_fn fn)
1014 struct kvm_pte_chain *pte_chain;
1015 struct hlist_node *node;
1016 struct kvm_mmu_page *parent_sp;
1019 if (!sp->multimapped && sp->parent_pte) {
1020 parent_sp = page_header(__pa(sp->parent_pte));
1021 fn(vcpu, parent_sp);
1022 mmu_parent_walk(vcpu, parent_sp, fn);
1025 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1026 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1027 if (!pte_chain->parent_ptes[i])
1029 parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
1030 fn(vcpu, parent_sp);
1031 mmu_parent_walk(vcpu, parent_sp, fn);
1035 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
1038 struct kvm_mmu_page *sp = page_header(__pa(spte));
1040 index = spte - sp->spt;
1041 if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
1042 sp->unsync_children++;
1043 WARN_ON(!sp->unsync_children);
1046 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
1048 struct kvm_pte_chain *pte_chain;
1049 struct hlist_node *node;
1052 if (!sp->parent_pte)
1055 if (!sp->multimapped) {
1056 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
1060 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1061 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1062 if (!pte_chain->parent_ptes[i])
1064 kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
1068 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1070 kvm_mmu_update_parents_unsync(sp);
1074 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
1075 struct kvm_mmu_page *sp)
1077 mmu_parent_walk(vcpu, sp, unsync_walk_fn);
1078 kvm_mmu_update_parents_unsync(sp);
1081 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1082 struct kvm_mmu_page *sp)
1086 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1087 sp->spt[i] = shadow_trap_nonpresent_pte;
1090 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1091 struct kvm_mmu_page *sp)
1096 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1100 #define KVM_PAGE_ARRAY_NR 16
1102 struct kvm_mmu_pages {
1103 struct mmu_page_and_offset {
1104 struct kvm_mmu_page *sp;
1106 } page[KVM_PAGE_ARRAY_NR];
1110 #define for_each_unsync_children(bitmap, idx) \
1111 for (idx = find_first_bit(bitmap, 512); \
1113 idx = find_next_bit(bitmap, 512, idx+1))
1115 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1121 for (i=0; i < pvec->nr; i++)
1122 if (pvec->page[i].sp == sp)
1125 pvec->page[pvec->nr].sp = sp;
1126 pvec->page[pvec->nr].idx = idx;
1128 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1131 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1132 struct kvm_mmu_pages *pvec)
1134 int i, ret, nr_unsync_leaf = 0;
1136 for_each_unsync_children(sp->unsync_child_bitmap, i) {
1137 u64 ent = sp->spt[i];
1139 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1140 struct kvm_mmu_page *child;
1141 child = page_header(ent & PT64_BASE_ADDR_MASK);
1143 if (child->unsync_children) {
1144 if (mmu_pages_add(pvec, child, i))
1147 ret = __mmu_unsync_walk(child, pvec);
1149 __clear_bit(i, sp->unsync_child_bitmap);
1151 nr_unsync_leaf += ret;
1156 if (child->unsync) {
1158 if (mmu_pages_add(pvec, child, i))
1164 if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1165 sp->unsync_children = 0;
1167 return nr_unsync_leaf;
1170 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1171 struct kvm_mmu_pages *pvec)
1173 if (!sp->unsync_children)
1176 mmu_pages_add(pvec, sp, 0);
1177 return __mmu_unsync_walk(sp, pvec);
1180 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1183 struct hlist_head *bucket;
1184 struct kvm_mmu_page *sp;
1185 struct hlist_node *node;
1187 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1188 index = kvm_page_table_hashfn(gfn);
1189 bucket = &kvm->arch.mmu_page_hash[index];
1190 hlist_for_each_entry(sp, node, bucket, hash_link)
1191 if (sp->gfn == gfn && !sp->role.direct
1192 && !sp->role.invalid) {
1193 pgprintk("%s: found role %x\n",
1194 __func__, sp->role.word);
1200 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1202 WARN_ON(!sp->unsync);
1204 --kvm->stat.mmu_unsync;
1207 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1209 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1211 if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1212 kvm_mmu_zap_page(vcpu->kvm, sp);
1216 trace_kvm_mmu_sync_page(sp);
1217 if (rmap_write_protect(vcpu->kvm, sp->gfn))
1218 kvm_flush_remote_tlbs(vcpu->kvm);
1219 kvm_unlink_unsync_page(vcpu->kvm, sp);
1220 if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1221 kvm_mmu_zap_page(vcpu->kvm, sp);
1225 kvm_mmu_flush_tlb(vcpu);
1229 struct mmu_page_path {
1230 struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1231 unsigned int idx[PT64_ROOT_LEVEL-1];
1234 #define for_each_sp(pvec, sp, parents, i) \
1235 for (i = mmu_pages_next(&pvec, &parents, -1), \
1236 sp = pvec.page[i].sp; \
1237 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1238 i = mmu_pages_next(&pvec, &parents, i))
1240 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1241 struct mmu_page_path *parents,
1246 for (n = i+1; n < pvec->nr; n++) {
1247 struct kvm_mmu_page *sp = pvec->page[n].sp;
1249 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1250 parents->idx[0] = pvec->page[n].idx;
1254 parents->parent[sp->role.level-2] = sp;
1255 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1261 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1263 struct kvm_mmu_page *sp;
1264 unsigned int level = 0;
1267 unsigned int idx = parents->idx[level];
1269 sp = parents->parent[level];
1273 --sp->unsync_children;
1274 WARN_ON((int)sp->unsync_children < 0);
1275 __clear_bit(idx, sp->unsync_child_bitmap);
1277 } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1280 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1281 struct mmu_page_path *parents,
1282 struct kvm_mmu_pages *pvec)
1284 parents->parent[parent->role.level-1] = NULL;
1288 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1289 struct kvm_mmu_page *parent)
1292 struct kvm_mmu_page *sp;
1293 struct mmu_page_path parents;
1294 struct kvm_mmu_pages pages;
1296 kvm_mmu_pages_init(parent, &parents, &pages);
1297 while (mmu_unsync_walk(parent, &pages)) {
1300 for_each_sp(pages, sp, parents, i)
1301 protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1304 kvm_flush_remote_tlbs(vcpu->kvm);
1306 for_each_sp(pages, sp, parents, i) {
1307 kvm_sync_page(vcpu, sp);
1308 mmu_pages_clear_parents(&parents);
1310 cond_resched_lock(&vcpu->kvm->mmu_lock);
1311 kvm_mmu_pages_init(parent, &parents, &pages);
1315 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1323 union kvm_mmu_page_role role;
1326 struct hlist_head *bucket;
1327 struct kvm_mmu_page *sp;
1328 struct hlist_node *node, *tmp;
1330 role = vcpu->arch.mmu.base_role;
1332 role.direct = direct;
1333 role.access = access;
1334 if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1335 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1336 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1337 role.quadrant = quadrant;
1339 index = kvm_page_table_hashfn(gfn);
1340 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1341 hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1342 if (sp->gfn == gfn) {
1344 if (kvm_sync_page(vcpu, sp))
1347 if (sp->role.word != role.word)
1350 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1351 if (sp->unsync_children) {
1352 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1353 kvm_mmu_mark_parents_unsync(vcpu, sp);
1355 trace_kvm_mmu_get_page(sp, false);
1358 ++vcpu->kvm->stat.mmu_cache_miss;
1359 sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1364 hlist_add_head(&sp->hash_link, bucket);
1366 if (rmap_write_protect(vcpu->kvm, gfn))
1367 kvm_flush_remote_tlbs(vcpu->kvm);
1368 account_shadowed(vcpu->kvm, gfn);
1370 if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1371 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1373 nonpaging_prefetch_page(vcpu, sp);
1374 trace_kvm_mmu_get_page(sp, true);
1378 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1379 struct kvm_vcpu *vcpu, u64 addr)
1381 iterator->addr = addr;
1382 iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1383 iterator->level = vcpu->arch.mmu.shadow_root_level;
1384 if (iterator->level == PT32E_ROOT_LEVEL) {
1385 iterator->shadow_addr
1386 = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1387 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1389 if (!iterator->shadow_addr)
1390 iterator->level = 0;
1394 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1396 if (iterator->level < PT_PAGE_TABLE_LEVEL)
1399 if (iterator->level == PT_PAGE_TABLE_LEVEL)
1400 if (is_large_pte(*iterator->sptep))
1403 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1404 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1408 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1410 iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1414 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1415 struct kvm_mmu_page *sp)
1423 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1426 if (is_shadow_present_pte(ent)) {
1427 if (!is_last_spte(ent, sp->role.level)) {
1428 ent &= PT64_BASE_ADDR_MASK;
1429 mmu_page_remove_parent_pte(page_header(ent),
1432 if (is_large_pte(ent))
1434 rmap_remove(kvm, &pt[i]);
1437 pt[i] = shadow_trap_nonpresent_pte;
1441 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1443 mmu_page_remove_parent_pte(sp, parent_pte);
1446 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1449 struct kvm_vcpu *vcpu;
1451 kvm_for_each_vcpu(i, vcpu, kvm)
1452 vcpu->arch.last_pte_updated = NULL;
1455 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1459 while (sp->multimapped || sp->parent_pte) {
1460 if (!sp->multimapped)
1461 parent_pte = sp->parent_pte;
1463 struct kvm_pte_chain *chain;
1465 chain = container_of(sp->parent_ptes.first,
1466 struct kvm_pte_chain, link);
1467 parent_pte = chain->parent_ptes[0];
1469 BUG_ON(!parent_pte);
1470 kvm_mmu_put_page(sp, parent_pte);
1471 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1475 static int mmu_zap_unsync_children(struct kvm *kvm,
1476 struct kvm_mmu_page *parent)
1479 struct mmu_page_path parents;
1480 struct kvm_mmu_pages pages;
1482 if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1485 kvm_mmu_pages_init(parent, &parents, &pages);
1486 while (mmu_unsync_walk(parent, &pages)) {
1487 struct kvm_mmu_page *sp;
1489 for_each_sp(pages, sp, parents, i) {
1490 kvm_mmu_zap_page(kvm, sp);
1491 mmu_pages_clear_parents(&parents);
1494 kvm_mmu_pages_init(parent, &parents, &pages);
1500 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1504 trace_kvm_mmu_zap_page(sp);
1505 ++kvm->stat.mmu_shadow_zapped;
1506 ret = mmu_zap_unsync_children(kvm, sp);
1507 kvm_mmu_page_unlink_children(kvm, sp);
1508 kvm_mmu_unlink_parents(kvm, sp);
1509 kvm_flush_remote_tlbs(kvm);
1510 if (!sp->role.invalid && !sp->role.direct)
1511 unaccount_shadowed(kvm, sp->gfn);
1513 kvm_unlink_unsync_page(kvm, sp);
1514 if (!sp->root_count) {
1515 hlist_del(&sp->hash_link);
1516 kvm_mmu_free_page(kvm, sp);
1518 sp->role.invalid = 1;
1519 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1520 kvm_reload_remote_mmus(kvm);
1522 kvm_mmu_reset_last_pte_updated(kvm);
1527 * Changing the number of mmu pages allocated to the vm
1528 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1530 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1534 used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1535 used_pages = max(0, used_pages);
1538 * If we set the number of mmu pages to be smaller be than the
1539 * number of actived pages , we must to free some mmu pages before we
1543 if (used_pages > kvm_nr_mmu_pages) {
1544 while (used_pages > kvm_nr_mmu_pages) {
1545 struct kvm_mmu_page *page;
1547 page = container_of(kvm->arch.active_mmu_pages.prev,
1548 struct kvm_mmu_page, link);
1549 kvm_mmu_zap_page(kvm, page);
1552 kvm->arch.n_free_mmu_pages = 0;
1555 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1556 - kvm->arch.n_alloc_mmu_pages;
1558 kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1561 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1564 struct hlist_head *bucket;
1565 struct kvm_mmu_page *sp;
1566 struct hlist_node *node, *n;
1569 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1571 index = kvm_page_table_hashfn(gfn);
1572 bucket = &kvm->arch.mmu_page_hash[index];
1573 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1574 if (sp->gfn == gfn && !sp->role.direct) {
1575 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1578 if (kvm_mmu_zap_page(kvm, sp))
1584 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1587 struct hlist_head *bucket;
1588 struct kvm_mmu_page *sp;
1589 struct hlist_node *node, *nn;
1591 index = kvm_page_table_hashfn(gfn);
1592 bucket = &kvm->arch.mmu_page_hash[index];
1593 hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1594 if (sp->gfn == gfn && !sp->role.direct
1595 && !sp->role.invalid) {
1596 pgprintk("%s: zap %lx %x\n",
1597 __func__, gfn, sp->role.word);
1598 kvm_mmu_zap_page(kvm, sp);
1603 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1605 int slot = memslot_id(kvm, gfn);
1606 struct kvm_mmu_page *sp = page_header(__pa(pte));
1608 __set_bit(slot, sp->slot_bitmap);
1611 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1616 if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1619 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1620 if (pt[i] == shadow_notrap_nonpresent_pte)
1621 __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1625 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1629 gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
1631 if (gpa == UNMAPPED_GVA)
1634 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1640 * The function is based on mtrr_type_lookup() in
1641 * arch/x86/kernel/cpu/mtrr/generic.c
1643 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1648 u8 prev_match, curr_match;
1649 int num_var_ranges = KVM_NR_VAR_MTRR;
1651 if (!mtrr_state->enabled)
1654 /* Make end inclusive end, instead of exclusive */
1657 /* Look in fixed ranges. Just return the type as per start */
1658 if (mtrr_state->have_fixed && (start < 0x100000)) {
1661 if (start < 0x80000) {
1663 idx += (start >> 16);
1664 return mtrr_state->fixed_ranges[idx];
1665 } else if (start < 0xC0000) {
1667 idx += ((start - 0x80000) >> 14);
1668 return mtrr_state->fixed_ranges[idx];
1669 } else if (start < 0x1000000) {
1671 idx += ((start - 0xC0000) >> 12);
1672 return mtrr_state->fixed_ranges[idx];
1677 * Look in variable ranges
1678 * Look of multiple ranges matching this address and pick type
1679 * as per MTRR precedence
1681 if (!(mtrr_state->enabled & 2))
1682 return mtrr_state->def_type;
1685 for (i = 0; i < num_var_ranges; ++i) {
1686 unsigned short start_state, end_state;
1688 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1691 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1692 (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1693 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1694 (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1696 start_state = ((start & mask) == (base & mask));
1697 end_state = ((end & mask) == (base & mask));
1698 if (start_state != end_state)
1701 if ((start & mask) != (base & mask))
1704 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1705 if (prev_match == 0xFF) {
1706 prev_match = curr_match;
1710 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1711 curr_match == MTRR_TYPE_UNCACHABLE)
1712 return MTRR_TYPE_UNCACHABLE;
1714 if ((prev_match == MTRR_TYPE_WRBACK &&
1715 curr_match == MTRR_TYPE_WRTHROUGH) ||
1716 (prev_match == MTRR_TYPE_WRTHROUGH &&
1717 curr_match == MTRR_TYPE_WRBACK)) {
1718 prev_match = MTRR_TYPE_WRTHROUGH;
1719 curr_match = MTRR_TYPE_WRTHROUGH;
1722 if (prev_match != curr_match)
1723 return MTRR_TYPE_UNCACHABLE;
1726 if (prev_match != 0xFF)
1729 return mtrr_state->def_type;
1732 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1736 mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1737 (gfn << PAGE_SHIFT) + PAGE_SIZE);
1738 if (mtrr == 0xfe || mtrr == 0xff)
1739 mtrr = MTRR_TYPE_WRBACK;
1742 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1744 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1747 struct hlist_head *bucket;
1748 struct kvm_mmu_page *s;
1749 struct hlist_node *node, *n;
1751 trace_kvm_mmu_unsync_page(sp);
1752 index = kvm_page_table_hashfn(sp->gfn);
1753 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1754 /* don't unsync if pagetable is shadowed with multiple roles */
1755 hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1756 if (s->gfn != sp->gfn || s->role.direct)
1758 if (s->role.word != sp->role.word)
1761 ++vcpu->kvm->stat.mmu_unsync;
1764 kvm_mmu_mark_parents_unsync(vcpu, sp);
1766 mmu_convert_notrap(sp);
1770 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1773 struct kvm_mmu_page *shadow;
1775 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1777 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1781 if (can_unsync && oos_shadow)
1782 return kvm_unsync_page(vcpu, shadow);
1788 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1789 unsigned pte_access, int user_fault,
1790 int write_fault, int dirty, int level,
1791 gfn_t gfn, pfn_t pfn, bool speculative,
1792 bool can_unsync, bool reset_host_protection)
1798 * We don't set the accessed bit, since we sometimes want to see
1799 * whether the guest actually used the pte (in order to detect
1802 spte = shadow_base_present_pte | shadow_dirty_mask;
1804 spte |= shadow_accessed_mask;
1806 pte_access &= ~ACC_WRITE_MASK;
1807 if (pte_access & ACC_EXEC_MASK)
1808 spte |= shadow_x_mask;
1810 spte |= shadow_nx_mask;
1811 if (pte_access & ACC_USER_MASK)
1812 spte |= shadow_user_mask;
1813 if (level > PT_PAGE_TABLE_LEVEL)
1814 spte |= PT_PAGE_SIZE_MASK;
1816 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1817 kvm_is_mmio_pfn(pfn));
1819 if (reset_host_protection)
1820 spte |= SPTE_HOST_WRITEABLE;
1822 spte |= (u64)pfn << PAGE_SHIFT;
1824 if ((pte_access & ACC_WRITE_MASK)
1825 || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1827 if (level > PT_PAGE_TABLE_LEVEL &&
1828 has_wrprotected_page(vcpu->kvm, gfn, level)) {
1830 spte = shadow_trap_nonpresent_pte;
1834 spte |= PT_WRITABLE_MASK;
1837 * Optimization: for pte sync, if spte was writable the hash
1838 * lookup is unnecessary (and expensive). Write protection
1839 * is responsibility of mmu_get_page / kvm_sync_page.
1840 * Same reasoning can be applied to dirty page accounting.
1842 if (!can_unsync && is_writable_pte(*sptep))
1845 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1846 pgprintk("%s: found shadow page for %lx, marking ro\n",
1849 pte_access &= ~ACC_WRITE_MASK;
1850 if (is_writable_pte(spte))
1851 spte &= ~PT_WRITABLE_MASK;
1855 if (pte_access & ACC_WRITE_MASK)
1856 mark_page_dirty(vcpu->kvm, gfn);
1859 __set_spte(sptep, spte);
1863 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1864 unsigned pt_access, unsigned pte_access,
1865 int user_fault, int write_fault, int dirty,
1866 int *ptwrite, int level, gfn_t gfn,
1867 pfn_t pfn, bool speculative,
1868 bool reset_host_protection)
1870 int was_rmapped = 0;
1871 int was_writable = is_writable_pte(*sptep);
1874 pgprintk("%s: spte %llx access %x write_fault %d"
1875 " user_fault %d gfn %lx\n",
1876 __func__, *sptep, pt_access,
1877 write_fault, user_fault, gfn);
1879 if (is_rmap_spte(*sptep)) {
1881 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1882 * the parent of the now unreachable PTE.
1884 if (level > PT_PAGE_TABLE_LEVEL &&
1885 !is_large_pte(*sptep)) {
1886 struct kvm_mmu_page *child;
1889 child = page_header(pte & PT64_BASE_ADDR_MASK);
1890 mmu_page_remove_parent_pte(child, sptep);
1891 } else if (pfn != spte_to_pfn(*sptep)) {
1892 pgprintk("hfn old %lx new %lx\n",
1893 spte_to_pfn(*sptep), pfn);
1894 rmap_remove(vcpu->kvm, sptep);
1899 if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1900 dirty, level, gfn, pfn, speculative, true,
1901 reset_host_protection)) {
1904 kvm_x86_ops->tlb_flush(vcpu);
1907 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1908 pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1909 is_large_pte(*sptep)? "2MB" : "4kB",
1910 *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1912 if (!was_rmapped && is_large_pte(*sptep))
1913 ++vcpu->kvm->stat.lpages;
1915 page_header_update_slot(vcpu->kvm, sptep, gfn);
1917 rmap_count = rmap_add(vcpu, sptep, gfn);
1918 kvm_release_pfn_clean(pfn);
1919 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1920 rmap_recycle(vcpu, sptep, gfn);
1923 kvm_release_pfn_dirty(pfn);
1925 kvm_release_pfn_clean(pfn);
1928 vcpu->arch.last_pte_updated = sptep;
1929 vcpu->arch.last_pte_gfn = gfn;
1933 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1937 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1938 int level, gfn_t gfn, pfn_t pfn)
1940 struct kvm_shadow_walk_iterator iterator;
1941 struct kvm_mmu_page *sp;
1945 for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1946 if (iterator.level == level) {
1947 mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1948 0, write, 1, &pt_write,
1949 level, gfn, pfn, false, true);
1950 ++vcpu->stat.pf_fixed;
1954 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1955 pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1956 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1958 1, ACC_ALL, iterator.sptep);
1960 pgprintk("nonpaging_map: ENOMEM\n");
1961 kvm_release_pfn_clean(pfn);
1965 __set_spte(iterator.sptep,
1967 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1968 | shadow_user_mask | shadow_x_mask);
1974 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1979 unsigned long mmu_seq;
1981 level = mapping_level(vcpu, gfn);
1984 * This path builds a PAE pagetable - so we can map 2mb pages at
1985 * maximum. Therefore check if the level is larger than that.
1987 if (level > PT_DIRECTORY_LEVEL)
1988 level = PT_DIRECTORY_LEVEL;
1990 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
1992 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1994 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1997 if (is_error_pfn(pfn)) {
1998 kvm_release_pfn_clean(pfn);
2002 spin_lock(&vcpu->kvm->mmu_lock);
2003 if (mmu_notifier_retry(vcpu, mmu_seq))
2005 kvm_mmu_free_some_pages(vcpu);
2006 r = __direct_map(vcpu, v, write, level, gfn, pfn);
2007 spin_unlock(&vcpu->kvm->mmu_lock);
2013 spin_unlock(&vcpu->kvm->mmu_lock);
2014 kvm_release_pfn_clean(pfn);
2019 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2022 struct kvm_mmu_page *sp;
2024 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2026 spin_lock(&vcpu->kvm->mmu_lock);
2027 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2028 hpa_t root = vcpu->arch.mmu.root_hpa;
2030 sp = page_header(root);
2032 if (!sp->root_count && sp->role.invalid)
2033 kvm_mmu_zap_page(vcpu->kvm, sp);
2034 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2035 spin_unlock(&vcpu->kvm->mmu_lock);
2038 for (i = 0; i < 4; ++i) {
2039 hpa_t root = vcpu->arch.mmu.pae_root[i];
2042 root &= PT64_BASE_ADDR_MASK;
2043 sp = page_header(root);
2045 if (!sp->root_count && sp->role.invalid)
2046 kvm_mmu_zap_page(vcpu->kvm, sp);
2048 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2050 spin_unlock(&vcpu->kvm->mmu_lock);
2051 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2054 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2058 if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2059 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2066 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2070 struct kvm_mmu_page *sp;
2074 root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2076 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2077 hpa_t root = vcpu->arch.mmu.root_hpa;
2079 ASSERT(!VALID_PAGE(root));
2082 if (mmu_check_root(vcpu, root_gfn))
2084 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2085 PT64_ROOT_LEVEL, direct,
2087 root = __pa(sp->spt);
2089 vcpu->arch.mmu.root_hpa = root;
2092 direct = !is_paging(vcpu);
2095 for (i = 0; i < 4; ++i) {
2096 hpa_t root = vcpu->arch.mmu.pae_root[i];
2098 ASSERT(!VALID_PAGE(root));
2099 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2100 pdptr = kvm_pdptr_read(vcpu, i);
2101 if (!is_present_gpte(pdptr)) {
2102 vcpu->arch.mmu.pae_root[i] = 0;
2105 root_gfn = pdptr >> PAGE_SHIFT;
2106 } else if (vcpu->arch.mmu.root_level == 0)
2108 if (mmu_check_root(vcpu, root_gfn))
2110 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2111 PT32_ROOT_LEVEL, direct,
2113 root = __pa(sp->spt);
2115 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2117 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2121 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2124 struct kvm_mmu_page *sp;
2126 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2128 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2129 hpa_t root = vcpu->arch.mmu.root_hpa;
2130 sp = page_header(root);
2131 mmu_sync_children(vcpu, sp);
2134 for (i = 0; i < 4; ++i) {
2135 hpa_t root = vcpu->arch.mmu.pae_root[i];
2137 if (root && VALID_PAGE(root)) {
2138 root &= PT64_BASE_ADDR_MASK;
2139 sp = page_header(root);
2140 mmu_sync_children(vcpu, sp);
2145 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2147 spin_lock(&vcpu->kvm->mmu_lock);
2148 mmu_sync_roots(vcpu);
2149 spin_unlock(&vcpu->kvm->mmu_lock);
2152 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2153 u32 access, u32 *error)
2160 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2166 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2167 r = mmu_topup_memory_caches(vcpu);
2172 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2174 gfn = gva >> PAGE_SHIFT;
2176 return nonpaging_map(vcpu, gva & PAGE_MASK,
2177 error_code & PFERR_WRITE_MASK, gfn);
2180 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2186 gfn_t gfn = gpa >> PAGE_SHIFT;
2187 unsigned long mmu_seq;
2190 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2192 r = mmu_topup_memory_caches(vcpu);
2196 level = mapping_level(vcpu, gfn);
2198 gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2200 mmu_seq = vcpu->kvm->mmu_notifier_seq;
2202 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2203 if (is_error_pfn(pfn)) {
2204 kvm_release_pfn_clean(pfn);
2207 spin_lock(&vcpu->kvm->mmu_lock);
2208 if (mmu_notifier_retry(vcpu, mmu_seq))
2210 kvm_mmu_free_some_pages(vcpu);
2211 r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2213 spin_unlock(&vcpu->kvm->mmu_lock);
2218 spin_unlock(&vcpu->kvm->mmu_lock);
2219 kvm_release_pfn_clean(pfn);
2223 static void nonpaging_free(struct kvm_vcpu *vcpu)
2225 mmu_free_roots(vcpu);
2228 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2230 struct kvm_mmu *context = &vcpu->arch.mmu;
2232 context->new_cr3 = nonpaging_new_cr3;
2233 context->page_fault = nonpaging_page_fault;
2234 context->gva_to_gpa = nonpaging_gva_to_gpa;
2235 context->free = nonpaging_free;
2236 context->prefetch_page = nonpaging_prefetch_page;
2237 context->sync_page = nonpaging_sync_page;
2238 context->invlpg = nonpaging_invlpg;
2239 context->root_level = 0;
2240 context->shadow_root_level = PT32E_ROOT_LEVEL;
2241 context->root_hpa = INVALID_PAGE;
2245 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2247 ++vcpu->stat.tlb_flush;
2248 kvm_x86_ops->tlb_flush(vcpu);
2251 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2253 pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2254 mmu_free_roots(vcpu);
2257 static void inject_page_fault(struct kvm_vcpu *vcpu,
2261 kvm_inject_page_fault(vcpu, addr, err_code);
2264 static void paging_free(struct kvm_vcpu *vcpu)
2266 nonpaging_free(vcpu);
2269 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2273 bit7 = (gpte >> 7) & 1;
2274 return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2278 #include "paging_tmpl.h"
2282 #include "paging_tmpl.h"
2285 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2287 struct kvm_mmu *context = &vcpu->arch.mmu;
2288 int maxphyaddr = cpuid_maxphyaddr(vcpu);
2289 u64 exb_bit_rsvd = 0;
2292 exb_bit_rsvd = rsvd_bits(63, 63);
2294 case PT32_ROOT_LEVEL:
2295 /* no rsvd bits for 2 level 4K page table entries */
2296 context->rsvd_bits_mask[0][1] = 0;
2297 context->rsvd_bits_mask[0][0] = 0;
2298 if (is_cpuid_PSE36())
2299 /* 36bits PSE 4MB page */
2300 context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2302 /* 32 bits PSE 4MB page */
2303 context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2304 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2306 case PT32E_ROOT_LEVEL:
2307 context->rsvd_bits_mask[0][2] =
2308 rsvd_bits(maxphyaddr, 63) |
2309 rsvd_bits(7, 8) | rsvd_bits(1, 2); /* PDPTE */
2310 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2311 rsvd_bits(maxphyaddr, 62); /* PDE */
2312 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2313 rsvd_bits(maxphyaddr, 62); /* PTE */
2314 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2315 rsvd_bits(maxphyaddr, 62) |
2316 rsvd_bits(13, 20); /* large page */
2317 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2319 case PT64_ROOT_LEVEL:
2320 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2321 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2322 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2323 rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2324 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2325 rsvd_bits(maxphyaddr, 51);
2326 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2327 rsvd_bits(maxphyaddr, 51);
2328 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2329 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2330 rsvd_bits(maxphyaddr, 51) |
2332 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2333 rsvd_bits(maxphyaddr, 51) |
2334 rsvd_bits(13, 20); /* large page */
2335 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[1][0];
2340 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2342 struct kvm_mmu *context = &vcpu->arch.mmu;
2344 ASSERT(is_pae(vcpu));
2345 context->new_cr3 = paging_new_cr3;
2346 context->page_fault = paging64_page_fault;
2347 context->gva_to_gpa = paging64_gva_to_gpa;
2348 context->prefetch_page = paging64_prefetch_page;
2349 context->sync_page = paging64_sync_page;
2350 context->invlpg = paging64_invlpg;
2351 context->free = paging_free;
2352 context->root_level = level;
2353 context->shadow_root_level = level;
2354 context->root_hpa = INVALID_PAGE;
2358 static int paging64_init_context(struct kvm_vcpu *vcpu)
2360 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2361 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2364 static int paging32_init_context(struct kvm_vcpu *vcpu)
2366 struct kvm_mmu *context = &vcpu->arch.mmu;
2368 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2369 context->new_cr3 = paging_new_cr3;
2370 context->page_fault = paging32_page_fault;
2371 context->gva_to_gpa = paging32_gva_to_gpa;
2372 context->free = paging_free;
2373 context->prefetch_page = paging32_prefetch_page;
2374 context->sync_page = paging32_sync_page;
2375 context->invlpg = paging32_invlpg;
2376 context->root_level = PT32_ROOT_LEVEL;
2377 context->shadow_root_level = PT32E_ROOT_LEVEL;
2378 context->root_hpa = INVALID_PAGE;
2382 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2384 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2385 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2388 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2390 struct kvm_mmu *context = &vcpu->arch.mmu;
2392 context->new_cr3 = nonpaging_new_cr3;
2393 context->page_fault = tdp_page_fault;
2394 context->free = nonpaging_free;
2395 context->prefetch_page = nonpaging_prefetch_page;
2396 context->sync_page = nonpaging_sync_page;
2397 context->invlpg = nonpaging_invlpg;
2398 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2399 context->root_hpa = INVALID_PAGE;
2401 if (!is_paging(vcpu)) {
2402 context->gva_to_gpa = nonpaging_gva_to_gpa;
2403 context->root_level = 0;
2404 } else if (is_long_mode(vcpu)) {
2405 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2406 context->gva_to_gpa = paging64_gva_to_gpa;
2407 context->root_level = PT64_ROOT_LEVEL;
2408 } else if (is_pae(vcpu)) {
2409 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2410 context->gva_to_gpa = paging64_gva_to_gpa;
2411 context->root_level = PT32E_ROOT_LEVEL;
2413 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2414 context->gva_to_gpa = paging32_gva_to_gpa;
2415 context->root_level = PT32_ROOT_LEVEL;
2421 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2426 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2428 if (!is_paging(vcpu))
2429 r = nonpaging_init_context(vcpu);
2430 else if (is_long_mode(vcpu))
2431 r = paging64_init_context(vcpu);
2432 else if (is_pae(vcpu))
2433 r = paging32E_init_context(vcpu);
2435 r = paging32_init_context(vcpu);
2437 vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2442 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2444 vcpu->arch.update_pte.pfn = bad_pfn;
2447 return init_kvm_tdp_mmu(vcpu);
2449 return init_kvm_softmmu(vcpu);
2452 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2455 if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2456 vcpu->arch.mmu.free(vcpu);
2457 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2461 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2463 destroy_kvm_mmu(vcpu);
2464 return init_kvm_mmu(vcpu);
2466 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2468 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2472 r = mmu_topup_memory_caches(vcpu);
2475 spin_lock(&vcpu->kvm->mmu_lock);
2476 kvm_mmu_free_some_pages(vcpu);
2477 r = mmu_alloc_roots(vcpu);
2478 mmu_sync_roots(vcpu);
2479 spin_unlock(&vcpu->kvm->mmu_lock);
2482 /* set_cr3() should ensure TLB has been flushed */
2483 kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2487 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2489 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2491 mmu_free_roots(vcpu);
2494 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2495 struct kvm_mmu_page *sp,
2499 struct kvm_mmu_page *child;
2502 if (is_shadow_present_pte(pte)) {
2503 if (is_last_spte(pte, sp->role.level))
2504 rmap_remove(vcpu->kvm, spte);
2506 child = page_header(pte & PT64_BASE_ADDR_MASK);
2507 mmu_page_remove_parent_pte(child, spte);
2510 __set_spte(spte, shadow_trap_nonpresent_pte);
2511 if (is_large_pte(pte))
2512 --vcpu->kvm->stat.lpages;
2515 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2516 struct kvm_mmu_page *sp,
2520 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2521 ++vcpu->kvm->stat.mmu_pde_zapped;
2525 ++vcpu->kvm->stat.mmu_pte_updated;
2526 if (sp->role.glevels == PT32_ROOT_LEVEL)
2527 paging32_update_pte(vcpu, sp, spte, new);
2529 paging64_update_pte(vcpu, sp, spte, new);
2532 static bool need_remote_flush(u64 old, u64 new)
2534 if (!is_shadow_present_pte(old))
2536 if (!is_shadow_present_pte(new))
2538 if ((old ^ new) & PT64_BASE_ADDR_MASK)
2540 old ^= PT64_NX_MASK;
2541 new ^= PT64_NX_MASK;
2542 return (old & ~new & PT64_PERM_MASK) != 0;
2545 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2547 if (need_remote_flush(old, new))
2548 kvm_flush_remote_tlbs(vcpu->kvm);
2550 kvm_mmu_flush_tlb(vcpu);
2553 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2555 u64 *spte = vcpu->arch.last_pte_updated;
2557 return !!(spte && (*spte & shadow_accessed_mask));
2560 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2561 const u8 *new, int bytes)
2568 if (bytes != 4 && bytes != 8)
2572 * Assume that the pte write on a page table of the same type
2573 * as the current vcpu paging mode. This is nearly always true
2574 * (might be false while changing modes). Note it is verified later
2578 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2579 if ((bytes == 4) && (gpa % 4 == 0)) {
2580 r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2583 memcpy((void *)&gpte + (gpa % 8), new, 4);
2584 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2585 memcpy((void *)&gpte, new, 8);
2588 if ((bytes == 4) && (gpa % 4 == 0))
2589 memcpy((void *)&gpte, new, 4);
2591 if (!is_present_gpte(gpte))
2593 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2595 vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2597 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2599 if (is_error_pfn(pfn)) {
2600 kvm_release_pfn_clean(pfn);
2603 vcpu->arch.update_pte.gfn = gfn;
2604 vcpu->arch.update_pte.pfn = pfn;
2607 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2609 u64 *spte = vcpu->arch.last_pte_updated;
2612 && vcpu->arch.last_pte_gfn == gfn
2613 && shadow_accessed_mask
2614 && !(*spte & shadow_accessed_mask)
2615 && is_shadow_present_pte(*spte))
2616 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2619 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2620 const u8 *new, int bytes,
2621 bool guest_initiated)
2623 gfn_t gfn = gpa >> PAGE_SHIFT;
2624 struct kvm_mmu_page *sp;
2625 struct hlist_node *node, *n;
2626 struct hlist_head *bucket;
2630 unsigned offset = offset_in_page(gpa);
2632 unsigned page_offset;
2633 unsigned misaligned;
2640 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2641 mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2642 spin_lock(&vcpu->kvm->mmu_lock);
2643 kvm_mmu_access_page(vcpu, gfn);
2644 kvm_mmu_free_some_pages(vcpu);
2645 ++vcpu->kvm->stat.mmu_pte_write;
2646 kvm_mmu_audit(vcpu, "pre pte write");
2647 if (guest_initiated) {
2648 if (gfn == vcpu->arch.last_pt_write_gfn
2649 && !last_updated_pte_accessed(vcpu)) {
2650 ++vcpu->arch.last_pt_write_count;
2651 if (vcpu->arch.last_pt_write_count >= 3)
2654 vcpu->arch.last_pt_write_gfn = gfn;
2655 vcpu->arch.last_pt_write_count = 1;
2656 vcpu->arch.last_pte_updated = NULL;
2659 index = kvm_page_table_hashfn(gfn);
2660 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2661 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2662 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2664 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2665 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2666 misaligned |= bytes < 4;
2667 if (misaligned || flooded) {
2669 * Misaligned accesses are too much trouble to fix
2670 * up; also, they usually indicate a page is not used
2673 * If we're seeing too many writes to a page,
2674 * it may no longer be a page table, or we may be
2675 * forking, in which case it is better to unmap the
2678 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2679 gpa, bytes, sp->role.word);
2680 if (kvm_mmu_zap_page(vcpu->kvm, sp))
2682 ++vcpu->kvm->stat.mmu_flooded;
2685 page_offset = offset;
2686 level = sp->role.level;
2688 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2689 page_offset <<= 1; /* 32->64 */
2691 * A 32-bit pde maps 4MB while the shadow pdes map
2692 * only 2MB. So we need to double the offset again
2693 * and zap two pdes instead of one.
2695 if (level == PT32_ROOT_LEVEL) {
2696 page_offset &= ~7; /* kill rounding error */
2700 quadrant = page_offset >> PAGE_SHIFT;
2701 page_offset &= ~PAGE_MASK;
2702 if (quadrant != sp->role.quadrant)
2705 spte = &sp->spt[page_offset / sizeof(*spte)];
2706 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2708 r = kvm_read_guest_atomic(vcpu->kvm,
2709 gpa & ~(u64)(pte_size - 1),
2711 new = (const void *)&gentry;
2717 mmu_pte_write_zap_pte(vcpu, sp, spte);
2719 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2720 mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2724 kvm_mmu_audit(vcpu, "post pte write");
2725 spin_unlock(&vcpu->kvm->mmu_lock);
2726 if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2727 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2728 vcpu->arch.update_pte.pfn = bad_pfn;
2732 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2740 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2742 spin_lock(&vcpu->kvm->mmu_lock);
2743 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2744 spin_unlock(&vcpu->kvm->mmu_lock);
2747 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2749 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2751 while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES &&
2752 !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2753 struct kvm_mmu_page *sp;
2755 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2756 struct kvm_mmu_page, link);
2757 kvm_mmu_zap_page(vcpu->kvm, sp);
2758 ++vcpu->kvm->stat.mmu_recycled;
2762 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2765 enum emulation_result er;
2767 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2776 r = mmu_topup_memory_caches(vcpu);
2780 er = emulate_instruction(vcpu, cr2, error_code, 0);
2785 case EMULATE_DO_MMIO:
2786 ++vcpu->stat.mmio_exits;
2789 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2790 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
2791 vcpu->run->internal.ndata = 0;
2799 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2801 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2803 vcpu->arch.mmu.invlpg(vcpu, gva);
2804 kvm_mmu_flush_tlb(vcpu);
2805 ++vcpu->stat.invlpg;
2807 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2809 void kvm_enable_tdp(void)
2813 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2815 void kvm_disable_tdp(void)
2817 tdp_enabled = false;
2819 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2821 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2823 free_page((unsigned long)vcpu->arch.mmu.pae_root);
2826 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2834 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2835 * Therefore we need to allocate shadow page tables in the first
2836 * 4GB of memory, which happens to fit the DMA32 zone.
2838 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2842 vcpu->arch.mmu.pae_root = page_address(page);
2843 for (i = 0; i < 4; ++i)
2844 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2849 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2852 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2854 return alloc_mmu_pages(vcpu);
2857 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2860 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2862 return init_kvm_mmu(vcpu);
2865 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2869 destroy_kvm_mmu(vcpu);
2870 free_mmu_pages(vcpu);
2871 mmu_free_memory_caches(vcpu);
2874 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2876 struct kvm_mmu_page *sp;
2878 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2882 if (!test_bit(slot, sp->slot_bitmap))
2886 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2888 if (pt[i] & PT_WRITABLE_MASK)
2889 pt[i] &= ~PT_WRITABLE_MASK;
2891 kvm_flush_remote_tlbs(kvm);
2894 void kvm_mmu_zap_all(struct kvm *kvm)
2896 struct kvm_mmu_page *sp, *node;
2898 spin_lock(&kvm->mmu_lock);
2899 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2900 if (kvm_mmu_zap_page(kvm, sp))
2901 node = container_of(kvm->arch.active_mmu_pages.next,
2902 struct kvm_mmu_page, link);
2903 spin_unlock(&kvm->mmu_lock);
2905 kvm_flush_remote_tlbs(kvm);
2908 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2910 struct kvm_mmu_page *page;
2912 page = container_of(kvm->arch.active_mmu_pages.prev,
2913 struct kvm_mmu_page, link);
2914 kvm_mmu_zap_page(kvm, page);
2917 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2920 struct kvm *kvm_freed = NULL;
2921 int cache_count = 0;
2923 spin_lock(&kvm_lock);
2925 list_for_each_entry(kvm, &vm_list, vm_list) {
2928 idx = srcu_read_lock(&kvm->srcu);
2929 spin_lock(&kvm->mmu_lock);
2930 npages = kvm->arch.n_alloc_mmu_pages -
2931 kvm->arch.n_free_mmu_pages;
2932 cache_count += npages;
2933 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2934 kvm_mmu_remove_one_alloc_mmu_page(kvm);
2940 spin_unlock(&kvm->mmu_lock);
2941 srcu_read_unlock(&kvm->srcu, idx);
2944 list_move_tail(&kvm_freed->vm_list, &vm_list);
2946 spin_unlock(&kvm_lock);
2951 static struct shrinker mmu_shrinker = {
2952 .shrink = mmu_shrink,
2953 .seeks = DEFAULT_SEEKS * 10,
2956 static void mmu_destroy_caches(void)
2958 if (pte_chain_cache)
2959 kmem_cache_destroy(pte_chain_cache);
2960 if (rmap_desc_cache)
2961 kmem_cache_destroy(rmap_desc_cache);
2962 if (mmu_page_header_cache)
2963 kmem_cache_destroy(mmu_page_header_cache);
2966 void kvm_mmu_module_exit(void)
2968 mmu_destroy_caches();
2969 unregister_shrinker(&mmu_shrinker);
2972 int kvm_mmu_module_init(void)
2974 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2975 sizeof(struct kvm_pte_chain),
2977 if (!pte_chain_cache)
2979 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2980 sizeof(struct kvm_rmap_desc),
2982 if (!rmap_desc_cache)
2985 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2986 sizeof(struct kvm_mmu_page),
2988 if (!mmu_page_header_cache)
2991 register_shrinker(&mmu_shrinker);
2996 mmu_destroy_caches();
3001 * Caculate mmu pages needed for kvm.
3003 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3006 unsigned int nr_mmu_pages;
3007 unsigned int nr_pages = 0;
3008 struct kvm_memslots *slots;
3010 slots = rcu_dereference(kvm->memslots);
3011 for (i = 0; i < slots->nmemslots; i++)
3012 nr_pages += slots->memslots[i].npages;
3014 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3015 nr_mmu_pages = max(nr_mmu_pages,
3016 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3018 return nr_mmu_pages;
3021 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3024 if (len > buffer->len)
3029 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3034 ret = pv_mmu_peek_buffer(buffer, len);
3039 buffer->processed += len;
3043 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3044 gpa_t addr, gpa_t value)
3049 if (!is_long_mode(vcpu) && !is_pae(vcpu))
3052 r = mmu_topup_memory_caches(vcpu);
3056 if (!emulator_write_phys(vcpu, addr, &value, bytes))
3062 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3064 kvm_set_cr3(vcpu, vcpu->arch.cr3);
3068 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3070 spin_lock(&vcpu->kvm->mmu_lock);
3071 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3072 spin_unlock(&vcpu->kvm->mmu_lock);
3076 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3077 struct kvm_pv_mmu_op_buffer *buffer)
3079 struct kvm_mmu_op_header *header;
3081 header = pv_mmu_peek_buffer(buffer, sizeof *header);
3084 switch (header->op) {
3085 case KVM_MMU_OP_WRITE_PTE: {
3086 struct kvm_mmu_op_write_pte *wpte;
3088 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3091 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3094 case KVM_MMU_OP_FLUSH_TLB: {
3095 struct kvm_mmu_op_flush_tlb *ftlb;
3097 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3100 return kvm_pv_mmu_flush_tlb(vcpu);
3102 case KVM_MMU_OP_RELEASE_PT: {
3103 struct kvm_mmu_op_release_pt *rpt;
3105 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3108 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3114 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3115 gpa_t addr, unsigned long *ret)
3118 struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3120 buffer->ptr = buffer->buf;
3121 buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3122 buffer->processed = 0;
3124 r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3128 while (buffer->len) {
3129 r = kvm_pv_mmu_op_one(vcpu, buffer);
3138 *ret = buffer->processed;
3142 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3144 struct kvm_shadow_walk_iterator iterator;
3147 spin_lock(&vcpu->kvm->mmu_lock);
3148 for_each_shadow_entry(vcpu, addr, iterator) {
3149 sptes[iterator.level-1] = *iterator.sptep;
3151 if (!is_shadow_present_pte(*iterator.sptep))
3154 spin_unlock(&vcpu->kvm->mmu_lock);
3158 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3162 static const char *audit_msg;
3164 static gva_t canonicalize(gva_t gva)
3166 #ifdef CONFIG_X86_64
3167 gva = (long long)(gva << 16) >> 16;
3173 typedef void (*inspect_spte_fn) (struct kvm *kvm, struct kvm_mmu_page *sp,
3176 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3181 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3182 u64 ent = sp->spt[i];
3184 if (is_shadow_present_pte(ent)) {
3185 if (!is_last_spte(ent, sp->role.level)) {
3186 struct kvm_mmu_page *child;
3187 child = page_header(ent & PT64_BASE_ADDR_MASK);
3188 __mmu_spte_walk(kvm, child, fn);
3190 fn(kvm, sp, &sp->spt[i]);
3195 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3198 struct kvm_mmu_page *sp;
3200 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3202 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3203 hpa_t root = vcpu->arch.mmu.root_hpa;
3204 sp = page_header(root);
3205 __mmu_spte_walk(vcpu->kvm, sp, fn);
3208 for (i = 0; i < 4; ++i) {
3209 hpa_t root = vcpu->arch.mmu.pae_root[i];
3211 if (root && VALID_PAGE(root)) {
3212 root &= PT64_BASE_ADDR_MASK;
3213 sp = page_header(root);
3214 __mmu_spte_walk(vcpu->kvm, sp, fn);
3220 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3221 gva_t va, int level)
3223 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3225 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3227 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3230 if (ent == shadow_trap_nonpresent_pte)
3233 va = canonicalize(va);
3234 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3235 audit_mappings_page(vcpu, ent, va, level - 1);
3237 gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, va, NULL);
3238 gfn_t gfn = gpa >> PAGE_SHIFT;
3239 pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3240 hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3242 if (is_error_pfn(pfn)) {
3243 kvm_release_pfn_clean(pfn);
3247 if (is_shadow_present_pte(ent)
3248 && (ent & PT64_BASE_ADDR_MASK) != hpa)
3249 printk(KERN_ERR "xx audit error: (%s) levels %d"
3250 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3251 audit_msg, vcpu->arch.mmu.root_level,
3253 is_shadow_present_pte(ent));
3254 else if (ent == shadow_notrap_nonpresent_pte
3255 && !is_error_hpa(hpa))
3256 printk(KERN_ERR "audit: (%s) notrap shadow,"
3257 " valid guest gva %lx\n", audit_msg, va);
3258 kvm_release_pfn_clean(pfn);
3264 static void audit_mappings(struct kvm_vcpu *vcpu)
3268 if (vcpu->arch.mmu.root_level == 4)
3269 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3271 for (i = 0; i < 4; ++i)
3272 if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3273 audit_mappings_page(vcpu,
3274 vcpu->arch.mmu.pae_root[i],
3279 static int count_rmaps(struct kvm_vcpu *vcpu)
3284 idx = srcu_read_lock(&kvm->srcu);
3285 slots = rcu_dereference(kvm->memslots);
3286 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3287 struct kvm_memory_slot *m = &slots->memslots[i];
3288 struct kvm_rmap_desc *d;
3290 for (j = 0; j < m->npages; ++j) {
3291 unsigned long *rmapp = &m->rmap[j];
3295 if (!(*rmapp & 1)) {
3299 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3301 for (k = 0; k < RMAP_EXT; ++k)
3310 srcu_read_unlock(&kvm->srcu, idx);
3314 void inspect_spte_has_rmap(struct kvm *kvm, struct kvm_mmu_page *sp, u64 *sptep)
3316 unsigned long *rmapp;
3317 struct kvm_mmu_page *rev_sp;
3320 if (*sptep & PT_WRITABLE_MASK) {
3321 rev_sp = page_header(__pa(sptep));
3322 gfn = rev_sp->gfns[sptep - rev_sp->spt];
3324 if (!gfn_to_memslot(kvm, gfn)) {
3325 if (!printk_ratelimit())
3327 printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3329 printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3330 audit_msg, sptep - rev_sp->spt,
3336 rmapp = gfn_to_rmap(kvm, rev_sp->gfns[sptep - rev_sp->spt],
3337 is_large_pte(*sptep));
3339 if (!printk_ratelimit())
3341 printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3349 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3351 mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3354 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3356 struct kvm_mmu_page *sp;
3359 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3362 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3365 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3368 if (!(ent & PT_PRESENT_MASK))
3370 if (!(ent & PT_WRITABLE_MASK))
3372 inspect_spte_has_rmap(vcpu->kvm, sp, &pt[i]);
3378 static void audit_rmap(struct kvm_vcpu *vcpu)
3380 check_writable_mappings_rmap(vcpu);
3384 static void audit_write_protection(struct kvm_vcpu *vcpu)
3386 struct kvm_mmu_page *sp;
3387 struct kvm_memory_slot *slot;
3388 unsigned long *rmapp;
3392 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3393 if (sp->role.direct)
3398 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3399 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3400 rmapp = &slot->rmap[gfn - slot->base_gfn];
3402 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3404 if (*spte & PT_WRITABLE_MASK)
3405 printk(KERN_ERR "%s: (%s) shadow page has "
3406 "writable mappings: gfn %lx role %x\n",
3407 __func__, audit_msg, sp->gfn,
3409 spte = rmap_next(vcpu->kvm, rmapp, spte);
3414 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3421 audit_write_protection(vcpu);
3422 if (strcmp("pre pte write", audit_msg) != 0)
3423 audit_mappings(vcpu);
3424 audit_writable_sptes_have_rmaps(vcpu);