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>
30 #include <linux/hugetlb.h>
31 #include <linux/compiler.h>
34 #include <asm/cmpxchg.h>
38 * When setting this variable to true it enables Two-Dimensional-Paging
39 * where the hardware walks 2 page tables:
40 * 1. the guest-virtual to guest-physical
41 * 2. while doing 1. it walks guest-physical to host-physical
42 * If the hardware supports that we don't need to do shadow paging.
44 bool tdp_enabled
= false;
51 static void kvm_mmu_audit(struct kvm_vcpu
*vcpu
, const char *msg
);
53 static void kvm_mmu_audit(struct kvm_vcpu
*vcpu
, const char *msg
) {}
58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
63 #define pgprintk(x...) do { } while (0)
64 #define rmap_printk(x...) do { } while (0)
68 #if defined(MMU_DEBUG) || defined(AUDIT)
70 module_param(dbg
, bool, 0644);
74 #define ASSERT(x) do { } while (0)
78 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
79 __FILE__, __LINE__, #x); \
83 #define PT_FIRST_AVAIL_BITS_SHIFT 9
84 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
86 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
88 #define PT64_LEVEL_BITS 9
90 #define PT64_LEVEL_SHIFT(level) \
91 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
93 #define PT64_LEVEL_MASK(level) \
94 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
96 #define PT64_INDEX(address, level)\
97 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
100 #define PT32_LEVEL_BITS 10
102 #define PT32_LEVEL_SHIFT(level) \
103 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
105 #define PT32_LEVEL_MASK(level) \
106 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
108 #define PT32_INDEX(address, level)\
109 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
112 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
113 #define PT64_DIR_BASE_ADDR_MASK \
114 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
116 #define PT32_BASE_ADDR_MASK PAGE_MASK
117 #define PT32_DIR_BASE_ADDR_MASK \
118 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
120 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
123 #define PFERR_PRESENT_MASK (1U << 0)
124 #define PFERR_WRITE_MASK (1U << 1)
125 #define PFERR_USER_MASK (1U << 2)
126 #define PFERR_FETCH_MASK (1U << 4)
128 #define PT_DIRECTORY_LEVEL 2
129 #define PT_PAGE_TABLE_LEVEL 1
133 #define ACC_EXEC_MASK 1
134 #define ACC_WRITE_MASK PT_WRITABLE_MASK
135 #define ACC_USER_MASK PT_USER_MASK
136 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
138 struct kvm_pv_mmu_op_buffer
{
142 char buf
[512] __aligned(sizeof(long));
145 struct kvm_rmap_desc
{
146 u64
*shadow_ptes
[RMAP_EXT
];
147 struct kvm_rmap_desc
*more
;
150 static struct kmem_cache
*pte_chain_cache
;
151 static struct kmem_cache
*rmap_desc_cache
;
152 static struct kmem_cache
*mmu_page_header_cache
;
154 static u64 __read_mostly shadow_trap_nonpresent_pte
;
155 static u64 __read_mostly shadow_notrap_nonpresent_pte
;
156 static u64 __read_mostly shadow_base_present_pte
;
157 static u64 __read_mostly shadow_nx_mask
;
158 static u64 __read_mostly shadow_x_mask
; /* mutual exclusive with nx_mask */
159 static u64 __read_mostly shadow_user_mask
;
160 static u64 __read_mostly shadow_accessed_mask
;
161 static u64 __read_mostly shadow_dirty_mask
;
163 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte
, u64 notrap_pte
)
165 shadow_trap_nonpresent_pte
= trap_pte
;
166 shadow_notrap_nonpresent_pte
= notrap_pte
;
168 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes
);
170 void kvm_mmu_set_base_ptes(u64 base_pte
)
172 shadow_base_present_pte
= base_pte
;
174 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes
);
176 void kvm_mmu_set_mask_ptes(u64 user_mask
, u64 accessed_mask
,
177 u64 dirty_mask
, u64 nx_mask
, u64 x_mask
)
179 shadow_user_mask
= user_mask
;
180 shadow_accessed_mask
= accessed_mask
;
181 shadow_dirty_mask
= dirty_mask
;
182 shadow_nx_mask
= nx_mask
;
183 shadow_x_mask
= x_mask
;
185 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes
);
187 static int is_write_protection(struct kvm_vcpu
*vcpu
)
189 return vcpu
->arch
.cr0
& X86_CR0_WP
;
192 static int is_cpuid_PSE36(void)
197 static int is_nx(struct kvm_vcpu
*vcpu
)
199 return vcpu
->arch
.shadow_efer
& EFER_NX
;
202 static int is_present_pte(unsigned long pte
)
204 return pte
& PT_PRESENT_MASK
;
207 static int is_shadow_present_pte(u64 pte
)
209 return pte
!= shadow_trap_nonpresent_pte
210 && pte
!= shadow_notrap_nonpresent_pte
;
213 static int is_large_pte(u64 pte
)
215 return pte
& PT_PAGE_SIZE_MASK
;
218 static int is_writeble_pte(unsigned long pte
)
220 return pte
& PT_WRITABLE_MASK
;
223 static int is_dirty_pte(unsigned long pte
)
225 return pte
& shadow_dirty_mask
;
228 static int is_rmap_pte(u64 pte
)
230 return is_shadow_present_pte(pte
);
233 static pfn_t
spte_to_pfn(u64 pte
)
235 return (pte
& PT64_BASE_ADDR_MASK
) >> PAGE_SHIFT
;
238 static gfn_t
pse36_gfn_delta(u32 gpte
)
240 int shift
= 32 - PT32_DIR_PSE36_SHIFT
- PAGE_SHIFT
;
242 return (gpte
& PT32_DIR_PSE36_MASK
) << shift
;
245 static void set_shadow_pte(u64
*sptep
, u64 spte
)
248 set_64bit((unsigned long *)sptep
, spte
);
250 set_64bit((unsigned long long *)sptep
, spte
);
254 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache
*cache
,
255 struct kmem_cache
*base_cache
, int min
)
259 if (cache
->nobjs
>= min
)
261 while (cache
->nobjs
< ARRAY_SIZE(cache
->objects
)) {
262 obj
= kmem_cache_zalloc(base_cache
, GFP_KERNEL
);
265 cache
->objects
[cache
->nobjs
++] = obj
;
270 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache
*mc
)
273 kfree(mc
->objects
[--mc
->nobjs
]);
276 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache
*cache
,
281 if (cache
->nobjs
>= min
)
283 while (cache
->nobjs
< ARRAY_SIZE(cache
->objects
)) {
284 page
= alloc_page(GFP_KERNEL
);
287 set_page_private(page
, 0);
288 cache
->objects
[cache
->nobjs
++] = page_address(page
);
293 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache
*mc
)
296 free_page((unsigned long)mc
->objects
[--mc
->nobjs
]);
299 static int mmu_topup_memory_caches(struct kvm_vcpu
*vcpu
)
303 r
= mmu_topup_memory_cache(&vcpu
->arch
.mmu_pte_chain_cache
,
307 r
= mmu_topup_memory_cache(&vcpu
->arch
.mmu_rmap_desc_cache
,
311 r
= mmu_topup_memory_cache_page(&vcpu
->arch
.mmu_page_cache
, 8);
314 r
= mmu_topup_memory_cache(&vcpu
->arch
.mmu_page_header_cache
,
315 mmu_page_header_cache
, 4);
320 static void mmu_free_memory_caches(struct kvm_vcpu
*vcpu
)
322 mmu_free_memory_cache(&vcpu
->arch
.mmu_pte_chain_cache
);
323 mmu_free_memory_cache(&vcpu
->arch
.mmu_rmap_desc_cache
);
324 mmu_free_memory_cache_page(&vcpu
->arch
.mmu_page_cache
);
325 mmu_free_memory_cache(&vcpu
->arch
.mmu_page_header_cache
);
328 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache
*mc
,
334 p
= mc
->objects
[--mc
->nobjs
];
339 static struct kvm_pte_chain
*mmu_alloc_pte_chain(struct kvm_vcpu
*vcpu
)
341 return mmu_memory_cache_alloc(&vcpu
->arch
.mmu_pte_chain_cache
,
342 sizeof(struct kvm_pte_chain
));
345 static void mmu_free_pte_chain(struct kvm_pte_chain
*pc
)
350 static struct kvm_rmap_desc
*mmu_alloc_rmap_desc(struct kvm_vcpu
*vcpu
)
352 return mmu_memory_cache_alloc(&vcpu
->arch
.mmu_rmap_desc_cache
,
353 sizeof(struct kvm_rmap_desc
));
356 static void mmu_free_rmap_desc(struct kvm_rmap_desc
*rd
)
362 * Return the pointer to the largepage write count for a given
363 * gfn, handling slots that are not large page aligned.
365 static int *slot_largepage_idx(gfn_t gfn
, struct kvm_memory_slot
*slot
)
369 idx
= (gfn
/ KVM_PAGES_PER_HPAGE
) -
370 (slot
->base_gfn
/ KVM_PAGES_PER_HPAGE
);
371 return &slot
->lpage_info
[idx
].write_count
;
374 static void account_shadowed(struct kvm
*kvm
, gfn_t gfn
)
378 write_count
= slot_largepage_idx(gfn
, gfn_to_memslot(kvm
, gfn
));
382 static void unaccount_shadowed(struct kvm
*kvm
, gfn_t gfn
)
386 write_count
= slot_largepage_idx(gfn
, gfn_to_memslot(kvm
, gfn
));
388 WARN_ON(*write_count
< 0);
391 static int has_wrprotected_page(struct kvm
*kvm
, gfn_t gfn
)
393 struct kvm_memory_slot
*slot
= gfn_to_memslot(kvm
, gfn
);
397 largepage_idx
= slot_largepage_idx(gfn
, slot
);
398 return *largepage_idx
;
404 static int host_largepage_backed(struct kvm
*kvm
, gfn_t gfn
)
406 struct vm_area_struct
*vma
;
409 addr
= gfn_to_hva(kvm
, gfn
);
410 if (kvm_is_error_hva(addr
))
413 vma
= find_vma(current
->mm
, addr
);
414 if (vma
&& is_vm_hugetlb_page(vma
))
420 static int is_largepage_backed(struct kvm_vcpu
*vcpu
, gfn_t large_gfn
)
422 struct kvm_memory_slot
*slot
;
424 if (has_wrprotected_page(vcpu
->kvm
, large_gfn
))
427 if (!host_largepage_backed(vcpu
->kvm
, large_gfn
))
430 slot
= gfn_to_memslot(vcpu
->kvm
, large_gfn
);
431 if (slot
&& slot
->dirty_bitmap
)
438 * Take gfn and return the reverse mapping to it.
439 * Note: gfn must be unaliased before this function get called
442 static unsigned long *gfn_to_rmap(struct kvm
*kvm
, gfn_t gfn
, int lpage
)
444 struct kvm_memory_slot
*slot
;
447 slot
= gfn_to_memslot(kvm
, gfn
);
449 return &slot
->rmap
[gfn
- slot
->base_gfn
];
451 idx
= (gfn
/ KVM_PAGES_PER_HPAGE
) -
452 (slot
->base_gfn
/ KVM_PAGES_PER_HPAGE
);
454 return &slot
->lpage_info
[idx
].rmap_pde
;
458 * Reverse mapping data structures:
460 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
461 * that points to page_address(page).
463 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
464 * containing more mappings.
466 static void rmap_add(struct kvm_vcpu
*vcpu
, u64
*spte
, gfn_t gfn
, int lpage
)
468 struct kvm_mmu_page
*sp
;
469 struct kvm_rmap_desc
*desc
;
470 unsigned long *rmapp
;
473 if (!is_rmap_pte(*spte
))
475 gfn
= unalias_gfn(vcpu
->kvm
, gfn
);
476 sp
= page_header(__pa(spte
));
477 sp
->gfns
[spte
- sp
->spt
] = gfn
;
478 rmapp
= gfn_to_rmap(vcpu
->kvm
, gfn
, lpage
);
480 rmap_printk("rmap_add: %p %llx 0->1\n", spte
, *spte
);
481 *rmapp
= (unsigned long)spte
;
482 } else if (!(*rmapp
& 1)) {
483 rmap_printk("rmap_add: %p %llx 1->many\n", spte
, *spte
);
484 desc
= mmu_alloc_rmap_desc(vcpu
);
485 desc
->shadow_ptes
[0] = (u64
*)*rmapp
;
486 desc
->shadow_ptes
[1] = spte
;
487 *rmapp
= (unsigned long)desc
| 1;
489 rmap_printk("rmap_add: %p %llx many->many\n", spte
, *spte
);
490 desc
= (struct kvm_rmap_desc
*)(*rmapp
& ~1ul);
491 while (desc
->shadow_ptes
[RMAP_EXT
-1] && desc
->more
)
493 if (desc
->shadow_ptes
[RMAP_EXT
-1]) {
494 desc
->more
= mmu_alloc_rmap_desc(vcpu
);
497 for (i
= 0; desc
->shadow_ptes
[i
]; ++i
)
499 desc
->shadow_ptes
[i
] = spte
;
503 static void rmap_desc_remove_entry(unsigned long *rmapp
,
504 struct kvm_rmap_desc
*desc
,
506 struct kvm_rmap_desc
*prev_desc
)
510 for (j
= RMAP_EXT
- 1; !desc
->shadow_ptes
[j
] && j
> i
; --j
)
512 desc
->shadow_ptes
[i
] = desc
->shadow_ptes
[j
];
513 desc
->shadow_ptes
[j
] = NULL
;
516 if (!prev_desc
&& !desc
->more
)
517 *rmapp
= (unsigned long)desc
->shadow_ptes
[0];
520 prev_desc
->more
= desc
->more
;
522 *rmapp
= (unsigned long)desc
->more
| 1;
523 mmu_free_rmap_desc(desc
);
526 static void rmap_remove(struct kvm
*kvm
, u64
*spte
)
528 struct kvm_rmap_desc
*desc
;
529 struct kvm_rmap_desc
*prev_desc
;
530 struct kvm_mmu_page
*sp
;
532 unsigned long *rmapp
;
535 if (!is_rmap_pte(*spte
))
537 sp
= page_header(__pa(spte
));
538 pfn
= spte_to_pfn(*spte
);
539 if (*spte
& shadow_accessed_mask
)
540 kvm_set_pfn_accessed(pfn
);
541 if (is_writeble_pte(*spte
))
542 kvm_release_pfn_dirty(pfn
);
544 kvm_release_pfn_clean(pfn
);
545 rmapp
= gfn_to_rmap(kvm
, sp
->gfns
[spte
- sp
->spt
], is_large_pte(*spte
));
547 printk(KERN_ERR
"rmap_remove: %p %llx 0->BUG\n", spte
, *spte
);
549 } else if (!(*rmapp
& 1)) {
550 rmap_printk("rmap_remove: %p %llx 1->0\n", spte
, *spte
);
551 if ((u64
*)*rmapp
!= spte
) {
552 printk(KERN_ERR
"rmap_remove: %p %llx 1->BUG\n",
558 rmap_printk("rmap_remove: %p %llx many->many\n", spte
, *spte
);
559 desc
= (struct kvm_rmap_desc
*)(*rmapp
& ~1ul);
562 for (i
= 0; i
< RMAP_EXT
&& desc
->shadow_ptes
[i
]; ++i
)
563 if (desc
->shadow_ptes
[i
] == spte
) {
564 rmap_desc_remove_entry(rmapp
,
576 static u64
*rmap_next(struct kvm
*kvm
, unsigned long *rmapp
, u64
*spte
)
578 struct kvm_rmap_desc
*desc
;
579 struct kvm_rmap_desc
*prev_desc
;
585 else if (!(*rmapp
& 1)) {
587 return (u64
*)*rmapp
;
590 desc
= (struct kvm_rmap_desc
*)(*rmapp
& ~1ul);
594 for (i
= 0; i
< RMAP_EXT
&& desc
->shadow_ptes
[i
]; ++i
) {
595 if (prev_spte
== spte
)
596 return desc
->shadow_ptes
[i
];
597 prev_spte
= desc
->shadow_ptes
[i
];
604 static void rmap_write_protect(struct kvm
*kvm
, u64 gfn
)
606 unsigned long *rmapp
;
608 int write_protected
= 0;
610 gfn
= unalias_gfn(kvm
, gfn
);
611 rmapp
= gfn_to_rmap(kvm
, gfn
, 0);
613 spte
= rmap_next(kvm
, rmapp
, NULL
);
616 BUG_ON(!(*spte
& PT_PRESENT_MASK
));
617 rmap_printk("rmap_write_protect: spte %p %llx\n", spte
, *spte
);
618 if (is_writeble_pte(*spte
)) {
619 set_shadow_pte(spte
, *spte
& ~PT_WRITABLE_MASK
);
622 spte
= rmap_next(kvm
, rmapp
, spte
);
624 if (write_protected
) {
627 spte
= rmap_next(kvm
, rmapp
, NULL
);
628 pfn
= spte_to_pfn(*spte
);
629 kvm_set_pfn_dirty(pfn
);
632 /* check for huge page mappings */
633 rmapp
= gfn_to_rmap(kvm
, gfn
, 1);
634 spte
= rmap_next(kvm
, rmapp
, NULL
);
637 BUG_ON(!(*spte
& PT_PRESENT_MASK
));
638 BUG_ON((*spte
& (PT_PAGE_SIZE_MASK
|PT_PRESENT_MASK
)) != (PT_PAGE_SIZE_MASK
|PT_PRESENT_MASK
));
639 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte
, *spte
, gfn
);
640 if (is_writeble_pte(*spte
)) {
641 rmap_remove(kvm
, spte
);
643 set_shadow_pte(spte
, shadow_trap_nonpresent_pte
);
647 spte
= rmap_next(kvm
, rmapp
, spte
);
651 kvm_flush_remote_tlbs(kvm
);
653 account_shadowed(kvm
, gfn
);
656 static int kvm_unmap_rmapp(struct kvm
*kvm
, unsigned long *rmapp
)
659 int need_tlb_flush
= 0;
661 while ((spte
= rmap_next(kvm
, rmapp
, NULL
))) {
662 BUG_ON(!(*spte
& PT_PRESENT_MASK
));
663 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte
, *spte
);
664 rmap_remove(kvm
, spte
);
665 set_shadow_pte(spte
, shadow_trap_nonpresent_pte
);
668 return need_tlb_flush
;
671 static int kvm_handle_hva(struct kvm
*kvm
, unsigned long hva
,
672 int (*handler
)(struct kvm
*kvm
, unsigned long *rmapp
))
678 * If mmap_sem isn't taken, we can look the memslots with only
679 * the mmu_lock by skipping over the slots with userspace_addr == 0.
681 for (i
= 0; i
< kvm
->nmemslots
; i
++) {
682 struct kvm_memory_slot
*memslot
= &kvm
->memslots
[i
];
683 unsigned long start
= memslot
->userspace_addr
;
686 /* mmu_lock protects userspace_addr */
690 end
= start
+ (memslot
->npages
<< PAGE_SHIFT
);
691 if (hva
>= start
&& hva
< end
) {
692 gfn_t gfn_offset
= (hva
- start
) >> PAGE_SHIFT
;
693 retval
|= handler(kvm
, &memslot
->rmap
[gfn_offset
]);
694 retval
|= handler(kvm
,
695 &memslot
->lpage_info
[
697 KVM_PAGES_PER_HPAGE
].rmap_pde
);
704 int kvm_unmap_hva(struct kvm
*kvm
, unsigned long hva
)
706 return kvm_handle_hva(kvm
, hva
, kvm_unmap_rmapp
);
709 static int kvm_age_rmapp(struct kvm
*kvm
, unsigned long *rmapp
)
714 spte
= rmap_next(kvm
, rmapp
, NULL
);
718 BUG_ON(!(_spte
& PT_PRESENT_MASK
));
719 _young
= _spte
& PT_ACCESSED_MASK
;
722 clear_bit(PT_ACCESSED_SHIFT
, (unsigned long *)spte
);
724 spte
= rmap_next(kvm
, rmapp
, spte
);
729 int kvm_age_hva(struct kvm
*kvm
, unsigned long hva
)
731 return kvm_handle_hva(kvm
, hva
, kvm_age_rmapp
);
735 static int is_empty_shadow_page(u64
*spt
)
740 for (pos
= spt
, end
= pos
+ PAGE_SIZE
/ sizeof(u64
); pos
!= end
; pos
++)
741 if (is_shadow_present_pte(*pos
)) {
742 printk(KERN_ERR
"%s: %p %llx\n", __func__
,
750 static void kvm_mmu_free_page(struct kvm
*kvm
, struct kvm_mmu_page
*sp
)
752 ASSERT(is_empty_shadow_page(sp
->spt
));
754 __free_page(virt_to_page(sp
->spt
));
755 __free_page(virt_to_page(sp
->gfns
));
757 ++kvm
->arch
.n_free_mmu_pages
;
760 static unsigned kvm_page_table_hashfn(gfn_t gfn
)
762 return gfn
& ((1 << KVM_MMU_HASH_SHIFT
) - 1);
765 static struct kvm_mmu_page
*kvm_mmu_alloc_page(struct kvm_vcpu
*vcpu
,
768 struct kvm_mmu_page
*sp
;
770 sp
= mmu_memory_cache_alloc(&vcpu
->arch
.mmu_page_header_cache
, sizeof *sp
);
771 sp
->spt
= mmu_memory_cache_alloc(&vcpu
->arch
.mmu_page_cache
, PAGE_SIZE
);
772 sp
->gfns
= mmu_memory_cache_alloc(&vcpu
->arch
.mmu_page_cache
, PAGE_SIZE
);
773 set_page_private(virt_to_page(sp
->spt
), (unsigned long)sp
);
774 list_add(&sp
->link
, &vcpu
->kvm
->arch
.active_mmu_pages
);
775 ASSERT(is_empty_shadow_page(sp
->spt
));
778 sp
->parent_pte
= parent_pte
;
779 --vcpu
->kvm
->arch
.n_free_mmu_pages
;
783 static void mmu_page_add_parent_pte(struct kvm_vcpu
*vcpu
,
784 struct kvm_mmu_page
*sp
, u64
*parent_pte
)
786 struct kvm_pte_chain
*pte_chain
;
787 struct hlist_node
*node
;
792 if (!sp
->multimapped
) {
793 u64
*old
= sp
->parent_pte
;
796 sp
->parent_pte
= parent_pte
;
800 pte_chain
= mmu_alloc_pte_chain(vcpu
);
801 INIT_HLIST_HEAD(&sp
->parent_ptes
);
802 hlist_add_head(&pte_chain
->link
, &sp
->parent_ptes
);
803 pte_chain
->parent_ptes
[0] = old
;
805 hlist_for_each_entry(pte_chain
, node
, &sp
->parent_ptes
, link
) {
806 if (pte_chain
->parent_ptes
[NR_PTE_CHAIN_ENTRIES
-1])
808 for (i
= 0; i
< NR_PTE_CHAIN_ENTRIES
; ++i
)
809 if (!pte_chain
->parent_ptes
[i
]) {
810 pte_chain
->parent_ptes
[i
] = parent_pte
;
814 pte_chain
= mmu_alloc_pte_chain(vcpu
);
816 hlist_add_head(&pte_chain
->link
, &sp
->parent_ptes
);
817 pte_chain
->parent_ptes
[0] = parent_pte
;
820 static void mmu_page_remove_parent_pte(struct kvm_mmu_page
*sp
,
823 struct kvm_pte_chain
*pte_chain
;
824 struct hlist_node
*node
;
827 if (!sp
->multimapped
) {
828 BUG_ON(sp
->parent_pte
!= parent_pte
);
829 sp
->parent_pte
= NULL
;
832 hlist_for_each_entry(pte_chain
, node
, &sp
->parent_ptes
, link
)
833 for (i
= 0; i
< NR_PTE_CHAIN_ENTRIES
; ++i
) {
834 if (!pte_chain
->parent_ptes
[i
])
836 if (pte_chain
->parent_ptes
[i
] != parent_pte
)
838 while (i
+ 1 < NR_PTE_CHAIN_ENTRIES
839 && pte_chain
->parent_ptes
[i
+ 1]) {
840 pte_chain
->parent_ptes
[i
]
841 = pte_chain
->parent_ptes
[i
+ 1];
844 pte_chain
->parent_ptes
[i
] = NULL
;
846 hlist_del(&pte_chain
->link
);
847 mmu_free_pte_chain(pte_chain
);
848 if (hlist_empty(&sp
->parent_ptes
)) {
850 sp
->parent_pte
= NULL
;
858 static void nonpaging_prefetch_page(struct kvm_vcpu
*vcpu
,
859 struct kvm_mmu_page
*sp
)
863 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
)
864 sp
->spt
[i
] = shadow_trap_nonpresent_pte
;
867 static struct kvm_mmu_page
*kvm_mmu_lookup_page(struct kvm
*kvm
, gfn_t gfn
)
870 struct hlist_head
*bucket
;
871 struct kvm_mmu_page
*sp
;
872 struct hlist_node
*node
;
874 pgprintk("%s: looking for gfn %lx\n", __func__
, gfn
);
875 index
= kvm_page_table_hashfn(gfn
);
876 bucket
= &kvm
->arch
.mmu_page_hash
[index
];
877 hlist_for_each_entry(sp
, node
, bucket
, hash_link
)
878 if (sp
->gfn
== gfn
&& !sp
->role
.metaphysical
879 && !sp
->role
.invalid
) {
880 pgprintk("%s: found role %x\n",
881 __func__
, sp
->role
.word
);
887 static struct kvm_mmu_page
*kvm_mmu_get_page(struct kvm_vcpu
*vcpu
,
895 union kvm_mmu_page_role role
;
898 struct hlist_head
*bucket
;
899 struct kvm_mmu_page
*sp
;
900 struct hlist_node
*node
;
903 role
.glevels
= vcpu
->arch
.mmu
.root_level
;
905 role
.metaphysical
= metaphysical
;
906 role
.access
= access
;
907 if (vcpu
->arch
.mmu
.root_level
<= PT32_ROOT_LEVEL
) {
908 quadrant
= gaddr
>> (PAGE_SHIFT
+ (PT64_PT_BITS
* level
));
909 quadrant
&= (1 << ((PT32_PT_BITS
- PT64_PT_BITS
) * level
)) - 1;
910 role
.quadrant
= quadrant
;
912 pgprintk("%s: looking gfn %lx role %x\n", __func__
,
914 index
= kvm_page_table_hashfn(gfn
);
915 bucket
= &vcpu
->kvm
->arch
.mmu_page_hash
[index
];
916 hlist_for_each_entry(sp
, node
, bucket
, hash_link
)
917 if (sp
->gfn
== gfn
&& sp
->role
.word
== role
.word
) {
918 mmu_page_add_parent_pte(vcpu
, sp
, parent_pte
);
919 pgprintk("%s: found\n", __func__
);
922 ++vcpu
->kvm
->stat
.mmu_cache_miss
;
923 sp
= kvm_mmu_alloc_page(vcpu
, parent_pte
);
926 pgprintk("%s: adding gfn %lx role %x\n", __func__
, gfn
, role
.word
);
929 hlist_add_head(&sp
->hash_link
, bucket
);
931 rmap_write_protect(vcpu
->kvm
, gfn
);
932 if (shadow_trap_nonpresent_pte
!= shadow_notrap_nonpresent_pte
)
933 vcpu
->arch
.mmu
.prefetch_page(vcpu
, sp
);
935 nonpaging_prefetch_page(vcpu
, sp
);
939 static void kvm_mmu_page_unlink_children(struct kvm
*kvm
,
940 struct kvm_mmu_page
*sp
)
948 if (sp
->role
.level
== PT_PAGE_TABLE_LEVEL
) {
949 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
) {
950 if (is_shadow_present_pte(pt
[i
]))
951 rmap_remove(kvm
, &pt
[i
]);
952 pt
[i
] = shadow_trap_nonpresent_pte
;
954 kvm_flush_remote_tlbs(kvm
);
958 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
) {
961 if (is_shadow_present_pte(ent
)) {
962 if (!is_large_pte(ent
)) {
963 ent
&= PT64_BASE_ADDR_MASK
;
964 mmu_page_remove_parent_pte(page_header(ent
),
968 rmap_remove(kvm
, &pt
[i
]);
971 pt
[i
] = shadow_trap_nonpresent_pte
;
973 kvm_flush_remote_tlbs(kvm
);
976 static void kvm_mmu_put_page(struct kvm_mmu_page
*sp
, u64
*parent_pte
)
978 mmu_page_remove_parent_pte(sp
, parent_pte
);
981 static void kvm_mmu_reset_last_pte_updated(struct kvm
*kvm
)
985 for (i
= 0; i
< KVM_MAX_VCPUS
; ++i
)
987 kvm
->vcpus
[i
]->arch
.last_pte_updated
= NULL
;
990 static void kvm_mmu_zap_page(struct kvm
*kvm
, struct kvm_mmu_page
*sp
)
994 ++kvm
->stat
.mmu_shadow_zapped
;
995 while (sp
->multimapped
|| sp
->parent_pte
) {
996 if (!sp
->multimapped
)
997 parent_pte
= sp
->parent_pte
;
999 struct kvm_pte_chain
*chain
;
1001 chain
= container_of(sp
->parent_ptes
.first
,
1002 struct kvm_pte_chain
, link
);
1003 parent_pte
= chain
->parent_ptes
[0];
1005 BUG_ON(!parent_pte
);
1006 kvm_mmu_put_page(sp
, parent_pte
);
1007 set_shadow_pte(parent_pte
, shadow_trap_nonpresent_pte
);
1009 kvm_mmu_page_unlink_children(kvm
, sp
);
1010 if (!sp
->root_count
) {
1011 if (!sp
->role
.metaphysical
&& !sp
->role
.invalid
)
1012 unaccount_shadowed(kvm
, sp
->gfn
);
1013 hlist_del(&sp
->hash_link
);
1014 kvm_mmu_free_page(kvm
, sp
);
1016 int invalid
= sp
->role
.invalid
;
1017 list_move(&sp
->link
, &kvm
->arch
.active_mmu_pages
);
1018 sp
->role
.invalid
= 1;
1019 kvm_reload_remote_mmus(kvm
);
1020 if (!sp
->role
.metaphysical
&& !invalid
)
1021 unaccount_shadowed(kvm
, sp
->gfn
);
1023 kvm_mmu_reset_last_pte_updated(kvm
);
1027 * Changing the number of mmu pages allocated to the vm
1028 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1030 void kvm_mmu_change_mmu_pages(struct kvm
*kvm
, unsigned int kvm_nr_mmu_pages
)
1033 * If we set the number of mmu pages to be smaller be than the
1034 * number of actived pages , we must to free some mmu pages before we
1038 if ((kvm
->arch
.n_alloc_mmu_pages
- kvm
->arch
.n_free_mmu_pages
) >
1040 int n_used_mmu_pages
= kvm
->arch
.n_alloc_mmu_pages
1041 - kvm
->arch
.n_free_mmu_pages
;
1043 while (n_used_mmu_pages
> kvm_nr_mmu_pages
) {
1044 struct kvm_mmu_page
*page
;
1046 page
= container_of(kvm
->arch
.active_mmu_pages
.prev
,
1047 struct kvm_mmu_page
, link
);
1048 kvm_mmu_zap_page(kvm
, page
);
1051 kvm
->arch
.n_free_mmu_pages
= 0;
1054 kvm
->arch
.n_free_mmu_pages
+= kvm_nr_mmu_pages
1055 - kvm
->arch
.n_alloc_mmu_pages
;
1057 kvm
->arch
.n_alloc_mmu_pages
= kvm_nr_mmu_pages
;
1060 static int kvm_mmu_unprotect_page(struct kvm
*kvm
, gfn_t gfn
)
1063 struct hlist_head
*bucket
;
1064 struct kvm_mmu_page
*sp
;
1065 struct hlist_node
*node
, *n
;
1068 pgprintk("%s: looking for gfn %lx\n", __func__
, gfn
);
1070 index
= kvm_page_table_hashfn(gfn
);
1071 bucket
= &kvm
->arch
.mmu_page_hash
[index
];
1072 hlist_for_each_entry_safe(sp
, node
, n
, bucket
, hash_link
)
1073 if (sp
->gfn
== gfn
&& !sp
->role
.metaphysical
) {
1074 pgprintk("%s: gfn %lx role %x\n", __func__
, gfn
,
1076 kvm_mmu_zap_page(kvm
, sp
);
1082 static void mmu_unshadow(struct kvm
*kvm
, gfn_t gfn
)
1084 struct kvm_mmu_page
*sp
;
1086 while ((sp
= kvm_mmu_lookup_page(kvm
, gfn
)) != NULL
) {
1087 pgprintk("%s: zap %lx %x\n", __func__
, gfn
, sp
->role
.word
);
1088 kvm_mmu_zap_page(kvm
, sp
);
1092 static void page_header_update_slot(struct kvm
*kvm
, void *pte
, gfn_t gfn
)
1094 int slot
= memslot_id(kvm
, gfn_to_memslot(kvm
, gfn
));
1095 struct kvm_mmu_page
*sp
= page_header(__pa(pte
));
1097 __set_bit(slot
, &sp
->slot_bitmap
);
1100 struct page
*gva_to_page(struct kvm_vcpu
*vcpu
, gva_t gva
)
1104 gpa_t gpa
= vcpu
->arch
.mmu
.gva_to_gpa(vcpu
, gva
);
1106 if (gpa
== UNMAPPED_GVA
)
1109 down_read(¤t
->mm
->mmap_sem
);
1110 page
= gfn_to_page(vcpu
->kvm
, gpa
>> PAGE_SHIFT
);
1111 up_read(¤t
->mm
->mmap_sem
);
1116 static void mmu_set_spte(struct kvm_vcpu
*vcpu
, u64
*shadow_pte
,
1117 unsigned pt_access
, unsigned pte_access
,
1118 int user_fault
, int write_fault
, int dirty
,
1119 int *ptwrite
, int largepage
, gfn_t gfn
,
1120 pfn_t pfn
, bool speculative
)
1123 int was_rmapped
= 0;
1124 int was_writeble
= is_writeble_pte(*shadow_pte
);
1126 pgprintk("%s: spte %llx access %x write_fault %d"
1127 " user_fault %d gfn %lx\n",
1128 __func__
, *shadow_pte
, pt_access
,
1129 write_fault
, user_fault
, gfn
);
1131 if (is_rmap_pte(*shadow_pte
)) {
1133 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1134 * the parent of the now unreachable PTE.
1136 if (largepage
&& !is_large_pte(*shadow_pte
)) {
1137 struct kvm_mmu_page
*child
;
1138 u64 pte
= *shadow_pte
;
1140 child
= page_header(pte
& PT64_BASE_ADDR_MASK
);
1141 mmu_page_remove_parent_pte(child
, shadow_pte
);
1142 } else if (pfn
!= spte_to_pfn(*shadow_pte
)) {
1143 pgprintk("hfn old %lx new %lx\n",
1144 spte_to_pfn(*shadow_pte
), pfn
);
1145 rmap_remove(vcpu
->kvm
, shadow_pte
);
1148 was_rmapped
= is_large_pte(*shadow_pte
);
1155 * We don't set the accessed bit, since we sometimes want to see
1156 * whether the guest actually used the pte (in order to detect
1159 spte
= shadow_base_present_pte
| shadow_dirty_mask
;
1161 pte_access
|= PT_ACCESSED_MASK
;
1163 pte_access
&= ~ACC_WRITE_MASK
;
1164 if (pte_access
& ACC_EXEC_MASK
)
1165 spte
|= shadow_x_mask
;
1167 spte
|= shadow_nx_mask
;
1168 if (pte_access
& ACC_USER_MASK
)
1169 spte
|= shadow_user_mask
;
1171 spte
|= PT_PAGE_SIZE_MASK
;
1173 spte
|= (u64
)pfn
<< PAGE_SHIFT
;
1175 if ((pte_access
& ACC_WRITE_MASK
)
1176 || (write_fault
&& !is_write_protection(vcpu
) && !user_fault
)) {
1177 struct kvm_mmu_page
*shadow
;
1179 spte
|= PT_WRITABLE_MASK
;
1181 shadow
= kvm_mmu_lookup_page(vcpu
->kvm
, gfn
);
1183 (largepage
&& has_wrprotected_page(vcpu
->kvm
, gfn
))) {
1184 pgprintk("%s: found shadow page for %lx, marking ro\n",
1186 pte_access
&= ~ACC_WRITE_MASK
;
1187 if (is_writeble_pte(spte
)) {
1188 spte
&= ~PT_WRITABLE_MASK
;
1189 kvm_x86_ops
->tlb_flush(vcpu
);
1196 if (pte_access
& ACC_WRITE_MASK
)
1197 mark_page_dirty(vcpu
->kvm
, gfn
);
1199 pgprintk("%s: setting spte %llx\n", __func__
, spte
);
1200 pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1201 (spte
&PT_PAGE_SIZE_MASK
)? "2MB" : "4kB",
1202 (spte
&PT_WRITABLE_MASK
)?"RW":"R", gfn
, spte
, shadow_pte
);
1203 set_shadow_pte(shadow_pte
, spte
);
1204 if (!was_rmapped
&& (spte
& PT_PAGE_SIZE_MASK
)
1205 && (spte
& PT_PRESENT_MASK
))
1206 ++vcpu
->kvm
->stat
.lpages
;
1208 page_header_update_slot(vcpu
->kvm
, shadow_pte
, gfn
);
1210 rmap_add(vcpu
, shadow_pte
, gfn
, largepage
);
1211 if (!is_rmap_pte(*shadow_pte
))
1212 kvm_release_pfn_clean(pfn
);
1215 kvm_release_pfn_dirty(pfn
);
1217 kvm_release_pfn_clean(pfn
);
1220 vcpu
->arch
.last_pte_updated
= shadow_pte
;
1221 vcpu
->arch
.last_pte_gfn
= gfn
;
1225 static void nonpaging_new_cr3(struct kvm_vcpu
*vcpu
)
1229 static int __direct_map(struct kvm_vcpu
*vcpu
, gpa_t v
, int write
,
1230 int largepage
, gfn_t gfn
, pfn_t pfn
,
1233 hpa_t table_addr
= vcpu
->arch
.mmu
.root_hpa
;
1237 u32 index
= PT64_INDEX(v
, level
);
1240 ASSERT(VALID_PAGE(table_addr
));
1241 table
= __va(table_addr
);
1244 mmu_set_spte(vcpu
, &table
[index
], ACC_ALL
, ACC_ALL
,
1245 0, write
, 1, &pt_write
, 0, gfn
, pfn
, false);
1249 if (largepage
&& level
== 2) {
1250 mmu_set_spte(vcpu
, &table
[index
], ACC_ALL
, ACC_ALL
,
1251 0, write
, 1, &pt_write
, 1, gfn
, pfn
, false);
1255 if (table
[index
] == shadow_trap_nonpresent_pte
) {
1256 struct kvm_mmu_page
*new_table
;
1259 pseudo_gfn
= (v
& PT64_DIR_BASE_ADDR_MASK
)
1261 new_table
= kvm_mmu_get_page(vcpu
, pseudo_gfn
,
1263 1, ACC_ALL
, &table
[index
]);
1265 pgprintk("nonpaging_map: ENOMEM\n");
1266 kvm_release_pfn_clean(pfn
);
1270 set_shadow_pte(&table
[index
],
1271 __pa(new_table
->spt
)
1272 | PT_PRESENT_MASK
| PT_WRITABLE_MASK
1273 | shadow_user_mask
| shadow_x_mask
);
1275 table_addr
= table
[index
] & PT64_BASE_ADDR_MASK
;
1279 static int nonpaging_map(struct kvm_vcpu
*vcpu
, gva_t v
, int write
, gfn_t gfn
)
1284 unsigned long mmu_seq
;
1286 down_read(¤t
->mm
->mmap_sem
);
1287 if (is_largepage_backed(vcpu
, gfn
& ~(KVM_PAGES_PER_HPAGE
-1))) {
1288 gfn
&= ~(KVM_PAGES_PER_HPAGE
-1);
1292 mmu_seq
= vcpu
->kvm
->mmu_notifier_seq
;
1293 /* implicit mb(), we'll read before PT lock is unlocked */
1294 pfn
= gfn_to_pfn(vcpu
->kvm
, gfn
);
1295 up_read(¤t
->mm
->mmap_sem
);
1298 if (is_error_pfn(pfn
)) {
1299 kvm_release_pfn_clean(pfn
);
1303 spin_lock(&vcpu
->kvm
->mmu_lock
);
1304 if (mmu_notifier_retry(vcpu
, mmu_seq
))
1306 kvm_mmu_free_some_pages(vcpu
);
1307 r
= __direct_map(vcpu
, v
, write
, largepage
, gfn
, pfn
,
1309 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1315 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1316 kvm_release_pfn_clean(pfn
);
1321 static void mmu_free_roots(struct kvm_vcpu
*vcpu
)
1324 struct kvm_mmu_page
*sp
;
1326 if (!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
))
1328 spin_lock(&vcpu
->kvm
->mmu_lock
);
1329 if (vcpu
->arch
.mmu
.shadow_root_level
== PT64_ROOT_LEVEL
) {
1330 hpa_t root
= vcpu
->arch
.mmu
.root_hpa
;
1332 sp
= page_header(root
);
1334 if (!sp
->root_count
&& sp
->role
.invalid
)
1335 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
1336 vcpu
->arch
.mmu
.root_hpa
= INVALID_PAGE
;
1337 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1340 for (i
= 0; i
< 4; ++i
) {
1341 hpa_t root
= vcpu
->arch
.mmu
.pae_root
[i
];
1344 root
&= PT64_BASE_ADDR_MASK
;
1345 sp
= page_header(root
);
1347 if (!sp
->root_count
&& sp
->role
.invalid
)
1348 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
1350 vcpu
->arch
.mmu
.pae_root
[i
] = INVALID_PAGE
;
1352 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1353 vcpu
->arch
.mmu
.root_hpa
= INVALID_PAGE
;
1356 static void mmu_alloc_roots(struct kvm_vcpu
*vcpu
)
1360 struct kvm_mmu_page
*sp
;
1361 int metaphysical
= 0;
1363 root_gfn
= vcpu
->arch
.cr3
>> PAGE_SHIFT
;
1365 if (vcpu
->arch
.mmu
.shadow_root_level
== PT64_ROOT_LEVEL
) {
1366 hpa_t root
= vcpu
->arch
.mmu
.root_hpa
;
1368 ASSERT(!VALID_PAGE(root
));
1371 sp
= kvm_mmu_get_page(vcpu
, root_gfn
, 0,
1372 PT64_ROOT_LEVEL
, metaphysical
,
1374 root
= __pa(sp
->spt
);
1376 vcpu
->arch
.mmu
.root_hpa
= root
;
1379 metaphysical
= !is_paging(vcpu
);
1382 for (i
= 0; i
< 4; ++i
) {
1383 hpa_t root
= vcpu
->arch
.mmu
.pae_root
[i
];
1385 ASSERT(!VALID_PAGE(root
));
1386 if (vcpu
->arch
.mmu
.root_level
== PT32E_ROOT_LEVEL
) {
1387 if (!is_present_pte(vcpu
->arch
.pdptrs
[i
])) {
1388 vcpu
->arch
.mmu
.pae_root
[i
] = 0;
1391 root_gfn
= vcpu
->arch
.pdptrs
[i
] >> PAGE_SHIFT
;
1392 } else if (vcpu
->arch
.mmu
.root_level
== 0)
1394 sp
= kvm_mmu_get_page(vcpu
, root_gfn
, i
<< 30,
1395 PT32_ROOT_LEVEL
, metaphysical
,
1397 root
= __pa(sp
->spt
);
1399 vcpu
->arch
.mmu
.pae_root
[i
] = root
| PT_PRESENT_MASK
;
1401 vcpu
->arch
.mmu
.root_hpa
= __pa(vcpu
->arch
.mmu
.pae_root
);
1404 static gpa_t
nonpaging_gva_to_gpa(struct kvm_vcpu
*vcpu
, gva_t vaddr
)
1409 static int nonpaging_page_fault(struct kvm_vcpu
*vcpu
, gva_t gva
,
1415 pgprintk("%s: gva %lx error %x\n", __func__
, gva
, error_code
);
1416 r
= mmu_topup_memory_caches(vcpu
);
1421 ASSERT(VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
1423 gfn
= gva
>> PAGE_SHIFT
;
1425 return nonpaging_map(vcpu
, gva
& PAGE_MASK
,
1426 error_code
& PFERR_WRITE_MASK
, gfn
);
1429 static int tdp_page_fault(struct kvm_vcpu
*vcpu
, gva_t gpa
,
1435 gfn_t gfn
= gpa
>> PAGE_SHIFT
;
1436 unsigned long mmu_seq
;
1439 ASSERT(VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
1441 r
= mmu_topup_memory_caches(vcpu
);
1445 down_read(¤t
->mm
->mmap_sem
);
1446 if (is_largepage_backed(vcpu
, gfn
& ~(KVM_PAGES_PER_HPAGE
-1))) {
1447 gfn
&= ~(KVM_PAGES_PER_HPAGE
-1);
1450 mmu_seq
= vcpu
->kvm
->mmu_notifier_seq
;
1451 /* implicit mb(), we'll read before PT lock is unlocked */
1452 pfn
= gfn_to_pfn(vcpu
->kvm
, gfn
);
1453 up_read(¤t
->mm
->mmap_sem
);
1454 if (is_error_pfn(pfn
)) {
1455 kvm_release_pfn_clean(pfn
);
1458 spin_lock(&vcpu
->kvm
->mmu_lock
);
1459 if (mmu_notifier_retry(vcpu
, mmu_seq
))
1461 kvm_mmu_free_some_pages(vcpu
);
1462 r
= __direct_map(vcpu
, gpa
, error_code
& PFERR_WRITE_MASK
,
1463 largepage
, gfn
, pfn
, kvm_x86_ops
->get_tdp_level());
1464 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1469 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1470 kvm_release_pfn_clean(pfn
);
1474 static void nonpaging_free(struct kvm_vcpu
*vcpu
)
1476 mmu_free_roots(vcpu
);
1479 static int nonpaging_init_context(struct kvm_vcpu
*vcpu
)
1481 struct kvm_mmu
*context
= &vcpu
->arch
.mmu
;
1483 context
->new_cr3
= nonpaging_new_cr3
;
1484 context
->page_fault
= nonpaging_page_fault
;
1485 context
->gva_to_gpa
= nonpaging_gva_to_gpa
;
1486 context
->free
= nonpaging_free
;
1487 context
->prefetch_page
= nonpaging_prefetch_page
;
1488 context
->root_level
= 0;
1489 context
->shadow_root_level
= PT32E_ROOT_LEVEL
;
1490 context
->root_hpa
= INVALID_PAGE
;
1494 void kvm_mmu_flush_tlb(struct kvm_vcpu
*vcpu
)
1496 ++vcpu
->stat
.tlb_flush
;
1497 kvm_x86_ops
->tlb_flush(vcpu
);
1500 static void paging_new_cr3(struct kvm_vcpu
*vcpu
)
1502 pgprintk("%s: cr3 %lx\n", __func__
, vcpu
->arch
.cr3
);
1503 mmu_free_roots(vcpu
);
1506 static void inject_page_fault(struct kvm_vcpu
*vcpu
,
1510 kvm_inject_page_fault(vcpu
, addr
, err_code
);
1513 static void paging_free(struct kvm_vcpu
*vcpu
)
1515 nonpaging_free(vcpu
);
1519 #include "paging_tmpl.h"
1523 #include "paging_tmpl.h"
1526 static int paging64_init_context_common(struct kvm_vcpu
*vcpu
, int level
)
1528 struct kvm_mmu
*context
= &vcpu
->arch
.mmu
;
1530 ASSERT(is_pae(vcpu
));
1531 context
->new_cr3
= paging_new_cr3
;
1532 context
->page_fault
= paging64_page_fault
;
1533 context
->gva_to_gpa
= paging64_gva_to_gpa
;
1534 context
->prefetch_page
= paging64_prefetch_page
;
1535 context
->free
= paging_free
;
1536 context
->root_level
= level
;
1537 context
->shadow_root_level
= level
;
1538 context
->root_hpa
= INVALID_PAGE
;
1542 static int paging64_init_context(struct kvm_vcpu
*vcpu
)
1544 return paging64_init_context_common(vcpu
, PT64_ROOT_LEVEL
);
1547 static int paging32_init_context(struct kvm_vcpu
*vcpu
)
1549 struct kvm_mmu
*context
= &vcpu
->arch
.mmu
;
1551 context
->new_cr3
= paging_new_cr3
;
1552 context
->page_fault
= paging32_page_fault
;
1553 context
->gva_to_gpa
= paging32_gva_to_gpa
;
1554 context
->free
= paging_free
;
1555 context
->prefetch_page
= paging32_prefetch_page
;
1556 context
->root_level
= PT32_ROOT_LEVEL
;
1557 context
->shadow_root_level
= PT32E_ROOT_LEVEL
;
1558 context
->root_hpa
= INVALID_PAGE
;
1562 static int paging32E_init_context(struct kvm_vcpu
*vcpu
)
1564 return paging64_init_context_common(vcpu
, PT32E_ROOT_LEVEL
);
1567 static int init_kvm_tdp_mmu(struct kvm_vcpu
*vcpu
)
1569 struct kvm_mmu
*context
= &vcpu
->arch
.mmu
;
1571 context
->new_cr3
= nonpaging_new_cr3
;
1572 context
->page_fault
= tdp_page_fault
;
1573 context
->free
= nonpaging_free
;
1574 context
->prefetch_page
= nonpaging_prefetch_page
;
1575 context
->shadow_root_level
= kvm_x86_ops
->get_tdp_level();
1576 context
->root_hpa
= INVALID_PAGE
;
1578 if (!is_paging(vcpu
)) {
1579 context
->gva_to_gpa
= nonpaging_gva_to_gpa
;
1580 context
->root_level
= 0;
1581 } else if (is_long_mode(vcpu
)) {
1582 context
->gva_to_gpa
= paging64_gva_to_gpa
;
1583 context
->root_level
= PT64_ROOT_LEVEL
;
1584 } else if (is_pae(vcpu
)) {
1585 context
->gva_to_gpa
= paging64_gva_to_gpa
;
1586 context
->root_level
= PT32E_ROOT_LEVEL
;
1588 context
->gva_to_gpa
= paging32_gva_to_gpa
;
1589 context
->root_level
= PT32_ROOT_LEVEL
;
1595 static int init_kvm_softmmu(struct kvm_vcpu
*vcpu
)
1598 ASSERT(!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
1600 if (!is_paging(vcpu
))
1601 return nonpaging_init_context(vcpu
);
1602 else if (is_long_mode(vcpu
))
1603 return paging64_init_context(vcpu
);
1604 else if (is_pae(vcpu
))
1605 return paging32E_init_context(vcpu
);
1607 return paging32_init_context(vcpu
);
1610 static int init_kvm_mmu(struct kvm_vcpu
*vcpu
)
1612 vcpu
->arch
.update_pte
.pfn
= bad_pfn
;
1615 return init_kvm_tdp_mmu(vcpu
);
1617 return init_kvm_softmmu(vcpu
);
1620 static void destroy_kvm_mmu(struct kvm_vcpu
*vcpu
)
1623 if (VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
)) {
1624 vcpu
->arch
.mmu
.free(vcpu
);
1625 vcpu
->arch
.mmu
.root_hpa
= INVALID_PAGE
;
1629 int kvm_mmu_reset_context(struct kvm_vcpu
*vcpu
)
1631 destroy_kvm_mmu(vcpu
);
1632 return init_kvm_mmu(vcpu
);
1634 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context
);
1636 int kvm_mmu_load(struct kvm_vcpu
*vcpu
)
1640 r
= mmu_topup_memory_caches(vcpu
);
1643 spin_lock(&vcpu
->kvm
->mmu_lock
);
1644 kvm_mmu_free_some_pages(vcpu
);
1645 mmu_alloc_roots(vcpu
);
1646 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1647 kvm_x86_ops
->set_cr3(vcpu
, vcpu
->arch
.mmu
.root_hpa
);
1648 kvm_mmu_flush_tlb(vcpu
);
1652 EXPORT_SYMBOL_GPL(kvm_mmu_load
);
1654 void kvm_mmu_unload(struct kvm_vcpu
*vcpu
)
1656 mmu_free_roots(vcpu
);
1659 static void mmu_pte_write_zap_pte(struct kvm_vcpu
*vcpu
,
1660 struct kvm_mmu_page
*sp
,
1664 struct kvm_mmu_page
*child
;
1667 if (is_shadow_present_pte(pte
)) {
1668 if (sp
->role
.level
== PT_PAGE_TABLE_LEVEL
||
1670 rmap_remove(vcpu
->kvm
, spte
);
1672 child
= page_header(pte
& PT64_BASE_ADDR_MASK
);
1673 mmu_page_remove_parent_pte(child
, spte
);
1676 set_shadow_pte(spte
, shadow_trap_nonpresent_pte
);
1677 if (is_large_pte(pte
))
1678 --vcpu
->kvm
->stat
.lpages
;
1681 static void mmu_pte_write_new_pte(struct kvm_vcpu
*vcpu
,
1682 struct kvm_mmu_page
*sp
,
1686 if (sp
->role
.level
!= PT_PAGE_TABLE_LEVEL
) {
1687 if (!vcpu
->arch
.update_pte
.largepage
||
1688 sp
->role
.glevels
== PT32_ROOT_LEVEL
) {
1689 ++vcpu
->kvm
->stat
.mmu_pde_zapped
;
1694 ++vcpu
->kvm
->stat
.mmu_pte_updated
;
1695 if (sp
->role
.glevels
== PT32_ROOT_LEVEL
)
1696 paging32_update_pte(vcpu
, sp
, spte
, new);
1698 paging64_update_pte(vcpu
, sp
, spte
, new);
1701 static bool need_remote_flush(u64 old
, u64
new)
1703 if (!is_shadow_present_pte(old
))
1705 if (!is_shadow_present_pte(new))
1707 if ((old
^ new) & PT64_BASE_ADDR_MASK
)
1709 old
^= PT64_NX_MASK
;
1710 new ^= PT64_NX_MASK
;
1711 return (old
& ~new & PT64_PERM_MASK
) != 0;
1714 static void mmu_pte_write_flush_tlb(struct kvm_vcpu
*vcpu
, u64 old
, u64
new)
1716 if (need_remote_flush(old
, new))
1717 kvm_flush_remote_tlbs(vcpu
->kvm
);
1719 kvm_mmu_flush_tlb(vcpu
);
1722 static bool last_updated_pte_accessed(struct kvm_vcpu
*vcpu
)
1724 u64
*spte
= vcpu
->arch
.last_pte_updated
;
1726 return !!(spte
&& (*spte
& shadow_accessed_mask
));
1729 static void mmu_guess_page_from_pte_write(struct kvm_vcpu
*vcpu
, gpa_t gpa
,
1730 const u8
*new, int bytes
)
1737 vcpu
->arch
.update_pte
.largepage
= 0;
1739 if (bytes
!= 4 && bytes
!= 8)
1743 * Assume that the pte write on a page table of the same type
1744 * as the current vcpu paging mode. This is nearly always true
1745 * (might be false while changing modes). Note it is verified later
1749 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1750 if ((bytes
== 4) && (gpa
% 4 == 0)) {
1751 r
= kvm_read_guest(vcpu
->kvm
, gpa
& ~(u64
)7, &gpte
, 8);
1754 memcpy((void *)&gpte
+ (gpa
% 8), new, 4);
1755 } else if ((bytes
== 8) && (gpa
% 8 == 0)) {
1756 memcpy((void *)&gpte
, new, 8);
1759 if ((bytes
== 4) && (gpa
% 4 == 0))
1760 memcpy((void *)&gpte
, new, 4);
1762 if (!is_present_pte(gpte
))
1764 gfn
= (gpte
& PT64_BASE_ADDR_MASK
) >> PAGE_SHIFT
;
1766 down_read(¤t
->mm
->mmap_sem
);
1767 if (is_large_pte(gpte
) && is_largepage_backed(vcpu
, gfn
)) {
1768 gfn
&= ~(KVM_PAGES_PER_HPAGE
-1);
1769 vcpu
->arch
.update_pte
.largepage
= 1;
1771 vcpu
->arch
.update_pte
.mmu_seq
= vcpu
->kvm
->mmu_notifier_seq
;
1772 /* implicit mb(), we'll read before PT lock is unlocked */
1773 pfn
= gfn_to_pfn(vcpu
->kvm
, gfn
);
1774 up_read(¤t
->mm
->mmap_sem
);
1776 if (is_error_pfn(pfn
)) {
1777 kvm_release_pfn_clean(pfn
);
1780 vcpu
->arch
.update_pte
.gfn
= gfn
;
1781 vcpu
->arch
.update_pte
.pfn
= pfn
;
1784 static void kvm_mmu_access_page(struct kvm_vcpu
*vcpu
, gfn_t gfn
)
1786 u64
*spte
= vcpu
->arch
.last_pte_updated
;
1789 && vcpu
->arch
.last_pte_gfn
== gfn
1790 && shadow_accessed_mask
1791 && !(*spte
& shadow_accessed_mask
)
1792 && is_shadow_present_pte(*spte
))
1793 set_bit(PT_ACCESSED_SHIFT
, (unsigned long *)spte
);
1796 void kvm_mmu_pte_write(struct kvm_vcpu
*vcpu
, gpa_t gpa
,
1797 const u8
*new, int bytes
)
1799 gfn_t gfn
= gpa
>> PAGE_SHIFT
;
1800 struct kvm_mmu_page
*sp
;
1801 struct hlist_node
*node
, *n
;
1802 struct hlist_head
*bucket
;
1806 unsigned offset
= offset_in_page(gpa
);
1808 unsigned page_offset
;
1809 unsigned misaligned
;
1816 pgprintk("%s: gpa %llx bytes %d\n", __func__
, gpa
, bytes
);
1817 mmu_guess_page_from_pte_write(vcpu
, gpa
, new, bytes
);
1818 spin_lock(&vcpu
->kvm
->mmu_lock
);
1819 kvm_mmu_access_page(vcpu
, gfn
);
1820 kvm_mmu_free_some_pages(vcpu
);
1821 ++vcpu
->kvm
->stat
.mmu_pte_write
;
1822 kvm_mmu_audit(vcpu
, "pre pte write");
1823 if (gfn
== vcpu
->arch
.last_pt_write_gfn
1824 && !last_updated_pte_accessed(vcpu
)) {
1825 ++vcpu
->arch
.last_pt_write_count
;
1826 if (vcpu
->arch
.last_pt_write_count
>= 3)
1829 vcpu
->arch
.last_pt_write_gfn
= gfn
;
1830 vcpu
->arch
.last_pt_write_count
= 1;
1831 vcpu
->arch
.last_pte_updated
= NULL
;
1833 index
= kvm_page_table_hashfn(gfn
);
1834 bucket
= &vcpu
->kvm
->arch
.mmu_page_hash
[index
];
1835 hlist_for_each_entry_safe(sp
, node
, n
, bucket
, hash_link
) {
1836 if (sp
->gfn
!= gfn
|| sp
->role
.metaphysical
)
1838 pte_size
= sp
->role
.glevels
== PT32_ROOT_LEVEL
? 4 : 8;
1839 misaligned
= (offset
^ (offset
+ bytes
- 1)) & ~(pte_size
- 1);
1840 misaligned
|= bytes
< 4;
1841 if (misaligned
|| flooded
) {
1843 * Misaligned accesses are too much trouble to fix
1844 * up; also, they usually indicate a page is not used
1847 * If we're seeing too many writes to a page,
1848 * it may no longer be a page table, or we may be
1849 * forking, in which case it is better to unmap the
1852 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1853 gpa
, bytes
, sp
->role
.word
);
1854 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
1855 ++vcpu
->kvm
->stat
.mmu_flooded
;
1858 page_offset
= offset
;
1859 level
= sp
->role
.level
;
1861 if (sp
->role
.glevels
== PT32_ROOT_LEVEL
) {
1862 page_offset
<<= 1; /* 32->64 */
1864 * A 32-bit pde maps 4MB while the shadow pdes map
1865 * only 2MB. So we need to double the offset again
1866 * and zap two pdes instead of one.
1868 if (level
== PT32_ROOT_LEVEL
) {
1869 page_offset
&= ~7; /* kill rounding error */
1873 quadrant
= page_offset
>> PAGE_SHIFT
;
1874 page_offset
&= ~PAGE_MASK
;
1875 if (quadrant
!= sp
->role
.quadrant
)
1878 spte
= &sp
->spt
[page_offset
/ sizeof(*spte
)];
1879 if ((gpa
& (pte_size
- 1)) || (bytes
< pte_size
)) {
1881 r
= kvm_read_guest_atomic(vcpu
->kvm
,
1882 gpa
& ~(u64
)(pte_size
- 1),
1884 new = (const void *)&gentry
;
1890 mmu_pte_write_zap_pte(vcpu
, sp
, spte
);
1892 mmu_pte_write_new_pte(vcpu
, sp
, spte
, new);
1893 mmu_pte_write_flush_tlb(vcpu
, entry
, *spte
);
1897 kvm_mmu_audit(vcpu
, "post pte write");
1898 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1899 if (!is_error_pfn(vcpu
->arch
.update_pte
.pfn
)) {
1900 kvm_release_pfn_clean(vcpu
->arch
.update_pte
.pfn
);
1901 vcpu
->arch
.update_pte
.pfn
= bad_pfn
;
1905 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu
*vcpu
, gva_t gva
)
1910 gpa
= vcpu
->arch
.mmu
.gva_to_gpa(vcpu
, gva
);
1912 spin_lock(&vcpu
->kvm
->mmu_lock
);
1913 r
= kvm_mmu_unprotect_page(vcpu
->kvm
, gpa
>> PAGE_SHIFT
);
1914 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1917 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt
);
1919 void __kvm_mmu_free_some_pages(struct kvm_vcpu
*vcpu
)
1921 while (vcpu
->kvm
->arch
.n_free_mmu_pages
< KVM_REFILL_PAGES
) {
1922 struct kvm_mmu_page
*sp
;
1924 sp
= container_of(vcpu
->kvm
->arch
.active_mmu_pages
.prev
,
1925 struct kvm_mmu_page
, link
);
1926 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
1927 ++vcpu
->kvm
->stat
.mmu_recycled
;
1931 int kvm_mmu_page_fault(struct kvm_vcpu
*vcpu
, gva_t cr2
, u32 error_code
)
1934 enum emulation_result er
;
1936 r
= vcpu
->arch
.mmu
.page_fault(vcpu
, cr2
, error_code
);
1945 r
= mmu_topup_memory_caches(vcpu
);
1949 er
= emulate_instruction(vcpu
, vcpu
->run
, cr2
, error_code
, 0);
1954 case EMULATE_DO_MMIO
:
1955 ++vcpu
->stat
.mmio_exits
;
1958 kvm_report_emulation_failure(vcpu
, "pagetable");
1966 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault
);
1968 void kvm_enable_tdp(void)
1972 EXPORT_SYMBOL_GPL(kvm_enable_tdp
);
1974 void kvm_disable_tdp(void)
1976 tdp_enabled
= false;
1978 EXPORT_SYMBOL_GPL(kvm_disable_tdp
);
1980 static void free_mmu_pages(struct kvm_vcpu
*vcpu
)
1982 struct kvm_mmu_page
*sp
;
1984 while (!list_empty(&vcpu
->kvm
->arch
.active_mmu_pages
)) {
1985 sp
= container_of(vcpu
->kvm
->arch
.active_mmu_pages
.next
,
1986 struct kvm_mmu_page
, link
);
1987 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
1990 free_page((unsigned long)vcpu
->arch
.mmu
.pae_root
);
1993 static int alloc_mmu_pages(struct kvm_vcpu
*vcpu
)
2000 if (vcpu
->kvm
->arch
.n_requested_mmu_pages
)
2001 vcpu
->kvm
->arch
.n_free_mmu_pages
=
2002 vcpu
->kvm
->arch
.n_requested_mmu_pages
;
2004 vcpu
->kvm
->arch
.n_free_mmu_pages
=
2005 vcpu
->kvm
->arch
.n_alloc_mmu_pages
;
2007 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2008 * Therefore we need to allocate shadow page tables in the first
2009 * 4GB of memory, which happens to fit the DMA32 zone.
2011 page
= alloc_page(GFP_KERNEL
| __GFP_DMA32
);
2014 vcpu
->arch
.mmu
.pae_root
= page_address(page
);
2015 for (i
= 0; i
< 4; ++i
)
2016 vcpu
->arch
.mmu
.pae_root
[i
] = INVALID_PAGE
;
2021 free_mmu_pages(vcpu
);
2025 int kvm_mmu_create(struct kvm_vcpu
*vcpu
)
2028 ASSERT(!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
2030 return alloc_mmu_pages(vcpu
);
2033 int kvm_mmu_setup(struct kvm_vcpu
*vcpu
)
2036 ASSERT(!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
2038 return init_kvm_mmu(vcpu
);
2041 void kvm_mmu_destroy(struct kvm_vcpu
*vcpu
)
2045 destroy_kvm_mmu(vcpu
);
2046 free_mmu_pages(vcpu
);
2047 mmu_free_memory_caches(vcpu
);
2050 void kvm_mmu_slot_remove_write_access(struct kvm
*kvm
, int slot
)
2052 struct kvm_mmu_page
*sp
;
2054 list_for_each_entry(sp
, &kvm
->arch
.active_mmu_pages
, link
) {
2058 if (!test_bit(slot
, &sp
->slot_bitmap
))
2062 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
)
2064 if (pt
[i
] & PT_WRITABLE_MASK
)
2065 pt
[i
] &= ~PT_WRITABLE_MASK
;
2069 void kvm_mmu_zap_all(struct kvm
*kvm
)
2071 struct kvm_mmu_page
*sp
, *node
;
2073 spin_lock(&kvm
->mmu_lock
);
2074 list_for_each_entry_safe(sp
, node
, &kvm
->arch
.active_mmu_pages
, link
)
2075 kvm_mmu_zap_page(kvm
, sp
);
2076 spin_unlock(&kvm
->mmu_lock
);
2078 kvm_flush_remote_tlbs(kvm
);
2081 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm
*kvm
)
2083 struct kvm_mmu_page
*page
;
2085 page
= container_of(kvm
->arch
.active_mmu_pages
.prev
,
2086 struct kvm_mmu_page
, link
);
2087 kvm_mmu_zap_page(kvm
, page
);
2090 static int mmu_shrink(int nr_to_scan
, gfp_t gfp_mask
)
2093 struct kvm
*kvm_freed
= NULL
;
2094 int cache_count
= 0;
2096 spin_lock(&kvm_lock
);
2098 list_for_each_entry(kvm
, &vm_list
, vm_list
) {
2101 if (!down_read_trylock(&kvm
->slots_lock
))
2103 spin_lock(&kvm
->mmu_lock
);
2104 npages
= kvm
->arch
.n_alloc_mmu_pages
-
2105 kvm
->arch
.n_free_mmu_pages
;
2106 cache_count
+= npages
;
2107 if (!kvm_freed
&& nr_to_scan
> 0 && npages
> 0) {
2108 kvm_mmu_remove_one_alloc_mmu_page(kvm
);
2114 spin_unlock(&kvm
->mmu_lock
);
2115 up_read(&kvm
->slots_lock
);
2118 list_move_tail(&kvm_freed
->vm_list
, &vm_list
);
2120 spin_unlock(&kvm_lock
);
2125 static struct shrinker mmu_shrinker
= {
2126 .shrink
= mmu_shrink
,
2127 .seeks
= DEFAULT_SEEKS
* 10,
2130 static void mmu_destroy_caches(void)
2132 if (pte_chain_cache
)
2133 kmem_cache_destroy(pte_chain_cache
);
2134 if (rmap_desc_cache
)
2135 kmem_cache_destroy(rmap_desc_cache
);
2136 if (mmu_page_header_cache
)
2137 kmem_cache_destroy(mmu_page_header_cache
);
2140 void kvm_mmu_module_exit(void)
2142 mmu_destroy_caches();
2143 unregister_shrinker(&mmu_shrinker
);
2146 int kvm_mmu_module_init(void)
2148 pte_chain_cache
= kmem_cache_create("kvm_pte_chain",
2149 sizeof(struct kvm_pte_chain
),
2151 if (!pte_chain_cache
)
2153 rmap_desc_cache
= kmem_cache_create("kvm_rmap_desc",
2154 sizeof(struct kvm_rmap_desc
),
2156 if (!rmap_desc_cache
)
2159 mmu_page_header_cache
= kmem_cache_create("kvm_mmu_page_header",
2160 sizeof(struct kvm_mmu_page
),
2162 if (!mmu_page_header_cache
)
2165 register_shrinker(&mmu_shrinker
);
2170 mmu_destroy_caches();
2175 * Caculate mmu pages needed for kvm.
2177 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm
*kvm
)
2180 unsigned int nr_mmu_pages
;
2181 unsigned int nr_pages
= 0;
2183 for (i
= 0; i
< kvm
->nmemslots
; i
++)
2184 nr_pages
+= kvm
->memslots
[i
].npages
;
2186 nr_mmu_pages
= nr_pages
* KVM_PERMILLE_MMU_PAGES
/ 1000;
2187 nr_mmu_pages
= max(nr_mmu_pages
,
2188 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES
);
2190 return nr_mmu_pages
;
2193 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer
*buffer
,
2196 if (len
> buffer
->len
)
2201 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer
*buffer
,
2206 ret
= pv_mmu_peek_buffer(buffer
, len
);
2211 buffer
->processed
+= len
;
2215 static int kvm_pv_mmu_write(struct kvm_vcpu
*vcpu
,
2216 gpa_t addr
, gpa_t value
)
2221 if (!is_long_mode(vcpu
) && !is_pae(vcpu
))
2224 r
= mmu_topup_memory_caches(vcpu
);
2228 if (!emulator_write_phys(vcpu
, addr
, &value
, bytes
))
2234 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu
*vcpu
)
2236 kvm_x86_ops
->tlb_flush(vcpu
);
2240 static int kvm_pv_mmu_release_pt(struct kvm_vcpu
*vcpu
, gpa_t addr
)
2242 spin_lock(&vcpu
->kvm
->mmu_lock
);
2243 mmu_unshadow(vcpu
->kvm
, addr
>> PAGE_SHIFT
);
2244 spin_unlock(&vcpu
->kvm
->mmu_lock
);
2248 static int kvm_pv_mmu_op_one(struct kvm_vcpu
*vcpu
,
2249 struct kvm_pv_mmu_op_buffer
*buffer
)
2251 struct kvm_mmu_op_header
*header
;
2253 header
= pv_mmu_peek_buffer(buffer
, sizeof *header
);
2256 switch (header
->op
) {
2257 case KVM_MMU_OP_WRITE_PTE
: {
2258 struct kvm_mmu_op_write_pte
*wpte
;
2260 wpte
= pv_mmu_read_buffer(buffer
, sizeof *wpte
);
2263 return kvm_pv_mmu_write(vcpu
, wpte
->pte_phys
,
2266 case KVM_MMU_OP_FLUSH_TLB
: {
2267 struct kvm_mmu_op_flush_tlb
*ftlb
;
2269 ftlb
= pv_mmu_read_buffer(buffer
, sizeof *ftlb
);
2272 return kvm_pv_mmu_flush_tlb(vcpu
);
2274 case KVM_MMU_OP_RELEASE_PT
: {
2275 struct kvm_mmu_op_release_pt
*rpt
;
2277 rpt
= pv_mmu_read_buffer(buffer
, sizeof *rpt
);
2280 return kvm_pv_mmu_release_pt(vcpu
, rpt
->pt_phys
);
2286 int kvm_pv_mmu_op(struct kvm_vcpu
*vcpu
, unsigned long bytes
,
2287 gpa_t addr
, unsigned long *ret
)
2290 struct kvm_pv_mmu_op_buffer buffer
;
2292 buffer
.ptr
= buffer
.buf
;
2293 buffer
.len
= min_t(unsigned long, bytes
, sizeof buffer
.buf
);
2294 buffer
.processed
= 0;
2296 r
= kvm_read_guest(vcpu
->kvm
, addr
, buffer
.buf
, buffer
.len
);
2300 while (buffer
.len
) {
2301 r
= kvm_pv_mmu_op_one(vcpu
, &buffer
);
2310 *ret
= buffer
.processed
;
2316 static const char *audit_msg
;
2318 static gva_t
canonicalize(gva_t gva
)
2320 #ifdef CONFIG_X86_64
2321 gva
= (long long)(gva
<< 16) >> 16;
2326 static void audit_mappings_page(struct kvm_vcpu
*vcpu
, u64 page_pte
,
2327 gva_t va
, int level
)
2329 u64
*pt
= __va(page_pte
& PT64_BASE_ADDR_MASK
);
2331 gva_t va_delta
= 1ul << (PAGE_SHIFT
+ 9 * (level
- 1));
2333 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
, va
+= va_delta
) {
2336 if (ent
== shadow_trap_nonpresent_pte
)
2339 va
= canonicalize(va
);
2341 if (ent
== shadow_notrap_nonpresent_pte
)
2342 printk(KERN_ERR
"audit: (%s) nontrapping pte"
2343 " in nonleaf level: levels %d gva %lx"
2344 " level %d pte %llx\n", audit_msg
,
2345 vcpu
->arch
.mmu
.root_level
, va
, level
, ent
);
2347 audit_mappings_page(vcpu
, ent
, va
, level
- 1);
2349 gpa_t gpa
= vcpu
->arch
.mmu
.gva_to_gpa(vcpu
, va
);
2350 hpa_t hpa
= (hpa_t
)gpa_to_pfn(vcpu
, gpa
) << PAGE_SHIFT
;
2352 if (is_shadow_present_pte(ent
)
2353 && (ent
& PT64_BASE_ADDR_MASK
) != hpa
)
2354 printk(KERN_ERR
"xx audit error: (%s) levels %d"
2355 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2356 audit_msg
, vcpu
->arch
.mmu
.root_level
,
2358 is_shadow_present_pte(ent
));
2359 else if (ent
== shadow_notrap_nonpresent_pte
2360 && !is_error_hpa(hpa
))
2361 printk(KERN_ERR
"audit: (%s) notrap shadow,"
2362 " valid guest gva %lx\n", audit_msg
, va
);
2363 kvm_release_pfn_clean(pfn
);
2369 static void audit_mappings(struct kvm_vcpu
*vcpu
)
2373 if (vcpu
->arch
.mmu
.root_level
== 4)
2374 audit_mappings_page(vcpu
, vcpu
->arch
.mmu
.root_hpa
, 0, 4);
2376 for (i
= 0; i
< 4; ++i
)
2377 if (vcpu
->arch
.mmu
.pae_root
[i
] & PT_PRESENT_MASK
)
2378 audit_mappings_page(vcpu
,
2379 vcpu
->arch
.mmu
.pae_root
[i
],
2384 static int count_rmaps(struct kvm_vcpu
*vcpu
)
2389 for (i
= 0; i
< KVM_MEMORY_SLOTS
; ++i
) {
2390 struct kvm_memory_slot
*m
= &vcpu
->kvm
->memslots
[i
];
2391 struct kvm_rmap_desc
*d
;
2393 for (j
= 0; j
< m
->npages
; ++j
) {
2394 unsigned long *rmapp
= &m
->rmap
[j
];
2398 if (!(*rmapp
& 1)) {
2402 d
= (struct kvm_rmap_desc
*)(*rmapp
& ~1ul);
2404 for (k
= 0; k
< RMAP_EXT
; ++k
)
2405 if (d
->shadow_ptes
[k
])
2416 static int count_writable_mappings(struct kvm_vcpu
*vcpu
)
2419 struct kvm_mmu_page
*sp
;
2422 list_for_each_entry(sp
, &vcpu
->kvm
->arch
.active_mmu_pages
, link
) {
2425 if (sp
->role
.level
!= PT_PAGE_TABLE_LEVEL
)
2428 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
) {
2431 if (!(ent
& PT_PRESENT_MASK
))
2433 if (!(ent
& PT_WRITABLE_MASK
))
2441 static void audit_rmap(struct kvm_vcpu
*vcpu
)
2443 int n_rmap
= count_rmaps(vcpu
);
2444 int n_actual
= count_writable_mappings(vcpu
);
2446 if (n_rmap
!= n_actual
)
2447 printk(KERN_ERR
"%s: (%s) rmap %d actual %d\n",
2448 __func__
, audit_msg
, n_rmap
, n_actual
);
2451 static void audit_write_protection(struct kvm_vcpu
*vcpu
)
2453 struct kvm_mmu_page
*sp
;
2454 struct kvm_memory_slot
*slot
;
2455 unsigned long *rmapp
;
2458 list_for_each_entry(sp
, &vcpu
->kvm
->arch
.active_mmu_pages
, link
) {
2459 if (sp
->role
.metaphysical
)
2462 slot
= gfn_to_memslot(vcpu
->kvm
, sp
->gfn
);
2463 gfn
= unalias_gfn(vcpu
->kvm
, sp
->gfn
);
2464 rmapp
= &slot
->rmap
[gfn
- slot
->base_gfn
];
2466 printk(KERN_ERR
"%s: (%s) shadow page has writable"
2467 " mappings: gfn %lx role %x\n",
2468 __func__
, audit_msg
, sp
->gfn
,
2473 static void kvm_mmu_audit(struct kvm_vcpu
*vcpu
, const char *msg
)
2480 audit_write_protection(vcpu
);
2481 audit_mappings(vcpu
);