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.
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>
32 #include <asm/cmpxchg.h>
40 static void kvm_mmu_audit(struct kvm_vcpu
*vcpu
, const char *msg
);
42 static void kvm_mmu_audit(struct kvm_vcpu
*vcpu
, const char *msg
) {}
47 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
48 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
52 #define pgprintk(x...) do { } while (0)
53 #define rmap_printk(x...) do { } while (0)
57 #if defined(MMU_DEBUG) || defined(AUDIT)
62 #define ASSERT(x) do { } while (0)
66 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
67 __FILE__, __LINE__, #x); \
71 #define PT64_PT_BITS 9
72 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
73 #define PT32_PT_BITS 10
74 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
76 #define PT_WRITABLE_SHIFT 1
78 #define PT_PRESENT_MASK (1ULL << 0)
79 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
80 #define PT_USER_MASK (1ULL << 2)
81 #define PT_PWT_MASK (1ULL << 3)
82 #define PT_PCD_MASK (1ULL << 4)
83 #define PT_ACCESSED_MASK (1ULL << 5)
84 #define PT_DIRTY_MASK (1ULL << 6)
85 #define PT_PAGE_SIZE_MASK (1ULL << 7)
86 #define PT_PAT_MASK (1ULL << 7)
87 #define PT_GLOBAL_MASK (1ULL << 8)
88 #define PT64_NX_SHIFT 63
89 #define PT64_NX_MASK (1ULL << PT64_NX_SHIFT)
91 #define PT_PAT_SHIFT 7
92 #define PT_DIR_PAT_SHIFT 12
93 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
95 #define PT32_DIR_PSE36_SIZE 4
96 #define PT32_DIR_PSE36_SHIFT 13
97 #define PT32_DIR_PSE36_MASK \
98 (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
101 #define PT_FIRST_AVAIL_BITS_SHIFT 9
102 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
104 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
106 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
108 #define PT64_LEVEL_BITS 9
110 #define PT64_LEVEL_SHIFT(level) \
111 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
113 #define PT64_LEVEL_MASK(level) \
114 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
116 #define PT64_INDEX(address, level)\
117 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
120 #define PT32_LEVEL_BITS 10
122 #define PT32_LEVEL_SHIFT(level) \
123 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
125 #define PT32_LEVEL_MASK(level) \
126 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
128 #define PT32_INDEX(address, level)\
129 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
132 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
133 #define PT64_DIR_BASE_ADDR_MASK \
134 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
136 #define PT32_BASE_ADDR_MASK PAGE_MASK
137 #define PT32_DIR_BASE_ADDR_MASK \
138 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
140 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
143 #define PFERR_PRESENT_MASK (1U << 0)
144 #define PFERR_WRITE_MASK (1U << 1)
145 #define PFERR_USER_MASK (1U << 2)
146 #define PFERR_FETCH_MASK (1U << 4)
148 #define PT64_ROOT_LEVEL 4
149 #define PT32_ROOT_LEVEL 2
150 #define PT32E_ROOT_LEVEL 3
152 #define PT_DIRECTORY_LEVEL 2
153 #define PT_PAGE_TABLE_LEVEL 1
157 #define ACC_EXEC_MASK 1
158 #define ACC_WRITE_MASK PT_WRITABLE_MASK
159 #define ACC_USER_MASK PT_USER_MASK
160 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
162 struct kvm_rmap_desc
{
163 u64
*shadow_ptes
[RMAP_EXT
];
164 struct kvm_rmap_desc
*more
;
167 static struct kmem_cache
*pte_chain_cache
;
168 static struct kmem_cache
*rmap_desc_cache
;
169 static struct kmem_cache
*mmu_page_header_cache
;
171 static u64 __read_mostly shadow_trap_nonpresent_pte
;
172 static u64 __read_mostly shadow_notrap_nonpresent_pte
;
174 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte
, u64 notrap_pte
)
176 shadow_trap_nonpresent_pte
= trap_pte
;
177 shadow_notrap_nonpresent_pte
= notrap_pte
;
179 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes
);
181 static int is_write_protection(struct kvm_vcpu
*vcpu
)
183 return vcpu
->arch
.cr0
& X86_CR0_WP
;
186 static int is_cpuid_PSE36(void)
191 static int is_nx(struct kvm_vcpu
*vcpu
)
193 return vcpu
->arch
.shadow_efer
& EFER_NX
;
196 static int is_present_pte(unsigned long pte
)
198 return pte
& PT_PRESENT_MASK
;
201 static int is_shadow_present_pte(u64 pte
)
203 pte
&= ~PT_SHADOW_IO_MARK
;
204 return pte
!= shadow_trap_nonpresent_pte
205 && pte
!= shadow_notrap_nonpresent_pte
;
208 static int is_writeble_pte(unsigned long pte
)
210 return pte
& PT_WRITABLE_MASK
;
213 static int is_dirty_pte(unsigned long pte
)
215 return pte
& PT_DIRTY_MASK
;
218 static int is_io_pte(unsigned long pte
)
220 return pte
& PT_SHADOW_IO_MARK
;
223 static int is_rmap_pte(u64 pte
)
225 return is_shadow_present_pte(pte
);
228 static gfn_t
pse36_gfn_delta(u32 gpte
)
230 int shift
= 32 - PT32_DIR_PSE36_SHIFT
- PAGE_SHIFT
;
232 return (gpte
& PT32_DIR_PSE36_MASK
) << shift
;
235 static void set_shadow_pte(u64
*sptep
, u64 spte
)
238 set_64bit((unsigned long *)sptep
, spte
);
240 set_64bit((unsigned long long *)sptep
, spte
);
244 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache
*cache
,
245 struct kmem_cache
*base_cache
, int min
)
249 if (cache
->nobjs
>= min
)
251 while (cache
->nobjs
< ARRAY_SIZE(cache
->objects
)) {
252 obj
= kmem_cache_zalloc(base_cache
, GFP_KERNEL
);
255 cache
->objects
[cache
->nobjs
++] = obj
;
260 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache
*mc
)
263 kfree(mc
->objects
[--mc
->nobjs
]);
266 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache
*cache
,
271 if (cache
->nobjs
>= min
)
273 while (cache
->nobjs
< ARRAY_SIZE(cache
->objects
)) {
274 page
= alloc_page(GFP_KERNEL
);
277 set_page_private(page
, 0);
278 cache
->objects
[cache
->nobjs
++] = page_address(page
);
283 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache
*mc
)
286 free_page((unsigned long)mc
->objects
[--mc
->nobjs
]);
289 static int mmu_topup_memory_caches(struct kvm_vcpu
*vcpu
)
293 r
= mmu_topup_memory_cache(&vcpu
->arch
.mmu_pte_chain_cache
,
297 r
= mmu_topup_memory_cache(&vcpu
->arch
.mmu_rmap_desc_cache
,
301 r
= mmu_topup_memory_cache_page(&vcpu
->arch
.mmu_page_cache
, 8);
304 r
= mmu_topup_memory_cache(&vcpu
->arch
.mmu_page_header_cache
,
305 mmu_page_header_cache
, 4);
310 static void mmu_free_memory_caches(struct kvm_vcpu
*vcpu
)
312 mmu_free_memory_cache(&vcpu
->arch
.mmu_pte_chain_cache
);
313 mmu_free_memory_cache(&vcpu
->arch
.mmu_rmap_desc_cache
);
314 mmu_free_memory_cache_page(&vcpu
->arch
.mmu_page_cache
);
315 mmu_free_memory_cache(&vcpu
->arch
.mmu_page_header_cache
);
318 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache
*mc
,
324 p
= mc
->objects
[--mc
->nobjs
];
329 static struct kvm_pte_chain
*mmu_alloc_pte_chain(struct kvm_vcpu
*vcpu
)
331 return mmu_memory_cache_alloc(&vcpu
->arch
.mmu_pte_chain_cache
,
332 sizeof(struct kvm_pte_chain
));
335 static void mmu_free_pte_chain(struct kvm_pte_chain
*pc
)
340 static struct kvm_rmap_desc
*mmu_alloc_rmap_desc(struct kvm_vcpu
*vcpu
)
342 return mmu_memory_cache_alloc(&vcpu
->arch
.mmu_rmap_desc_cache
,
343 sizeof(struct kvm_rmap_desc
));
346 static void mmu_free_rmap_desc(struct kvm_rmap_desc
*rd
)
352 * Take gfn and return the reverse mapping to it.
353 * Note: gfn must be unaliased before this function get called
356 static unsigned long *gfn_to_rmap(struct kvm
*kvm
, gfn_t gfn
)
358 struct kvm_memory_slot
*slot
;
360 slot
= gfn_to_memslot(kvm
, gfn
);
361 return &slot
->rmap
[gfn
- slot
->base_gfn
];
365 * Reverse mapping data structures:
367 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
368 * that points to page_address(page).
370 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
371 * containing more mappings.
373 static void rmap_add(struct kvm_vcpu
*vcpu
, u64
*spte
, gfn_t gfn
)
375 struct kvm_mmu_page
*sp
;
376 struct kvm_rmap_desc
*desc
;
377 unsigned long *rmapp
;
380 if (!is_rmap_pte(*spte
))
382 gfn
= unalias_gfn(vcpu
->kvm
, gfn
);
383 sp
= page_header(__pa(spte
));
384 sp
->gfns
[spte
- sp
->spt
] = gfn
;
385 rmapp
= gfn_to_rmap(vcpu
->kvm
, gfn
);
387 rmap_printk("rmap_add: %p %llx 0->1\n", spte
, *spte
);
388 *rmapp
= (unsigned long)spte
;
389 } else if (!(*rmapp
& 1)) {
390 rmap_printk("rmap_add: %p %llx 1->many\n", spte
, *spte
);
391 desc
= mmu_alloc_rmap_desc(vcpu
);
392 desc
->shadow_ptes
[0] = (u64
*)*rmapp
;
393 desc
->shadow_ptes
[1] = spte
;
394 *rmapp
= (unsigned long)desc
| 1;
396 rmap_printk("rmap_add: %p %llx many->many\n", spte
, *spte
);
397 desc
= (struct kvm_rmap_desc
*)(*rmapp
& ~1ul);
398 while (desc
->shadow_ptes
[RMAP_EXT
-1] && desc
->more
)
400 if (desc
->shadow_ptes
[RMAP_EXT
-1]) {
401 desc
->more
= mmu_alloc_rmap_desc(vcpu
);
404 for (i
= 0; desc
->shadow_ptes
[i
]; ++i
)
406 desc
->shadow_ptes
[i
] = spte
;
410 static void rmap_desc_remove_entry(unsigned long *rmapp
,
411 struct kvm_rmap_desc
*desc
,
413 struct kvm_rmap_desc
*prev_desc
)
417 for (j
= RMAP_EXT
- 1; !desc
->shadow_ptes
[j
] && j
> i
; --j
)
419 desc
->shadow_ptes
[i
] = desc
->shadow_ptes
[j
];
420 desc
->shadow_ptes
[j
] = NULL
;
423 if (!prev_desc
&& !desc
->more
)
424 *rmapp
= (unsigned long)desc
->shadow_ptes
[0];
427 prev_desc
->more
= desc
->more
;
429 *rmapp
= (unsigned long)desc
->more
| 1;
430 mmu_free_rmap_desc(desc
);
433 static void rmap_remove(struct kvm
*kvm
, u64
*spte
)
435 struct kvm_rmap_desc
*desc
;
436 struct kvm_rmap_desc
*prev_desc
;
437 struct kvm_mmu_page
*sp
;
439 unsigned long *rmapp
;
442 if (!is_rmap_pte(*spte
))
444 sp
= page_header(__pa(spte
));
445 page
= pfn_to_page((*spte
& PT64_BASE_ADDR_MASK
) >> PAGE_SHIFT
);
446 mark_page_accessed(page
);
447 if (is_writeble_pte(*spte
))
448 kvm_release_page_dirty(page
);
450 kvm_release_page_clean(page
);
451 rmapp
= gfn_to_rmap(kvm
, sp
->gfns
[spte
- sp
->spt
]);
453 printk(KERN_ERR
"rmap_remove: %p %llx 0->BUG\n", spte
, *spte
);
455 } else if (!(*rmapp
& 1)) {
456 rmap_printk("rmap_remove: %p %llx 1->0\n", spte
, *spte
);
457 if ((u64
*)*rmapp
!= spte
) {
458 printk(KERN_ERR
"rmap_remove: %p %llx 1->BUG\n",
464 rmap_printk("rmap_remove: %p %llx many->many\n", spte
, *spte
);
465 desc
= (struct kvm_rmap_desc
*)(*rmapp
& ~1ul);
468 for (i
= 0; i
< RMAP_EXT
&& desc
->shadow_ptes
[i
]; ++i
)
469 if (desc
->shadow_ptes
[i
] == spte
) {
470 rmap_desc_remove_entry(rmapp
,
482 static u64
*rmap_next(struct kvm
*kvm
, unsigned long *rmapp
, u64
*spte
)
484 struct kvm_rmap_desc
*desc
;
485 struct kvm_rmap_desc
*prev_desc
;
491 else if (!(*rmapp
& 1)) {
493 return (u64
*)*rmapp
;
496 desc
= (struct kvm_rmap_desc
*)(*rmapp
& ~1ul);
500 for (i
= 0; i
< RMAP_EXT
&& desc
->shadow_ptes
[i
]; ++i
) {
501 if (prev_spte
== spte
)
502 return desc
->shadow_ptes
[i
];
503 prev_spte
= desc
->shadow_ptes
[i
];
510 static void rmap_write_protect(struct kvm
*kvm
, u64 gfn
)
512 unsigned long *rmapp
;
514 int write_protected
= 0;
516 gfn
= unalias_gfn(kvm
, gfn
);
517 rmapp
= gfn_to_rmap(kvm
, gfn
);
519 spte
= rmap_next(kvm
, rmapp
, NULL
);
522 BUG_ON(!(*spte
& PT_PRESENT_MASK
));
523 rmap_printk("rmap_write_protect: spte %p %llx\n", spte
, *spte
);
524 if (is_writeble_pte(*spte
)) {
525 set_shadow_pte(spte
, *spte
& ~PT_WRITABLE_MASK
);
528 spte
= rmap_next(kvm
, rmapp
, spte
);
531 kvm_flush_remote_tlbs(kvm
);
535 static int is_empty_shadow_page(u64
*spt
)
540 for (pos
= spt
, end
= pos
+ PAGE_SIZE
/ sizeof(u64
); pos
!= end
; pos
++)
541 if ((*pos
& ~PT_SHADOW_IO_MARK
) != shadow_trap_nonpresent_pte
) {
542 printk(KERN_ERR
"%s: %p %llx\n", __FUNCTION__
,
550 static void kvm_mmu_free_page(struct kvm
*kvm
, struct kvm_mmu_page
*sp
)
552 ASSERT(is_empty_shadow_page(sp
->spt
));
554 __free_page(virt_to_page(sp
->spt
));
555 __free_page(virt_to_page(sp
->gfns
));
557 ++kvm
->arch
.n_free_mmu_pages
;
560 static unsigned kvm_page_table_hashfn(gfn_t gfn
)
565 static struct kvm_mmu_page
*kvm_mmu_alloc_page(struct kvm_vcpu
*vcpu
,
568 struct kvm_mmu_page
*sp
;
570 sp
= mmu_memory_cache_alloc(&vcpu
->arch
.mmu_page_header_cache
, sizeof *sp
);
571 sp
->spt
= mmu_memory_cache_alloc(&vcpu
->arch
.mmu_page_cache
, PAGE_SIZE
);
572 sp
->gfns
= mmu_memory_cache_alloc(&vcpu
->arch
.mmu_page_cache
, PAGE_SIZE
);
573 set_page_private(virt_to_page(sp
->spt
), (unsigned long)sp
);
574 list_add(&sp
->link
, &vcpu
->kvm
->arch
.active_mmu_pages
);
575 ASSERT(is_empty_shadow_page(sp
->spt
));
578 sp
->parent_pte
= parent_pte
;
579 --vcpu
->kvm
->arch
.n_free_mmu_pages
;
583 static void mmu_page_add_parent_pte(struct kvm_vcpu
*vcpu
,
584 struct kvm_mmu_page
*sp
, u64
*parent_pte
)
586 struct kvm_pte_chain
*pte_chain
;
587 struct hlist_node
*node
;
592 if (!sp
->multimapped
) {
593 u64
*old
= sp
->parent_pte
;
596 sp
->parent_pte
= parent_pte
;
600 pte_chain
= mmu_alloc_pte_chain(vcpu
);
601 INIT_HLIST_HEAD(&sp
->parent_ptes
);
602 hlist_add_head(&pte_chain
->link
, &sp
->parent_ptes
);
603 pte_chain
->parent_ptes
[0] = old
;
605 hlist_for_each_entry(pte_chain
, node
, &sp
->parent_ptes
, link
) {
606 if (pte_chain
->parent_ptes
[NR_PTE_CHAIN_ENTRIES
-1])
608 for (i
= 0; i
< NR_PTE_CHAIN_ENTRIES
; ++i
)
609 if (!pte_chain
->parent_ptes
[i
]) {
610 pte_chain
->parent_ptes
[i
] = parent_pte
;
614 pte_chain
= mmu_alloc_pte_chain(vcpu
);
616 hlist_add_head(&pte_chain
->link
, &sp
->parent_ptes
);
617 pte_chain
->parent_ptes
[0] = parent_pte
;
620 static void mmu_page_remove_parent_pte(struct kvm_mmu_page
*sp
,
623 struct kvm_pte_chain
*pte_chain
;
624 struct hlist_node
*node
;
627 if (!sp
->multimapped
) {
628 BUG_ON(sp
->parent_pte
!= parent_pte
);
629 sp
->parent_pte
= NULL
;
632 hlist_for_each_entry(pte_chain
, node
, &sp
->parent_ptes
, link
)
633 for (i
= 0; i
< NR_PTE_CHAIN_ENTRIES
; ++i
) {
634 if (!pte_chain
->parent_ptes
[i
])
636 if (pte_chain
->parent_ptes
[i
] != parent_pte
)
638 while (i
+ 1 < NR_PTE_CHAIN_ENTRIES
639 && pte_chain
->parent_ptes
[i
+ 1]) {
640 pte_chain
->parent_ptes
[i
]
641 = pte_chain
->parent_ptes
[i
+ 1];
644 pte_chain
->parent_ptes
[i
] = NULL
;
646 hlist_del(&pte_chain
->link
);
647 mmu_free_pte_chain(pte_chain
);
648 if (hlist_empty(&sp
->parent_ptes
)) {
650 sp
->parent_pte
= NULL
;
658 static struct kvm_mmu_page
*kvm_mmu_lookup_page(struct kvm
*kvm
, gfn_t gfn
)
661 struct hlist_head
*bucket
;
662 struct kvm_mmu_page
*sp
;
663 struct hlist_node
*node
;
665 pgprintk("%s: looking for gfn %lx\n", __FUNCTION__
, gfn
);
666 index
= kvm_page_table_hashfn(gfn
) % KVM_NUM_MMU_PAGES
;
667 bucket
= &kvm
->arch
.mmu_page_hash
[index
];
668 hlist_for_each_entry(sp
, node
, bucket
, hash_link
)
669 if (sp
->gfn
== gfn
&& !sp
->role
.metaphysical
) {
670 pgprintk("%s: found role %x\n",
671 __FUNCTION__
, sp
->role
.word
);
677 static struct kvm_mmu_page
*kvm_mmu_get_page(struct kvm_vcpu
*vcpu
,
685 union kvm_mmu_page_role role
;
688 struct hlist_head
*bucket
;
689 struct kvm_mmu_page
*sp
;
690 struct hlist_node
*node
;
693 role
.glevels
= vcpu
->arch
.mmu
.root_level
;
695 role
.metaphysical
= metaphysical
;
696 role
.access
= access
;
697 if (vcpu
->arch
.mmu
.root_level
<= PT32_ROOT_LEVEL
) {
698 quadrant
= gaddr
>> (PAGE_SHIFT
+ (PT64_PT_BITS
* level
));
699 quadrant
&= (1 << ((PT32_PT_BITS
- PT64_PT_BITS
) * level
)) - 1;
700 role
.quadrant
= quadrant
;
702 pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__
,
704 index
= kvm_page_table_hashfn(gfn
) % KVM_NUM_MMU_PAGES
;
705 bucket
= &vcpu
->kvm
->arch
.mmu_page_hash
[index
];
706 hlist_for_each_entry(sp
, node
, bucket
, hash_link
)
707 if (sp
->gfn
== gfn
&& sp
->role
.word
== role
.word
) {
708 mmu_page_add_parent_pte(vcpu
, sp
, parent_pte
);
709 pgprintk("%s: found\n", __FUNCTION__
);
712 ++vcpu
->kvm
->stat
.mmu_cache_miss
;
713 sp
= kvm_mmu_alloc_page(vcpu
, parent_pte
);
716 pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__
, gfn
, role
.word
);
719 hlist_add_head(&sp
->hash_link
, bucket
);
720 vcpu
->arch
.mmu
.prefetch_page(vcpu
, sp
);
722 rmap_write_protect(vcpu
->kvm
, gfn
);
726 static void kvm_mmu_page_unlink_children(struct kvm
*kvm
,
727 struct kvm_mmu_page
*sp
)
735 if (sp
->role
.level
== PT_PAGE_TABLE_LEVEL
) {
736 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
) {
737 if (is_shadow_present_pte(pt
[i
]))
738 rmap_remove(kvm
, &pt
[i
]);
739 pt
[i
] = shadow_trap_nonpresent_pte
;
741 kvm_flush_remote_tlbs(kvm
);
745 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
) {
748 pt
[i
] = shadow_trap_nonpresent_pte
;
749 if (!is_shadow_present_pte(ent
))
751 ent
&= PT64_BASE_ADDR_MASK
;
752 mmu_page_remove_parent_pte(page_header(ent
), &pt
[i
]);
754 kvm_flush_remote_tlbs(kvm
);
757 static void kvm_mmu_put_page(struct kvm_mmu_page
*sp
, u64
*parent_pte
)
759 mmu_page_remove_parent_pte(sp
, parent_pte
);
762 static void kvm_mmu_reset_last_pte_updated(struct kvm
*kvm
)
766 for (i
= 0; i
< KVM_MAX_VCPUS
; ++i
)
768 kvm
->vcpus
[i
]->arch
.last_pte_updated
= NULL
;
771 static void kvm_mmu_zap_page(struct kvm
*kvm
, struct kvm_mmu_page
*sp
)
775 ++kvm
->stat
.mmu_shadow_zapped
;
776 while (sp
->multimapped
|| sp
->parent_pte
) {
777 if (!sp
->multimapped
)
778 parent_pte
= sp
->parent_pte
;
780 struct kvm_pte_chain
*chain
;
782 chain
= container_of(sp
->parent_ptes
.first
,
783 struct kvm_pte_chain
, link
);
784 parent_pte
= chain
->parent_ptes
[0];
787 kvm_mmu_put_page(sp
, parent_pte
);
788 set_shadow_pte(parent_pte
, shadow_trap_nonpresent_pte
);
790 kvm_mmu_page_unlink_children(kvm
, sp
);
791 if (!sp
->root_count
) {
792 hlist_del(&sp
->hash_link
);
793 kvm_mmu_free_page(kvm
, sp
);
795 list_move(&sp
->link
, &kvm
->arch
.active_mmu_pages
);
796 kvm_mmu_reset_last_pte_updated(kvm
);
800 * Changing the number of mmu pages allocated to the vm
801 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
803 void kvm_mmu_change_mmu_pages(struct kvm
*kvm
, unsigned int kvm_nr_mmu_pages
)
806 * If we set the number of mmu pages to be smaller be than the
807 * number of actived pages , we must to free some mmu pages before we
811 if ((kvm
->arch
.n_alloc_mmu_pages
- kvm
->arch
.n_free_mmu_pages
) >
813 int n_used_mmu_pages
= kvm
->arch
.n_alloc_mmu_pages
814 - kvm
->arch
.n_free_mmu_pages
;
816 while (n_used_mmu_pages
> kvm_nr_mmu_pages
) {
817 struct kvm_mmu_page
*page
;
819 page
= container_of(kvm
->arch
.active_mmu_pages
.prev
,
820 struct kvm_mmu_page
, link
);
821 kvm_mmu_zap_page(kvm
, page
);
824 kvm
->arch
.n_free_mmu_pages
= 0;
827 kvm
->arch
.n_free_mmu_pages
+= kvm_nr_mmu_pages
828 - kvm
->arch
.n_alloc_mmu_pages
;
830 kvm
->arch
.n_alloc_mmu_pages
= kvm_nr_mmu_pages
;
833 static int kvm_mmu_unprotect_page(struct kvm
*kvm
, gfn_t gfn
)
836 struct hlist_head
*bucket
;
837 struct kvm_mmu_page
*sp
;
838 struct hlist_node
*node
, *n
;
841 pgprintk("%s: looking for gfn %lx\n", __FUNCTION__
, gfn
);
843 index
= kvm_page_table_hashfn(gfn
) % KVM_NUM_MMU_PAGES
;
844 bucket
= &kvm
->arch
.mmu_page_hash
[index
];
845 hlist_for_each_entry_safe(sp
, node
, n
, bucket
, hash_link
)
846 if (sp
->gfn
== gfn
&& !sp
->role
.metaphysical
) {
847 pgprintk("%s: gfn %lx role %x\n", __FUNCTION__
, gfn
,
849 kvm_mmu_zap_page(kvm
, sp
);
855 static void mmu_unshadow(struct kvm
*kvm
, gfn_t gfn
)
857 struct kvm_mmu_page
*sp
;
859 while ((sp
= kvm_mmu_lookup_page(kvm
, gfn
)) != NULL
) {
860 pgprintk("%s: zap %lx %x\n", __FUNCTION__
, gfn
, sp
->role
.word
);
861 kvm_mmu_zap_page(kvm
, sp
);
865 static void page_header_update_slot(struct kvm
*kvm
, void *pte
, gfn_t gfn
)
867 int slot
= memslot_id(kvm
, gfn_to_memslot(kvm
, gfn
));
868 struct kvm_mmu_page
*sp
= page_header(__pa(pte
));
870 __set_bit(slot
, &sp
->slot_bitmap
);
873 struct page
*gva_to_page(struct kvm_vcpu
*vcpu
, gva_t gva
)
877 gpa_t gpa
= vcpu
->arch
.mmu
.gva_to_gpa(vcpu
, gva
);
879 if (gpa
== UNMAPPED_GVA
)
882 down_read(¤t
->mm
->mmap_sem
);
883 page
= gfn_to_page(vcpu
->kvm
, gpa
>> PAGE_SHIFT
);
884 up_read(¤t
->mm
->mmap_sem
);
889 static void mmu_set_spte(struct kvm_vcpu
*vcpu
, u64
*shadow_pte
,
890 unsigned pt_access
, unsigned pte_access
,
891 int user_fault
, int write_fault
, int dirty
,
892 int *ptwrite
, gfn_t gfn
, struct page
*page
)
896 int was_writeble
= is_writeble_pte(*shadow_pte
);
897 hfn_t host_pfn
= (*shadow_pte
& PT64_BASE_ADDR_MASK
) >> PAGE_SHIFT
;
899 pgprintk("%s: spte %llx access %x write_fault %d"
900 " user_fault %d gfn %lx\n",
901 __FUNCTION__
, *shadow_pte
, pt_access
,
902 write_fault
, user_fault
, gfn
);
904 if (is_rmap_pte(*shadow_pte
)) {
905 if (host_pfn
!= page_to_pfn(page
)) {
906 pgprintk("hfn old %lx new %lx\n",
907 host_pfn
, page_to_pfn(page
));
908 rmap_remove(vcpu
->kvm
, shadow_pte
);
915 * We don't set the accessed bit, since we sometimes want to see
916 * whether the guest actually used the pte (in order to detect
919 spte
= PT_PRESENT_MASK
| PT_DIRTY_MASK
;
921 pte_access
&= ~ACC_WRITE_MASK
;
922 if (!(pte_access
& ACC_EXEC_MASK
))
923 spte
|= PT64_NX_MASK
;
925 spte
|= PT_PRESENT_MASK
;
926 if (pte_access
& ACC_USER_MASK
)
927 spte
|= PT_USER_MASK
;
929 if (is_error_page(page
)) {
930 set_shadow_pte(shadow_pte
,
931 shadow_trap_nonpresent_pte
| PT_SHADOW_IO_MARK
);
932 kvm_release_page_clean(page
);
936 spte
|= page_to_phys(page
);
938 if ((pte_access
& ACC_WRITE_MASK
)
939 || (write_fault
&& !is_write_protection(vcpu
) && !user_fault
)) {
940 struct kvm_mmu_page
*shadow
;
942 spte
|= PT_WRITABLE_MASK
;
944 mmu_unshadow(vcpu
->kvm
, gfn
);
948 shadow
= kvm_mmu_lookup_page(vcpu
->kvm
, gfn
);
950 pgprintk("%s: found shadow page for %lx, marking ro\n",
952 pte_access
&= ~ACC_WRITE_MASK
;
953 if (is_writeble_pte(spte
)) {
954 spte
&= ~PT_WRITABLE_MASK
;
955 kvm_x86_ops
->tlb_flush(vcpu
);
964 if (pte_access
& ACC_WRITE_MASK
)
965 mark_page_dirty(vcpu
->kvm
, gfn
);
967 pgprintk("%s: setting spte %llx\n", __FUNCTION__
, spte
);
968 set_shadow_pte(shadow_pte
, spte
);
969 page_header_update_slot(vcpu
->kvm
, shadow_pte
, gfn
);
971 rmap_add(vcpu
, shadow_pte
, gfn
);
972 if (!is_rmap_pte(*shadow_pte
))
973 kvm_release_page_clean(page
);
976 kvm_release_page_dirty(page
);
978 kvm_release_page_clean(page
);
980 if (!ptwrite
|| !*ptwrite
)
981 vcpu
->arch
.last_pte_updated
= shadow_pte
;
984 static void nonpaging_new_cr3(struct kvm_vcpu
*vcpu
)
988 static int __nonpaging_map(struct kvm_vcpu
*vcpu
, gva_t v
, int write
,
989 gfn_t gfn
, struct page
*page
)
991 int level
= PT32E_ROOT_LEVEL
;
992 hpa_t table_addr
= vcpu
->arch
.mmu
.root_hpa
;
996 u32 index
= PT64_INDEX(v
, level
);
999 ASSERT(VALID_PAGE(table_addr
));
1000 table
= __va(table_addr
);
1003 mmu_set_spte(vcpu
, &table
[index
], ACC_ALL
, ACC_ALL
,
1004 0, write
, 1, &pt_write
, gfn
, page
);
1005 return pt_write
|| is_io_pte(table
[index
]);
1008 if (table
[index
] == shadow_trap_nonpresent_pte
) {
1009 struct kvm_mmu_page
*new_table
;
1012 pseudo_gfn
= (v
& PT64_DIR_BASE_ADDR_MASK
)
1014 new_table
= kvm_mmu_get_page(vcpu
, pseudo_gfn
,
1016 1, ACC_ALL
, &table
[index
]);
1018 pgprintk("nonpaging_map: ENOMEM\n");
1019 kvm_release_page_clean(page
);
1023 table
[index
] = __pa(new_table
->spt
) | PT_PRESENT_MASK
1024 | PT_WRITABLE_MASK
| PT_USER_MASK
;
1026 table_addr
= table
[index
] & PT64_BASE_ADDR_MASK
;
1030 static int nonpaging_map(struct kvm_vcpu
*vcpu
, gva_t v
, int write
, gfn_t gfn
)
1036 down_read(&vcpu
->kvm
->slots_lock
);
1038 down_read(¤t
->mm
->mmap_sem
);
1039 page
= gfn_to_page(vcpu
->kvm
, gfn
);
1040 up_read(¤t
->mm
->mmap_sem
);
1042 spin_lock(&vcpu
->kvm
->mmu_lock
);
1043 kvm_mmu_free_some_pages(vcpu
);
1044 r
= __nonpaging_map(vcpu
, v
, write
, gfn
, page
);
1045 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1047 up_read(&vcpu
->kvm
->slots_lock
);
1053 static void nonpaging_prefetch_page(struct kvm_vcpu
*vcpu
,
1054 struct kvm_mmu_page
*sp
)
1058 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
)
1059 sp
->spt
[i
] = shadow_trap_nonpresent_pte
;
1062 static void mmu_free_roots(struct kvm_vcpu
*vcpu
)
1065 struct kvm_mmu_page
*sp
;
1067 if (!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
))
1069 spin_lock(&vcpu
->kvm
->mmu_lock
);
1070 #ifdef CONFIG_X86_64
1071 if (vcpu
->arch
.mmu
.shadow_root_level
== PT64_ROOT_LEVEL
) {
1072 hpa_t root
= vcpu
->arch
.mmu
.root_hpa
;
1074 sp
= page_header(root
);
1076 vcpu
->arch
.mmu
.root_hpa
= INVALID_PAGE
;
1077 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1081 for (i
= 0; i
< 4; ++i
) {
1082 hpa_t root
= vcpu
->arch
.mmu
.pae_root
[i
];
1085 root
&= PT64_BASE_ADDR_MASK
;
1086 sp
= page_header(root
);
1089 vcpu
->arch
.mmu
.pae_root
[i
] = INVALID_PAGE
;
1091 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1092 vcpu
->arch
.mmu
.root_hpa
= INVALID_PAGE
;
1095 static void mmu_alloc_roots(struct kvm_vcpu
*vcpu
)
1099 struct kvm_mmu_page
*sp
;
1101 root_gfn
= vcpu
->arch
.cr3
>> PAGE_SHIFT
;
1103 #ifdef CONFIG_X86_64
1104 if (vcpu
->arch
.mmu
.shadow_root_level
== PT64_ROOT_LEVEL
) {
1105 hpa_t root
= vcpu
->arch
.mmu
.root_hpa
;
1107 ASSERT(!VALID_PAGE(root
));
1108 sp
= kvm_mmu_get_page(vcpu
, root_gfn
, 0,
1109 PT64_ROOT_LEVEL
, 0, ACC_ALL
, NULL
);
1110 root
= __pa(sp
->spt
);
1112 vcpu
->arch
.mmu
.root_hpa
= root
;
1116 for (i
= 0; i
< 4; ++i
) {
1117 hpa_t root
= vcpu
->arch
.mmu
.pae_root
[i
];
1119 ASSERT(!VALID_PAGE(root
));
1120 if (vcpu
->arch
.mmu
.root_level
== PT32E_ROOT_LEVEL
) {
1121 if (!is_present_pte(vcpu
->arch
.pdptrs
[i
])) {
1122 vcpu
->arch
.mmu
.pae_root
[i
] = 0;
1125 root_gfn
= vcpu
->arch
.pdptrs
[i
] >> PAGE_SHIFT
;
1126 } else if (vcpu
->arch
.mmu
.root_level
== 0)
1128 sp
= kvm_mmu_get_page(vcpu
, root_gfn
, i
<< 30,
1129 PT32_ROOT_LEVEL
, !is_paging(vcpu
),
1131 root
= __pa(sp
->spt
);
1133 vcpu
->arch
.mmu
.pae_root
[i
] = root
| PT_PRESENT_MASK
;
1135 vcpu
->arch
.mmu
.root_hpa
= __pa(vcpu
->arch
.mmu
.pae_root
);
1138 static gpa_t
nonpaging_gva_to_gpa(struct kvm_vcpu
*vcpu
, gva_t vaddr
)
1143 static int nonpaging_page_fault(struct kvm_vcpu
*vcpu
, gva_t gva
,
1149 pgprintk("%s: gva %lx error %x\n", __FUNCTION__
, gva
, error_code
);
1150 r
= mmu_topup_memory_caches(vcpu
);
1155 ASSERT(VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
1157 gfn
= gva
>> PAGE_SHIFT
;
1159 return nonpaging_map(vcpu
, gva
& PAGE_MASK
,
1160 error_code
& PFERR_WRITE_MASK
, gfn
);
1163 static void nonpaging_free(struct kvm_vcpu
*vcpu
)
1165 mmu_free_roots(vcpu
);
1168 static int nonpaging_init_context(struct kvm_vcpu
*vcpu
)
1170 struct kvm_mmu
*context
= &vcpu
->arch
.mmu
;
1172 context
->new_cr3
= nonpaging_new_cr3
;
1173 context
->page_fault
= nonpaging_page_fault
;
1174 context
->gva_to_gpa
= nonpaging_gva_to_gpa
;
1175 context
->free
= nonpaging_free
;
1176 context
->prefetch_page
= nonpaging_prefetch_page
;
1177 context
->root_level
= 0;
1178 context
->shadow_root_level
= PT32E_ROOT_LEVEL
;
1179 context
->root_hpa
= INVALID_PAGE
;
1183 void kvm_mmu_flush_tlb(struct kvm_vcpu
*vcpu
)
1185 ++vcpu
->stat
.tlb_flush
;
1186 kvm_x86_ops
->tlb_flush(vcpu
);
1189 static void paging_new_cr3(struct kvm_vcpu
*vcpu
)
1191 pgprintk("%s: cr3 %lx\n", __FUNCTION__
, vcpu
->arch
.cr3
);
1192 mmu_free_roots(vcpu
);
1195 static void inject_page_fault(struct kvm_vcpu
*vcpu
,
1199 kvm_inject_page_fault(vcpu
, addr
, err_code
);
1202 static void paging_free(struct kvm_vcpu
*vcpu
)
1204 nonpaging_free(vcpu
);
1208 #include "paging_tmpl.h"
1212 #include "paging_tmpl.h"
1215 static int paging64_init_context_common(struct kvm_vcpu
*vcpu
, int level
)
1217 struct kvm_mmu
*context
= &vcpu
->arch
.mmu
;
1219 ASSERT(is_pae(vcpu
));
1220 context
->new_cr3
= paging_new_cr3
;
1221 context
->page_fault
= paging64_page_fault
;
1222 context
->gva_to_gpa
= paging64_gva_to_gpa
;
1223 context
->prefetch_page
= paging64_prefetch_page
;
1224 context
->free
= paging_free
;
1225 context
->root_level
= level
;
1226 context
->shadow_root_level
= level
;
1227 context
->root_hpa
= INVALID_PAGE
;
1231 static int paging64_init_context(struct kvm_vcpu
*vcpu
)
1233 return paging64_init_context_common(vcpu
, PT64_ROOT_LEVEL
);
1236 static int paging32_init_context(struct kvm_vcpu
*vcpu
)
1238 struct kvm_mmu
*context
= &vcpu
->arch
.mmu
;
1240 context
->new_cr3
= paging_new_cr3
;
1241 context
->page_fault
= paging32_page_fault
;
1242 context
->gva_to_gpa
= paging32_gva_to_gpa
;
1243 context
->free
= paging_free
;
1244 context
->prefetch_page
= paging32_prefetch_page
;
1245 context
->root_level
= PT32_ROOT_LEVEL
;
1246 context
->shadow_root_level
= PT32E_ROOT_LEVEL
;
1247 context
->root_hpa
= INVALID_PAGE
;
1251 static int paging32E_init_context(struct kvm_vcpu
*vcpu
)
1253 return paging64_init_context_common(vcpu
, PT32E_ROOT_LEVEL
);
1256 static int init_kvm_mmu(struct kvm_vcpu
*vcpu
)
1259 ASSERT(!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
1261 if (!is_paging(vcpu
))
1262 return nonpaging_init_context(vcpu
);
1263 else if (is_long_mode(vcpu
))
1264 return paging64_init_context(vcpu
);
1265 else if (is_pae(vcpu
))
1266 return paging32E_init_context(vcpu
);
1268 return paging32_init_context(vcpu
);
1271 static void destroy_kvm_mmu(struct kvm_vcpu
*vcpu
)
1274 if (VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
)) {
1275 vcpu
->arch
.mmu
.free(vcpu
);
1276 vcpu
->arch
.mmu
.root_hpa
= INVALID_PAGE
;
1280 int kvm_mmu_reset_context(struct kvm_vcpu
*vcpu
)
1282 destroy_kvm_mmu(vcpu
);
1283 return init_kvm_mmu(vcpu
);
1285 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context
);
1287 int kvm_mmu_load(struct kvm_vcpu
*vcpu
)
1291 r
= mmu_topup_memory_caches(vcpu
);
1294 spin_lock(&vcpu
->kvm
->mmu_lock
);
1295 kvm_mmu_free_some_pages(vcpu
);
1296 mmu_alloc_roots(vcpu
);
1297 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1298 kvm_x86_ops
->set_cr3(vcpu
, vcpu
->arch
.mmu
.root_hpa
);
1299 kvm_mmu_flush_tlb(vcpu
);
1303 EXPORT_SYMBOL_GPL(kvm_mmu_load
);
1305 void kvm_mmu_unload(struct kvm_vcpu
*vcpu
)
1307 mmu_free_roots(vcpu
);
1310 static void mmu_pte_write_zap_pte(struct kvm_vcpu
*vcpu
,
1311 struct kvm_mmu_page
*sp
,
1315 struct kvm_mmu_page
*child
;
1318 if (is_shadow_present_pte(pte
)) {
1319 if (sp
->role
.level
== PT_PAGE_TABLE_LEVEL
)
1320 rmap_remove(vcpu
->kvm
, spte
);
1322 child
= page_header(pte
& PT64_BASE_ADDR_MASK
);
1323 mmu_page_remove_parent_pte(child
, spte
);
1326 set_shadow_pte(spte
, shadow_trap_nonpresent_pte
);
1329 static void mmu_pte_write_new_pte(struct kvm_vcpu
*vcpu
,
1330 struct kvm_mmu_page
*sp
,
1332 const void *new, int bytes
,
1335 if (sp
->role
.level
!= PT_PAGE_TABLE_LEVEL
) {
1336 ++vcpu
->kvm
->stat
.mmu_pde_zapped
;
1340 ++vcpu
->kvm
->stat
.mmu_pte_updated
;
1341 if (sp
->role
.glevels
== PT32_ROOT_LEVEL
)
1342 paging32_update_pte(vcpu
, sp
, spte
, new, bytes
, offset_in_pte
);
1344 paging64_update_pte(vcpu
, sp
, spte
, new, bytes
, offset_in_pte
);
1347 static bool need_remote_flush(u64 old
, u64
new)
1349 if (!is_shadow_present_pte(old
))
1351 if (!is_shadow_present_pte(new))
1353 if ((old
^ new) & PT64_BASE_ADDR_MASK
)
1355 old
^= PT64_NX_MASK
;
1356 new ^= PT64_NX_MASK
;
1357 return (old
& ~new & PT64_PERM_MASK
) != 0;
1360 static void mmu_pte_write_flush_tlb(struct kvm_vcpu
*vcpu
, u64 old
, u64
new)
1362 if (need_remote_flush(old
, new))
1363 kvm_flush_remote_tlbs(vcpu
->kvm
);
1365 kvm_mmu_flush_tlb(vcpu
);
1368 static bool last_updated_pte_accessed(struct kvm_vcpu
*vcpu
)
1370 u64
*spte
= vcpu
->arch
.last_pte_updated
;
1372 return !!(spte
&& (*spte
& PT_ACCESSED_MASK
));
1375 static void mmu_guess_page_from_pte_write(struct kvm_vcpu
*vcpu
, gpa_t gpa
,
1376 const u8
*new, int bytes
)
1383 if (bytes
!= 4 && bytes
!= 8)
1387 * Assume that the pte write on a page table of the same type
1388 * as the current vcpu paging mode. This is nearly always true
1389 * (might be false while changing modes). Note it is verified later
1393 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1394 if ((bytes
== 4) && (gpa
% 4 == 0)) {
1395 r
= kvm_read_guest(vcpu
->kvm
, gpa
& ~(u64
)7, &gpte
, 8);
1398 memcpy((void *)&gpte
+ (gpa
% 8), new, 4);
1399 } else if ((bytes
== 8) && (gpa
% 8 == 0)) {
1400 memcpy((void *)&gpte
, new, 8);
1403 if ((bytes
== 4) && (gpa
% 4 == 0))
1404 memcpy((void *)&gpte
, new, 4);
1406 if (!is_present_pte(gpte
))
1408 gfn
= (gpte
& PT64_BASE_ADDR_MASK
) >> PAGE_SHIFT
;
1410 down_read(¤t
->mm
->mmap_sem
);
1411 page
= gfn_to_page(vcpu
->kvm
, gfn
);
1412 up_read(¤t
->mm
->mmap_sem
);
1414 vcpu
->arch
.update_pte
.gfn
= gfn
;
1415 vcpu
->arch
.update_pte
.page
= page
;
1418 void kvm_mmu_pte_write(struct kvm_vcpu
*vcpu
, gpa_t gpa
,
1419 const u8
*new, int bytes
)
1421 gfn_t gfn
= gpa
>> PAGE_SHIFT
;
1422 struct kvm_mmu_page
*sp
;
1423 struct hlist_node
*node
, *n
;
1424 struct hlist_head
*bucket
;
1428 unsigned offset
= offset_in_page(gpa
);
1430 unsigned page_offset
;
1431 unsigned misaligned
;
1437 pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__
, gpa
, bytes
);
1438 mmu_guess_page_from_pte_write(vcpu
, gpa
, new, bytes
);
1439 spin_lock(&vcpu
->kvm
->mmu_lock
);
1440 kvm_mmu_free_some_pages(vcpu
);
1441 ++vcpu
->kvm
->stat
.mmu_pte_write
;
1442 kvm_mmu_audit(vcpu
, "pre pte write");
1443 if (gfn
== vcpu
->arch
.last_pt_write_gfn
1444 && !last_updated_pte_accessed(vcpu
)) {
1445 ++vcpu
->arch
.last_pt_write_count
;
1446 if (vcpu
->arch
.last_pt_write_count
>= 3)
1449 vcpu
->arch
.last_pt_write_gfn
= gfn
;
1450 vcpu
->arch
.last_pt_write_count
= 1;
1451 vcpu
->arch
.last_pte_updated
= NULL
;
1453 index
= kvm_page_table_hashfn(gfn
) % KVM_NUM_MMU_PAGES
;
1454 bucket
= &vcpu
->kvm
->arch
.mmu_page_hash
[index
];
1455 hlist_for_each_entry_safe(sp
, node
, n
, bucket
, hash_link
) {
1456 if (sp
->gfn
!= gfn
|| sp
->role
.metaphysical
)
1458 pte_size
= sp
->role
.glevels
== PT32_ROOT_LEVEL
? 4 : 8;
1459 misaligned
= (offset
^ (offset
+ bytes
- 1)) & ~(pte_size
- 1);
1460 misaligned
|= bytes
< 4;
1461 if (misaligned
|| flooded
) {
1463 * Misaligned accesses are too much trouble to fix
1464 * up; also, they usually indicate a page is not used
1467 * If we're seeing too many writes to a page,
1468 * it may no longer be a page table, or we may be
1469 * forking, in which case it is better to unmap the
1472 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1473 gpa
, bytes
, sp
->role
.word
);
1474 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
1475 ++vcpu
->kvm
->stat
.mmu_flooded
;
1478 page_offset
= offset
;
1479 level
= sp
->role
.level
;
1481 if (sp
->role
.glevels
== PT32_ROOT_LEVEL
) {
1482 page_offset
<<= 1; /* 32->64 */
1484 * A 32-bit pde maps 4MB while the shadow pdes map
1485 * only 2MB. So we need to double the offset again
1486 * and zap two pdes instead of one.
1488 if (level
== PT32_ROOT_LEVEL
) {
1489 page_offset
&= ~7; /* kill rounding error */
1493 quadrant
= page_offset
>> PAGE_SHIFT
;
1494 page_offset
&= ~PAGE_MASK
;
1495 if (quadrant
!= sp
->role
.quadrant
)
1498 spte
= &sp
->spt
[page_offset
/ sizeof(*spte
)];
1501 mmu_pte_write_zap_pte(vcpu
, sp
, spte
);
1502 mmu_pte_write_new_pte(vcpu
, sp
, spte
, new, bytes
,
1503 page_offset
& (pte_size
- 1));
1504 mmu_pte_write_flush_tlb(vcpu
, entry
, *spte
);
1508 kvm_mmu_audit(vcpu
, "post pte write");
1509 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1510 if (vcpu
->arch
.update_pte
.page
) {
1511 kvm_release_page_clean(vcpu
->arch
.update_pte
.page
);
1512 vcpu
->arch
.update_pte
.page
= NULL
;
1516 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu
*vcpu
, gva_t gva
)
1521 down_read(&vcpu
->kvm
->slots_lock
);
1522 gpa
= vcpu
->arch
.mmu
.gva_to_gpa(vcpu
, gva
);
1523 up_read(&vcpu
->kvm
->slots_lock
);
1525 spin_lock(&vcpu
->kvm
->mmu_lock
);
1526 r
= kvm_mmu_unprotect_page(vcpu
->kvm
, gpa
>> PAGE_SHIFT
);
1527 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1531 void __kvm_mmu_free_some_pages(struct kvm_vcpu
*vcpu
)
1533 while (vcpu
->kvm
->arch
.n_free_mmu_pages
< KVM_REFILL_PAGES
) {
1534 struct kvm_mmu_page
*sp
;
1536 sp
= container_of(vcpu
->kvm
->arch
.active_mmu_pages
.prev
,
1537 struct kvm_mmu_page
, link
);
1538 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
1539 ++vcpu
->kvm
->stat
.mmu_recycled
;
1543 int kvm_mmu_page_fault(struct kvm_vcpu
*vcpu
, gva_t cr2
, u32 error_code
)
1546 enum emulation_result er
;
1548 r
= vcpu
->arch
.mmu
.page_fault(vcpu
, cr2
, error_code
);
1557 r
= mmu_topup_memory_caches(vcpu
);
1561 er
= emulate_instruction(vcpu
, vcpu
->run
, cr2
, error_code
, 0);
1566 case EMULATE_DO_MMIO
:
1567 ++vcpu
->stat
.mmio_exits
;
1570 kvm_report_emulation_failure(vcpu
, "pagetable");
1578 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault
);
1580 static void free_mmu_pages(struct kvm_vcpu
*vcpu
)
1582 struct kvm_mmu_page
*sp
;
1584 while (!list_empty(&vcpu
->kvm
->arch
.active_mmu_pages
)) {
1585 sp
= container_of(vcpu
->kvm
->arch
.active_mmu_pages
.next
,
1586 struct kvm_mmu_page
, link
);
1587 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
1589 free_page((unsigned long)vcpu
->arch
.mmu
.pae_root
);
1592 static int alloc_mmu_pages(struct kvm_vcpu
*vcpu
)
1599 if (vcpu
->kvm
->arch
.n_requested_mmu_pages
)
1600 vcpu
->kvm
->arch
.n_free_mmu_pages
=
1601 vcpu
->kvm
->arch
.n_requested_mmu_pages
;
1603 vcpu
->kvm
->arch
.n_free_mmu_pages
=
1604 vcpu
->kvm
->arch
.n_alloc_mmu_pages
;
1606 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1607 * Therefore we need to allocate shadow page tables in the first
1608 * 4GB of memory, which happens to fit the DMA32 zone.
1610 page
= alloc_page(GFP_KERNEL
| __GFP_DMA32
);
1613 vcpu
->arch
.mmu
.pae_root
= page_address(page
);
1614 for (i
= 0; i
< 4; ++i
)
1615 vcpu
->arch
.mmu
.pae_root
[i
] = INVALID_PAGE
;
1620 free_mmu_pages(vcpu
);
1624 int kvm_mmu_create(struct kvm_vcpu
*vcpu
)
1627 ASSERT(!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
1629 return alloc_mmu_pages(vcpu
);
1632 int kvm_mmu_setup(struct kvm_vcpu
*vcpu
)
1635 ASSERT(!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
1637 return init_kvm_mmu(vcpu
);
1640 void kvm_mmu_destroy(struct kvm_vcpu
*vcpu
)
1644 destroy_kvm_mmu(vcpu
);
1645 free_mmu_pages(vcpu
);
1646 mmu_free_memory_caches(vcpu
);
1649 void kvm_mmu_slot_remove_write_access(struct kvm
*kvm
, int slot
)
1651 struct kvm_mmu_page
*sp
;
1653 list_for_each_entry(sp
, &kvm
->arch
.active_mmu_pages
, link
) {
1657 if (!test_bit(slot
, &sp
->slot_bitmap
))
1661 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
)
1663 if (pt
[i
] & PT_WRITABLE_MASK
)
1664 pt
[i
] &= ~PT_WRITABLE_MASK
;
1668 void kvm_mmu_zap_all(struct kvm
*kvm
)
1670 struct kvm_mmu_page
*sp
, *node
;
1672 spin_lock(&kvm
->mmu_lock
);
1673 list_for_each_entry_safe(sp
, node
, &kvm
->arch
.active_mmu_pages
, link
)
1674 kvm_mmu_zap_page(kvm
, sp
);
1675 spin_unlock(&kvm
->mmu_lock
);
1677 kvm_flush_remote_tlbs(kvm
);
1680 void kvm_mmu_module_exit(void)
1682 if (pte_chain_cache
)
1683 kmem_cache_destroy(pte_chain_cache
);
1684 if (rmap_desc_cache
)
1685 kmem_cache_destroy(rmap_desc_cache
);
1686 if (mmu_page_header_cache
)
1687 kmem_cache_destroy(mmu_page_header_cache
);
1690 int kvm_mmu_module_init(void)
1692 pte_chain_cache
= kmem_cache_create("kvm_pte_chain",
1693 sizeof(struct kvm_pte_chain
),
1695 if (!pte_chain_cache
)
1697 rmap_desc_cache
= kmem_cache_create("kvm_rmap_desc",
1698 sizeof(struct kvm_rmap_desc
),
1700 if (!rmap_desc_cache
)
1703 mmu_page_header_cache
= kmem_cache_create("kvm_mmu_page_header",
1704 sizeof(struct kvm_mmu_page
),
1706 if (!mmu_page_header_cache
)
1712 kvm_mmu_module_exit();
1717 * Caculate mmu pages needed for kvm.
1719 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm
*kvm
)
1722 unsigned int nr_mmu_pages
;
1723 unsigned int nr_pages
= 0;
1725 for (i
= 0; i
< kvm
->nmemslots
; i
++)
1726 nr_pages
+= kvm
->memslots
[i
].npages
;
1728 nr_mmu_pages
= nr_pages
* KVM_PERMILLE_MMU_PAGES
/ 1000;
1729 nr_mmu_pages
= max(nr_mmu_pages
,
1730 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES
);
1732 return nr_mmu_pages
;
1737 static const char *audit_msg
;
1739 static gva_t
canonicalize(gva_t gva
)
1741 #ifdef CONFIG_X86_64
1742 gva
= (long long)(gva
<< 16) >> 16;
1747 static void audit_mappings_page(struct kvm_vcpu
*vcpu
, u64 page_pte
,
1748 gva_t va
, int level
)
1750 u64
*pt
= __va(page_pte
& PT64_BASE_ADDR_MASK
);
1752 gva_t va_delta
= 1ul << (PAGE_SHIFT
+ 9 * (level
- 1));
1754 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
, va
+= va_delta
) {
1757 if (ent
== shadow_trap_nonpresent_pte
)
1760 va
= canonicalize(va
);
1762 if (ent
== shadow_notrap_nonpresent_pte
)
1763 printk(KERN_ERR
"audit: (%s) nontrapping pte"
1764 " in nonleaf level: levels %d gva %lx"
1765 " level %d pte %llx\n", audit_msg
,
1766 vcpu
->arch
.mmu
.root_level
, va
, level
, ent
);
1768 audit_mappings_page(vcpu
, ent
, va
, level
- 1);
1770 gpa_t gpa
= vcpu
->arch
.mmu
.gva_to_gpa(vcpu
, va
);
1771 struct page
*page
= gpa_to_page(vcpu
, gpa
);
1772 hpa_t hpa
= page_to_phys(page
);
1774 if (is_shadow_present_pte(ent
)
1775 && (ent
& PT64_BASE_ADDR_MASK
) != hpa
)
1776 printk(KERN_ERR
"xx audit error: (%s) levels %d"
1777 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1778 audit_msg
, vcpu
->arch
.mmu
.root_level
,
1780 is_shadow_present_pte(ent
));
1781 else if (ent
== shadow_notrap_nonpresent_pte
1782 && !is_error_hpa(hpa
))
1783 printk(KERN_ERR
"audit: (%s) notrap shadow,"
1784 " valid guest gva %lx\n", audit_msg
, va
);
1785 kvm_release_page_clean(page
);
1791 static void audit_mappings(struct kvm_vcpu
*vcpu
)
1795 if (vcpu
->arch
.mmu
.root_level
== 4)
1796 audit_mappings_page(vcpu
, vcpu
->arch
.mmu
.root_hpa
, 0, 4);
1798 for (i
= 0; i
< 4; ++i
)
1799 if (vcpu
->arch
.mmu
.pae_root
[i
] & PT_PRESENT_MASK
)
1800 audit_mappings_page(vcpu
,
1801 vcpu
->arch
.mmu
.pae_root
[i
],
1806 static int count_rmaps(struct kvm_vcpu
*vcpu
)
1811 for (i
= 0; i
< KVM_MEMORY_SLOTS
; ++i
) {
1812 struct kvm_memory_slot
*m
= &vcpu
->kvm
->memslots
[i
];
1813 struct kvm_rmap_desc
*d
;
1815 for (j
= 0; j
< m
->npages
; ++j
) {
1816 unsigned long *rmapp
= &m
->rmap
[j
];
1820 if (!(*rmapp
& 1)) {
1824 d
= (struct kvm_rmap_desc
*)(*rmapp
& ~1ul);
1826 for (k
= 0; k
< RMAP_EXT
; ++k
)
1827 if (d
->shadow_ptes
[k
])
1838 static int count_writable_mappings(struct kvm_vcpu
*vcpu
)
1841 struct kvm_mmu_page
*sp
;
1844 list_for_each_entry(sp
, &vcpu
->kvm
->arch
.active_mmu_pages
, link
) {
1847 if (sp
->role
.level
!= PT_PAGE_TABLE_LEVEL
)
1850 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
) {
1853 if (!(ent
& PT_PRESENT_MASK
))
1855 if (!(ent
& PT_WRITABLE_MASK
))
1863 static void audit_rmap(struct kvm_vcpu
*vcpu
)
1865 int n_rmap
= count_rmaps(vcpu
);
1866 int n_actual
= count_writable_mappings(vcpu
);
1868 if (n_rmap
!= n_actual
)
1869 printk(KERN_ERR
"%s: (%s) rmap %d actual %d\n",
1870 __FUNCTION__
, audit_msg
, n_rmap
, n_actual
);
1873 static void audit_write_protection(struct kvm_vcpu
*vcpu
)
1875 struct kvm_mmu_page
*sp
;
1876 struct kvm_memory_slot
*slot
;
1877 unsigned long *rmapp
;
1880 list_for_each_entry(sp
, &vcpu
->kvm
->arch
.active_mmu_pages
, link
) {
1881 if (sp
->role
.metaphysical
)
1884 slot
= gfn_to_memslot(vcpu
->kvm
, sp
->gfn
);
1885 gfn
= unalias_gfn(vcpu
->kvm
, sp
->gfn
);
1886 rmapp
= &slot
->rmap
[gfn
- slot
->base_gfn
];
1888 printk(KERN_ERR
"%s: (%s) shadow page has writable"
1889 " mappings: gfn %lx role %x\n",
1890 __FUNCTION__
, audit_msg
, sp
->gfn
,
1895 static void kvm_mmu_audit(struct kvm_vcpu
*vcpu
, const char *msg
)
1902 audit_write_protection(vcpu
);
1903 audit_mappings(vcpu
);