KVM: Lazy guest cr3 switching
[linux-2.6.git] / drivers / kvm / mmu.c
1 /*
2  * Kernel-based Virtual Machine driver for Linux
3  *
4  * This module enables machines with Intel VT-x extensions to run virtual
5  * machines without emulation or binary translation.
6  *
7  * MMU support
8  *
9  * Copyright (C) 2006 Qumranet, Inc.
10  *
11  * Authors:
12  *   Yaniv Kamay  <yaniv@qumranet.com>
13  *   Avi Kivity   <avi@qumranet.com>
14  *
15  * This work is licensed under the terms of the GNU GPL, version 2.  See
16  * the COPYING file in the top-level directory.
17  *
18  */
19 #include <linux/types.h>
20 #include <linux/string.h>
21 #include <asm/page.h>
22 #include <linux/mm.h>
23 #include <linux/highmem.h>
24 #include <linux/module.h>
25 #include <asm/cmpxchg.h>
26
27 #include "vmx.h"
28 #include "kvm.h"
29
30 #undef MMU_DEBUG
31
32 #undef AUDIT
33
34 #ifdef AUDIT
35 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
36 #else
37 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
38 #endif
39
40 #ifdef MMU_DEBUG
41
42 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
43 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
44
45 #else
46
47 #define pgprintk(x...) do { } while (0)
48 #define rmap_printk(x...) do { } while (0)
49
50 #endif
51
52 #if defined(MMU_DEBUG) || defined(AUDIT)
53 static int dbg = 1;
54 #endif
55
56 #ifndef MMU_DEBUG
57 #define ASSERT(x) do { } while (0)
58 #else
59 #define ASSERT(x)                                                       \
60         if (!(x)) {                                                     \
61                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
62                        __FILE__, __LINE__, #x);                         \
63         }
64 #endif
65
66 #define PT64_PT_BITS 9
67 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
68 #define PT32_PT_BITS 10
69 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
70
71 #define PT_WRITABLE_SHIFT 1
72
73 #define PT_PRESENT_MASK (1ULL << 0)
74 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
75 #define PT_USER_MASK (1ULL << 2)
76 #define PT_PWT_MASK (1ULL << 3)
77 #define PT_PCD_MASK (1ULL << 4)
78 #define PT_ACCESSED_MASK (1ULL << 5)
79 #define PT_DIRTY_MASK (1ULL << 6)
80 #define PT_PAGE_SIZE_MASK (1ULL << 7)
81 #define PT_PAT_MASK (1ULL << 7)
82 #define PT_GLOBAL_MASK (1ULL << 8)
83 #define PT64_NX_MASK (1ULL << 63)
84
85 #define PT_PAT_SHIFT 7
86 #define PT_DIR_PAT_SHIFT 12
87 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
88
89 #define PT32_DIR_PSE36_SIZE 4
90 #define PT32_DIR_PSE36_SHIFT 13
91 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
92
93
94 #define PT_FIRST_AVAIL_BITS_SHIFT 9
95 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
96
97 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
98
99 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
100
101 #define PT64_LEVEL_BITS 9
102
103 #define PT64_LEVEL_SHIFT(level) \
104                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
105
106 #define PT64_LEVEL_MASK(level) \
107                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
108
109 #define PT64_INDEX(address, level)\
110         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
111
112
113 #define PT32_LEVEL_BITS 10
114
115 #define PT32_LEVEL_SHIFT(level) \
116                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
117
118 #define PT32_LEVEL_MASK(level) \
119                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
120
121 #define PT32_INDEX(address, level)\
122         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
123
124
125 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
126 #define PT64_DIR_BASE_ADDR_MASK \
127         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
128
129 #define PT32_BASE_ADDR_MASK PAGE_MASK
130 #define PT32_DIR_BASE_ADDR_MASK \
131         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
132
133
134 #define PFERR_PRESENT_MASK (1U << 0)
135 #define PFERR_WRITE_MASK (1U << 1)
136 #define PFERR_USER_MASK (1U << 2)
137 #define PFERR_FETCH_MASK (1U << 4)
138
139 #define PT64_ROOT_LEVEL 4
140 #define PT32_ROOT_LEVEL 2
141 #define PT32E_ROOT_LEVEL 3
142
143 #define PT_DIRECTORY_LEVEL 2
144 #define PT_PAGE_TABLE_LEVEL 1
145
146 #define RMAP_EXT 4
147
148 struct kvm_rmap_desc {
149         u64 *shadow_ptes[RMAP_EXT];
150         struct kvm_rmap_desc *more;
151 };
152
153 static struct kmem_cache *pte_chain_cache;
154 static struct kmem_cache *rmap_desc_cache;
155 static struct kmem_cache *mmu_page_cache;
156 static struct kmem_cache *mmu_page_header_cache;
157
158 static int is_write_protection(struct kvm_vcpu *vcpu)
159 {
160         return vcpu->cr0 & CR0_WP_MASK;
161 }
162
163 static int is_cpuid_PSE36(void)
164 {
165         return 1;
166 }
167
168 static int is_nx(struct kvm_vcpu *vcpu)
169 {
170         return vcpu->shadow_efer & EFER_NX;
171 }
172
173 static int is_present_pte(unsigned long pte)
174 {
175         return pte & PT_PRESENT_MASK;
176 }
177
178 static int is_writeble_pte(unsigned long pte)
179 {
180         return pte & PT_WRITABLE_MASK;
181 }
182
183 static int is_io_pte(unsigned long pte)
184 {
185         return pte & PT_SHADOW_IO_MARK;
186 }
187
188 static int is_rmap_pte(u64 pte)
189 {
190         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
191                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
192 }
193
194 static void set_shadow_pte(u64 *sptep, u64 spte)
195 {
196 #ifdef CONFIG_X86_64
197         set_64bit((unsigned long *)sptep, spte);
198 #else
199         set_64bit((unsigned long long *)sptep, spte);
200 #endif
201 }
202
203 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
204                                   struct kmem_cache *base_cache, int min,
205                                   gfp_t gfp_flags)
206 {
207         void *obj;
208
209         if (cache->nobjs >= min)
210                 return 0;
211         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
212                 obj = kmem_cache_zalloc(base_cache, gfp_flags);
213                 if (!obj)
214                         return -ENOMEM;
215                 cache->objects[cache->nobjs++] = obj;
216         }
217         return 0;
218 }
219
220 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
221 {
222         while (mc->nobjs)
223                 kfree(mc->objects[--mc->nobjs]);
224 }
225
226 static int __mmu_topup_memory_caches(struct kvm_vcpu *vcpu, gfp_t gfp_flags)
227 {
228         int r;
229
230         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
231                                    pte_chain_cache, 4, gfp_flags);
232         if (r)
233                 goto out;
234         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
235                                    rmap_desc_cache, 1, gfp_flags);
236         if (r)
237                 goto out;
238         r = mmu_topup_memory_cache(&vcpu->mmu_page_cache,
239                                    mmu_page_cache, 4, gfp_flags);
240         if (r)
241                 goto out;
242         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
243                                    mmu_page_header_cache, 4, gfp_flags);
244 out:
245         return r;
246 }
247
248 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
249 {
250         int r;
251
252         r = __mmu_topup_memory_caches(vcpu, GFP_NOWAIT);
253         if (r < 0) {
254                 spin_unlock(&vcpu->kvm->lock);
255                 kvm_arch_ops->vcpu_put(vcpu);
256                 r = __mmu_topup_memory_caches(vcpu, GFP_KERNEL);
257                 kvm_arch_ops->vcpu_load(vcpu);
258                 spin_lock(&vcpu->kvm->lock);
259         }
260         return r;
261 }
262
263 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
264 {
265         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
266         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
267         mmu_free_memory_cache(&vcpu->mmu_page_cache);
268         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
269 }
270
271 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
272                                     size_t size)
273 {
274         void *p;
275
276         BUG_ON(!mc->nobjs);
277         p = mc->objects[--mc->nobjs];
278         memset(p, 0, size);
279         return p;
280 }
281
282 static void mmu_memory_cache_free(struct kvm_mmu_memory_cache *mc, void *obj)
283 {
284         if (mc->nobjs < KVM_NR_MEM_OBJS)
285                 mc->objects[mc->nobjs++] = obj;
286         else
287                 kfree(obj);
288 }
289
290 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
291 {
292         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
293                                       sizeof(struct kvm_pte_chain));
294 }
295
296 static void mmu_free_pte_chain(struct kvm_vcpu *vcpu,
297                                struct kvm_pte_chain *pc)
298 {
299         mmu_memory_cache_free(&vcpu->mmu_pte_chain_cache, pc);
300 }
301
302 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
303 {
304         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
305                                       sizeof(struct kvm_rmap_desc));
306 }
307
308 static void mmu_free_rmap_desc(struct kvm_vcpu *vcpu,
309                                struct kvm_rmap_desc *rd)
310 {
311         mmu_memory_cache_free(&vcpu->mmu_rmap_desc_cache, rd);
312 }
313
314 /*
315  * Reverse mapping data structures:
316  *
317  * If page->private bit zero is zero, then page->private points to the
318  * shadow page table entry that points to page_address(page).
319  *
320  * If page->private bit zero is one, (then page->private & ~1) points
321  * to a struct kvm_rmap_desc containing more mappings.
322  */
323 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
324 {
325         struct page *page;
326         struct kvm_rmap_desc *desc;
327         int i;
328
329         if (!is_rmap_pte(*spte))
330                 return;
331         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
332         if (!page_private(page)) {
333                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
334                 set_page_private(page,(unsigned long)spte);
335         } else if (!(page_private(page) & 1)) {
336                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
337                 desc = mmu_alloc_rmap_desc(vcpu);
338                 desc->shadow_ptes[0] = (u64 *)page_private(page);
339                 desc->shadow_ptes[1] = spte;
340                 set_page_private(page,(unsigned long)desc | 1);
341         } else {
342                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
343                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
344                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
345                         desc = desc->more;
346                 if (desc->shadow_ptes[RMAP_EXT-1]) {
347                         desc->more = mmu_alloc_rmap_desc(vcpu);
348                         desc = desc->more;
349                 }
350                 for (i = 0; desc->shadow_ptes[i]; ++i)
351                         ;
352                 desc->shadow_ptes[i] = spte;
353         }
354 }
355
356 static void rmap_desc_remove_entry(struct kvm_vcpu *vcpu,
357                                    struct page *page,
358                                    struct kvm_rmap_desc *desc,
359                                    int i,
360                                    struct kvm_rmap_desc *prev_desc)
361 {
362         int j;
363
364         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
365                 ;
366         desc->shadow_ptes[i] = desc->shadow_ptes[j];
367         desc->shadow_ptes[j] = NULL;
368         if (j != 0)
369                 return;
370         if (!prev_desc && !desc->more)
371                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
372         else
373                 if (prev_desc)
374                         prev_desc->more = desc->more;
375                 else
376                         set_page_private(page,(unsigned long)desc->more | 1);
377         mmu_free_rmap_desc(vcpu, desc);
378 }
379
380 static void rmap_remove(struct kvm_vcpu *vcpu, u64 *spte)
381 {
382         struct page *page;
383         struct kvm_rmap_desc *desc;
384         struct kvm_rmap_desc *prev_desc;
385         int i;
386
387         if (!is_rmap_pte(*spte))
388                 return;
389         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
390         if (!page_private(page)) {
391                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
392                 BUG();
393         } else if (!(page_private(page) & 1)) {
394                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
395                 if ((u64 *)page_private(page) != spte) {
396                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
397                                spte, *spte);
398                         BUG();
399                 }
400                 set_page_private(page,0);
401         } else {
402                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
403                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
404                 prev_desc = NULL;
405                 while (desc) {
406                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
407                                 if (desc->shadow_ptes[i] == spte) {
408                                         rmap_desc_remove_entry(vcpu, page,
409                                                                desc, i,
410                                                                prev_desc);
411                                         return;
412                                 }
413                         prev_desc = desc;
414                         desc = desc->more;
415                 }
416                 BUG();
417         }
418 }
419
420 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
421 {
422         struct kvm *kvm = vcpu->kvm;
423         struct page *page;
424         struct kvm_rmap_desc *desc;
425         u64 *spte;
426
427         page = gfn_to_page(kvm, gfn);
428         BUG_ON(!page);
429
430         while (page_private(page)) {
431                 if (!(page_private(page) & 1))
432                         spte = (u64 *)page_private(page);
433                 else {
434                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
435                         spte = desc->shadow_ptes[0];
436                 }
437                 BUG_ON(!spte);
438                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
439                        != page_to_pfn(page));
440                 BUG_ON(!(*spte & PT_PRESENT_MASK));
441                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
442                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
443                 rmap_remove(vcpu, spte);
444                 kvm_arch_ops->tlb_flush(vcpu);
445                 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
446         }
447 }
448
449 #ifdef MMU_DEBUG
450 static int is_empty_shadow_page(u64 *spt)
451 {
452         u64 *pos;
453         u64 *end;
454
455         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
456                 if (*pos != 0) {
457                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
458                                pos, *pos);
459                         return 0;
460                 }
461         return 1;
462 }
463 #endif
464
465 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu,
466                               struct kvm_mmu_page *page_head)
467 {
468         ASSERT(is_empty_shadow_page(page_head->spt));
469         list_del(&page_head->link);
470         mmu_memory_cache_free(&vcpu->mmu_page_cache, page_head->spt);
471         mmu_memory_cache_free(&vcpu->mmu_page_header_cache, page_head);
472         ++vcpu->kvm->n_free_mmu_pages;
473 }
474
475 static unsigned kvm_page_table_hashfn(gfn_t gfn)
476 {
477         return gfn;
478 }
479
480 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
481                                                u64 *parent_pte)
482 {
483         struct kvm_mmu_page *page;
484
485         if (!vcpu->kvm->n_free_mmu_pages)
486                 return NULL;
487
488         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
489                                       sizeof *page);
490         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
491         set_page_private(virt_to_page(page->spt), (unsigned long)page);
492         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
493         ASSERT(is_empty_shadow_page(page->spt));
494         page->slot_bitmap = 0;
495         page->multimapped = 0;
496         page->parent_pte = parent_pte;
497         --vcpu->kvm->n_free_mmu_pages;
498         return page;
499 }
500
501 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
502                                     struct kvm_mmu_page *page, u64 *parent_pte)
503 {
504         struct kvm_pte_chain *pte_chain;
505         struct hlist_node *node;
506         int i;
507
508         if (!parent_pte)
509                 return;
510         if (!page->multimapped) {
511                 u64 *old = page->parent_pte;
512
513                 if (!old) {
514                         page->parent_pte = parent_pte;
515                         return;
516                 }
517                 page->multimapped = 1;
518                 pte_chain = mmu_alloc_pte_chain(vcpu);
519                 INIT_HLIST_HEAD(&page->parent_ptes);
520                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
521                 pte_chain->parent_ptes[0] = old;
522         }
523         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
524                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
525                         continue;
526                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
527                         if (!pte_chain->parent_ptes[i]) {
528                                 pte_chain->parent_ptes[i] = parent_pte;
529                                 return;
530                         }
531         }
532         pte_chain = mmu_alloc_pte_chain(vcpu);
533         BUG_ON(!pte_chain);
534         hlist_add_head(&pte_chain->link, &page->parent_ptes);
535         pte_chain->parent_ptes[0] = parent_pte;
536 }
537
538 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
539                                        struct kvm_mmu_page *page,
540                                        u64 *parent_pte)
541 {
542         struct kvm_pte_chain *pte_chain;
543         struct hlist_node *node;
544         int i;
545
546         if (!page->multimapped) {
547                 BUG_ON(page->parent_pte != parent_pte);
548                 page->parent_pte = NULL;
549                 return;
550         }
551         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
552                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
553                         if (!pte_chain->parent_ptes[i])
554                                 break;
555                         if (pte_chain->parent_ptes[i] != parent_pte)
556                                 continue;
557                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
558                                 && pte_chain->parent_ptes[i + 1]) {
559                                 pte_chain->parent_ptes[i]
560                                         = pte_chain->parent_ptes[i + 1];
561                                 ++i;
562                         }
563                         pte_chain->parent_ptes[i] = NULL;
564                         if (i == 0) {
565                                 hlist_del(&pte_chain->link);
566                                 mmu_free_pte_chain(vcpu, pte_chain);
567                                 if (hlist_empty(&page->parent_ptes)) {
568                                         page->multimapped = 0;
569                                         page->parent_pte = NULL;
570                                 }
571                         }
572                         return;
573                 }
574         BUG();
575 }
576
577 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
578                                                 gfn_t gfn)
579 {
580         unsigned index;
581         struct hlist_head *bucket;
582         struct kvm_mmu_page *page;
583         struct hlist_node *node;
584
585         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
586         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
587         bucket = &vcpu->kvm->mmu_page_hash[index];
588         hlist_for_each_entry(page, node, bucket, hash_link)
589                 if (page->gfn == gfn && !page->role.metaphysical) {
590                         pgprintk("%s: found role %x\n",
591                                  __FUNCTION__, page->role.word);
592                         return page;
593                 }
594         return NULL;
595 }
596
597 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
598                                              gfn_t gfn,
599                                              gva_t gaddr,
600                                              unsigned level,
601                                              int metaphysical,
602                                              unsigned hugepage_access,
603                                              u64 *parent_pte)
604 {
605         union kvm_mmu_page_role role;
606         unsigned index;
607         unsigned quadrant;
608         struct hlist_head *bucket;
609         struct kvm_mmu_page *page;
610         struct hlist_node *node;
611
612         role.word = 0;
613         role.glevels = vcpu->mmu.root_level;
614         role.level = level;
615         role.metaphysical = metaphysical;
616         role.hugepage_access = hugepage_access;
617         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
618                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
619                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
620                 role.quadrant = quadrant;
621         }
622         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
623                  gfn, role.word);
624         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
625         bucket = &vcpu->kvm->mmu_page_hash[index];
626         hlist_for_each_entry(page, node, bucket, hash_link)
627                 if (page->gfn == gfn && page->role.word == role.word) {
628                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
629                         pgprintk("%s: found\n", __FUNCTION__);
630                         return page;
631                 }
632         page = kvm_mmu_alloc_page(vcpu, parent_pte);
633         if (!page)
634                 return page;
635         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
636         page->gfn = gfn;
637         page->role = role;
638         hlist_add_head(&page->hash_link, bucket);
639         if (!metaphysical)
640                 rmap_write_protect(vcpu, gfn);
641         return page;
642 }
643
644 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
645                                          struct kvm_mmu_page *page)
646 {
647         unsigned i;
648         u64 *pt;
649         u64 ent;
650
651         pt = page->spt;
652
653         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
654                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
655                         if (pt[i] & PT_PRESENT_MASK)
656                                 rmap_remove(vcpu, &pt[i]);
657                         pt[i] = 0;
658                 }
659                 kvm_arch_ops->tlb_flush(vcpu);
660                 return;
661         }
662
663         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
664                 ent = pt[i];
665
666                 pt[i] = 0;
667                 if (!(ent & PT_PRESENT_MASK))
668                         continue;
669                 ent &= PT64_BASE_ADDR_MASK;
670                 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
671         }
672 }
673
674 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
675                              struct kvm_mmu_page *page,
676                              u64 *parent_pte)
677 {
678         mmu_page_remove_parent_pte(vcpu, page, parent_pte);
679 }
680
681 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
682                              struct kvm_mmu_page *page)
683 {
684         u64 *parent_pte;
685
686         while (page->multimapped || page->parent_pte) {
687                 if (!page->multimapped)
688                         parent_pte = page->parent_pte;
689                 else {
690                         struct kvm_pte_chain *chain;
691
692                         chain = container_of(page->parent_ptes.first,
693                                              struct kvm_pte_chain, link);
694                         parent_pte = chain->parent_ptes[0];
695                 }
696                 BUG_ON(!parent_pte);
697                 kvm_mmu_put_page(vcpu, page, parent_pte);
698                 set_shadow_pte(parent_pte, 0);
699         }
700         kvm_mmu_page_unlink_children(vcpu, page);
701         if (!page->root_count) {
702                 hlist_del(&page->hash_link);
703                 kvm_mmu_free_page(vcpu, page);
704         } else
705                 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
706 }
707
708 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
709 {
710         unsigned index;
711         struct hlist_head *bucket;
712         struct kvm_mmu_page *page;
713         struct hlist_node *node, *n;
714         int r;
715
716         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
717         r = 0;
718         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
719         bucket = &vcpu->kvm->mmu_page_hash[index];
720         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
721                 if (page->gfn == gfn && !page->role.metaphysical) {
722                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
723                                  page->role.word);
724                         kvm_mmu_zap_page(vcpu, page);
725                         r = 1;
726                 }
727         return r;
728 }
729
730 static void mmu_unshadow(struct kvm_vcpu *vcpu, gfn_t gfn)
731 {
732         struct kvm_mmu_page *page;
733
734         while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
735                 pgprintk("%s: zap %lx %x\n",
736                          __FUNCTION__, gfn, page->role.word);
737                 kvm_mmu_zap_page(vcpu, page);
738         }
739 }
740
741 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
742 {
743         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
744         struct kvm_mmu_page *page_head = page_header(__pa(pte));
745
746         __set_bit(slot, &page_head->slot_bitmap);
747 }
748
749 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
750 {
751         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
752
753         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
754 }
755
756 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
757 {
758         struct page *page;
759
760         ASSERT((gpa & HPA_ERR_MASK) == 0);
761         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
762         if (!page)
763                 return gpa | HPA_ERR_MASK;
764         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
765                 | (gpa & (PAGE_SIZE-1));
766 }
767
768 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
769 {
770         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
771
772         if (gpa == UNMAPPED_GVA)
773                 return UNMAPPED_GVA;
774         return gpa_to_hpa(vcpu, gpa);
775 }
776
777 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
778 {
779         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
780
781         if (gpa == UNMAPPED_GVA)
782                 return NULL;
783         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
784 }
785
786 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
787 {
788 }
789
790 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
791 {
792         int level = PT32E_ROOT_LEVEL;
793         hpa_t table_addr = vcpu->mmu.root_hpa;
794
795         for (; ; level--) {
796                 u32 index = PT64_INDEX(v, level);
797                 u64 *table;
798                 u64 pte;
799
800                 ASSERT(VALID_PAGE(table_addr));
801                 table = __va(table_addr);
802
803                 if (level == 1) {
804                         pte = table[index];
805                         if (is_present_pte(pte) && is_writeble_pte(pte))
806                                 return 0;
807                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
808                         page_header_update_slot(vcpu->kvm, table, v);
809                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
810                                                                 PT_USER_MASK;
811                         rmap_add(vcpu, &table[index]);
812                         return 0;
813                 }
814
815                 if (table[index] == 0) {
816                         struct kvm_mmu_page *new_table;
817                         gfn_t pseudo_gfn;
818
819                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
820                                 >> PAGE_SHIFT;
821                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
822                                                      v, level - 1,
823                                                      1, 0, &table[index]);
824                         if (!new_table) {
825                                 pgprintk("nonpaging_map: ENOMEM\n");
826                                 return -ENOMEM;
827                         }
828
829                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
830                                 | PT_WRITABLE_MASK | PT_USER_MASK;
831                 }
832                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
833         }
834 }
835
836 static void mmu_free_roots(struct kvm_vcpu *vcpu)
837 {
838         int i;
839         struct kvm_mmu_page *page;
840
841 #ifdef CONFIG_X86_64
842         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
843                 hpa_t root = vcpu->mmu.root_hpa;
844
845                 ASSERT(VALID_PAGE(root));
846                 page = page_header(root);
847                 --page->root_count;
848                 vcpu->mmu.root_hpa = INVALID_PAGE;
849                 return;
850         }
851 #endif
852         for (i = 0; i < 4; ++i) {
853                 hpa_t root = vcpu->mmu.pae_root[i];
854
855                 if (root) {
856                         ASSERT(VALID_PAGE(root));
857                         root &= PT64_BASE_ADDR_MASK;
858                         page = page_header(root);
859                         --page->root_count;
860                 }
861                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
862         }
863         vcpu->mmu.root_hpa = INVALID_PAGE;
864 }
865
866 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
867 {
868         int i;
869         gfn_t root_gfn;
870         struct kvm_mmu_page *page;
871
872         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
873
874 #ifdef CONFIG_X86_64
875         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
876                 hpa_t root = vcpu->mmu.root_hpa;
877
878                 ASSERT(!VALID_PAGE(root));
879                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
880                                         PT64_ROOT_LEVEL, 0, 0, NULL);
881                 root = __pa(page->spt);
882                 ++page->root_count;
883                 vcpu->mmu.root_hpa = root;
884                 return;
885         }
886 #endif
887         for (i = 0; i < 4; ++i) {
888                 hpa_t root = vcpu->mmu.pae_root[i];
889
890                 ASSERT(!VALID_PAGE(root));
891                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
892                         if (!is_present_pte(vcpu->pdptrs[i])) {
893                                 vcpu->mmu.pae_root[i] = 0;
894                                 continue;
895                         }
896                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
897                 } else if (vcpu->mmu.root_level == 0)
898                         root_gfn = 0;
899                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
900                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
901                                         0, NULL);
902                 root = __pa(page->spt);
903                 ++page->root_count;
904                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
905         }
906         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
907 }
908
909 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
910 {
911         return vaddr;
912 }
913
914 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
915                                u32 error_code)
916 {
917         gpa_t addr = gva;
918         hpa_t paddr;
919         int r;
920
921         r = mmu_topup_memory_caches(vcpu);
922         if (r)
923                 return r;
924
925         ASSERT(vcpu);
926         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
927
928
929         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
930
931         if (is_error_hpa(paddr))
932                 return 1;
933
934         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
935 }
936
937 static void nonpaging_free(struct kvm_vcpu *vcpu)
938 {
939         mmu_free_roots(vcpu);
940 }
941
942 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
943 {
944         struct kvm_mmu *context = &vcpu->mmu;
945
946         context->new_cr3 = nonpaging_new_cr3;
947         context->page_fault = nonpaging_page_fault;
948         context->gva_to_gpa = nonpaging_gva_to_gpa;
949         context->free = nonpaging_free;
950         context->root_level = 0;
951         context->shadow_root_level = PT32E_ROOT_LEVEL;
952         context->root_hpa = INVALID_PAGE;
953         return 0;
954 }
955
956 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
957 {
958         ++vcpu->stat.tlb_flush;
959         kvm_arch_ops->tlb_flush(vcpu);
960 }
961
962 static void paging_new_cr3(struct kvm_vcpu *vcpu)
963 {
964         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
965         mmu_free_roots(vcpu);
966 }
967
968 static void inject_page_fault(struct kvm_vcpu *vcpu,
969                               u64 addr,
970                               u32 err_code)
971 {
972         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
973 }
974
975 static void paging_free(struct kvm_vcpu *vcpu)
976 {
977         nonpaging_free(vcpu);
978 }
979
980 #define PTTYPE 64
981 #include "paging_tmpl.h"
982 #undef PTTYPE
983
984 #define PTTYPE 32
985 #include "paging_tmpl.h"
986 #undef PTTYPE
987
988 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
989 {
990         struct kvm_mmu *context = &vcpu->mmu;
991
992         ASSERT(is_pae(vcpu));
993         context->new_cr3 = paging_new_cr3;
994         context->page_fault = paging64_page_fault;
995         context->gva_to_gpa = paging64_gva_to_gpa;
996         context->free = paging_free;
997         context->root_level = level;
998         context->shadow_root_level = level;
999         context->root_hpa = INVALID_PAGE;
1000         return 0;
1001 }
1002
1003 static int paging64_init_context(struct kvm_vcpu *vcpu)
1004 {
1005         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1006 }
1007
1008 static int paging32_init_context(struct kvm_vcpu *vcpu)
1009 {
1010         struct kvm_mmu *context = &vcpu->mmu;
1011
1012         context->new_cr3 = paging_new_cr3;
1013         context->page_fault = paging32_page_fault;
1014         context->gva_to_gpa = paging32_gva_to_gpa;
1015         context->free = paging_free;
1016         context->root_level = PT32_ROOT_LEVEL;
1017         context->shadow_root_level = PT32E_ROOT_LEVEL;
1018         context->root_hpa = INVALID_PAGE;
1019         return 0;
1020 }
1021
1022 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1023 {
1024         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1025 }
1026
1027 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1028 {
1029         ASSERT(vcpu);
1030         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1031
1032         if (!is_paging(vcpu))
1033                 return nonpaging_init_context(vcpu);
1034         else if (is_long_mode(vcpu))
1035                 return paging64_init_context(vcpu);
1036         else if (is_pae(vcpu))
1037                 return paging32E_init_context(vcpu);
1038         else
1039                 return paging32_init_context(vcpu);
1040 }
1041
1042 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1043 {
1044         ASSERT(vcpu);
1045         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1046                 vcpu->mmu.free(vcpu);
1047                 vcpu->mmu.root_hpa = INVALID_PAGE;
1048         }
1049 }
1050
1051 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1052 {
1053         destroy_kvm_mmu(vcpu);
1054         return init_kvm_mmu(vcpu);
1055 }
1056
1057 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1058 {
1059         int r;
1060
1061         spin_lock(&vcpu->kvm->lock);
1062         r = mmu_topup_memory_caches(vcpu);
1063         if (r)
1064                 goto out;
1065         mmu_alloc_roots(vcpu);
1066         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1067         kvm_mmu_flush_tlb(vcpu);
1068 out:
1069         spin_unlock(&vcpu->kvm->lock);
1070         return r;
1071 }
1072 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1073
1074 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1075 {
1076         mmu_free_roots(vcpu);
1077 }
1078
1079 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1080                                   struct kvm_mmu_page *page,
1081                                   u64 *spte)
1082 {
1083         u64 pte;
1084         struct kvm_mmu_page *child;
1085
1086         pte = *spte;
1087         if (is_present_pte(pte)) {
1088                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1089                         rmap_remove(vcpu, spte);
1090                 else {
1091                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1092                         mmu_page_remove_parent_pte(vcpu, child, spte);
1093                 }
1094         }
1095         *spte = 0;
1096 }
1097
1098 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1099                                   struct kvm_mmu_page *page,
1100                                   u64 *spte,
1101                                   const void *new, int bytes)
1102 {
1103         if (page->role.level != PT_PAGE_TABLE_LEVEL)
1104                 return;
1105
1106         if (page->role.glevels == PT32_ROOT_LEVEL)
1107                 paging32_update_pte(vcpu, page, spte, new, bytes);
1108         else
1109                 paging64_update_pte(vcpu, page, spte, new, bytes);
1110 }
1111
1112 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1113                        const u8 *old, const u8 *new, int bytes)
1114 {
1115         gfn_t gfn = gpa >> PAGE_SHIFT;
1116         struct kvm_mmu_page *page;
1117         struct hlist_node *node, *n;
1118         struct hlist_head *bucket;
1119         unsigned index;
1120         u64 *spte;
1121         unsigned offset = offset_in_page(gpa);
1122         unsigned pte_size;
1123         unsigned page_offset;
1124         unsigned misaligned;
1125         unsigned quadrant;
1126         int level;
1127         int flooded = 0;
1128         int npte;
1129
1130         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1131         if (gfn == vcpu->last_pt_write_gfn) {
1132                 ++vcpu->last_pt_write_count;
1133                 if (vcpu->last_pt_write_count >= 3)
1134                         flooded = 1;
1135         } else {
1136                 vcpu->last_pt_write_gfn = gfn;
1137                 vcpu->last_pt_write_count = 1;
1138         }
1139         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1140         bucket = &vcpu->kvm->mmu_page_hash[index];
1141         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1142                 if (page->gfn != gfn || page->role.metaphysical)
1143                         continue;
1144                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1145                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1146                 misaligned |= bytes < 4;
1147                 if (misaligned || flooded) {
1148                         /*
1149                          * Misaligned accesses are too much trouble to fix
1150                          * up; also, they usually indicate a page is not used
1151                          * as a page table.
1152                          *
1153                          * If we're seeing too many writes to a page,
1154                          * it may no longer be a page table, or we may be
1155                          * forking, in which case it is better to unmap the
1156                          * page.
1157                          */
1158                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1159                                  gpa, bytes, page->role.word);
1160                         kvm_mmu_zap_page(vcpu, page);
1161                         continue;
1162                 }
1163                 page_offset = offset;
1164                 level = page->role.level;
1165                 npte = 1;
1166                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1167                         page_offset <<= 1;      /* 32->64 */
1168                         /*
1169                          * A 32-bit pde maps 4MB while the shadow pdes map
1170                          * only 2MB.  So we need to double the offset again
1171                          * and zap two pdes instead of one.
1172                          */
1173                         if (level == PT32_ROOT_LEVEL) {
1174                                 page_offset &= ~7; /* kill rounding error */
1175                                 page_offset <<= 1;
1176                                 npte = 2;
1177                         }
1178                         quadrant = page_offset >> PAGE_SHIFT;
1179                         page_offset &= ~PAGE_MASK;
1180                         if (quadrant != page->role.quadrant)
1181                                 continue;
1182                 }
1183                 spte = &page->spt[page_offset / sizeof(*spte)];
1184                 while (npte--) {
1185                         mmu_pte_write_zap_pte(vcpu, page, spte);
1186                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes);
1187                         ++spte;
1188                 }
1189         }
1190 }
1191
1192 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1193 {
1194         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1195
1196         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1197 }
1198
1199 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1200 {
1201         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1202                 struct kvm_mmu_page *page;
1203
1204                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1205                                     struct kvm_mmu_page, link);
1206                 kvm_mmu_zap_page(vcpu, page);
1207         }
1208 }
1209 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1210
1211 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1212 {
1213         struct kvm_mmu_page *page;
1214
1215         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1216                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1217                                     struct kvm_mmu_page, link);
1218                 kvm_mmu_zap_page(vcpu, page);
1219         }
1220         free_page((unsigned long)vcpu->mmu.pae_root);
1221 }
1222
1223 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1224 {
1225         struct page *page;
1226         int i;
1227
1228         ASSERT(vcpu);
1229
1230         vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1231
1232         /*
1233          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1234          * Therefore we need to allocate shadow page tables in the first
1235          * 4GB of memory, which happens to fit the DMA32 zone.
1236          */
1237         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1238         if (!page)
1239                 goto error_1;
1240         vcpu->mmu.pae_root = page_address(page);
1241         for (i = 0; i < 4; ++i)
1242                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1243
1244         return 0;
1245
1246 error_1:
1247         free_mmu_pages(vcpu);
1248         return -ENOMEM;
1249 }
1250
1251 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1252 {
1253         ASSERT(vcpu);
1254         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1255
1256         return alloc_mmu_pages(vcpu);
1257 }
1258
1259 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1260 {
1261         ASSERT(vcpu);
1262         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1263
1264         return init_kvm_mmu(vcpu);
1265 }
1266
1267 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1268 {
1269         ASSERT(vcpu);
1270
1271         destroy_kvm_mmu(vcpu);
1272         free_mmu_pages(vcpu);
1273         mmu_free_memory_caches(vcpu);
1274 }
1275
1276 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1277 {
1278         struct kvm *kvm = vcpu->kvm;
1279         struct kvm_mmu_page *page;
1280
1281         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1282                 int i;
1283                 u64 *pt;
1284
1285                 if (!test_bit(slot, &page->slot_bitmap))
1286                         continue;
1287
1288                 pt = page->spt;
1289                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1290                         /* avoid RMW */
1291                         if (pt[i] & PT_WRITABLE_MASK) {
1292                                 rmap_remove(vcpu, &pt[i]);
1293                                 pt[i] &= ~PT_WRITABLE_MASK;
1294                         }
1295         }
1296 }
1297
1298 void kvm_mmu_zap_all(struct kvm_vcpu *vcpu)
1299 {
1300         destroy_kvm_mmu(vcpu);
1301
1302         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1303                 struct kvm_mmu_page *page;
1304
1305                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1306                                     struct kvm_mmu_page, link);
1307                 kvm_mmu_zap_page(vcpu, page);
1308         }
1309
1310         mmu_free_memory_caches(vcpu);
1311         kvm_arch_ops->tlb_flush(vcpu);
1312         init_kvm_mmu(vcpu);
1313 }
1314
1315 void kvm_mmu_module_exit(void)
1316 {
1317         if (pte_chain_cache)
1318                 kmem_cache_destroy(pte_chain_cache);
1319         if (rmap_desc_cache)
1320                 kmem_cache_destroy(rmap_desc_cache);
1321         if (mmu_page_cache)
1322                 kmem_cache_destroy(mmu_page_cache);
1323         if (mmu_page_header_cache)
1324                 kmem_cache_destroy(mmu_page_header_cache);
1325 }
1326
1327 int kvm_mmu_module_init(void)
1328 {
1329         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1330                                             sizeof(struct kvm_pte_chain),
1331                                             0, 0, NULL, NULL);
1332         if (!pte_chain_cache)
1333                 goto nomem;
1334         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1335                                             sizeof(struct kvm_rmap_desc),
1336                                             0, 0, NULL, NULL);
1337         if (!rmap_desc_cache)
1338                 goto nomem;
1339
1340         mmu_page_cache = kmem_cache_create("kvm_mmu_page",
1341                                            PAGE_SIZE,
1342                                            PAGE_SIZE, 0, NULL, NULL);
1343         if (!mmu_page_cache)
1344                 goto nomem;
1345
1346         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1347                                                   sizeof(struct kvm_mmu_page),
1348                                                   0, 0, NULL, NULL);
1349         if (!mmu_page_header_cache)
1350                 goto nomem;
1351
1352         return 0;
1353
1354 nomem:
1355         kvm_mmu_module_exit();
1356         return -ENOMEM;
1357 }
1358
1359 #ifdef AUDIT
1360
1361 static const char *audit_msg;
1362
1363 static gva_t canonicalize(gva_t gva)
1364 {
1365 #ifdef CONFIG_X86_64
1366         gva = (long long)(gva << 16) >> 16;
1367 #endif
1368         return gva;
1369 }
1370
1371 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1372                                 gva_t va, int level)
1373 {
1374         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1375         int i;
1376         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1377
1378         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1379                 u64 ent = pt[i];
1380
1381                 if (!(ent & PT_PRESENT_MASK))
1382                         continue;
1383
1384                 va = canonicalize(va);
1385                 if (level > 1)
1386                         audit_mappings_page(vcpu, ent, va, level - 1);
1387                 else {
1388                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1389                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1390
1391                         if ((ent & PT_PRESENT_MASK)
1392                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1393                                 printk(KERN_ERR "audit error: (%s) levels %d"
1394                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1395                                        audit_msg, vcpu->mmu.root_level,
1396                                        va, gpa, hpa, ent);
1397                 }
1398         }
1399 }
1400
1401 static void audit_mappings(struct kvm_vcpu *vcpu)
1402 {
1403         unsigned i;
1404
1405         if (vcpu->mmu.root_level == 4)
1406                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1407         else
1408                 for (i = 0; i < 4; ++i)
1409                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1410                                 audit_mappings_page(vcpu,
1411                                                     vcpu->mmu.pae_root[i],
1412                                                     i << 30,
1413                                                     2);
1414 }
1415
1416 static int count_rmaps(struct kvm_vcpu *vcpu)
1417 {
1418         int nmaps = 0;
1419         int i, j, k;
1420
1421         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1422                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1423                 struct kvm_rmap_desc *d;
1424
1425                 for (j = 0; j < m->npages; ++j) {
1426                         struct page *page = m->phys_mem[j];
1427
1428                         if (!page->private)
1429                                 continue;
1430                         if (!(page->private & 1)) {
1431                                 ++nmaps;
1432                                 continue;
1433                         }
1434                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1435                         while (d) {
1436                                 for (k = 0; k < RMAP_EXT; ++k)
1437                                         if (d->shadow_ptes[k])
1438                                                 ++nmaps;
1439                                         else
1440                                                 break;
1441                                 d = d->more;
1442                         }
1443                 }
1444         }
1445         return nmaps;
1446 }
1447
1448 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1449 {
1450         int nmaps = 0;
1451         struct kvm_mmu_page *page;
1452         int i;
1453
1454         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1455                 u64 *pt = page->spt;
1456
1457                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1458                         continue;
1459
1460                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1461                         u64 ent = pt[i];
1462
1463                         if (!(ent & PT_PRESENT_MASK))
1464                                 continue;
1465                         if (!(ent & PT_WRITABLE_MASK))
1466                                 continue;
1467                         ++nmaps;
1468                 }
1469         }
1470         return nmaps;
1471 }
1472
1473 static void audit_rmap(struct kvm_vcpu *vcpu)
1474 {
1475         int n_rmap = count_rmaps(vcpu);
1476         int n_actual = count_writable_mappings(vcpu);
1477
1478         if (n_rmap != n_actual)
1479                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1480                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1481 }
1482
1483 static void audit_write_protection(struct kvm_vcpu *vcpu)
1484 {
1485         struct kvm_mmu_page *page;
1486
1487         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1488                 hfn_t hfn;
1489                 struct page *pg;
1490
1491                 if (page->role.metaphysical)
1492                         continue;
1493
1494                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1495                         >> PAGE_SHIFT;
1496                 pg = pfn_to_page(hfn);
1497                 if (pg->private)
1498                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1499                                " mappings: gfn %lx role %x\n",
1500                                __FUNCTION__, audit_msg, page->gfn,
1501                                page->role.word);
1502         }
1503 }
1504
1505 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1506 {
1507         int olddbg = dbg;
1508
1509         dbg = 0;
1510         audit_msg = msg;
1511         audit_rmap(vcpu);
1512         audit_write_protection(vcpu);
1513         audit_mappings(vcpu);
1514         dbg = olddbg;
1515 }
1516
1517 #endif