Merge branch 'pxa-devel' into pxa
[linux-2.6] / arch / x86 / 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
20 #include "vmx.h"
21 #include "mmu.h"
22
23 #include <linux/kvm_host.h>
24 #include <linux/types.h>
25 #include <linux/string.h>
26 #include <linux/mm.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>
32
33 #include <asm/page.h>
34 #include <asm/cmpxchg.h>
35 #include <asm/io.h>
36
37 /*
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.
43  */
44 bool tdp_enabled = false;
45
46 #undef MMU_DEBUG
47
48 #undef AUDIT
49
50 #ifdef AUDIT
51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
52 #else
53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
54 #endif
55
56 #ifdef MMU_DEBUG
57
58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
60
61 #else
62
63 #define pgprintk(x...) do { } while (0)
64 #define rmap_printk(x...) do { } while (0)
65
66 #endif
67
68 #if defined(MMU_DEBUG) || defined(AUDIT)
69 static int dbg = 1;
70 #endif
71
72 #ifndef MMU_DEBUG
73 #define ASSERT(x) do { } while (0)
74 #else
75 #define ASSERT(x)                                                       \
76         if (!(x)) {                                                     \
77                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
78                        __FILE__, __LINE__, #x);                         \
79         }
80 #endif
81
82 #define PT_FIRST_AVAIL_BITS_SHIFT 9
83 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
84
85 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
86
87 #define PT64_LEVEL_BITS 9
88
89 #define PT64_LEVEL_SHIFT(level) \
90                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
91
92 #define PT64_LEVEL_MASK(level) \
93                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
94
95 #define PT64_INDEX(address, level)\
96         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
97
98
99 #define PT32_LEVEL_BITS 10
100
101 #define PT32_LEVEL_SHIFT(level) \
102                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
103
104 #define PT32_LEVEL_MASK(level) \
105                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
106
107 #define PT32_INDEX(address, level)\
108         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
109
110
111 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
112 #define PT64_DIR_BASE_ADDR_MASK \
113         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
114
115 #define PT32_BASE_ADDR_MASK PAGE_MASK
116 #define PT32_DIR_BASE_ADDR_MASK \
117         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
118
119 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
120                         | PT64_NX_MASK)
121
122 #define PFERR_PRESENT_MASK (1U << 0)
123 #define PFERR_WRITE_MASK (1U << 1)
124 #define PFERR_USER_MASK (1U << 2)
125 #define PFERR_FETCH_MASK (1U << 4)
126
127 #define PT_DIRECTORY_LEVEL 2
128 #define PT_PAGE_TABLE_LEVEL 1
129
130 #define RMAP_EXT 4
131
132 #define ACC_EXEC_MASK    1
133 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
134 #define ACC_USER_MASK    PT_USER_MASK
135 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
136
137 struct kvm_pv_mmu_op_buffer {
138         void *ptr;
139         unsigned len;
140         unsigned processed;
141         char buf[512] __aligned(sizeof(long));
142 };
143
144 struct kvm_rmap_desc {
145         u64 *shadow_ptes[RMAP_EXT];
146         struct kvm_rmap_desc *more;
147 };
148
149 static struct kmem_cache *pte_chain_cache;
150 static struct kmem_cache *rmap_desc_cache;
151 static struct kmem_cache *mmu_page_header_cache;
152
153 static u64 __read_mostly shadow_trap_nonpresent_pte;
154 static u64 __read_mostly shadow_notrap_nonpresent_pte;
155 static u64 __read_mostly shadow_base_present_pte;
156 static u64 __read_mostly shadow_nx_mask;
157 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
158 static u64 __read_mostly shadow_user_mask;
159 static u64 __read_mostly shadow_accessed_mask;
160 static u64 __read_mostly shadow_dirty_mask;
161
162 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
163 {
164         shadow_trap_nonpresent_pte = trap_pte;
165         shadow_notrap_nonpresent_pte = notrap_pte;
166 }
167 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
168
169 void kvm_mmu_set_base_ptes(u64 base_pte)
170 {
171         shadow_base_present_pte = base_pte;
172 }
173 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
174
175 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
176                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
177 {
178         shadow_user_mask = user_mask;
179         shadow_accessed_mask = accessed_mask;
180         shadow_dirty_mask = dirty_mask;
181         shadow_nx_mask = nx_mask;
182         shadow_x_mask = x_mask;
183 }
184 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
185
186 static int is_write_protection(struct kvm_vcpu *vcpu)
187 {
188         return vcpu->arch.cr0 & X86_CR0_WP;
189 }
190
191 static int is_cpuid_PSE36(void)
192 {
193         return 1;
194 }
195
196 static int is_nx(struct kvm_vcpu *vcpu)
197 {
198         return vcpu->arch.shadow_efer & EFER_NX;
199 }
200
201 static int is_present_pte(unsigned long pte)
202 {
203         return pte & PT_PRESENT_MASK;
204 }
205
206 static int is_shadow_present_pte(u64 pte)
207 {
208         return pte != shadow_trap_nonpresent_pte
209                 && pte != shadow_notrap_nonpresent_pte;
210 }
211
212 static int is_large_pte(u64 pte)
213 {
214         return pte & PT_PAGE_SIZE_MASK;
215 }
216
217 static int is_writeble_pte(unsigned long pte)
218 {
219         return pte & PT_WRITABLE_MASK;
220 }
221
222 static int is_dirty_pte(unsigned long pte)
223 {
224         return pte & shadow_dirty_mask;
225 }
226
227 static int is_rmap_pte(u64 pte)
228 {
229         return is_shadow_present_pte(pte);
230 }
231
232 static pfn_t spte_to_pfn(u64 pte)
233 {
234         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
235 }
236
237 static gfn_t pse36_gfn_delta(u32 gpte)
238 {
239         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
240
241         return (gpte & PT32_DIR_PSE36_MASK) << shift;
242 }
243
244 static void set_shadow_pte(u64 *sptep, u64 spte)
245 {
246 #ifdef CONFIG_X86_64
247         set_64bit((unsigned long *)sptep, spte);
248 #else
249         set_64bit((unsigned long long *)sptep, spte);
250 #endif
251 }
252
253 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
254                                   struct kmem_cache *base_cache, int min)
255 {
256         void *obj;
257
258         if (cache->nobjs >= min)
259                 return 0;
260         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
261                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
262                 if (!obj)
263                         return -ENOMEM;
264                 cache->objects[cache->nobjs++] = obj;
265         }
266         return 0;
267 }
268
269 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
270 {
271         while (mc->nobjs)
272                 kfree(mc->objects[--mc->nobjs]);
273 }
274
275 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
276                                        int min)
277 {
278         struct page *page;
279
280         if (cache->nobjs >= min)
281                 return 0;
282         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
283                 page = alloc_page(GFP_KERNEL);
284                 if (!page)
285                         return -ENOMEM;
286                 set_page_private(page, 0);
287                 cache->objects[cache->nobjs++] = page_address(page);
288         }
289         return 0;
290 }
291
292 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
293 {
294         while (mc->nobjs)
295                 free_page((unsigned long)mc->objects[--mc->nobjs]);
296 }
297
298 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
299 {
300         int r;
301
302         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
303                                    pte_chain_cache, 4);
304         if (r)
305                 goto out;
306         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
307                                    rmap_desc_cache, 1);
308         if (r)
309                 goto out;
310         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
311         if (r)
312                 goto out;
313         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
314                                    mmu_page_header_cache, 4);
315 out:
316         return r;
317 }
318
319 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
320 {
321         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
322         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
323         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
324         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
325 }
326
327 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
328                                     size_t size)
329 {
330         void *p;
331
332         BUG_ON(!mc->nobjs);
333         p = mc->objects[--mc->nobjs];
334         memset(p, 0, size);
335         return p;
336 }
337
338 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
339 {
340         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
341                                       sizeof(struct kvm_pte_chain));
342 }
343
344 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
345 {
346         kfree(pc);
347 }
348
349 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
350 {
351         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
352                                       sizeof(struct kvm_rmap_desc));
353 }
354
355 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
356 {
357         kfree(rd);
358 }
359
360 /*
361  * Return the pointer to the largepage write count for a given
362  * gfn, handling slots that are not large page aligned.
363  */
364 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
365 {
366         unsigned long idx;
367
368         idx = (gfn / KVM_PAGES_PER_HPAGE) -
369               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
370         return &slot->lpage_info[idx].write_count;
371 }
372
373 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
374 {
375         int *write_count;
376
377         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
378         *write_count += 1;
379 }
380
381 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
382 {
383         int *write_count;
384
385         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
386         *write_count -= 1;
387         WARN_ON(*write_count < 0);
388 }
389
390 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
391 {
392         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
393         int *largepage_idx;
394
395         if (slot) {
396                 largepage_idx = slot_largepage_idx(gfn, slot);
397                 return *largepage_idx;
398         }
399
400         return 1;
401 }
402
403 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
404 {
405         struct vm_area_struct *vma;
406         unsigned long addr;
407
408         addr = gfn_to_hva(kvm, gfn);
409         if (kvm_is_error_hva(addr))
410                 return 0;
411
412         vma = find_vma(current->mm, addr);
413         if (vma && is_vm_hugetlb_page(vma))
414                 return 1;
415
416         return 0;
417 }
418
419 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
420 {
421         struct kvm_memory_slot *slot;
422
423         if (has_wrprotected_page(vcpu->kvm, large_gfn))
424                 return 0;
425
426         if (!host_largepage_backed(vcpu->kvm, large_gfn))
427                 return 0;
428
429         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
430         if (slot && slot->dirty_bitmap)
431                 return 0;
432
433         return 1;
434 }
435
436 /*
437  * Take gfn and return the reverse mapping to it.
438  * Note: gfn must be unaliased before this function get called
439  */
440
441 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
442 {
443         struct kvm_memory_slot *slot;
444         unsigned long idx;
445
446         slot = gfn_to_memslot(kvm, gfn);
447         if (!lpage)
448                 return &slot->rmap[gfn - slot->base_gfn];
449
450         idx = (gfn / KVM_PAGES_PER_HPAGE) -
451               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
452
453         return &slot->lpage_info[idx].rmap_pde;
454 }
455
456 /*
457  * Reverse mapping data structures:
458  *
459  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
460  * that points to page_address(page).
461  *
462  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
463  * containing more mappings.
464  */
465 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
466 {
467         struct kvm_mmu_page *sp;
468         struct kvm_rmap_desc *desc;
469         unsigned long *rmapp;
470         int i;
471
472         if (!is_rmap_pte(*spte))
473                 return;
474         gfn = unalias_gfn(vcpu->kvm, gfn);
475         sp = page_header(__pa(spte));
476         sp->gfns[spte - sp->spt] = gfn;
477         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
478         if (!*rmapp) {
479                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
480                 *rmapp = (unsigned long)spte;
481         } else if (!(*rmapp & 1)) {
482                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
483                 desc = mmu_alloc_rmap_desc(vcpu);
484                 desc->shadow_ptes[0] = (u64 *)*rmapp;
485                 desc->shadow_ptes[1] = spte;
486                 *rmapp = (unsigned long)desc | 1;
487         } else {
488                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
489                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
490                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
491                         desc = desc->more;
492                 if (desc->shadow_ptes[RMAP_EXT-1]) {
493                         desc->more = mmu_alloc_rmap_desc(vcpu);
494                         desc = desc->more;
495                 }
496                 for (i = 0; desc->shadow_ptes[i]; ++i)
497                         ;
498                 desc->shadow_ptes[i] = spte;
499         }
500 }
501
502 static void rmap_desc_remove_entry(unsigned long *rmapp,
503                                    struct kvm_rmap_desc *desc,
504                                    int i,
505                                    struct kvm_rmap_desc *prev_desc)
506 {
507         int j;
508
509         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
510                 ;
511         desc->shadow_ptes[i] = desc->shadow_ptes[j];
512         desc->shadow_ptes[j] = NULL;
513         if (j != 0)
514                 return;
515         if (!prev_desc && !desc->more)
516                 *rmapp = (unsigned long)desc->shadow_ptes[0];
517         else
518                 if (prev_desc)
519                         prev_desc->more = desc->more;
520                 else
521                         *rmapp = (unsigned long)desc->more | 1;
522         mmu_free_rmap_desc(desc);
523 }
524
525 static void rmap_remove(struct kvm *kvm, u64 *spte)
526 {
527         struct kvm_rmap_desc *desc;
528         struct kvm_rmap_desc *prev_desc;
529         struct kvm_mmu_page *sp;
530         pfn_t pfn;
531         unsigned long *rmapp;
532         int i;
533
534         if (!is_rmap_pte(*spte))
535                 return;
536         sp = page_header(__pa(spte));
537         pfn = spte_to_pfn(*spte);
538         if (*spte & shadow_accessed_mask)
539                 kvm_set_pfn_accessed(pfn);
540         if (is_writeble_pte(*spte))
541                 kvm_release_pfn_dirty(pfn);
542         else
543                 kvm_release_pfn_clean(pfn);
544         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
545         if (!*rmapp) {
546                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
547                 BUG();
548         } else if (!(*rmapp & 1)) {
549                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
550                 if ((u64 *)*rmapp != spte) {
551                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
552                                spte, *spte);
553                         BUG();
554                 }
555                 *rmapp = 0;
556         } else {
557                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
558                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
559                 prev_desc = NULL;
560                 while (desc) {
561                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
562                                 if (desc->shadow_ptes[i] == spte) {
563                                         rmap_desc_remove_entry(rmapp,
564                                                                desc, i,
565                                                                prev_desc);
566                                         return;
567                                 }
568                         prev_desc = desc;
569                         desc = desc->more;
570                 }
571                 BUG();
572         }
573 }
574
575 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
576 {
577         struct kvm_rmap_desc *desc;
578         struct kvm_rmap_desc *prev_desc;
579         u64 *prev_spte;
580         int i;
581
582         if (!*rmapp)
583                 return NULL;
584         else if (!(*rmapp & 1)) {
585                 if (!spte)
586                         return (u64 *)*rmapp;
587                 return NULL;
588         }
589         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
590         prev_desc = NULL;
591         prev_spte = NULL;
592         while (desc) {
593                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
594                         if (prev_spte == spte)
595                                 return desc->shadow_ptes[i];
596                         prev_spte = desc->shadow_ptes[i];
597                 }
598                 desc = desc->more;
599         }
600         return NULL;
601 }
602
603 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
604 {
605         unsigned long *rmapp;
606         u64 *spte;
607         int write_protected = 0;
608
609         gfn = unalias_gfn(kvm, gfn);
610         rmapp = gfn_to_rmap(kvm, gfn, 0);
611
612         spte = rmap_next(kvm, rmapp, NULL);
613         while (spte) {
614                 BUG_ON(!spte);
615                 BUG_ON(!(*spte & PT_PRESENT_MASK));
616                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
617                 if (is_writeble_pte(*spte)) {
618                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
619                         write_protected = 1;
620                 }
621                 spte = rmap_next(kvm, rmapp, spte);
622         }
623         if (write_protected) {
624                 pfn_t pfn;
625
626                 spte = rmap_next(kvm, rmapp, NULL);
627                 pfn = spte_to_pfn(*spte);
628                 kvm_set_pfn_dirty(pfn);
629         }
630
631         /* check for huge page mappings */
632         rmapp = gfn_to_rmap(kvm, gfn, 1);
633         spte = rmap_next(kvm, rmapp, NULL);
634         while (spte) {
635                 BUG_ON(!spte);
636                 BUG_ON(!(*spte & PT_PRESENT_MASK));
637                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
638                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
639                 if (is_writeble_pte(*spte)) {
640                         rmap_remove(kvm, spte);
641                         --kvm->stat.lpages;
642                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
643                         spte = NULL;
644                         write_protected = 1;
645                 }
646                 spte = rmap_next(kvm, rmapp, spte);
647         }
648
649         if (write_protected)
650                 kvm_flush_remote_tlbs(kvm);
651
652         account_shadowed(kvm, gfn);
653 }
654
655 #ifdef MMU_DEBUG
656 static int is_empty_shadow_page(u64 *spt)
657 {
658         u64 *pos;
659         u64 *end;
660
661         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
662                 if (is_shadow_present_pte(*pos)) {
663                         printk(KERN_ERR "%s: %p %llx\n", __func__,
664                                pos, *pos);
665                         return 0;
666                 }
667         return 1;
668 }
669 #endif
670
671 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
672 {
673         ASSERT(is_empty_shadow_page(sp->spt));
674         list_del(&sp->link);
675         __free_page(virt_to_page(sp->spt));
676         __free_page(virt_to_page(sp->gfns));
677         kfree(sp);
678         ++kvm->arch.n_free_mmu_pages;
679 }
680
681 static unsigned kvm_page_table_hashfn(gfn_t gfn)
682 {
683         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
684 }
685
686 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
687                                                u64 *parent_pte)
688 {
689         struct kvm_mmu_page *sp;
690
691         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
692         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
693         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
694         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
695         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
696         ASSERT(is_empty_shadow_page(sp->spt));
697         sp->slot_bitmap = 0;
698         sp->multimapped = 0;
699         sp->parent_pte = parent_pte;
700         --vcpu->kvm->arch.n_free_mmu_pages;
701         return sp;
702 }
703
704 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
705                                     struct kvm_mmu_page *sp, u64 *parent_pte)
706 {
707         struct kvm_pte_chain *pte_chain;
708         struct hlist_node *node;
709         int i;
710
711         if (!parent_pte)
712                 return;
713         if (!sp->multimapped) {
714                 u64 *old = sp->parent_pte;
715
716                 if (!old) {
717                         sp->parent_pte = parent_pte;
718                         return;
719                 }
720                 sp->multimapped = 1;
721                 pte_chain = mmu_alloc_pte_chain(vcpu);
722                 INIT_HLIST_HEAD(&sp->parent_ptes);
723                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
724                 pte_chain->parent_ptes[0] = old;
725         }
726         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
727                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
728                         continue;
729                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
730                         if (!pte_chain->parent_ptes[i]) {
731                                 pte_chain->parent_ptes[i] = parent_pte;
732                                 return;
733                         }
734         }
735         pte_chain = mmu_alloc_pte_chain(vcpu);
736         BUG_ON(!pte_chain);
737         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
738         pte_chain->parent_ptes[0] = parent_pte;
739 }
740
741 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
742                                        u64 *parent_pte)
743 {
744         struct kvm_pte_chain *pte_chain;
745         struct hlist_node *node;
746         int i;
747
748         if (!sp->multimapped) {
749                 BUG_ON(sp->parent_pte != parent_pte);
750                 sp->parent_pte = NULL;
751                 return;
752         }
753         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
754                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
755                         if (!pte_chain->parent_ptes[i])
756                                 break;
757                         if (pte_chain->parent_ptes[i] != parent_pte)
758                                 continue;
759                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
760                                 && pte_chain->parent_ptes[i + 1]) {
761                                 pte_chain->parent_ptes[i]
762                                         = pte_chain->parent_ptes[i + 1];
763                                 ++i;
764                         }
765                         pte_chain->parent_ptes[i] = NULL;
766                         if (i == 0) {
767                                 hlist_del(&pte_chain->link);
768                                 mmu_free_pte_chain(pte_chain);
769                                 if (hlist_empty(&sp->parent_ptes)) {
770                                         sp->multimapped = 0;
771                                         sp->parent_pte = NULL;
772                                 }
773                         }
774                         return;
775                 }
776         BUG();
777 }
778
779 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
780 {
781         unsigned index;
782         struct hlist_head *bucket;
783         struct kvm_mmu_page *sp;
784         struct hlist_node *node;
785
786         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
787         index = kvm_page_table_hashfn(gfn);
788         bucket = &kvm->arch.mmu_page_hash[index];
789         hlist_for_each_entry(sp, node, bucket, hash_link)
790                 if (sp->gfn == gfn && !sp->role.metaphysical
791                     && !sp->role.invalid) {
792                         pgprintk("%s: found role %x\n",
793                                  __func__, sp->role.word);
794                         return sp;
795                 }
796         return NULL;
797 }
798
799 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
800                                              gfn_t gfn,
801                                              gva_t gaddr,
802                                              unsigned level,
803                                              int metaphysical,
804                                              unsigned access,
805                                              u64 *parent_pte)
806 {
807         union kvm_mmu_page_role role;
808         unsigned index;
809         unsigned quadrant;
810         struct hlist_head *bucket;
811         struct kvm_mmu_page *sp;
812         struct hlist_node *node;
813
814         role.word = 0;
815         role.glevels = vcpu->arch.mmu.root_level;
816         role.level = level;
817         role.metaphysical = metaphysical;
818         role.access = access;
819         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
820                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
821                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
822                 role.quadrant = quadrant;
823         }
824         pgprintk("%s: looking gfn %lx role %x\n", __func__,
825                  gfn, role.word);
826         index = kvm_page_table_hashfn(gfn);
827         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
828         hlist_for_each_entry(sp, node, bucket, hash_link)
829                 if (sp->gfn == gfn && sp->role.word == role.word) {
830                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
831                         pgprintk("%s: found\n", __func__);
832                         return sp;
833                 }
834         ++vcpu->kvm->stat.mmu_cache_miss;
835         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
836         if (!sp)
837                 return sp;
838         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
839         sp->gfn = gfn;
840         sp->role = role;
841         hlist_add_head(&sp->hash_link, bucket);
842         if (!metaphysical)
843                 rmap_write_protect(vcpu->kvm, gfn);
844         vcpu->arch.mmu.prefetch_page(vcpu, sp);
845         return sp;
846 }
847
848 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
849                                          struct kvm_mmu_page *sp)
850 {
851         unsigned i;
852         u64 *pt;
853         u64 ent;
854
855         pt = sp->spt;
856
857         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
858                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
859                         if (is_shadow_present_pte(pt[i]))
860                                 rmap_remove(kvm, &pt[i]);
861                         pt[i] = shadow_trap_nonpresent_pte;
862                 }
863                 kvm_flush_remote_tlbs(kvm);
864                 return;
865         }
866
867         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
868                 ent = pt[i];
869
870                 if (is_shadow_present_pte(ent)) {
871                         if (!is_large_pte(ent)) {
872                                 ent &= PT64_BASE_ADDR_MASK;
873                                 mmu_page_remove_parent_pte(page_header(ent),
874                                                            &pt[i]);
875                         } else {
876                                 --kvm->stat.lpages;
877                                 rmap_remove(kvm, &pt[i]);
878                         }
879                 }
880                 pt[i] = shadow_trap_nonpresent_pte;
881         }
882         kvm_flush_remote_tlbs(kvm);
883 }
884
885 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
886 {
887         mmu_page_remove_parent_pte(sp, parent_pte);
888 }
889
890 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
891 {
892         int i;
893
894         for (i = 0; i < KVM_MAX_VCPUS; ++i)
895                 if (kvm->vcpus[i])
896                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
897 }
898
899 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
900 {
901         u64 *parent_pte;
902
903         ++kvm->stat.mmu_shadow_zapped;
904         while (sp->multimapped || sp->parent_pte) {
905                 if (!sp->multimapped)
906                         parent_pte = sp->parent_pte;
907                 else {
908                         struct kvm_pte_chain *chain;
909
910                         chain = container_of(sp->parent_ptes.first,
911                                              struct kvm_pte_chain, link);
912                         parent_pte = chain->parent_ptes[0];
913                 }
914                 BUG_ON(!parent_pte);
915                 kvm_mmu_put_page(sp, parent_pte);
916                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
917         }
918         kvm_mmu_page_unlink_children(kvm, sp);
919         if (!sp->root_count) {
920                 if (!sp->role.metaphysical)
921                         unaccount_shadowed(kvm, sp->gfn);
922                 hlist_del(&sp->hash_link);
923                 kvm_mmu_free_page(kvm, sp);
924         } else {
925                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
926                 sp->role.invalid = 1;
927                 kvm_reload_remote_mmus(kvm);
928         }
929         kvm_mmu_reset_last_pte_updated(kvm);
930 }
931
932 /*
933  * Changing the number of mmu pages allocated to the vm
934  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
935  */
936 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
937 {
938         /*
939          * If we set the number of mmu pages to be smaller be than the
940          * number of actived pages , we must to free some mmu pages before we
941          * change the value
942          */
943
944         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
945             kvm_nr_mmu_pages) {
946                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
947                                        - kvm->arch.n_free_mmu_pages;
948
949                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
950                         struct kvm_mmu_page *page;
951
952                         page = container_of(kvm->arch.active_mmu_pages.prev,
953                                             struct kvm_mmu_page, link);
954                         kvm_mmu_zap_page(kvm, page);
955                         n_used_mmu_pages--;
956                 }
957                 kvm->arch.n_free_mmu_pages = 0;
958         }
959         else
960                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
961                                          - kvm->arch.n_alloc_mmu_pages;
962
963         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
964 }
965
966 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
967 {
968         unsigned index;
969         struct hlist_head *bucket;
970         struct kvm_mmu_page *sp;
971         struct hlist_node *node, *n;
972         int r;
973
974         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
975         r = 0;
976         index = kvm_page_table_hashfn(gfn);
977         bucket = &kvm->arch.mmu_page_hash[index];
978         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
979                 if (sp->gfn == gfn && !sp->role.metaphysical) {
980                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
981                                  sp->role.word);
982                         kvm_mmu_zap_page(kvm, sp);
983                         r = 1;
984                 }
985         return r;
986 }
987
988 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
989 {
990         struct kvm_mmu_page *sp;
991
992         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
993                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
994                 kvm_mmu_zap_page(kvm, sp);
995         }
996 }
997
998 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
999 {
1000         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1001         struct kvm_mmu_page *sp = page_header(__pa(pte));
1002
1003         __set_bit(slot, &sp->slot_bitmap);
1004 }
1005
1006 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1007 {
1008         struct page *page;
1009
1010         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1011
1012         if (gpa == UNMAPPED_GVA)
1013                 return NULL;
1014
1015         down_read(&current->mm->mmap_sem);
1016         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1017         up_read(&current->mm->mmap_sem);
1018
1019         return page;
1020 }
1021
1022 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1023                          unsigned pt_access, unsigned pte_access,
1024                          int user_fault, int write_fault, int dirty,
1025                          int *ptwrite, int largepage, gfn_t gfn,
1026                          pfn_t pfn, bool speculative)
1027 {
1028         u64 spte;
1029         int was_rmapped = 0;
1030         int was_writeble = is_writeble_pte(*shadow_pte);
1031
1032         pgprintk("%s: spte %llx access %x write_fault %d"
1033                  " user_fault %d gfn %lx\n",
1034                  __func__, *shadow_pte, pt_access,
1035                  write_fault, user_fault, gfn);
1036
1037         if (is_rmap_pte(*shadow_pte)) {
1038                 /*
1039                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1040                  * the parent of the now unreachable PTE.
1041                  */
1042                 if (largepage && !is_large_pte(*shadow_pte)) {
1043                         struct kvm_mmu_page *child;
1044                         u64 pte = *shadow_pte;
1045
1046                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1047                         mmu_page_remove_parent_pte(child, shadow_pte);
1048                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1049                         pgprintk("hfn old %lx new %lx\n",
1050                                  spte_to_pfn(*shadow_pte), pfn);
1051                         rmap_remove(vcpu->kvm, shadow_pte);
1052                 } else {
1053                         if (largepage)
1054                                 was_rmapped = is_large_pte(*shadow_pte);
1055                         else
1056                                 was_rmapped = 1;
1057                 }
1058         }
1059
1060         /*
1061          * We don't set the accessed bit, since we sometimes want to see
1062          * whether the guest actually used the pte (in order to detect
1063          * demand paging).
1064          */
1065         spte = shadow_base_present_pte | shadow_dirty_mask;
1066         if (!speculative)
1067                 pte_access |= PT_ACCESSED_MASK;
1068         if (!dirty)
1069                 pte_access &= ~ACC_WRITE_MASK;
1070         if (pte_access & ACC_EXEC_MASK)
1071                 spte |= shadow_x_mask;
1072         else
1073                 spte |= shadow_nx_mask;
1074         if (pte_access & ACC_USER_MASK)
1075                 spte |= shadow_user_mask;
1076         if (largepage)
1077                 spte |= PT_PAGE_SIZE_MASK;
1078
1079         spte |= (u64)pfn << PAGE_SHIFT;
1080
1081         if ((pte_access & ACC_WRITE_MASK)
1082             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1083                 struct kvm_mmu_page *shadow;
1084
1085                 spte |= PT_WRITABLE_MASK;
1086
1087                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1088                 if (shadow ||
1089                    (largepage && has_wrprotected_page(vcpu->kvm, gfn))) {
1090                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1091                                  __func__, gfn);
1092                         pte_access &= ~ACC_WRITE_MASK;
1093                         if (is_writeble_pte(spte)) {
1094                                 spte &= ~PT_WRITABLE_MASK;
1095                                 kvm_x86_ops->tlb_flush(vcpu);
1096                         }
1097                         if (write_fault)
1098                                 *ptwrite = 1;
1099                 }
1100         }
1101
1102         if (pte_access & ACC_WRITE_MASK)
1103                 mark_page_dirty(vcpu->kvm, gfn);
1104
1105         pgprintk("%s: setting spte %llx\n", __func__, spte);
1106         pgprintk("instantiating %s PTE (%s) at %d (%llx) addr %llx\n",
1107                  (spte&PT_PAGE_SIZE_MASK)? "2MB" : "4kB",
1108                  (spte&PT_WRITABLE_MASK)?"RW":"R", gfn, spte, shadow_pte);
1109         set_shadow_pte(shadow_pte, spte);
1110         if (!was_rmapped && (spte & PT_PAGE_SIZE_MASK)
1111             && (spte & PT_PRESENT_MASK))
1112                 ++vcpu->kvm->stat.lpages;
1113
1114         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1115         if (!was_rmapped) {
1116                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1117                 if (!is_rmap_pte(*shadow_pte))
1118                         kvm_release_pfn_clean(pfn);
1119         } else {
1120                 if (was_writeble)
1121                         kvm_release_pfn_dirty(pfn);
1122                 else
1123                         kvm_release_pfn_clean(pfn);
1124         }
1125         if (!ptwrite || !*ptwrite)
1126                 vcpu->arch.last_pte_updated = shadow_pte;
1127 }
1128
1129 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1130 {
1131 }
1132
1133 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1134                            int largepage, gfn_t gfn, pfn_t pfn,
1135                            int level)
1136 {
1137         hpa_t table_addr = vcpu->arch.mmu.root_hpa;
1138         int pt_write = 0;
1139
1140         for (; ; level--) {
1141                 u32 index = PT64_INDEX(v, level);
1142                 u64 *table;
1143
1144                 ASSERT(VALID_PAGE(table_addr));
1145                 table = __va(table_addr);
1146
1147                 if (level == 1) {
1148                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1149                                      0, write, 1, &pt_write, 0, gfn, pfn, false);
1150                         return pt_write;
1151                 }
1152
1153                 if (largepage && level == 2) {
1154                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1155                                      0, write, 1, &pt_write, 1, gfn, pfn, false);
1156                         return pt_write;
1157                 }
1158
1159                 if (table[index] == shadow_trap_nonpresent_pte) {
1160                         struct kvm_mmu_page *new_table;
1161                         gfn_t pseudo_gfn;
1162
1163                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1164                                 >> PAGE_SHIFT;
1165                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1166                                                      v, level - 1,
1167                                                      1, ACC_ALL, &table[index]);
1168                         if (!new_table) {
1169                                 pgprintk("nonpaging_map: ENOMEM\n");
1170                                 kvm_release_pfn_clean(pfn);
1171                                 return -ENOMEM;
1172                         }
1173
1174                         table[index] = __pa(new_table->spt)
1175                                 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1176                                 | shadow_user_mask | shadow_x_mask;
1177                 }
1178                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1179         }
1180 }
1181
1182 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1183 {
1184         int r;
1185         int largepage = 0;
1186         pfn_t pfn;
1187
1188         down_read(&current->mm->mmap_sem);
1189         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1190                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1191                 largepage = 1;
1192         }
1193
1194         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1195         up_read(&current->mm->mmap_sem);
1196
1197         /* mmio */
1198         if (is_error_pfn(pfn)) {
1199                 kvm_release_pfn_clean(pfn);
1200                 return 1;
1201         }
1202
1203         spin_lock(&vcpu->kvm->mmu_lock);
1204         kvm_mmu_free_some_pages(vcpu);
1205         r = __direct_map(vcpu, v, write, largepage, gfn, pfn,
1206                          PT32E_ROOT_LEVEL);
1207         spin_unlock(&vcpu->kvm->mmu_lock);
1208
1209
1210         return r;
1211 }
1212
1213
1214 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1215                                     struct kvm_mmu_page *sp)
1216 {
1217         int i;
1218
1219         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1220                 sp->spt[i] = shadow_trap_nonpresent_pte;
1221 }
1222
1223 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1224 {
1225         int i;
1226         struct kvm_mmu_page *sp;
1227
1228         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1229                 return;
1230         spin_lock(&vcpu->kvm->mmu_lock);
1231         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1232                 hpa_t root = vcpu->arch.mmu.root_hpa;
1233
1234                 sp = page_header(root);
1235                 --sp->root_count;
1236                 if (!sp->root_count && sp->role.invalid)
1237                         kvm_mmu_zap_page(vcpu->kvm, sp);
1238                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1239                 spin_unlock(&vcpu->kvm->mmu_lock);
1240                 return;
1241         }
1242         for (i = 0; i < 4; ++i) {
1243                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1244
1245                 if (root) {
1246                         root &= PT64_BASE_ADDR_MASK;
1247                         sp = page_header(root);
1248                         --sp->root_count;
1249                         if (!sp->root_count && sp->role.invalid)
1250                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1251                 }
1252                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1253         }
1254         spin_unlock(&vcpu->kvm->mmu_lock);
1255         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1256 }
1257
1258 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1259 {
1260         int i;
1261         gfn_t root_gfn;
1262         struct kvm_mmu_page *sp;
1263         int metaphysical = 0;
1264
1265         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1266
1267         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1268                 hpa_t root = vcpu->arch.mmu.root_hpa;
1269
1270                 ASSERT(!VALID_PAGE(root));
1271                 if (tdp_enabled)
1272                         metaphysical = 1;
1273                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1274                                       PT64_ROOT_LEVEL, metaphysical,
1275                                       ACC_ALL, NULL);
1276                 root = __pa(sp->spt);
1277                 ++sp->root_count;
1278                 vcpu->arch.mmu.root_hpa = root;
1279                 return;
1280         }
1281         metaphysical = !is_paging(vcpu);
1282         if (tdp_enabled)
1283                 metaphysical = 1;
1284         for (i = 0; i < 4; ++i) {
1285                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1286
1287                 ASSERT(!VALID_PAGE(root));
1288                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1289                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1290                                 vcpu->arch.mmu.pae_root[i] = 0;
1291                                 continue;
1292                         }
1293                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1294                 } else if (vcpu->arch.mmu.root_level == 0)
1295                         root_gfn = 0;
1296                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1297                                       PT32_ROOT_LEVEL, metaphysical,
1298                                       ACC_ALL, NULL);
1299                 root = __pa(sp->spt);
1300                 ++sp->root_count;
1301                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1302         }
1303         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1304 }
1305
1306 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1307 {
1308         return vaddr;
1309 }
1310
1311 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1312                                 u32 error_code)
1313 {
1314         gfn_t gfn;
1315         int r;
1316
1317         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1318         r = mmu_topup_memory_caches(vcpu);
1319         if (r)
1320                 return r;
1321
1322         ASSERT(vcpu);
1323         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1324
1325         gfn = gva >> PAGE_SHIFT;
1326
1327         return nonpaging_map(vcpu, gva & PAGE_MASK,
1328                              error_code & PFERR_WRITE_MASK, gfn);
1329 }
1330
1331 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1332                                 u32 error_code)
1333 {
1334         pfn_t pfn;
1335         int r;
1336         int largepage = 0;
1337         gfn_t gfn = gpa >> PAGE_SHIFT;
1338
1339         ASSERT(vcpu);
1340         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1341
1342         r = mmu_topup_memory_caches(vcpu);
1343         if (r)
1344                 return r;
1345
1346         down_read(&current->mm->mmap_sem);
1347         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1348                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1349                 largepage = 1;
1350         }
1351         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1352         up_read(&current->mm->mmap_sem);
1353         if (is_error_pfn(pfn)) {
1354                 kvm_release_pfn_clean(pfn);
1355                 return 1;
1356         }
1357         spin_lock(&vcpu->kvm->mmu_lock);
1358         kvm_mmu_free_some_pages(vcpu);
1359         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1360                          largepage, gfn, pfn, kvm_x86_ops->get_tdp_level());
1361         spin_unlock(&vcpu->kvm->mmu_lock);
1362
1363         return r;
1364 }
1365
1366 static void nonpaging_free(struct kvm_vcpu *vcpu)
1367 {
1368         mmu_free_roots(vcpu);
1369 }
1370
1371 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1372 {
1373         struct kvm_mmu *context = &vcpu->arch.mmu;
1374
1375         context->new_cr3 = nonpaging_new_cr3;
1376         context->page_fault = nonpaging_page_fault;
1377         context->gva_to_gpa = nonpaging_gva_to_gpa;
1378         context->free = nonpaging_free;
1379         context->prefetch_page = nonpaging_prefetch_page;
1380         context->root_level = 0;
1381         context->shadow_root_level = PT32E_ROOT_LEVEL;
1382         context->root_hpa = INVALID_PAGE;
1383         return 0;
1384 }
1385
1386 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1387 {
1388         ++vcpu->stat.tlb_flush;
1389         kvm_x86_ops->tlb_flush(vcpu);
1390 }
1391
1392 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1393 {
1394         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1395         mmu_free_roots(vcpu);
1396 }
1397
1398 static void inject_page_fault(struct kvm_vcpu *vcpu,
1399                               u64 addr,
1400                               u32 err_code)
1401 {
1402         kvm_inject_page_fault(vcpu, addr, err_code);
1403 }
1404
1405 static void paging_free(struct kvm_vcpu *vcpu)
1406 {
1407         nonpaging_free(vcpu);
1408 }
1409
1410 #define PTTYPE 64
1411 #include "paging_tmpl.h"
1412 #undef PTTYPE
1413
1414 #define PTTYPE 32
1415 #include "paging_tmpl.h"
1416 #undef PTTYPE
1417
1418 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1419 {
1420         struct kvm_mmu *context = &vcpu->arch.mmu;
1421
1422         ASSERT(is_pae(vcpu));
1423         context->new_cr3 = paging_new_cr3;
1424         context->page_fault = paging64_page_fault;
1425         context->gva_to_gpa = paging64_gva_to_gpa;
1426         context->prefetch_page = paging64_prefetch_page;
1427         context->free = paging_free;
1428         context->root_level = level;
1429         context->shadow_root_level = level;
1430         context->root_hpa = INVALID_PAGE;
1431         return 0;
1432 }
1433
1434 static int paging64_init_context(struct kvm_vcpu *vcpu)
1435 {
1436         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1437 }
1438
1439 static int paging32_init_context(struct kvm_vcpu *vcpu)
1440 {
1441         struct kvm_mmu *context = &vcpu->arch.mmu;
1442
1443         context->new_cr3 = paging_new_cr3;
1444         context->page_fault = paging32_page_fault;
1445         context->gva_to_gpa = paging32_gva_to_gpa;
1446         context->free = paging_free;
1447         context->prefetch_page = paging32_prefetch_page;
1448         context->root_level = PT32_ROOT_LEVEL;
1449         context->shadow_root_level = PT32E_ROOT_LEVEL;
1450         context->root_hpa = INVALID_PAGE;
1451         return 0;
1452 }
1453
1454 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1455 {
1456         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1457 }
1458
1459 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1460 {
1461         struct kvm_mmu *context = &vcpu->arch.mmu;
1462
1463         context->new_cr3 = nonpaging_new_cr3;
1464         context->page_fault = tdp_page_fault;
1465         context->free = nonpaging_free;
1466         context->prefetch_page = nonpaging_prefetch_page;
1467         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
1468         context->root_hpa = INVALID_PAGE;
1469
1470         if (!is_paging(vcpu)) {
1471                 context->gva_to_gpa = nonpaging_gva_to_gpa;
1472                 context->root_level = 0;
1473         } else if (is_long_mode(vcpu)) {
1474                 context->gva_to_gpa = paging64_gva_to_gpa;
1475                 context->root_level = PT64_ROOT_LEVEL;
1476         } else if (is_pae(vcpu)) {
1477                 context->gva_to_gpa = paging64_gva_to_gpa;
1478                 context->root_level = PT32E_ROOT_LEVEL;
1479         } else {
1480                 context->gva_to_gpa = paging32_gva_to_gpa;
1481                 context->root_level = PT32_ROOT_LEVEL;
1482         }
1483
1484         return 0;
1485 }
1486
1487 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1488 {
1489         ASSERT(vcpu);
1490         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1491
1492         if (!is_paging(vcpu))
1493                 return nonpaging_init_context(vcpu);
1494         else if (is_long_mode(vcpu))
1495                 return paging64_init_context(vcpu);
1496         else if (is_pae(vcpu))
1497                 return paging32E_init_context(vcpu);
1498         else
1499                 return paging32_init_context(vcpu);
1500 }
1501
1502 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1503 {
1504         vcpu->arch.update_pte.pfn = bad_pfn;
1505
1506         if (tdp_enabled)
1507                 return init_kvm_tdp_mmu(vcpu);
1508         else
1509                 return init_kvm_softmmu(vcpu);
1510 }
1511
1512 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1513 {
1514         ASSERT(vcpu);
1515         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1516                 vcpu->arch.mmu.free(vcpu);
1517                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1518         }
1519 }
1520
1521 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1522 {
1523         destroy_kvm_mmu(vcpu);
1524         return init_kvm_mmu(vcpu);
1525 }
1526 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1527
1528 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1529 {
1530         int r;
1531
1532         r = mmu_topup_memory_caches(vcpu);
1533         if (r)
1534                 goto out;
1535         spin_lock(&vcpu->kvm->mmu_lock);
1536         kvm_mmu_free_some_pages(vcpu);
1537         mmu_alloc_roots(vcpu);
1538         spin_unlock(&vcpu->kvm->mmu_lock);
1539         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1540         kvm_mmu_flush_tlb(vcpu);
1541 out:
1542         return r;
1543 }
1544 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1545
1546 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1547 {
1548         mmu_free_roots(vcpu);
1549 }
1550
1551 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1552                                   struct kvm_mmu_page *sp,
1553                                   u64 *spte)
1554 {
1555         u64 pte;
1556         struct kvm_mmu_page *child;
1557
1558         pte = *spte;
1559         if (is_shadow_present_pte(pte)) {
1560                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
1561                     is_large_pte(pte))
1562                         rmap_remove(vcpu->kvm, spte);
1563                 else {
1564                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1565                         mmu_page_remove_parent_pte(child, spte);
1566                 }
1567         }
1568         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1569         if (is_large_pte(pte))
1570                 --vcpu->kvm->stat.lpages;
1571 }
1572
1573 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1574                                   struct kvm_mmu_page *sp,
1575                                   u64 *spte,
1576                                   const void *new)
1577 {
1578         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1579                 if (!vcpu->arch.update_pte.largepage ||
1580                     sp->role.glevels == PT32_ROOT_LEVEL) {
1581                         ++vcpu->kvm->stat.mmu_pde_zapped;
1582                         return;
1583                 }
1584         }
1585
1586         ++vcpu->kvm->stat.mmu_pte_updated;
1587         if (sp->role.glevels == PT32_ROOT_LEVEL)
1588                 paging32_update_pte(vcpu, sp, spte, new);
1589         else
1590                 paging64_update_pte(vcpu, sp, spte, new);
1591 }
1592
1593 static bool need_remote_flush(u64 old, u64 new)
1594 {
1595         if (!is_shadow_present_pte(old))
1596                 return false;
1597         if (!is_shadow_present_pte(new))
1598                 return true;
1599         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1600                 return true;
1601         old ^= PT64_NX_MASK;
1602         new ^= PT64_NX_MASK;
1603         return (old & ~new & PT64_PERM_MASK) != 0;
1604 }
1605
1606 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1607 {
1608         if (need_remote_flush(old, new))
1609                 kvm_flush_remote_tlbs(vcpu->kvm);
1610         else
1611                 kvm_mmu_flush_tlb(vcpu);
1612 }
1613
1614 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1615 {
1616         u64 *spte = vcpu->arch.last_pte_updated;
1617
1618         return !!(spte && (*spte & shadow_accessed_mask));
1619 }
1620
1621 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1622                                           const u8 *new, int bytes)
1623 {
1624         gfn_t gfn;
1625         int r;
1626         u64 gpte = 0;
1627         pfn_t pfn;
1628
1629         vcpu->arch.update_pte.largepage = 0;
1630
1631         if (bytes != 4 && bytes != 8)
1632                 return;
1633
1634         /*
1635          * Assume that the pte write on a page table of the same type
1636          * as the current vcpu paging mode.  This is nearly always true
1637          * (might be false while changing modes).  Note it is verified later
1638          * by update_pte().
1639          */
1640         if (is_pae(vcpu)) {
1641                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1642                 if ((bytes == 4) && (gpa % 4 == 0)) {
1643                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1644                         if (r)
1645                                 return;
1646                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1647                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1648                         memcpy((void *)&gpte, new, 8);
1649                 }
1650         } else {
1651                 if ((bytes == 4) && (gpa % 4 == 0))
1652                         memcpy((void *)&gpte, new, 4);
1653         }
1654         if (!is_present_pte(gpte))
1655                 return;
1656         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1657
1658         down_read(&current->mm->mmap_sem);
1659         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
1660                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1661                 vcpu->arch.update_pte.largepage = 1;
1662         }
1663         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1664         up_read(&current->mm->mmap_sem);
1665
1666         if (is_error_pfn(pfn)) {
1667                 kvm_release_pfn_clean(pfn);
1668                 return;
1669         }
1670         vcpu->arch.update_pte.gfn = gfn;
1671         vcpu->arch.update_pte.pfn = pfn;
1672 }
1673
1674 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1675                        const u8 *new, int bytes)
1676 {
1677         gfn_t gfn = gpa >> PAGE_SHIFT;
1678         struct kvm_mmu_page *sp;
1679         struct hlist_node *node, *n;
1680         struct hlist_head *bucket;
1681         unsigned index;
1682         u64 entry, gentry;
1683         u64 *spte;
1684         unsigned offset = offset_in_page(gpa);
1685         unsigned pte_size;
1686         unsigned page_offset;
1687         unsigned misaligned;
1688         unsigned quadrant;
1689         int level;
1690         int flooded = 0;
1691         int npte;
1692         int r;
1693
1694         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
1695         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1696         spin_lock(&vcpu->kvm->mmu_lock);
1697         kvm_mmu_free_some_pages(vcpu);
1698         ++vcpu->kvm->stat.mmu_pte_write;
1699         kvm_mmu_audit(vcpu, "pre pte write");
1700         if (gfn == vcpu->arch.last_pt_write_gfn
1701             && !last_updated_pte_accessed(vcpu)) {
1702                 ++vcpu->arch.last_pt_write_count;
1703                 if (vcpu->arch.last_pt_write_count >= 3)
1704                         flooded = 1;
1705         } else {
1706                 vcpu->arch.last_pt_write_gfn = gfn;
1707                 vcpu->arch.last_pt_write_count = 1;
1708                 vcpu->arch.last_pte_updated = NULL;
1709         }
1710         index = kvm_page_table_hashfn(gfn);
1711         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1712         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1713                 if (sp->gfn != gfn || sp->role.metaphysical)
1714                         continue;
1715                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1716                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1717                 misaligned |= bytes < 4;
1718                 if (misaligned || flooded) {
1719                         /*
1720                          * Misaligned accesses are too much trouble to fix
1721                          * up; also, they usually indicate a page is not used
1722                          * as a page table.
1723                          *
1724                          * If we're seeing too many writes to a page,
1725                          * it may no longer be a page table, or we may be
1726                          * forking, in which case it is better to unmap the
1727                          * page.
1728                          */
1729                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1730                                  gpa, bytes, sp->role.word);
1731                         kvm_mmu_zap_page(vcpu->kvm, sp);
1732                         ++vcpu->kvm->stat.mmu_flooded;
1733                         continue;
1734                 }
1735                 page_offset = offset;
1736                 level = sp->role.level;
1737                 npte = 1;
1738                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1739                         page_offset <<= 1;      /* 32->64 */
1740                         /*
1741                          * A 32-bit pde maps 4MB while the shadow pdes map
1742                          * only 2MB.  So we need to double the offset again
1743                          * and zap two pdes instead of one.
1744                          */
1745                         if (level == PT32_ROOT_LEVEL) {
1746                                 page_offset &= ~7; /* kill rounding error */
1747                                 page_offset <<= 1;
1748                                 npte = 2;
1749                         }
1750                         quadrant = page_offset >> PAGE_SHIFT;
1751                         page_offset &= ~PAGE_MASK;
1752                         if (quadrant != sp->role.quadrant)
1753                                 continue;
1754                 }
1755                 spte = &sp->spt[page_offset / sizeof(*spte)];
1756                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1757                         gentry = 0;
1758                         r = kvm_read_guest_atomic(vcpu->kvm,
1759                                                   gpa & ~(u64)(pte_size - 1),
1760                                                   &gentry, pte_size);
1761                         new = (const void *)&gentry;
1762                         if (r < 0)
1763                                 new = NULL;
1764                 }
1765                 while (npte--) {
1766                         entry = *spte;
1767                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1768                         if (new)
1769                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1770                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1771                         ++spte;
1772                 }
1773         }
1774         kvm_mmu_audit(vcpu, "post pte write");
1775         spin_unlock(&vcpu->kvm->mmu_lock);
1776         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
1777                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
1778                 vcpu->arch.update_pte.pfn = bad_pfn;
1779         }
1780 }
1781
1782 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1783 {
1784         gpa_t gpa;
1785         int r;
1786
1787         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1788
1789         spin_lock(&vcpu->kvm->mmu_lock);
1790         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1791         spin_unlock(&vcpu->kvm->mmu_lock);
1792         return r;
1793 }
1794
1795 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1796 {
1797         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1798                 struct kvm_mmu_page *sp;
1799
1800                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1801                                   struct kvm_mmu_page, link);
1802                 kvm_mmu_zap_page(vcpu->kvm, sp);
1803                 ++vcpu->kvm->stat.mmu_recycled;
1804         }
1805 }
1806
1807 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1808 {
1809         int r;
1810         enum emulation_result er;
1811
1812         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1813         if (r < 0)
1814                 goto out;
1815
1816         if (!r) {
1817                 r = 1;
1818                 goto out;
1819         }
1820
1821         r = mmu_topup_memory_caches(vcpu);
1822         if (r)
1823                 goto out;
1824
1825         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1826
1827         switch (er) {
1828         case EMULATE_DONE:
1829                 return 1;
1830         case EMULATE_DO_MMIO:
1831                 ++vcpu->stat.mmio_exits;
1832                 return 0;
1833         case EMULATE_FAIL:
1834                 kvm_report_emulation_failure(vcpu, "pagetable");
1835                 return 1;
1836         default:
1837                 BUG();
1838         }
1839 out:
1840         return r;
1841 }
1842 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1843
1844 void kvm_enable_tdp(void)
1845 {
1846         tdp_enabled = true;
1847 }
1848 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
1849
1850 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1851 {
1852         struct kvm_mmu_page *sp;
1853
1854         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1855                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1856                                   struct kvm_mmu_page, link);
1857                 kvm_mmu_zap_page(vcpu->kvm, sp);
1858                 cond_resched();
1859         }
1860         free_page((unsigned long)vcpu->arch.mmu.pae_root);
1861 }
1862
1863 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1864 {
1865         struct page *page;
1866         int i;
1867
1868         ASSERT(vcpu);
1869
1870         if (vcpu->kvm->arch.n_requested_mmu_pages)
1871                 vcpu->kvm->arch.n_free_mmu_pages =
1872                                         vcpu->kvm->arch.n_requested_mmu_pages;
1873         else
1874                 vcpu->kvm->arch.n_free_mmu_pages =
1875                                         vcpu->kvm->arch.n_alloc_mmu_pages;
1876         /*
1877          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1878          * Therefore we need to allocate shadow page tables in the first
1879          * 4GB of memory, which happens to fit the DMA32 zone.
1880          */
1881         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1882         if (!page)
1883                 goto error_1;
1884         vcpu->arch.mmu.pae_root = page_address(page);
1885         for (i = 0; i < 4; ++i)
1886                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1887
1888         return 0;
1889
1890 error_1:
1891         free_mmu_pages(vcpu);
1892         return -ENOMEM;
1893 }
1894
1895 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1896 {
1897         ASSERT(vcpu);
1898         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1899
1900         return alloc_mmu_pages(vcpu);
1901 }
1902
1903 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1904 {
1905         ASSERT(vcpu);
1906         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1907
1908         return init_kvm_mmu(vcpu);
1909 }
1910
1911 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1912 {
1913         ASSERT(vcpu);
1914
1915         destroy_kvm_mmu(vcpu);
1916         free_mmu_pages(vcpu);
1917         mmu_free_memory_caches(vcpu);
1918 }
1919
1920 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1921 {
1922         struct kvm_mmu_page *sp;
1923
1924         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1925                 int i;
1926                 u64 *pt;
1927
1928                 if (!test_bit(slot, &sp->slot_bitmap))
1929                         continue;
1930
1931                 pt = sp->spt;
1932                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1933                         /* avoid RMW */
1934                         if (pt[i] & PT_WRITABLE_MASK)
1935                                 pt[i] &= ~PT_WRITABLE_MASK;
1936         }
1937 }
1938
1939 void kvm_mmu_zap_all(struct kvm *kvm)
1940 {
1941         struct kvm_mmu_page *sp, *node;
1942
1943         spin_lock(&kvm->mmu_lock);
1944         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1945                 kvm_mmu_zap_page(kvm, sp);
1946         spin_unlock(&kvm->mmu_lock);
1947
1948         kvm_flush_remote_tlbs(kvm);
1949 }
1950
1951 void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
1952 {
1953         struct kvm_mmu_page *page;
1954
1955         page = container_of(kvm->arch.active_mmu_pages.prev,
1956                             struct kvm_mmu_page, link);
1957         kvm_mmu_zap_page(kvm, page);
1958 }
1959
1960 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
1961 {
1962         struct kvm *kvm;
1963         struct kvm *kvm_freed = NULL;
1964         int cache_count = 0;
1965
1966         spin_lock(&kvm_lock);
1967
1968         list_for_each_entry(kvm, &vm_list, vm_list) {
1969                 int npages;
1970
1971                 spin_lock(&kvm->mmu_lock);
1972                 npages = kvm->arch.n_alloc_mmu_pages -
1973                          kvm->arch.n_free_mmu_pages;
1974                 cache_count += npages;
1975                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
1976                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
1977                         cache_count--;
1978                         kvm_freed = kvm;
1979                 }
1980                 nr_to_scan--;
1981
1982                 spin_unlock(&kvm->mmu_lock);
1983         }
1984         if (kvm_freed)
1985                 list_move_tail(&kvm_freed->vm_list, &vm_list);
1986
1987         spin_unlock(&kvm_lock);
1988
1989         return cache_count;
1990 }
1991
1992 static struct shrinker mmu_shrinker = {
1993         .shrink = mmu_shrink,
1994         .seeks = DEFAULT_SEEKS * 10,
1995 };
1996
1997 static void mmu_destroy_caches(void)
1998 {
1999         if (pte_chain_cache)
2000                 kmem_cache_destroy(pte_chain_cache);
2001         if (rmap_desc_cache)
2002                 kmem_cache_destroy(rmap_desc_cache);
2003         if (mmu_page_header_cache)
2004                 kmem_cache_destroy(mmu_page_header_cache);
2005 }
2006
2007 void kvm_mmu_module_exit(void)
2008 {
2009         mmu_destroy_caches();
2010         unregister_shrinker(&mmu_shrinker);
2011 }
2012
2013 int kvm_mmu_module_init(void)
2014 {
2015         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2016                                             sizeof(struct kvm_pte_chain),
2017                                             0, 0, NULL);
2018         if (!pte_chain_cache)
2019                 goto nomem;
2020         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2021                                             sizeof(struct kvm_rmap_desc),
2022                                             0, 0, NULL);
2023         if (!rmap_desc_cache)
2024                 goto nomem;
2025
2026         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2027                                                   sizeof(struct kvm_mmu_page),
2028                                                   0, 0, NULL);
2029         if (!mmu_page_header_cache)
2030                 goto nomem;
2031
2032         register_shrinker(&mmu_shrinker);
2033
2034         return 0;
2035
2036 nomem:
2037         mmu_destroy_caches();
2038         return -ENOMEM;
2039 }
2040
2041 /*
2042  * Caculate mmu pages needed for kvm.
2043  */
2044 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2045 {
2046         int i;
2047         unsigned int nr_mmu_pages;
2048         unsigned int  nr_pages = 0;
2049
2050         for (i = 0; i < kvm->nmemslots; i++)
2051                 nr_pages += kvm->memslots[i].npages;
2052
2053         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2054         nr_mmu_pages = max(nr_mmu_pages,
2055                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2056
2057         return nr_mmu_pages;
2058 }
2059
2060 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2061                                 unsigned len)
2062 {
2063         if (len > buffer->len)
2064                 return NULL;
2065         return buffer->ptr;
2066 }
2067
2068 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2069                                 unsigned len)
2070 {
2071         void *ret;
2072
2073         ret = pv_mmu_peek_buffer(buffer, len);
2074         if (!ret)
2075                 return ret;
2076         buffer->ptr += len;
2077         buffer->len -= len;
2078         buffer->processed += len;
2079         return ret;
2080 }
2081
2082 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2083                              gpa_t addr, gpa_t value)
2084 {
2085         int bytes = 8;
2086         int r;
2087
2088         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2089                 bytes = 4;
2090
2091         r = mmu_topup_memory_caches(vcpu);
2092         if (r)
2093                 return r;
2094
2095         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2096                 return -EFAULT;
2097
2098         return 1;
2099 }
2100
2101 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2102 {
2103         kvm_x86_ops->tlb_flush(vcpu);
2104         return 1;
2105 }
2106
2107 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2108 {
2109         spin_lock(&vcpu->kvm->mmu_lock);
2110         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2111         spin_unlock(&vcpu->kvm->mmu_lock);
2112         return 1;
2113 }
2114
2115 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2116                              struct kvm_pv_mmu_op_buffer *buffer)
2117 {
2118         struct kvm_mmu_op_header *header;
2119
2120         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2121         if (!header)
2122                 return 0;
2123         switch (header->op) {
2124         case KVM_MMU_OP_WRITE_PTE: {
2125                 struct kvm_mmu_op_write_pte *wpte;
2126
2127                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2128                 if (!wpte)
2129                         return 0;
2130                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2131                                         wpte->pte_val);
2132         }
2133         case KVM_MMU_OP_FLUSH_TLB: {
2134                 struct kvm_mmu_op_flush_tlb *ftlb;
2135
2136                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2137                 if (!ftlb)
2138                         return 0;
2139                 return kvm_pv_mmu_flush_tlb(vcpu);
2140         }
2141         case KVM_MMU_OP_RELEASE_PT: {
2142                 struct kvm_mmu_op_release_pt *rpt;
2143
2144                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2145                 if (!rpt)
2146                         return 0;
2147                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2148         }
2149         default: return 0;
2150         }
2151 }
2152
2153 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2154                   gpa_t addr, unsigned long *ret)
2155 {
2156         int r;
2157         struct kvm_pv_mmu_op_buffer buffer;
2158
2159         buffer.ptr = buffer.buf;
2160         buffer.len = min_t(unsigned long, bytes, sizeof buffer.buf);
2161         buffer.processed = 0;
2162
2163         r = kvm_read_guest(vcpu->kvm, addr, buffer.buf, buffer.len);
2164         if (r)
2165                 goto out;
2166
2167         while (buffer.len) {
2168                 r = kvm_pv_mmu_op_one(vcpu, &buffer);
2169                 if (r < 0)
2170                         goto out;
2171                 if (r == 0)
2172                         break;
2173         }
2174
2175         r = 1;
2176 out:
2177         *ret = buffer.processed;
2178         return r;
2179 }
2180
2181 #ifdef AUDIT
2182
2183 static const char *audit_msg;
2184
2185 static gva_t canonicalize(gva_t gva)
2186 {
2187 #ifdef CONFIG_X86_64
2188         gva = (long long)(gva << 16) >> 16;
2189 #endif
2190         return gva;
2191 }
2192
2193 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2194                                 gva_t va, int level)
2195 {
2196         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2197         int i;
2198         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2199
2200         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2201                 u64 ent = pt[i];
2202
2203                 if (ent == shadow_trap_nonpresent_pte)
2204                         continue;
2205
2206                 va = canonicalize(va);
2207                 if (level > 1) {
2208                         if (ent == shadow_notrap_nonpresent_pte)
2209                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2210                                        " in nonleaf level: levels %d gva %lx"
2211                                        " level %d pte %llx\n", audit_msg,
2212                                        vcpu->arch.mmu.root_level, va, level, ent);
2213
2214                         audit_mappings_page(vcpu, ent, va, level - 1);
2215                 } else {
2216                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2217                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2218
2219                         if (is_shadow_present_pte(ent)
2220                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2221                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2222                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2223                                        audit_msg, vcpu->arch.mmu.root_level,
2224                                        va, gpa, hpa, ent,
2225                                        is_shadow_present_pte(ent));
2226                         else if (ent == shadow_notrap_nonpresent_pte
2227                                  && !is_error_hpa(hpa))
2228                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2229                                        " valid guest gva %lx\n", audit_msg, va);
2230                         kvm_release_pfn_clean(pfn);
2231
2232                 }
2233         }
2234 }
2235
2236 static void audit_mappings(struct kvm_vcpu *vcpu)
2237 {
2238         unsigned i;
2239
2240         if (vcpu->arch.mmu.root_level == 4)
2241                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2242         else
2243                 for (i = 0; i < 4; ++i)
2244                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2245                                 audit_mappings_page(vcpu,
2246                                                     vcpu->arch.mmu.pae_root[i],
2247                                                     i << 30,
2248                                                     2);
2249 }
2250
2251 static int count_rmaps(struct kvm_vcpu *vcpu)
2252 {
2253         int nmaps = 0;
2254         int i, j, k;
2255
2256         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2257                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2258                 struct kvm_rmap_desc *d;
2259
2260                 for (j = 0; j < m->npages; ++j) {
2261                         unsigned long *rmapp = &m->rmap[j];
2262
2263                         if (!*rmapp)
2264                                 continue;
2265                         if (!(*rmapp & 1)) {
2266                                 ++nmaps;
2267                                 continue;
2268                         }
2269                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2270                         while (d) {
2271                                 for (k = 0; k < RMAP_EXT; ++k)
2272                                         if (d->shadow_ptes[k])
2273                                                 ++nmaps;
2274                                         else
2275                                                 break;
2276                                 d = d->more;
2277                         }
2278                 }
2279         }
2280         return nmaps;
2281 }
2282
2283 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2284 {
2285         int nmaps = 0;
2286         struct kvm_mmu_page *sp;
2287         int i;
2288
2289         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2290                 u64 *pt = sp->spt;
2291
2292                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2293                         continue;
2294
2295                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2296                         u64 ent = pt[i];
2297
2298                         if (!(ent & PT_PRESENT_MASK))
2299                                 continue;
2300                         if (!(ent & PT_WRITABLE_MASK))
2301                                 continue;
2302                         ++nmaps;
2303                 }
2304         }
2305         return nmaps;
2306 }
2307
2308 static void audit_rmap(struct kvm_vcpu *vcpu)
2309 {
2310         int n_rmap = count_rmaps(vcpu);
2311         int n_actual = count_writable_mappings(vcpu);
2312
2313         if (n_rmap != n_actual)
2314                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2315                        __func__, audit_msg, n_rmap, n_actual);
2316 }
2317
2318 static void audit_write_protection(struct kvm_vcpu *vcpu)
2319 {
2320         struct kvm_mmu_page *sp;
2321         struct kvm_memory_slot *slot;
2322         unsigned long *rmapp;
2323         gfn_t gfn;
2324
2325         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2326                 if (sp->role.metaphysical)
2327                         continue;
2328
2329                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2330                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2331                 rmapp = &slot->rmap[gfn - slot->base_gfn];
2332                 if (*rmapp)
2333                         printk(KERN_ERR "%s: (%s) shadow page has writable"
2334                                " mappings: gfn %lx role %x\n",
2335                                __func__, audit_msg, sp->gfn,
2336                                sp->role.word);
2337         }
2338 }
2339
2340 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2341 {
2342         int olddbg = dbg;
2343
2344         dbg = 0;
2345         audit_msg = msg;
2346         audit_rmap(vcpu);
2347         audit_write_protection(vcpu);
2348         audit_mappings(vcpu);
2349         dbg = olddbg;
2350 }
2351
2352 #endif