8c2f580956d94c8b36f4f974f499da8327a0cfd9
[linux-2.6.git] / 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  * Copyright 2010 Red Hat, Inc. and/or its affilates.
11  *
12  * Authors:
13  *   Yaniv Kamay  <yaniv@qumranet.com>
14  *   Avi Kivity   <avi@qumranet.com>
15  *
16  * This work is licensed under the terms of the GNU GPL, version 2.  See
17  * the COPYING file in the top-level directory.
18  *
19  */
20
21 #include "mmu.h"
22 #include "x86.h"
23 #include "kvm_cache_regs.h"
24
25 #include <linux/kvm_host.h>
26 #include <linux/types.h>
27 #include <linux/string.h>
28 #include <linux/mm.h>
29 #include <linux/highmem.h>
30 #include <linux/module.h>
31 #include <linux/swap.h>
32 #include <linux/hugetlb.h>
33 #include <linux/compiler.h>
34 #include <linux/srcu.h>
35 #include <linux/slab.h>
36 #include <linux/uaccess.h>
37
38 #include <asm/page.h>
39 #include <asm/cmpxchg.h>
40 #include <asm/io.h>
41 #include <asm/vmx.h>
42
43 /*
44  * When setting this variable to true it enables Two-Dimensional-Paging
45  * where the hardware walks 2 page tables:
46  * 1. the guest-virtual to guest-physical
47  * 2. while doing 1. it walks guest-physical to host-physical
48  * If the hardware supports that we don't need to do shadow paging.
49  */
50 bool tdp_enabled = false;
51
52 #undef MMU_DEBUG
53
54 #undef AUDIT
55
56 #ifdef AUDIT
57 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
58 #else
59 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
60 #endif
61
62 #ifdef MMU_DEBUG
63
64 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
65 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
66
67 #else
68
69 #define pgprintk(x...) do { } while (0)
70 #define rmap_printk(x...) do { } while (0)
71
72 #endif
73
74 #if defined(MMU_DEBUG) || defined(AUDIT)
75 static int dbg = 0;
76 module_param(dbg, bool, 0644);
77 #endif
78
79 static int oos_shadow = 1;
80 module_param(oos_shadow, bool, 0644);
81
82 #ifndef MMU_DEBUG
83 #define ASSERT(x) do { } while (0)
84 #else
85 #define ASSERT(x)                                                       \
86         if (!(x)) {                                                     \
87                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
88                        __FILE__, __LINE__, #x);                         \
89         }
90 #endif
91
92 #define PT_FIRST_AVAIL_BITS_SHIFT 9
93 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
94
95 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
96
97 #define PT64_LEVEL_BITS 9
98
99 #define PT64_LEVEL_SHIFT(level) \
100                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
101
102 #define PT64_LEVEL_MASK(level) \
103                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
104
105 #define PT64_INDEX(address, level)\
106         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
107
108
109 #define PT32_LEVEL_BITS 10
110
111 #define PT32_LEVEL_SHIFT(level) \
112                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
113
114 #define PT32_LEVEL_MASK(level) \
115                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
116 #define PT32_LVL_OFFSET_MASK(level) \
117         (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
118                                                 * PT32_LEVEL_BITS))) - 1))
119
120 #define PT32_INDEX(address, level)\
121         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
122
123
124 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
125 #define PT64_DIR_BASE_ADDR_MASK \
126         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
127 #define PT64_LVL_ADDR_MASK(level) \
128         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
129                                                 * PT64_LEVEL_BITS))) - 1))
130 #define PT64_LVL_OFFSET_MASK(level) \
131         (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
132                                                 * PT64_LEVEL_BITS))) - 1))
133
134 #define PT32_BASE_ADDR_MASK PAGE_MASK
135 #define PT32_DIR_BASE_ADDR_MASK \
136         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
137 #define PT32_LVL_ADDR_MASK(level) \
138         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
139                                             * PT32_LEVEL_BITS))) - 1))
140
141 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
142                         | PT64_NX_MASK)
143
144 #define RMAP_EXT 4
145
146 #define ACC_EXEC_MASK    1
147 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
148 #define ACC_USER_MASK    PT_USER_MASK
149 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
150
151 #include <trace/events/kvm.h>
152
153 #define CREATE_TRACE_POINTS
154 #include "mmutrace.h"
155
156 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
157
158 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
159
160 struct kvm_rmap_desc {
161         u64 *sptes[RMAP_EXT];
162         struct kvm_rmap_desc *more;
163 };
164
165 struct kvm_shadow_walk_iterator {
166         u64 addr;
167         hpa_t shadow_addr;
168         int level;
169         u64 *sptep;
170         unsigned index;
171 };
172
173 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
174         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
175              shadow_walk_okay(&(_walker));                      \
176              shadow_walk_next(&(_walker)))
177
178 typedef void (*mmu_parent_walk_fn) (struct kvm_mmu_page *sp, u64 *spte);
179
180 static struct kmem_cache *pte_chain_cache;
181 static struct kmem_cache *rmap_desc_cache;
182 static struct kmem_cache *mmu_page_header_cache;
183
184 static u64 __read_mostly shadow_trap_nonpresent_pte;
185 static u64 __read_mostly shadow_notrap_nonpresent_pte;
186 static u64 __read_mostly shadow_base_present_pte;
187 static u64 __read_mostly shadow_nx_mask;
188 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
189 static u64 __read_mostly shadow_user_mask;
190 static u64 __read_mostly shadow_accessed_mask;
191 static u64 __read_mostly shadow_dirty_mask;
192
193 static inline u64 rsvd_bits(int s, int e)
194 {
195         return ((1ULL << (e - s + 1)) - 1) << s;
196 }
197
198 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
199 {
200         shadow_trap_nonpresent_pte = trap_pte;
201         shadow_notrap_nonpresent_pte = notrap_pte;
202 }
203 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
204
205 void kvm_mmu_set_base_ptes(u64 base_pte)
206 {
207         shadow_base_present_pte = base_pte;
208 }
209 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
210
211 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
212                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
213 {
214         shadow_user_mask = user_mask;
215         shadow_accessed_mask = accessed_mask;
216         shadow_dirty_mask = dirty_mask;
217         shadow_nx_mask = nx_mask;
218         shadow_x_mask = x_mask;
219 }
220 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
221
222 static bool is_write_protection(struct kvm_vcpu *vcpu)
223 {
224         return kvm_read_cr0_bits(vcpu, X86_CR0_WP);
225 }
226
227 static int is_cpuid_PSE36(void)
228 {
229         return 1;
230 }
231
232 static int is_nx(struct kvm_vcpu *vcpu)
233 {
234         return vcpu->arch.efer & EFER_NX;
235 }
236
237 static int is_shadow_present_pte(u64 pte)
238 {
239         return pte != shadow_trap_nonpresent_pte
240                 && pte != shadow_notrap_nonpresent_pte;
241 }
242
243 static int is_large_pte(u64 pte)
244 {
245         return pte & PT_PAGE_SIZE_MASK;
246 }
247
248 static int is_writable_pte(unsigned long pte)
249 {
250         return pte & PT_WRITABLE_MASK;
251 }
252
253 static int is_dirty_gpte(unsigned long pte)
254 {
255         return pte & PT_DIRTY_MASK;
256 }
257
258 static int is_rmap_spte(u64 pte)
259 {
260         return is_shadow_present_pte(pte);
261 }
262
263 static int is_last_spte(u64 pte, int level)
264 {
265         if (level == PT_PAGE_TABLE_LEVEL)
266                 return 1;
267         if (is_large_pte(pte))
268                 return 1;
269         return 0;
270 }
271
272 static pfn_t spte_to_pfn(u64 pte)
273 {
274         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
275 }
276
277 static gfn_t pse36_gfn_delta(u32 gpte)
278 {
279         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
280
281         return (gpte & PT32_DIR_PSE36_MASK) << shift;
282 }
283
284 static void __set_spte(u64 *sptep, u64 spte)
285 {
286 #ifdef CONFIG_X86_64
287         set_64bit((unsigned long *)sptep, spte);
288 #else
289         set_64bit((unsigned long long *)sptep, spte);
290 #endif
291 }
292
293 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
294                                   struct kmem_cache *base_cache, int min)
295 {
296         void *obj;
297
298         if (cache->nobjs >= min)
299                 return 0;
300         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
301                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
302                 if (!obj)
303                         return -ENOMEM;
304                 cache->objects[cache->nobjs++] = obj;
305         }
306         return 0;
307 }
308
309 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc,
310                                   struct kmem_cache *cache)
311 {
312         while (mc->nobjs)
313                 kmem_cache_free(cache, mc->objects[--mc->nobjs]);
314 }
315
316 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
317                                        int min)
318 {
319         struct page *page;
320
321         if (cache->nobjs >= min)
322                 return 0;
323         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
324                 page = alloc_page(GFP_KERNEL);
325                 if (!page)
326                         return -ENOMEM;
327                 cache->objects[cache->nobjs++] = page_address(page);
328         }
329         return 0;
330 }
331
332 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
333 {
334         while (mc->nobjs)
335                 free_page((unsigned long)mc->objects[--mc->nobjs]);
336 }
337
338 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
339 {
340         int r;
341
342         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
343                                    pte_chain_cache, 4);
344         if (r)
345                 goto out;
346         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
347                                    rmap_desc_cache, 4);
348         if (r)
349                 goto out;
350         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
351         if (r)
352                 goto out;
353         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
354                                    mmu_page_header_cache, 4);
355 out:
356         return r;
357 }
358
359 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
360 {
361         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache, pte_chain_cache);
362         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache, rmap_desc_cache);
363         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
364         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache,
365                                 mmu_page_header_cache);
366 }
367
368 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
369                                     size_t size)
370 {
371         void *p;
372
373         BUG_ON(!mc->nobjs);
374         p = mc->objects[--mc->nobjs];
375         return p;
376 }
377
378 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
379 {
380         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
381                                       sizeof(struct kvm_pte_chain));
382 }
383
384 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
385 {
386         kmem_cache_free(pte_chain_cache, pc);
387 }
388
389 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
390 {
391         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
392                                       sizeof(struct kvm_rmap_desc));
393 }
394
395 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
396 {
397         kmem_cache_free(rmap_desc_cache, rd);
398 }
399
400 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
401 {
402         if (!sp->role.direct)
403                 return sp->gfns[index];
404
405         return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
406 }
407
408 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
409 {
410         if (sp->role.direct)
411                 BUG_ON(gfn != kvm_mmu_page_get_gfn(sp, index));
412         else
413                 sp->gfns[index] = gfn;
414 }
415
416 /*
417  * Return the pointer to the largepage write count for a given
418  * gfn, handling slots that are not large page aligned.
419  */
420 static int *slot_largepage_idx(gfn_t gfn,
421                                struct kvm_memory_slot *slot,
422                                int level)
423 {
424         unsigned long idx;
425
426         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
427               (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
428         return &slot->lpage_info[level - 2][idx].write_count;
429 }
430
431 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
432 {
433         struct kvm_memory_slot *slot;
434         int *write_count;
435         int i;
436
437         gfn = unalias_gfn(kvm, gfn);
438
439         slot = gfn_to_memslot_unaliased(kvm, gfn);
440         for (i = PT_DIRECTORY_LEVEL;
441              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
442                 write_count   = slot_largepage_idx(gfn, slot, i);
443                 *write_count += 1;
444         }
445 }
446
447 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
448 {
449         struct kvm_memory_slot *slot;
450         int *write_count;
451         int i;
452
453         gfn = unalias_gfn(kvm, gfn);
454         slot = gfn_to_memslot_unaliased(kvm, gfn);
455         for (i = PT_DIRECTORY_LEVEL;
456              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
457                 write_count   = slot_largepage_idx(gfn, slot, i);
458                 *write_count -= 1;
459                 WARN_ON(*write_count < 0);
460         }
461 }
462
463 static int has_wrprotected_page(struct kvm *kvm,
464                                 gfn_t gfn,
465                                 int level)
466 {
467         struct kvm_memory_slot *slot;
468         int *largepage_idx;
469
470         gfn = unalias_gfn(kvm, gfn);
471         slot = gfn_to_memslot_unaliased(kvm, gfn);
472         if (slot) {
473                 largepage_idx = slot_largepage_idx(gfn, slot, level);
474                 return *largepage_idx;
475         }
476
477         return 1;
478 }
479
480 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
481 {
482         unsigned long page_size;
483         int i, ret = 0;
484
485         page_size = kvm_host_page_size(kvm, gfn);
486
487         for (i = PT_PAGE_TABLE_LEVEL;
488              i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
489                 if (page_size >= KVM_HPAGE_SIZE(i))
490                         ret = i;
491                 else
492                         break;
493         }
494
495         return ret;
496 }
497
498 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
499 {
500         struct kvm_memory_slot *slot;
501         int host_level, level, max_level;
502
503         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
504         if (slot && slot->dirty_bitmap)
505                 return PT_PAGE_TABLE_LEVEL;
506
507         host_level = host_mapping_level(vcpu->kvm, large_gfn);
508
509         if (host_level == PT_PAGE_TABLE_LEVEL)
510                 return host_level;
511
512         max_level = kvm_x86_ops->get_lpage_level() < host_level ?
513                 kvm_x86_ops->get_lpage_level() : host_level;
514
515         for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
516                 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
517                         break;
518
519         return level - 1;
520 }
521
522 /*
523  * Take gfn and return the reverse mapping to it.
524  * Note: gfn must be unaliased before this function get called
525  */
526
527 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
528 {
529         struct kvm_memory_slot *slot;
530         unsigned long idx;
531
532         slot = gfn_to_memslot(kvm, gfn);
533         if (likely(level == PT_PAGE_TABLE_LEVEL))
534                 return &slot->rmap[gfn - slot->base_gfn];
535
536         idx = (gfn / KVM_PAGES_PER_HPAGE(level)) -
537                 (slot->base_gfn / KVM_PAGES_PER_HPAGE(level));
538
539         return &slot->lpage_info[level - 2][idx].rmap_pde;
540 }
541
542 /*
543  * Reverse mapping data structures:
544  *
545  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
546  * that points to page_address(page).
547  *
548  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
549  * containing more mappings.
550  *
551  * Returns the number of rmap entries before the spte was added or zero if
552  * the spte was not added.
553  *
554  */
555 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
556 {
557         struct kvm_mmu_page *sp;
558         struct kvm_rmap_desc *desc;
559         unsigned long *rmapp;
560         int i, count = 0;
561
562         if (!is_rmap_spte(*spte))
563                 return count;
564         gfn = unalias_gfn(vcpu->kvm, gfn);
565         sp = page_header(__pa(spte));
566         kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
567         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
568         if (!*rmapp) {
569                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
570                 *rmapp = (unsigned long)spte;
571         } else if (!(*rmapp & 1)) {
572                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
573                 desc = mmu_alloc_rmap_desc(vcpu);
574                 desc->sptes[0] = (u64 *)*rmapp;
575                 desc->sptes[1] = spte;
576                 *rmapp = (unsigned long)desc | 1;
577         } else {
578                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
579                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
580                 while (desc->sptes[RMAP_EXT-1] && desc->more) {
581                         desc = desc->more;
582                         count += RMAP_EXT;
583                 }
584                 if (desc->sptes[RMAP_EXT-1]) {
585                         desc->more = mmu_alloc_rmap_desc(vcpu);
586                         desc = desc->more;
587                 }
588                 for (i = 0; desc->sptes[i]; ++i)
589                         ;
590                 desc->sptes[i] = spte;
591         }
592         return count;
593 }
594
595 static void rmap_desc_remove_entry(unsigned long *rmapp,
596                                    struct kvm_rmap_desc *desc,
597                                    int i,
598                                    struct kvm_rmap_desc *prev_desc)
599 {
600         int j;
601
602         for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
603                 ;
604         desc->sptes[i] = desc->sptes[j];
605         desc->sptes[j] = NULL;
606         if (j != 0)
607                 return;
608         if (!prev_desc && !desc->more)
609                 *rmapp = (unsigned long)desc->sptes[0];
610         else
611                 if (prev_desc)
612                         prev_desc->more = desc->more;
613                 else
614                         *rmapp = (unsigned long)desc->more | 1;
615         mmu_free_rmap_desc(desc);
616 }
617
618 static void rmap_remove(struct kvm *kvm, u64 *spte)
619 {
620         struct kvm_rmap_desc *desc;
621         struct kvm_rmap_desc *prev_desc;
622         struct kvm_mmu_page *sp;
623         pfn_t pfn;
624         gfn_t gfn;
625         unsigned long *rmapp;
626         int i;
627
628         if (!is_rmap_spte(*spte))
629                 return;
630         sp = page_header(__pa(spte));
631         pfn = spte_to_pfn(*spte);
632         if (*spte & shadow_accessed_mask)
633                 kvm_set_pfn_accessed(pfn);
634         if (is_writable_pte(*spte))
635                 kvm_set_pfn_dirty(pfn);
636         gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
637         rmapp = gfn_to_rmap(kvm, gfn, sp->role.level);
638         if (!*rmapp) {
639                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
640                 BUG();
641         } else if (!(*rmapp & 1)) {
642                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
643                 if ((u64 *)*rmapp != spte) {
644                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
645                                spte, *spte);
646                         BUG();
647                 }
648                 *rmapp = 0;
649         } else {
650                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
651                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
652                 prev_desc = NULL;
653                 while (desc) {
654                         for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
655                                 if (desc->sptes[i] == spte) {
656                                         rmap_desc_remove_entry(rmapp,
657                                                                desc, i,
658                                                                prev_desc);
659                                         return;
660                                 }
661                         prev_desc = desc;
662                         desc = desc->more;
663                 }
664                 pr_err("rmap_remove: %p %llx many->many\n", spte, *spte);
665                 BUG();
666         }
667 }
668
669 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
670 {
671         struct kvm_rmap_desc *desc;
672         u64 *prev_spte;
673         int i;
674
675         if (!*rmapp)
676                 return NULL;
677         else if (!(*rmapp & 1)) {
678                 if (!spte)
679                         return (u64 *)*rmapp;
680                 return NULL;
681         }
682         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
683         prev_spte = NULL;
684         while (desc) {
685                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
686                         if (prev_spte == spte)
687                                 return desc->sptes[i];
688                         prev_spte = desc->sptes[i];
689                 }
690                 desc = desc->more;
691         }
692         return NULL;
693 }
694
695 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
696 {
697         unsigned long *rmapp;
698         u64 *spte;
699         int i, write_protected = 0;
700
701         gfn = unalias_gfn(kvm, gfn);
702         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
703
704         spte = rmap_next(kvm, rmapp, NULL);
705         while (spte) {
706                 BUG_ON(!spte);
707                 BUG_ON(!(*spte & PT_PRESENT_MASK));
708                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
709                 if (is_writable_pte(*spte)) {
710                         __set_spte(spte, *spte & ~PT_WRITABLE_MASK);
711                         write_protected = 1;
712                 }
713                 spte = rmap_next(kvm, rmapp, spte);
714         }
715         if (write_protected) {
716                 pfn_t pfn;
717
718                 spte = rmap_next(kvm, rmapp, NULL);
719                 pfn = spte_to_pfn(*spte);
720                 kvm_set_pfn_dirty(pfn);
721         }
722
723         /* check for huge page mappings */
724         for (i = PT_DIRECTORY_LEVEL;
725              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
726                 rmapp = gfn_to_rmap(kvm, gfn, i);
727                 spte = rmap_next(kvm, rmapp, NULL);
728                 while (spte) {
729                         BUG_ON(!spte);
730                         BUG_ON(!(*spte & PT_PRESENT_MASK));
731                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
732                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
733                         if (is_writable_pte(*spte)) {
734                                 rmap_remove(kvm, spte);
735                                 --kvm->stat.lpages;
736                                 __set_spte(spte, shadow_trap_nonpresent_pte);
737                                 spte = NULL;
738                                 write_protected = 1;
739                         }
740                         spte = rmap_next(kvm, rmapp, spte);
741                 }
742         }
743
744         return write_protected;
745 }
746
747 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
748                            unsigned long data)
749 {
750         u64 *spte;
751         int need_tlb_flush = 0;
752
753         while ((spte = rmap_next(kvm, rmapp, NULL))) {
754                 BUG_ON(!(*spte & PT_PRESENT_MASK));
755                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
756                 rmap_remove(kvm, spte);
757                 __set_spte(spte, shadow_trap_nonpresent_pte);
758                 need_tlb_flush = 1;
759         }
760         return need_tlb_flush;
761 }
762
763 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
764                              unsigned long data)
765 {
766         int need_flush = 0;
767         u64 *spte, new_spte;
768         pte_t *ptep = (pte_t *)data;
769         pfn_t new_pfn;
770
771         WARN_ON(pte_huge(*ptep));
772         new_pfn = pte_pfn(*ptep);
773         spte = rmap_next(kvm, rmapp, NULL);
774         while (spte) {
775                 BUG_ON(!is_shadow_present_pte(*spte));
776                 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
777                 need_flush = 1;
778                 if (pte_write(*ptep)) {
779                         rmap_remove(kvm, spte);
780                         __set_spte(spte, shadow_trap_nonpresent_pte);
781                         spte = rmap_next(kvm, rmapp, NULL);
782                 } else {
783                         new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
784                         new_spte |= (u64)new_pfn << PAGE_SHIFT;
785
786                         new_spte &= ~PT_WRITABLE_MASK;
787                         new_spte &= ~SPTE_HOST_WRITEABLE;
788                         if (is_writable_pte(*spte))
789                                 kvm_set_pfn_dirty(spte_to_pfn(*spte));
790                         __set_spte(spte, new_spte);
791                         spte = rmap_next(kvm, rmapp, spte);
792                 }
793         }
794         if (need_flush)
795                 kvm_flush_remote_tlbs(kvm);
796
797         return 0;
798 }
799
800 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
801                           unsigned long data,
802                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
803                                          unsigned long data))
804 {
805         int i, j;
806         int ret;
807         int retval = 0;
808         struct kvm_memslots *slots;
809
810         slots = kvm_memslots(kvm);
811
812         for (i = 0; i < slots->nmemslots; i++) {
813                 struct kvm_memory_slot *memslot = &slots->memslots[i];
814                 unsigned long start = memslot->userspace_addr;
815                 unsigned long end;
816
817                 end = start + (memslot->npages << PAGE_SHIFT);
818                 if (hva >= start && hva < end) {
819                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
820
821                         ret = handler(kvm, &memslot->rmap[gfn_offset], data);
822
823                         for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
824                                 int idx = gfn_offset;
825                                 idx /= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL + j);
826                                 ret |= handler(kvm,
827                                         &memslot->lpage_info[j][idx].rmap_pde,
828                                         data);
829                         }
830                         trace_kvm_age_page(hva, memslot, ret);
831                         retval |= ret;
832                 }
833         }
834
835         return retval;
836 }
837
838 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
839 {
840         return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
841 }
842
843 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
844 {
845         kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
846 }
847
848 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
849                          unsigned long data)
850 {
851         u64 *spte;
852         int young = 0;
853
854         /*
855          * Emulate the accessed bit for EPT, by checking if this page has
856          * an EPT mapping, and clearing it if it does. On the next access,
857          * a new EPT mapping will be established.
858          * This has some overhead, but not as much as the cost of swapping
859          * out actively used pages or breaking up actively used hugepages.
860          */
861         if (!shadow_accessed_mask)
862                 return kvm_unmap_rmapp(kvm, rmapp, data);
863
864         spte = rmap_next(kvm, rmapp, NULL);
865         while (spte) {
866                 int _young;
867                 u64 _spte = *spte;
868                 BUG_ON(!(_spte & PT_PRESENT_MASK));
869                 _young = _spte & PT_ACCESSED_MASK;
870                 if (_young) {
871                         young = 1;
872                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
873                 }
874                 spte = rmap_next(kvm, rmapp, spte);
875         }
876         return young;
877 }
878
879 #define RMAP_RECYCLE_THRESHOLD 1000
880
881 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
882 {
883         unsigned long *rmapp;
884         struct kvm_mmu_page *sp;
885
886         sp = page_header(__pa(spte));
887
888         gfn = unalias_gfn(vcpu->kvm, gfn);
889         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
890
891         kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
892         kvm_flush_remote_tlbs(vcpu->kvm);
893 }
894
895 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
896 {
897         return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
898 }
899
900 #ifdef MMU_DEBUG
901 static int is_empty_shadow_page(u64 *spt)
902 {
903         u64 *pos;
904         u64 *end;
905
906         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
907                 if (is_shadow_present_pte(*pos)) {
908                         printk(KERN_ERR "%s: %p %llx\n", __func__,
909                                pos, *pos);
910                         return 0;
911                 }
912         return 1;
913 }
914 #endif
915
916 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
917 {
918         ASSERT(is_empty_shadow_page(sp->spt));
919         hlist_del(&sp->hash_link);
920         list_del(&sp->link);
921         __free_page(virt_to_page(sp->spt));
922         if (!sp->role.direct)
923                 __free_page(virt_to_page(sp->gfns));
924         kmem_cache_free(mmu_page_header_cache, sp);
925         ++kvm->arch.n_free_mmu_pages;
926 }
927
928 static unsigned kvm_page_table_hashfn(gfn_t gfn)
929 {
930         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
931 }
932
933 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
934                                                u64 *parent_pte, int direct)
935 {
936         struct kvm_mmu_page *sp;
937
938         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
939         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
940         if (!direct)
941                 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache,
942                                                   PAGE_SIZE);
943         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
944         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
945         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
946         sp->multimapped = 0;
947         sp->parent_pte = parent_pte;
948         --vcpu->kvm->arch.n_free_mmu_pages;
949         return sp;
950 }
951
952 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
953                                     struct kvm_mmu_page *sp, u64 *parent_pte)
954 {
955         struct kvm_pte_chain *pte_chain;
956         struct hlist_node *node;
957         int i;
958
959         if (!parent_pte)
960                 return;
961         if (!sp->multimapped) {
962                 u64 *old = sp->parent_pte;
963
964                 if (!old) {
965                         sp->parent_pte = parent_pte;
966                         return;
967                 }
968                 sp->multimapped = 1;
969                 pte_chain = mmu_alloc_pte_chain(vcpu);
970                 INIT_HLIST_HEAD(&sp->parent_ptes);
971                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
972                 pte_chain->parent_ptes[0] = old;
973         }
974         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
975                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
976                         continue;
977                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
978                         if (!pte_chain->parent_ptes[i]) {
979                                 pte_chain->parent_ptes[i] = parent_pte;
980                                 return;
981                         }
982         }
983         pte_chain = mmu_alloc_pte_chain(vcpu);
984         BUG_ON(!pte_chain);
985         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
986         pte_chain->parent_ptes[0] = parent_pte;
987 }
988
989 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
990                                        u64 *parent_pte)
991 {
992         struct kvm_pte_chain *pte_chain;
993         struct hlist_node *node;
994         int i;
995
996         if (!sp->multimapped) {
997                 BUG_ON(sp->parent_pte != parent_pte);
998                 sp->parent_pte = NULL;
999                 return;
1000         }
1001         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1002                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1003                         if (!pte_chain->parent_ptes[i])
1004                                 break;
1005                         if (pte_chain->parent_ptes[i] != parent_pte)
1006                                 continue;
1007                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
1008                                 && pte_chain->parent_ptes[i + 1]) {
1009                                 pte_chain->parent_ptes[i]
1010                                         = pte_chain->parent_ptes[i + 1];
1011                                 ++i;
1012                         }
1013                         pte_chain->parent_ptes[i] = NULL;
1014                         if (i == 0) {
1015                                 hlist_del(&pte_chain->link);
1016                                 mmu_free_pte_chain(pte_chain);
1017                                 if (hlist_empty(&sp->parent_ptes)) {
1018                                         sp->multimapped = 0;
1019                                         sp->parent_pte = NULL;
1020                                 }
1021                         }
1022                         return;
1023                 }
1024         BUG();
1025 }
1026
1027 static void mmu_parent_walk(struct kvm_mmu_page *sp, mmu_parent_walk_fn fn)
1028 {
1029         struct kvm_pte_chain *pte_chain;
1030         struct hlist_node *node;
1031         struct kvm_mmu_page *parent_sp;
1032         int i;
1033
1034         if (!sp->multimapped && sp->parent_pte) {
1035                 parent_sp = page_header(__pa(sp->parent_pte));
1036                 fn(parent_sp, sp->parent_pte);
1037                 return;
1038         }
1039
1040         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1041                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1042                         u64 *spte = pte_chain->parent_ptes[i];
1043
1044                         if (!spte)
1045                                 break;
1046                         parent_sp = page_header(__pa(spte));
1047                         fn(parent_sp, spte);
1048                 }
1049 }
1050
1051 static void mark_unsync(struct kvm_mmu_page *sp, u64 *spte);
1052 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1053 {
1054         mmu_parent_walk(sp, mark_unsync);
1055 }
1056
1057 static void mark_unsync(struct kvm_mmu_page *sp, u64 *spte)
1058 {
1059         unsigned int index;
1060
1061         index = spte - sp->spt;
1062         if (__test_and_set_bit(index, sp->unsync_child_bitmap))
1063                 return;
1064         if (sp->unsync_children++)
1065                 return;
1066         kvm_mmu_mark_parents_unsync(sp);
1067 }
1068
1069 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1070                                     struct kvm_mmu_page *sp)
1071 {
1072         int i;
1073
1074         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1075                 sp->spt[i] = shadow_trap_nonpresent_pte;
1076 }
1077
1078 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1079                                struct kvm_mmu_page *sp, bool clear_unsync)
1080 {
1081         return 1;
1082 }
1083
1084 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1085 {
1086 }
1087
1088 #define KVM_PAGE_ARRAY_NR 16
1089
1090 struct kvm_mmu_pages {
1091         struct mmu_page_and_offset {
1092                 struct kvm_mmu_page *sp;
1093                 unsigned int idx;
1094         } page[KVM_PAGE_ARRAY_NR];
1095         unsigned int nr;
1096 };
1097
1098 #define for_each_unsync_children(bitmap, idx)           \
1099         for (idx = find_first_bit(bitmap, 512);         \
1100              idx < 512;                                 \
1101              idx = find_next_bit(bitmap, 512, idx+1))
1102
1103 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1104                          int idx)
1105 {
1106         int i;
1107
1108         if (sp->unsync)
1109                 for (i=0; i < pvec->nr; i++)
1110                         if (pvec->page[i].sp == sp)
1111                                 return 0;
1112
1113         pvec->page[pvec->nr].sp = sp;
1114         pvec->page[pvec->nr].idx = idx;
1115         pvec->nr++;
1116         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1117 }
1118
1119 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1120                            struct kvm_mmu_pages *pvec)
1121 {
1122         int i, ret, nr_unsync_leaf = 0;
1123
1124         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1125                 struct kvm_mmu_page *child;
1126                 u64 ent = sp->spt[i];
1127
1128                 if (!is_shadow_present_pte(ent) || is_large_pte(ent))
1129                         goto clear_child_bitmap;
1130
1131                 child = page_header(ent & PT64_BASE_ADDR_MASK);
1132
1133                 if (child->unsync_children) {
1134                         if (mmu_pages_add(pvec, child, i))
1135                                 return -ENOSPC;
1136
1137                         ret = __mmu_unsync_walk(child, pvec);
1138                         if (!ret)
1139                                 goto clear_child_bitmap;
1140                         else if (ret > 0)
1141                                 nr_unsync_leaf += ret;
1142                         else
1143                                 return ret;
1144                 } else if (child->unsync) {
1145                         nr_unsync_leaf++;
1146                         if (mmu_pages_add(pvec, child, i))
1147                                 return -ENOSPC;
1148                 } else
1149                          goto clear_child_bitmap;
1150
1151                 continue;
1152
1153 clear_child_bitmap:
1154                 __clear_bit(i, sp->unsync_child_bitmap);
1155                 sp->unsync_children--;
1156                 WARN_ON((int)sp->unsync_children < 0);
1157         }
1158
1159
1160         return nr_unsync_leaf;
1161 }
1162
1163 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1164                            struct kvm_mmu_pages *pvec)
1165 {
1166         if (!sp->unsync_children)
1167                 return 0;
1168
1169         mmu_pages_add(pvec, sp, 0);
1170         return __mmu_unsync_walk(sp, pvec);
1171 }
1172
1173 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1174 {
1175         WARN_ON(!sp->unsync);
1176         trace_kvm_mmu_sync_page(sp);
1177         sp->unsync = 0;
1178         --kvm->stat.mmu_unsync;
1179 }
1180
1181 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1182                                     struct list_head *invalid_list);
1183 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1184                                     struct list_head *invalid_list);
1185
1186 #define for_each_gfn_sp(kvm, sp, gfn, pos)                              \
1187   hlist_for_each_entry(sp, pos,                                         \
1188    &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link)   \
1189         if ((sp)->gfn != (gfn)) {} else
1190
1191 #define for_each_gfn_indirect_valid_sp(kvm, sp, gfn, pos)               \
1192   hlist_for_each_entry(sp, pos,                                         \
1193    &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link)   \
1194                 if ((sp)->gfn != (gfn) || (sp)->role.direct ||          \
1195                         (sp)->role.invalid) {} else
1196
1197 /* @sp->gfn should be write-protected at the call site */
1198 static int __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1199                            struct list_head *invalid_list, bool clear_unsync)
1200 {
1201         if (sp->role.cr4_pae != !!is_pae(vcpu)) {
1202                 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1203                 return 1;
1204         }
1205
1206         if (clear_unsync)
1207                 kvm_unlink_unsync_page(vcpu->kvm, sp);
1208
1209         if (vcpu->arch.mmu.sync_page(vcpu, sp, clear_unsync)) {
1210                 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1211                 return 1;
1212         }
1213
1214         kvm_mmu_flush_tlb(vcpu);
1215         return 0;
1216 }
1217
1218 static int kvm_sync_page_transient(struct kvm_vcpu *vcpu,
1219                                    struct kvm_mmu_page *sp)
1220 {
1221         LIST_HEAD(invalid_list);
1222         int ret;
1223
1224         ret = __kvm_sync_page(vcpu, sp, &invalid_list, false);
1225         if (ret)
1226                 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1227
1228         return ret;
1229 }
1230
1231 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1232                          struct list_head *invalid_list)
1233 {
1234         return __kvm_sync_page(vcpu, sp, invalid_list, true);
1235 }
1236
1237 /* @gfn should be write-protected at the call site */
1238 static void kvm_sync_pages(struct kvm_vcpu *vcpu,  gfn_t gfn)
1239 {
1240         struct kvm_mmu_page *s;
1241         struct hlist_node *node;
1242         LIST_HEAD(invalid_list);
1243         bool flush = false;
1244
1245         for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1246                 if (!s->unsync)
1247                         continue;
1248
1249                 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1250                 if ((s->role.cr4_pae != !!is_pae(vcpu)) ||
1251                         (vcpu->arch.mmu.sync_page(vcpu, s, true))) {
1252                         kvm_mmu_prepare_zap_page(vcpu->kvm, s, &invalid_list);
1253                         continue;
1254                 }
1255                 kvm_unlink_unsync_page(vcpu->kvm, s);
1256                 flush = true;
1257         }
1258
1259         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1260         if (flush)
1261                 kvm_mmu_flush_tlb(vcpu);
1262 }
1263
1264 struct mmu_page_path {
1265         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1266         unsigned int idx[PT64_ROOT_LEVEL-1];
1267 };
1268
1269 #define for_each_sp(pvec, sp, parents, i)                       \
1270                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1271                         sp = pvec.page[i].sp;                   \
1272                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1273                         i = mmu_pages_next(&pvec, &parents, i))
1274
1275 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1276                           struct mmu_page_path *parents,
1277                           int i)
1278 {
1279         int n;
1280
1281         for (n = i+1; n < pvec->nr; n++) {
1282                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1283
1284                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1285                         parents->idx[0] = pvec->page[n].idx;
1286                         return n;
1287                 }
1288
1289                 parents->parent[sp->role.level-2] = sp;
1290                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1291         }
1292
1293         return n;
1294 }
1295
1296 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1297 {
1298         struct kvm_mmu_page *sp;
1299         unsigned int level = 0;
1300
1301         do {
1302                 unsigned int idx = parents->idx[level];
1303
1304                 sp = parents->parent[level];
1305                 if (!sp)
1306                         return;
1307
1308                 --sp->unsync_children;
1309                 WARN_ON((int)sp->unsync_children < 0);
1310                 __clear_bit(idx, sp->unsync_child_bitmap);
1311                 level++;
1312         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1313 }
1314
1315 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1316                                struct mmu_page_path *parents,
1317                                struct kvm_mmu_pages *pvec)
1318 {
1319         parents->parent[parent->role.level-1] = NULL;
1320         pvec->nr = 0;
1321 }
1322
1323 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1324                               struct kvm_mmu_page *parent)
1325 {
1326         int i;
1327         struct kvm_mmu_page *sp;
1328         struct mmu_page_path parents;
1329         struct kvm_mmu_pages pages;
1330         LIST_HEAD(invalid_list);
1331
1332         kvm_mmu_pages_init(parent, &parents, &pages);
1333         while (mmu_unsync_walk(parent, &pages)) {
1334                 int protected = 0;
1335
1336                 for_each_sp(pages, sp, parents, i)
1337                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1338
1339                 if (protected)
1340                         kvm_flush_remote_tlbs(vcpu->kvm);
1341
1342                 for_each_sp(pages, sp, parents, i) {
1343                         kvm_sync_page(vcpu, sp, &invalid_list);
1344                         mmu_pages_clear_parents(&parents);
1345                 }
1346                 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1347                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1348                 kvm_mmu_pages_init(parent, &parents, &pages);
1349         }
1350 }
1351
1352 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1353                                              gfn_t gfn,
1354                                              gva_t gaddr,
1355                                              unsigned level,
1356                                              int direct,
1357                                              unsigned access,
1358                                              u64 *parent_pte)
1359 {
1360         union kvm_mmu_page_role role;
1361         unsigned quadrant;
1362         struct kvm_mmu_page *sp;
1363         struct hlist_node *node;
1364         bool need_sync = false;
1365
1366         role = vcpu->arch.mmu.base_role;
1367         role.level = level;
1368         role.direct = direct;
1369         if (role.direct)
1370                 role.cr4_pae = 0;
1371         role.access = access;
1372         if (!tdp_enabled && vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1373                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1374                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1375                 role.quadrant = quadrant;
1376         }
1377         for_each_gfn_sp(vcpu->kvm, sp, gfn, node) {
1378                 if (!need_sync && sp->unsync)
1379                         need_sync = true;
1380
1381                 if (sp->role.word != role.word)
1382                         continue;
1383
1384                 if (sp->unsync && kvm_sync_page_transient(vcpu, sp))
1385                         break;
1386
1387                 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1388                 if (sp->unsync_children) {
1389                         set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1390                         kvm_mmu_mark_parents_unsync(sp);
1391                 } else if (sp->unsync)
1392                         kvm_mmu_mark_parents_unsync(sp);
1393
1394                 trace_kvm_mmu_get_page(sp, false);
1395                 return sp;
1396         }
1397         ++vcpu->kvm->stat.mmu_cache_miss;
1398         sp = kvm_mmu_alloc_page(vcpu, parent_pte, direct);
1399         if (!sp)
1400                 return sp;
1401         sp->gfn = gfn;
1402         sp->role = role;
1403         hlist_add_head(&sp->hash_link,
1404                 &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
1405         if (!direct) {
1406                 if (rmap_write_protect(vcpu->kvm, gfn))
1407                         kvm_flush_remote_tlbs(vcpu->kvm);
1408                 if (level > PT_PAGE_TABLE_LEVEL && need_sync)
1409                         kvm_sync_pages(vcpu, gfn);
1410
1411                 account_shadowed(vcpu->kvm, gfn);
1412         }
1413         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1414                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1415         else
1416                 nonpaging_prefetch_page(vcpu, sp);
1417         trace_kvm_mmu_get_page(sp, true);
1418         return sp;
1419 }
1420
1421 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1422                              struct kvm_vcpu *vcpu, u64 addr)
1423 {
1424         iterator->addr = addr;
1425         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1426         iterator->level = vcpu->arch.mmu.shadow_root_level;
1427         if (iterator->level == PT32E_ROOT_LEVEL) {
1428                 iterator->shadow_addr
1429                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1430                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1431                 --iterator->level;
1432                 if (!iterator->shadow_addr)
1433                         iterator->level = 0;
1434         }
1435 }
1436
1437 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1438 {
1439         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1440                 return false;
1441
1442         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1443                 if (is_large_pte(*iterator->sptep))
1444                         return false;
1445
1446         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1447         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1448         return true;
1449 }
1450
1451 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1452 {
1453         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1454         --iterator->level;
1455 }
1456
1457 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1458                                          struct kvm_mmu_page *sp)
1459 {
1460         unsigned i;
1461         u64 *pt;
1462         u64 ent;
1463
1464         pt = sp->spt;
1465
1466         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1467                 ent = pt[i];
1468
1469                 if (is_shadow_present_pte(ent)) {
1470                         if (!is_last_spte(ent, sp->role.level)) {
1471                                 ent &= PT64_BASE_ADDR_MASK;
1472                                 mmu_page_remove_parent_pte(page_header(ent),
1473                                                            &pt[i]);
1474                         } else {
1475                                 if (is_large_pte(ent))
1476                                         --kvm->stat.lpages;
1477                                 rmap_remove(kvm, &pt[i]);
1478                         }
1479                 }
1480                 pt[i] = shadow_trap_nonpresent_pte;
1481         }
1482 }
1483
1484 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1485 {
1486         mmu_page_remove_parent_pte(sp, parent_pte);
1487 }
1488
1489 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1490 {
1491         int i;
1492         struct kvm_vcpu *vcpu;
1493
1494         kvm_for_each_vcpu(i, vcpu, kvm)
1495                 vcpu->arch.last_pte_updated = NULL;
1496 }
1497
1498 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1499 {
1500         u64 *parent_pte;
1501
1502         while (sp->multimapped || sp->parent_pte) {
1503                 if (!sp->multimapped)
1504                         parent_pte = sp->parent_pte;
1505                 else {
1506                         struct kvm_pte_chain *chain;
1507
1508                         chain = container_of(sp->parent_ptes.first,
1509                                              struct kvm_pte_chain, link);
1510                         parent_pte = chain->parent_ptes[0];
1511                 }
1512                 BUG_ON(!parent_pte);
1513                 kvm_mmu_put_page(sp, parent_pte);
1514                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1515         }
1516 }
1517
1518 static int mmu_zap_unsync_children(struct kvm *kvm,
1519                                    struct kvm_mmu_page *parent,
1520                                    struct list_head *invalid_list)
1521 {
1522         int i, zapped = 0;
1523         struct mmu_page_path parents;
1524         struct kvm_mmu_pages pages;
1525
1526         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1527                 return 0;
1528
1529         kvm_mmu_pages_init(parent, &parents, &pages);
1530         while (mmu_unsync_walk(parent, &pages)) {
1531                 struct kvm_mmu_page *sp;
1532
1533                 for_each_sp(pages, sp, parents, i) {
1534                         kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
1535                         mmu_pages_clear_parents(&parents);
1536                         zapped++;
1537                 }
1538                 kvm_mmu_pages_init(parent, &parents, &pages);
1539         }
1540
1541         return zapped;
1542 }
1543
1544 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1545                                     struct list_head *invalid_list)
1546 {
1547         int ret;
1548
1549         trace_kvm_mmu_prepare_zap_page(sp);
1550         ++kvm->stat.mmu_shadow_zapped;
1551         ret = mmu_zap_unsync_children(kvm, sp, invalid_list);
1552         kvm_mmu_page_unlink_children(kvm, sp);
1553         kvm_mmu_unlink_parents(kvm, sp);
1554         if (!sp->role.invalid && !sp->role.direct)
1555                 unaccount_shadowed(kvm, sp->gfn);
1556         if (sp->unsync)
1557                 kvm_unlink_unsync_page(kvm, sp);
1558         if (!sp->root_count) {
1559                 /* Count self */
1560                 ret++;
1561                 list_move(&sp->link, invalid_list);
1562         } else {
1563                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1564                 kvm_reload_remote_mmus(kvm);
1565         }
1566
1567         sp->role.invalid = 1;
1568         kvm_mmu_reset_last_pte_updated(kvm);
1569         return ret;
1570 }
1571
1572 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1573                                     struct list_head *invalid_list)
1574 {
1575         struct kvm_mmu_page *sp;
1576
1577         if (list_empty(invalid_list))
1578                 return;
1579
1580         kvm_flush_remote_tlbs(kvm);
1581
1582         do {
1583                 sp = list_first_entry(invalid_list, struct kvm_mmu_page, link);
1584                 WARN_ON(!sp->role.invalid || sp->root_count);
1585                 kvm_mmu_free_page(kvm, sp);
1586         } while (!list_empty(invalid_list));
1587
1588 }
1589
1590 /*
1591  * Changing the number of mmu pages allocated to the vm
1592  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1593  */
1594 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1595 {
1596         int used_pages;
1597         LIST_HEAD(invalid_list);
1598
1599         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1600         used_pages = max(0, used_pages);
1601
1602         /*
1603          * If we set the number of mmu pages to be smaller be than the
1604          * number of actived pages , we must to free some mmu pages before we
1605          * change the value
1606          */
1607
1608         if (used_pages > kvm_nr_mmu_pages) {
1609                 while (used_pages > kvm_nr_mmu_pages &&
1610                         !list_empty(&kvm->arch.active_mmu_pages)) {
1611                         struct kvm_mmu_page *page;
1612
1613                         page = container_of(kvm->arch.active_mmu_pages.prev,
1614                                             struct kvm_mmu_page, link);
1615                         used_pages -= kvm_mmu_prepare_zap_page(kvm, page,
1616                                                                &invalid_list);
1617                 }
1618                 kvm_mmu_commit_zap_page(kvm, &invalid_list);
1619                 kvm_nr_mmu_pages = used_pages;
1620                 kvm->arch.n_free_mmu_pages = 0;
1621         }
1622         else
1623                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1624                                          - kvm->arch.n_alloc_mmu_pages;
1625
1626         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1627 }
1628
1629 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1630 {
1631         struct kvm_mmu_page *sp;
1632         struct hlist_node *node;
1633         LIST_HEAD(invalid_list);
1634         int r;
1635
1636         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1637         r = 0;
1638
1639         for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
1640                 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1641                          sp->role.word);
1642                 r = 1;
1643                 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
1644         }
1645         kvm_mmu_commit_zap_page(kvm, &invalid_list);
1646         return r;
1647 }
1648
1649 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1650 {
1651         struct kvm_mmu_page *sp;
1652         struct hlist_node *node;
1653         LIST_HEAD(invalid_list);
1654
1655         for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
1656                 pgprintk("%s: zap %lx %x\n",
1657                          __func__, gfn, sp->role.word);
1658                 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
1659         }
1660         kvm_mmu_commit_zap_page(kvm, &invalid_list);
1661 }
1662
1663 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1664 {
1665         int slot = memslot_id(kvm, gfn);
1666         struct kvm_mmu_page *sp = page_header(__pa(pte));
1667
1668         __set_bit(slot, sp->slot_bitmap);
1669 }
1670
1671 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1672 {
1673         int i;
1674         u64 *pt = sp->spt;
1675
1676         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1677                 return;
1678
1679         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1680                 if (pt[i] == shadow_notrap_nonpresent_pte)
1681                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1682         }
1683 }
1684
1685 /*
1686  * The function is based on mtrr_type_lookup() in
1687  * arch/x86/kernel/cpu/mtrr/generic.c
1688  */
1689 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1690                          u64 start, u64 end)
1691 {
1692         int i;
1693         u64 base, mask;
1694         u8 prev_match, curr_match;
1695         int num_var_ranges = KVM_NR_VAR_MTRR;
1696
1697         if (!mtrr_state->enabled)
1698                 return 0xFF;
1699
1700         /* Make end inclusive end, instead of exclusive */
1701         end--;
1702
1703         /* Look in fixed ranges. Just return the type as per start */
1704         if (mtrr_state->have_fixed && (start < 0x100000)) {
1705                 int idx;
1706
1707                 if (start < 0x80000) {
1708                         idx = 0;
1709                         idx += (start >> 16);
1710                         return mtrr_state->fixed_ranges[idx];
1711                 } else if (start < 0xC0000) {
1712                         idx = 1 * 8;
1713                         idx += ((start - 0x80000) >> 14);
1714                         return mtrr_state->fixed_ranges[idx];
1715                 } else if (start < 0x1000000) {
1716                         idx = 3 * 8;
1717                         idx += ((start - 0xC0000) >> 12);
1718                         return mtrr_state->fixed_ranges[idx];
1719                 }
1720         }
1721
1722         /*
1723          * Look in variable ranges
1724          * Look of multiple ranges matching this address and pick type
1725          * as per MTRR precedence
1726          */
1727         if (!(mtrr_state->enabled & 2))
1728                 return mtrr_state->def_type;
1729
1730         prev_match = 0xFF;
1731         for (i = 0; i < num_var_ranges; ++i) {
1732                 unsigned short start_state, end_state;
1733
1734                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1735                         continue;
1736
1737                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1738                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1739                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1740                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1741
1742                 start_state = ((start & mask) == (base & mask));
1743                 end_state = ((end & mask) == (base & mask));
1744                 if (start_state != end_state)
1745                         return 0xFE;
1746
1747                 if ((start & mask) != (base & mask))
1748                         continue;
1749
1750                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1751                 if (prev_match == 0xFF) {
1752                         prev_match = curr_match;
1753                         continue;
1754                 }
1755
1756                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1757                     curr_match == MTRR_TYPE_UNCACHABLE)
1758                         return MTRR_TYPE_UNCACHABLE;
1759
1760                 if ((prev_match == MTRR_TYPE_WRBACK &&
1761                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1762                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1763                      curr_match == MTRR_TYPE_WRBACK)) {
1764                         prev_match = MTRR_TYPE_WRTHROUGH;
1765                         curr_match = MTRR_TYPE_WRTHROUGH;
1766                 }
1767
1768                 if (prev_match != curr_match)
1769                         return MTRR_TYPE_UNCACHABLE;
1770         }
1771
1772         if (prev_match != 0xFF)
1773                 return prev_match;
1774
1775         return mtrr_state->def_type;
1776 }
1777
1778 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1779 {
1780         u8 mtrr;
1781
1782         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1783                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1784         if (mtrr == 0xfe || mtrr == 0xff)
1785                 mtrr = MTRR_TYPE_WRBACK;
1786         return mtrr;
1787 }
1788 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1789
1790 static void __kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1791 {
1792         trace_kvm_mmu_unsync_page(sp);
1793         ++vcpu->kvm->stat.mmu_unsync;
1794         sp->unsync = 1;
1795
1796         kvm_mmu_mark_parents_unsync(sp);
1797         mmu_convert_notrap(sp);
1798 }
1799
1800 static void kvm_unsync_pages(struct kvm_vcpu *vcpu,  gfn_t gfn)
1801 {
1802         struct kvm_mmu_page *s;
1803         struct hlist_node *node;
1804
1805         for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1806                 if (s->unsync)
1807                         continue;
1808                 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1809                 __kvm_unsync_page(vcpu, s);
1810         }
1811 }
1812
1813 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1814                                   bool can_unsync)
1815 {
1816         struct kvm_mmu_page *s;
1817         struct hlist_node *node;
1818         bool need_unsync = false;
1819
1820         for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1821                 if (s->role.level != PT_PAGE_TABLE_LEVEL)
1822                         return 1;
1823
1824                 if (!need_unsync && !s->unsync) {
1825                         if (!can_unsync || !oos_shadow)
1826                                 return 1;
1827                         need_unsync = true;
1828                 }
1829         }
1830         if (need_unsync)
1831                 kvm_unsync_pages(vcpu, gfn);
1832         return 0;
1833 }
1834
1835 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1836                     unsigned pte_access, int user_fault,
1837                     int write_fault, int dirty, int level,
1838                     gfn_t gfn, pfn_t pfn, bool speculative,
1839                     bool can_unsync, bool reset_host_protection)
1840 {
1841         u64 spte;
1842         int ret = 0;
1843
1844         /*
1845          * We don't set the accessed bit, since we sometimes want to see
1846          * whether the guest actually used the pte (in order to detect
1847          * demand paging).
1848          */
1849         spte = shadow_base_present_pte | shadow_dirty_mask;
1850         if (!speculative)
1851                 spte |= shadow_accessed_mask;
1852         if (!dirty)
1853                 pte_access &= ~ACC_WRITE_MASK;
1854         if (pte_access & ACC_EXEC_MASK)
1855                 spte |= shadow_x_mask;
1856         else
1857                 spte |= shadow_nx_mask;
1858         if (pte_access & ACC_USER_MASK)
1859                 spte |= shadow_user_mask;
1860         if (level > PT_PAGE_TABLE_LEVEL)
1861                 spte |= PT_PAGE_SIZE_MASK;
1862         if (tdp_enabled)
1863                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1864                         kvm_is_mmio_pfn(pfn));
1865
1866         if (reset_host_protection)
1867                 spte |= SPTE_HOST_WRITEABLE;
1868
1869         spte |= (u64)pfn << PAGE_SHIFT;
1870
1871         if ((pte_access & ACC_WRITE_MASK)
1872             || (!tdp_enabled && write_fault && !is_write_protection(vcpu)
1873                 && !user_fault)) {
1874
1875                 if (level > PT_PAGE_TABLE_LEVEL &&
1876                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1877                         ret = 1;
1878                         rmap_remove(vcpu->kvm, sptep);
1879                         spte = shadow_trap_nonpresent_pte;
1880                         goto set_pte;
1881                 }
1882
1883                 spte |= PT_WRITABLE_MASK;
1884
1885                 if (!tdp_enabled && !(pte_access & ACC_WRITE_MASK))
1886                         spte &= ~PT_USER_MASK;
1887
1888                 /*
1889                  * Optimization: for pte sync, if spte was writable the hash
1890                  * lookup is unnecessary (and expensive). Write protection
1891                  * is responsibility of mmu_get_page / kvm_sync_page.
1892                  * Same reasoning can be applied to dirty page accounting.
1893                  */
1894                 if (!can_unsync && is_writable_pte(*sptep))
1895                         goto set_pte;
1896
1897                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1898                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1899                                  __func__, gfn);
1900                         ret = 1;
1901                         pte_access &= ~ACC_WRITE_MASK;
1902                         if (is_writable_pte(spte))
1903                                 spte &= ~PT_WRITABLE_MASK;
1904                 }
1905         }
1906
1907         if (pte_access & ACC_WRITE_MASK)
1908                 mark_page_dirty(vcpu->kvm, gfn);
1909
1910 set_pte:
1911         __set_spte(sptep, spte);
1912         return ret;
1913 }
1914
1915 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1916                          unsigned pt_access, unsigned pte_access,
1917                          int user_fault, int write_fault, int dirty,
1918                          int *ptwrite, int level, gfn_t gfn,
1919                          pfn_t pfn, bool speculative,
1920                          bool reset_host_protection)
1921 {
1922         int was_rmapped = 0;
1923         int was_writable = is_writable_pte(*sptep);
1924         int rmap_count;
1925
1926         pgprintk("%s: spte %llx access %x write_fault %d"
1927                  " user_fault %d gfn %lx\n",
1928                  __func__, *sptep, pt_access,
1929                  write_fault, user_fault, gfn);
1930
1931         if (is_rmap_spte(*sptep)) {
1932                 /*
1933                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1934                  * the parent of the now unreachable PTE.
1935                  */
1936                 if (level > PT_PAGE_TABLE_LEVEL &&
1937                     !is_large_pte(*sptep)) {
1938                         struct kvm_mmu_page *child;
1939                         u64 pte = *sptep;
1940
1941                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1942                         mmu_page_remove_parent_pte(child, sptep);
1943                         __set_spte(sptep, shadow_trap_nonpresent_pte);
1944                         kvm_flush_remote_tlbs(vcpu->kvm);
1945                 } else if (pfn != spte_to_pfn(*sptep)) {
1946                         pgprintk("hfn old %lx new %lx\n",
1947                                  spte_to_pfn(*sptep), pfn);
1948                         rmap_remove(vcpu->kvm, sptep);
1949                         __set_spte(sptep, shadow_trap_nonpresent_pte);
1950                         kvm_flush_remote_tlbs(vcpu->kvm);
1951                 } else
1952                         was_rmapped = 1;
1953         }
1954
1955         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
1956                       dirty, level, gfn, pfn, speculative, true,
1957                       reset_host_protection)) {
1958                 if (write_fault)
1959                         *ptwrite = 1;
1960                 kvm_mmu_flush_tlb(vcpu);
1961         }
1962
1963         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
1964         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1965                  is_large_pte(*sptep)? "2MB" : "4kB",
1966                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
1967                  *sptep, sptep);
1968         if (!was_rmapped && is_large_pte(*sptep))
1969                 ++vcpu->kvm->stat.lpages;
1970
1971         page_header_update_slot(vcpu->kvm, sptep, gfn);
1972         if (!was_rmapped) {
1973                 rmap_count = rmap_add(vcpu, sptep, gfn);
1974                 kvm_release_pfn_clean(pfn);
1975                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
1976                         rmap_recycle(vcpu, sptep, gfn);
1977         } else {
1978                 if (was_writable)
1979                         kvm_release_pfn_dirty(pfn);
1980                 else
1981                         kvm_release_pfn_clean(pfn);
1982         }
1983         if (speculative) {
1984                 vcpu->arch.last_pte_updated = sptep;
1985                 vcpu->arch.last_pte_gfn = gfn;
1986         }
1987 }
1988
1989 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1990 {
1991 }
1992
1993 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1994                         int level, gfn_t gfn, pfn_t pfn)
1995 {
1996         struct kvm_shadow_walk_iterator iterator;
1997         struct kvm_mmu_page *sp;
1998         int pt_write = 0;
1999         gfn_t pseudo_gfn;
2000
2001         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
2002                 if (iterator.level == level) {
2003                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
2004                                      0, write, 1, &pt_write,
2005                                      level, gfn, pfn, false, true);
2006                         ++vcpu->stat.pf_fixed;
2007                         break;
2008                 }
2009
2010                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
2011                         u64 base_addr = iterator.addr;
2012
2013                         base_addr &= PT64_LVL_ADDR_MASK(iterator.level);
2014                         pseudo_gfn = base_addr >> PAGE_SHIFT;
2015                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
2016                                               iterator.level - 1,
2017                                               1, ACC_ALL, iterator.sptep);
2018                         if (!sp) {
2019                                 pgprintk("nonpaging_map: ENOMEM\n");
2020                                 kvm_release_pfn_clean(pfn);
2021                                 return -ENOMEM;
2022                         }
2023
2024                         __set_spte(iterator.sptep,
2025                                    __pa(sp->spt)
2026                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
2027                                    | shadow_user_mask | shadow_x_mask);
2028                 }
2029         }
2030         return pt_write;
2031 }
2032
2033 static void kvm_send_hwpoison_signal(struct kvm *kvm, gfn_t gfn)
2034 {
2035         char buf[1];
2036         void __user *hva;
2037         int r;
2038
2039         /* Touch the page, so send SIGBUS */
2040         hva = (void __user *)gfn_to_hva(kvm, gfn);
2041         r = copy_from_user(buf, hva, 1);
2042 }
2043
2044 static int kvm_handle_bad_page(struct kvm *kvm, gfn_t gfn, pfn_t pfn)
2045 {
2046         kvm_release_pfn_clean(pfn);
2047         if (is_hwpoison_pfn(pfn)) {
2048                 kvm_send_hwpoison_signal(kvm, gfn);
2049                 return 0;
2050         }
2051         return 1;
2052 }
2053
2054 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
2055 {
2056         int r;
2057         int level;
2058         pfn_t pfn;
2059         unsigned long mmu_seq;
2060
2061         level = mapping_level(vcpu, gfn);
2062
2063         /*
2064          * This path builds a PAE pagetable - so we can map 2mb pages at
2065          * maximum. Therefore check if the level is larger than that.
2066          */
2067         if (level > PT_DIRECTORY_LEVEL)
2068                 level = PT_DIRECTORY_LEVEL;
2069
2070         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2071
2072         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2073         smp_rmb();
2074         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2075
2076         /* mmio */
2077         if (is_error_pfn(pfn))
2078                 return kvm_handle_bad_page(vcpu->kvm, gfn, pfn);
2079
2080         spin_lock(&vcpu->kvm->mmu_lock);
2081         if (mmu_notifier_retry(vcpu, mmu_seq))
2082                 goto out_unlock;
2083         kvm_mmu_free_some_pages(vcpu);
2084         r = __direct_map(vcpu, v, write, level, gfn, pfn);
2085         spin_unlock(&vcpu->kvm->mmu_lock);
2086
2087
2088         return r;
2089
2090 out_unlock:
2091         spin_unlock(&vcpu->kvm->mmu_lock);
2092         kvm_release_pfn_clean(pfn);
2093         return 0;
2094 }
2095
2096
2097 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2098 {
2099         int i;
2100         struct kvm_mmu_page *sp;
2101         LIST_HEAD(invalid_list);
2102
2103         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2104                 return;
2105         spin_lock(&vcpu->kvm->mmu_lock);
2106         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2107                 hpa_t root = vcpu->arch.mmu.root_hpa;
2108
2109                 sp = page_header(root);
2110                 --sp->root_count;
2111                 if (!sp->root_count && sp->role.invalid) {
2112                         kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
2113                         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2114                 }
2115                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2116                 spin_unlock(&vcpu->kvm->mmu_lock);
2117                 return;
2118         }
2119         for (i = 0; i < 4; ++i) {
2120                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2121
2122                 if (root) {
2123                         root &= PT64_BASE_ADDR_MASK;
2124                         sp = page_header(root);
2125                         --sp->root_count;
2126                         if (!sp->root_count && sp->role.invalid)
2127                                 kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2128                                                          &invalid_list);
2129                 }
2130                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2131         }
2132         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2133         spin_unlock(&vcpu->kvm->mmu_lock);
2134         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2135 }
2136
2137 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2138 {
2139         int ret = 0;
2140
2141         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2142                 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2143                 ret = 1;
2144         }
2145
2146         return ret;
2147 }
2148
2149 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2150 {
2151         int i;
2152         gfn_t root_gfn;
2153         struct kvm_mmu_page *sp;
2154         int direct = 0;
2155         u64 pdptr;
2156
2157         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2158
2159         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2160                 hpa_t root = vcpu->arch.mmu.root_hpa;
2161
2162                 ASSERT(!VALID_PAGE(root));
2163                 if (mmu_check_root(vcpu, root_gfn))
2164                         return 1;
2165                 if (tdp_enabled) {
2166                         direct = 1;
2167                         root_gfn = 0;
2168                 }
2169                 spin_lock(&vcpu->kvm->mmu_lock);
2170                 kvm_mmu_free_some_pages(vcpu);
2171                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2172                                       PT64_ROOT_LEVEL, direct,
2173                                       ACC_ALL, NULL);
2174                 root = __pa(sp->spt);
2175                 ++sp->root_count;
2176                 spin_unlock(&vcpu->kvm->mmu_lock);
2177                 vcpu->arch.mmu.root_hpa = root;
2178                 return 0;
2179         }
2180         direct = !is_paging(vcpu);
2181         for (i = 0; i < 4; ++i) {
2182                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2183
2184                 ASSERT(!VALID_PAGE(root));
2185                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2186                         pdptr = kvm_pdptr_read(vcpu, i);
2187                         if (!is_present_gpte(pdptr)) {
2188                                 vcpu->arch.mmu.pae_root[i] = 0;
2189                                 continue;
2190                         }
2191                         root_gfn = pdptr >> PAGE_SHIFT;
2192                 } else if (vcpu->arch.mmu.root_level == 0)
2193                         root_gfn = 0;
2194                 if (mmu_check_root(vcpu, root_gfn))
2195                         return 1;
2196                 if (tdp_enabled) {
2197                         direct = 1;
2198                         root_gfn = i << 30;
2199                 }
2200                 spin_lock(&vcpu->kvm->mmu_lock);
2201                 kvm_mmu_free_some_pages(vcpu);
2202                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2203                                       PT32_ROOT_LEVEL, direct,
2204                                       ACC_ALL, NULL);
2205                 root = __pa(sp->spt);
2206                 ++sp->root_count;
2207                 spin_unlock(&vcpu->kvm->mmu_lock);
2208
2209                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2210         }
2211         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2212         return 0;
2213 }
2214
2215 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2216 {
2217         int i;
2218         struct kvm_mmu_page *sp;
2219
2220         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2221                 return;
2222         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2223                 hpa_t root = vcpu->arch.mmu.root_hpa;
2224                 sp = page_header(root);
2225                 mmu_sync_children(vcpu, sp);
2226                 return;
2227         }
2228         for (i = 0; i < 4; ++i) {
2229                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2230
2231                 if (root && VALID_PAGE(root)) {
2232                         root &= PT64_BASE_ADDR_MASK;
2233                         sp = page_header(root);
2234                         mmu_sync_children(vcpu, sp);
2235                 }
2236         }
2237 }
2238
2239 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2240 {
2241         spin_lock(&vcpu->kvm->mmu_lock);
2242         mmu_sync_roots(vcpu);
2243         spin_unlock(&vcpu->kvm->mmu_lock);
2244 }
2245
2246 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2247                                   u32 access, u32 *error)
2248 {
2249         if (error)
2250                 *error = 0;
2251         return vaddr;
2252 }
2253
2254 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2255                                 u32 error_code)
2256 {
2257         gfn_t gfn;
2258         int r;
2259
2260         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2261         r = mmu_topup_memory_caches(vcpu);
2262         if (r)
2263                 return r;
2264
2265         ASSERT(vcpu);
2266         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2267
2268         gfn = gva >> PAGE_SHIFT;
2269
2270         return nonpaging_map(vcpu, gva & PAGE_MASK,
2271                              error_code & PFERR_WRITE_MASK, gfn);
2272 }
2273
2274 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2275                                 u32 error_code)
2276 {
2277         pfn_t pfn;
2278         int r;
2279         int level;
2280         gfn_t gfn = gpa >> PAGE_SHIFT;
2281         unsigned long mmu_seq;
2282
2283         ASSERT(vcpu);
2284         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2285
2286         r = mmu_topup_memory_caches(vcpu);
2287         if (r)
2288                 return r;
2289
2290         level = mapping_level(vcpu, gfn);
2291
2292         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2293
2294         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2295         smp_rmb();
2296         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2297         if (is_error_pfn(pfn))
2298                 return kvm_handle_bad_page(vcpu->kvm, gfn, pfn);
2299         spin_lock(&vcpu->kvm->mmu_lock);
2300         if (mmu_notifier_retry(vcpu, mmu_seq))
2301                 goto out_unlock;
2302         kvm_mmu_free_some_pages(vcpu);
2303         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2304                          level, gfn, pfn);
2305         spin_unlock(&vcpu->kvm->mmu_lock);
2306
2307         return r;
2308
2309 out_unlock:
2310         spin_unlock(&vcpu->kvm->mmu_lock);
2311         kvm_release_pfn_clean(pfn);
2312         return 0;
2313 }
2314
2315 static void nonpaging_free(struct kvm_vcpu *vcpu)
2316 {
2317         mmu_free_roots(vcpu);
2318 }
2319
2320 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2321 {
2322         struct kvm_mmu *context = &vcpu->arch.mmu;
2323
2324         context->new_cr3 = nonpaging_new_cr3;
2325         context->page_fault = nonpaging_page_fault;
2326         context->gva_to_gpa = nonpaging_gva_to_gpa;
2327         context->free = nonpaging_free;
2328         context->prefetch_page = nonpaging_prefetch_page;
2329         context->sync_page = nonpaging_sync_page;
2330         context->invlpg = nonpaging_invlpg;
2331         context->root_level = 0;
2332         context->shadow_root_level = PT32E_ROOT_LEVEL;
2333         context->root_hpa = INVALID_PAGE;
2334         return 0;
2335 }
2336
2337 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2338 {
2339         ++vcpu->stat.tlb_flush;
2340         set_bit(KVM_REQ_TLB_FLUSH, &vcpu->requests);
2341 }
2342
2343 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2344 {
2345         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2346         mmu_free_roots(vcpu);
2347 }
2348
2349 static void inject_page_fault(struct kvm_vcpu *vcpu,
2350                               u64 addr,
2351                               u32 err_code)
2352 {
2353         kvm_inject_page_fault(vcpu, addr, err_code);
2354 }
2355
2356 static void paging_free(struct kvm_vcpu *vcpu)
2357 {
2358         nonpaging_free(vcpu);
2359 }
2360
2361 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2362 {
2363         int bit7;
2364
2365         bit7 = (gpte >> 7) & 1;
2366         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2367 }
2368
2369 #define PTTYPE 64
2370 #include "paging_tmpl.h"
2371 #undef PTTYPE
2372
2373 #define PTTYPE 32
2374 #include "paging_tmpl.h"
2375 #undef PTTYPE
2376
2377 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2378 {
2379         struct kvm_mmu *context = &vcpu->arch.mmu;
2380         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2381         u64 exb_bit_rsvd = 0;
2382
2383         if (!is_nx(vcpu))
2384                 exb_bit_rsvd = rsvd_bits(63, 63);
2385         switch (level) {
2386         case PT32_ROOT_LEVEL:
2387                 /* no rsvd bits for 2 level 4K page table entries */
2388                 context->rsvd_bits_mask[0][1] = 0;
2389                 context->rsvd_bits_mask[0][0] = 0;
2390                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2391
2392                 if (!is_pse(vcpu)) {
2393                         context->rsvd_bits_mask[1][1] = 0;
2394                         break;
2395                 }
2396
2397                 if (is_cpuid_PSE36())
2398                         /* 36bits PSE 4MB page */
2399                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2400                 else
2401                         /* 32 bits PSE 4MB page */
2402                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2403                 break;
2404         case PT32E_ROOT_LEVEL:
2405                 context->rsvd_bits_mask[0][2] =
2406                         rsvd_bits(maxphyaddr, 63) |
2407                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2408                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2409                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2410                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2411                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2412                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2413                         rsvd_bits(maxphyaddr, 62) |
2414                         rsvd_bits(13, 20);              /* large page */
2415                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2416                 break;
2417         case PT64_ROOT_LEVEL:
2418                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2419                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2420                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2421                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2422                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2423                         rsvd_bits(maxphyaddr, 51);
2424                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2425                         rsvd_bits(maxphyaddr, 51);
2426                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2427                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2428                         rsvd_bits(maxphyaddr, 51) |
2429                         rsvd_bits(13, 29);
2430                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2431                         rsvd_bits(maxphyaddr, 51) |
2432                         rsvd_bits(13, 20);              /* large page */
2433                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2434                 break;
2435         }
2436 }
2437
2438 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2439 {
2440         struct kvm_mmu *context = &vcpu->arch.mmu;
2441
2442         ASSERT(is_pae(vcpu));
2443         context->new_cr3 = paging_new_cr3;
2444         context->page_fault = paging64_page_fault;
2445         context->gva_to_gpa = paging64_gva_to_gpa;
2446         context->prefetch_page = paging64_prefetch_page;
2447         context->sync_page = paging64_sync_page;
2448         context->invlpg = paging64_invlpg;
2449         context->free = paging_free;
2450         context->root_level = level;
2451         context->shadow_root_level = level;
2452         context->root_hpa = INVALID_PAGE;
2453         return 0;
2454 }
2455
2456 static int paging64_init_context(struct kvm_vcpu *vcpu)
2457 {
2458         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2459         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2460 }
2461
2462 static int paging32_init_context(struct kvm_vcpu *vcpu)
2463 {
2464         struct kvm_mmu *context = &vcpu->arch.mmu;
2465
2466         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2467         context->new_cr3 = paging_new_cr3;
2468         context->page_fault = paging32_page_fault;
2469         context->gva_to_gpa = paging32_gva_to_gpa;
2470         context->free = paging_free;
2471         context->prefetch_page = paging32_prefetch_page;
2472         context->sync_page = paging32_sync_page;
2473         context->invlpg = paging32_invlpg;
2474         context->root_level = PT32_ROOT_LEVEL;
2475         context->shadow_root_level = PT32E_ROOT_LEVEL;
2476         context->root_hpa = INVALID_PAGE;
2477         return 0;
2478 }
2479
2480 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2481 {
2482         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2483         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2484 }
2485
2486 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2487 {
2488         struct kvm_mmu *context = &vcpu->arch.mmu;
2489
2490         context->new_cr3 = nonpaging_new_cr3;
2491         context->page_fault = tdp_page_fault;
2492         context->free = nonpaging_free;
2493         context->prefetch_page = nonpaging_prefetch_page;
2494         context->sync_page = nonpaging_sync_page;
2495         context->invlpg = nonpaging_invlpg;
2496         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2497         context->root_hpa = INVALID_PAGE;
2498
2499         if (!is_paging(vcpu)) {
2500                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2501                 context->root_level = 0;
2502         } else if (is_long_mode(vcpu)) {
2503                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2504                 context->gva_to_gpa = paging64_gva_to_gpa;
2505                 context->root_level = PT64_ROOT_LEVEL;
2506         } else if (is_pae(vcpu)) {
2507                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2508                 context->gva_to_gpa = paging64_gva_to_gpa;
2509                 context->root_level = PT32E_ROOT_LEVEL;
2510         } else {
2511                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2512                 context->gva_to_gpa = paging32_gva_to_gpa;
2513                 context->root_level = PT32_ROOT_LEVEL;
2514         }
2515
2516         return 0;
2517 }
2518
2519 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2520 {
2521         int r;
2522
2523         ASSERT(vcpu);
2524         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2525
2526         if (!is_paging(vcpu))
2527                 r = nonpaging_init_context(vcpu);
2528         else if (is_long_mode(vcpu))
2529                 r = paging64_init_context(vcpu);
2530         else if (is_pae(vcpu))
2531                 r = paging32E_init_context(vcpu);
2532         else
2533                 r = paging32_init_context(vcpu);
2534
2535         vcpu->arch.mmu.base_role.cr4_pae = !!is_pae(vcpu);
2536         vcpu->arch.mmu.base_role.cr0_wp = is_write_protection(vcpu);
2537
2538         return r;
2539 }
2540
2541 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2542 {
2543         vcpu->arch.update_pte.pfn = bad_pfn;
2544
2545         if (tdp_enabled)
2546                 return init_kvm_tdp_mmu(vcpu);
2547         else
2548                 return init_kvm_softmmu(vcpu);
2549 }
2550
2551 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2552 {
2553         ASSERT(vcpu);
2554         if (VALID_PAGE(vcpu->arch.mmu.root_hpa))
2555                 /* mmu.free() should set root_hpa = INVALID_PAGE */
2556                 vcpu->arch.mmu.free(vcpu);
2557 }
2558
2559 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2560 {
2561         destroy_kvm_mmu(vcpu);
2562         return init_kvm_mmu(vcpu);
2563 }
2564 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2565
2566 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2567 {
2568         int r;
2569
2570         r = mmu_topup_memory_caches(vcpu);
2571         if (r)
2572                 goto out;
2573         r = mmu_alloc_roots(vcpu);
2574         spin_lock(&vcpu->kvm->mmu_lock);
2575         mmu_sync_roots(vcpu);
2576         spin_unlock(&vcpu->kvm->mmu_lock);
2577         if (r)
2578                 goto out;
2579         /* set_cr3() should ensure TLB has been flushed */
2580         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2581 out:
2582         return r;
2583 }
2584 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2585
2586 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2587 {
2588         mmu_free_roots(vcpu);
2589 }
2590
2591 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2592                                   struct kvm_mmu_page *sp,
2593                                   u64 *spte)
2594 {
2595         u64 pte;
2596         struct kvm_mmu_page *child;
2597
2598         pte = *spte;
2599         if (is_shadow_present_pte(pte)) {
2600                 if (is_last_spte(pte, sp->role.level))
2601                         rmap_remove(vcpu->kvm, spte);
2602                 else {
2603                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2604                         mmu_page_remove_parent_pte(child, spte);
2605                 }
2606         }
2607         __set_spte(spte, shadow_trap_nonpresent_pte);
2608         if (is_large_pte(pte))
2609                 --vcpu->kvm->stat.lpages;
2610 }
2611
2612 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2613                                   struct kvm_mmu_page *sp,
2614                                   u64 *spte,
2615                                   const void *new)
2616 {
2617         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2618                 ++vcpu->kvm->stat.mmu_pde_zapped;
2619                 return;
2620         }
2621
2622         ++vcpu->kvm->stat.mmu_pte_updated;
2623         if (!sp->role.cr4_pae)
2624                 paging32_update_pte(vcpu, sp, spte, new);
2625         else
2626                 paging64_update_pte(vcpu, sp, spte, new);
2627 }
2628
2629 static bool need_remote_flush(u64 old, u64 new)
2630 {
2631         if (!is_shadow_present_pte(old))
2632                 return false;
2633         if (!is_shadow_present_pte(new))
2634                 return true;
2635         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2636                 return true;
2637         old ^= PT64_NX_MASK;
2638         new ^= PT64_NX_MASK;
2639         return (old & ~new & PT64_PERM_MASK) != 0;
2640 }
2641
2642 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, bool zap_page,
2643                                     bool remote_flush, bool local_flush)
2644 {
2645         if (zap_page)
2646                 return;
2647
2648         if (remote_flush)
2649                 kvm_flush_remote_tlbs(vcpu->kvm);
2650         else if (local_flush)
2651                 kvm_mmu_flush_tlb(vcpu);
2652 }
2653
2654 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2655 {
2656         u64 *spte = vcpu->arch.last_pte_updated;
2657
2658         return !!(spte && (*spte & shadow_accessed_mask));
2659 }
2660
2661 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2662                                           u64 gpte)
2663 {
2664         gfn_t gfn;
2665         pfn_t pfn;
2666
2667         if (!is_present_gpte(gpte))
2668                 return;
2669         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2670
2671         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2672         smp_rmb();
2673         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2674
2675         if (is_error_pfn(pfn)) {
2676                 kvm_release_pfn_clean(pfn);
2677                 return;
2678         }
2679         vcpu->arch.update_pte.gfn = gfn;
2680         vcpu->arch.update_pte.pfn = pfn;
2681 }
2682
2683 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2684 {
2685         u64 *spte = vcpu->arch.last_pte_updated;
2686
2687         if (spte
2688             && vcpu->arch.last_pte_gfn == gfn
2689             && shadow_accessed_mask
2690             && !(*spte & shadow_accessed_mask)
2691             && is_shadow_present_pte(*spte))
2692                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2693 }
2694
2695 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2696                        const u8 *new, int bytes,
2697                        bool guest_initiated)
2698 {
2699         gfn_t gfn = gpa >> PAGE_SHIFT;
2700         struct kvm_mmu_page *sp;
2701         struct hlist_node *node;
2702         LIST_HEAD(invalid_list);
2703         u64 entry, gentry;
2704         u64 *spte;
2705         unsigned offset = offset_in_page(gpa);
2706         unsigned pte_size;
2707         unsigned page_offset;
2708         unsigned misaligned;
2709         unsigned quadrant;
2710         int level;
2711         int flooded = 0;
2712         int npte;
2713         int r;
2714         int invlpg_counter;
2715         bool remote_flush, local_flush, zap_page;
2716
2717         zap_page = remote_flush = local_flush = false;
2718
2719         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2720
2721         invlpg_counter = atomic_read(&vcpu->kvm->arch.invlpg_counter);
2722
2723         /*
2724          * Assume that the pte write on a page table of the same type
2725          * as the current vcpu paging mode.  This is nearly always true
2726          * (might be false while changing modes).  Note it is verified later
2727          * by update_pte().
2728          */
2729         if ((is_pae(vcpu) && bytes == 4) || !new) {
2730                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2731                 if (is_pae(vcpu)) {
2732                         gpa &= ~(gpa_t)7;
2733                         bytes = 8;
2734                 }
2735                 r = kvm_read_guest(vcpu->kvm, gpa, &gentry, min(bytes, 8));
2736                 if (r)
2737                         gentry = 0;
2738                 new = (const u8 *)&gentry;
2739         }
2740
2741         switch (bytes) {
2742         case 4:
2743                 gentry = *(const u32 *)new;
2744                 break;
2745         case 8:
2746                 gentry = *(const u64 *)new;
2747                 break;
2748         default:
2749                 gentry = 0;
2750                 break;
2751         }
2752
2753         mmu_guess_page_from_pte_write(vcpu, gpa, gentry);
2754         spin_lock(&vcpu->kvm->mmu_lock);
2755         if (atomic_read(&vcpu->kvm->arch.invlpg_counter) != invlpg_counter)
2756                 gentry = 0;
2757         kvm_mmu_access_page(vcpu, gfn);
2758         kvm_mmu_free_some_pages(vcpu);
2759         ++vcpu->kvm->stat.mmu_pte_write;
2760         kvm_mmu_audit(vcpu, "pre pte write");
2761         if (guest_initiated) {
2762                 if (gfn == vcpu->arch.last_pt_write_gfn
2763                     && !last_updated_pte_accessed(vcpu)) {
2764                         ++vcpu->arch.last_pt_write_count;
2765                         if (vcpu->arch.last_pt_write_count >= 3)
2766                                 flooded = 1;
2767                 } else {
2768                         vcpu->arch.last_pt_write_gfn = gfn;
2769                         vcpu->arch.last_pt_write_count = 1;
2770                         vcpu->arch.last_pte_updated = NULL;
2771                 }
2772         }
2773
2774         for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn, node) {
2775                 pte_size = sp->role.cr4_pae ? 8 : 4;
2776                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2777                 misaligned |= bytes < 4;
2778                 if (misaligned || flooded) {
2779                         /*
2780                          * Misaligned accesses are too much trouble to fix
2781                          * up; also, they usually indicate a page is not used
2782                          * as a page table.
2783                          *
2784                          * If we're seeing too many writes to a page,
2785                          * it may no longer be a page table, or we may be
2786                          * forking, in which case it is better to unmap the
2787                          * page.
2788                          */
2789                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2790                                  gpa, bytes, sp->role.word);
2791                         zap_page |= !!kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2792                                                      &invalid_list);
2793                         ++vcpu->kvm->stat.mmu_flooded;
2794                         continue;
2795                 }
2796                 page_offset = offset;
2797                 level = sp->role.level;
2798                 npte = 1;
2799                 if (!sp->role.cr4_pae) {
2800                         page_offset <<= 1;      /* 32->64 */
2801                         /*
2802                          * A 32-bit pde maps 4MB while the shadow pdes map
2803                          * only 2MB.  So we need to double the offset again
2804                          * and zap two pdes instead of one.
2805                          */
2806                         if (level == PT32_ROOT_LEVEL) {
2807                                 page_offset &= ~7; /* kill rounding error */
2808                                 page_offset <<= 1;
2809                                 npte = 2;
2810                         }
2811                         quadrant = page_offset >> PAGE_SHIFT;
2812                         page_offset &= ~PAGE_MASK;
2813                         if (quadrant != sp->role.quadrant)
2814                                 continue;
2815                 }
2816                 local_flush = true;
2817                 spte = &sp->spt[page_offset / sizeof(*spte)];
2818                 while (npte--) {
2819                         entry = *spte;
2820                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2821                         if (gentry)
2822                                 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
2823                         if (!remote_flush && need_remote_flush(entry, *spte))
2824                                 remote_flush = true;
2825                         ++spte;
2826                 }
2827         }
2828         mmu_pte_write_flush_tlb(vcpu, zap_page, remote_flush, local_flush);
2829         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2830         kvm_mmu_audit(vcpu, "post pte write");
2831         spin_unlock(&vcpu->kvm->mmu_lock);
2832         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2833                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2834                 vcpu->arch.update_pte.pfn = bad_pfn;
2835         }
2836 }
2837
2838 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2839 {
2840         gpa_t gpa;
2841         int r;
2842
2843         if (tdp_enabled)
2844                 return 0;
2845
2846         gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2847
2848         spin_lock(&vcpu->kvm->mmu_lock);
2849         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2850         spin_unlock(&vcpu->kvm->mmu_lock);
2851         return r;
2852 }
2853 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2854
2855 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2856 {
2857         int free_pages;
2858         LIST_HEAD(invalid_list);
2859
2860         free_pages = vcpu->kvm->arch.n_free_mmu_pages;
2861         while (free_pages < KVM_REFILL_PAGES &&
2862                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2863                 struct kvm_mmu_page *sp;
2864
2865                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2866                                   struct kvm_mmu_page, link);
2867                 free_pages += kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2868                                                        &invalid_list);
2869                 ++vcpu->kvm->stat.mmu_recycled;
2870         }
2871         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2872 }
2873
2874 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2875 {
2876         int r;
2877         enum emulation_result er;
2878
2879         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2880         if (r < 0)
2881                 goto out;
2882
2883         if (!r) {
2884                 r = 1;
2885                 goto out;
2886         }
2887
2888         r = mmu_topup_memory_caches(vcpu);
2889         if (r)
2890                 goto out;
2891
2892         er = emulate_instruction(vcpu, cr2, error_code, 0);
2893
2894         switch (er) {
2895         case EMULATE_DONE:
2896                 return 1;
2897         case EMULATE_DO_MMIO:
2898                 ++vcpu->stat.mmio_exits;
2899                 /* fall through */
2900         case EMULATE_FAIL:
2901                 return 0;
2902         default:
2903                 BUG();
2904         }
2905 out:
2906         return r;
2907 }
2908 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2909
2910 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2911 {
2912         vcpu->arch.mmu.invlpg(vcpu, gva);
2913         kvm_mmu_flush_tlb(vcpu);
2914         ++vcpu->stat.invlpg;
2915 }
2916 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2917
2918 void kvm_enable_tdp(void)
2919 {
2920         tdp_enabled = true;
2921 }
2922 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2923
2924 void kvm_disable_tdp(void)
2925 {
2926         tdp_enabled = false;
2927 }
2928 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2929
2930 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2931 {
2932         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2933 }
2934
2935 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2936 {
2937         struct page *page;
2938         int i;
2939
2940         ASSERT(vcpu);
2941
2942         /*
2943          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2944          * Therefore we need to allocate shadow page tables in the first
2945          * 4GB of memory, which happens to fit the DMA32 zone.
2946          */
2947         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2948         if (!page)
2949                 return -ENOMEM;
2950
2951         vcpu->arch.mmu.pae_root = page_address(page);
2952         for (i = 0; i < 4; ++i)
2953                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2954
2955         return 0;
2956 }
2957
2958 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2959 {
2960         ASSERT(vcpu);
2961         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2962
2963         return alloc_mmu_pages(vcpu);
2964 }
2965
2966 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2967 {
2968         ASSERT(vcpu);
2969         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2970
2971         return init_kvm_mmu(vcpu);
2972 }
2973
2974 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2975 {
2976         ASSERT(vcpu);
2977
2978         destroy_kvm_mmu(vcpu);
2979         free_mmu_pages(vcpu);
2980         mmu_free_memory_caches(vcpu);
2981 }
2982
2983 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2984 {
2985         struct kvm_mmu_page *sp;
2986
2987         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2988                 int i;
2989                 u64 *pt;
2990
2991                 if (!test_bit(slot, sp->slot_bitmap))
2992                         continue;
2993
2994                 pt = sp->spt;
2995                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2996                         /* avoid RMW */
2997                         if (is_writable_pte(pt[i]))
2998                                 pt[i] &= ~PT_WRITABLE_MASK;
2999         }
3000         kvm_flush_remote_tlbs(kvm);
3001 }
3002
3003 void kvm_mmu_zap_all(struct kvm *kvm)
3004 {
3005         struct kvm_mmu_page *sp, *node;
3006         LIST_HEAD(invalid_list);
3007
3008         spin_lock(&kvm->mmu_lock);
3009 restart:
3010         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
3011                 if (kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list))
3012                         goto restart;
3013
3014         kvm_mmu_commit_zap_page(kvm, &invalid_list);
3015         spin_unlock(&kvm->mmu_lock);
3016 }
3017
3018 static int kvm_mmu_remove_some_alloc_mmu_pages(struct kvm *kvm,
3019                                                struct list_head *invalid_list)
3020 {
3021         struct kvm_mmu_page *page;
3022
3023         page = container_of(kvm->arch.active_mmu_pages.prev,
3024                             struct kvm_mmu_page, link);
3025         return kvm_mmu_prepare_zap_page(kvm, page, invalid_list);
3026 }
3027
3028 static int mmu_shrink(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
3029 {
3030         struct kvm *kvm;
3031         struct kvm *kvm_freed = NULL;
3032         int cache_count = 0;
3033
3034         spin_lock(&kvm_lock);
3035
3036         list_for_each_entry(kvm, &vm_list, vm_list) {
3037                 int npages, idx, freed_pages;
3038                 LIST_HEAD(invalid_list);
3039
3040                 idx = srcu_read_lock(&kvm->srcu);
3041                 spin_lock(&kvm->mmu_lock);
3042                 npages = kvm->arch.n_alloc_mmu_pages -
3043                          kvm->arch.n_free_mmu_pages;
3044                 cache_count += npages;
3045                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
3046                         freed_pages = kvm_mmu_remove_some_alloc_mmu_pages(kvm,
3047                                                           &invalid_list);
3048                         cache_count -= freed_pages;
3049                         kvm_freed = kvm;
3050                 }
3051                 nr_to_scan--;
3052
3053                 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3054                 spin_unlock(&kvm->mmu_lock);
3055                 srcu_read_unlock(&kvm->srcu, idx);
3056         }
3057         if (kvm_freed)
3058                 list_move_tail(&kvm_freed->vm_list, &vm_list);
3059
3060         spin_unlock(&kvm_lock);
3061
3062         return cache_count;
3063 }
3064
3065 static struct shrinker mmu_shrinker = {
3066         .shrink = mmu_shrink,
3067         .seeks = DEFAULT_SEEKS * 10,
3068 };
3069
3070 static void mmu_destroy_caches(void)
3071 {
3072         if (pte_chain_cache)
3073                 kmem_cache_destroy(pte_chain_cache);
3074         if (rmap_desc_cache)
3075                 kmem_cache_destroy(rmap_desc_cache);
3076         if (mmu_page_header_cache)
3077                 kmem_cache_destroy(mmu_page_header_cache);
3078 }
3079
3080 void kvm_mmu_module_exit(void)
3081 {
3082         mmu_destroy_caches();
3083         unregister_shrinker(&mmu_shrinker);
3084 }
3085
3086 int kvm_mmu_module_init(void)
3087 {
3088         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
3089                                             sizeof(struct kvm_pte_chain),
3090                                             0, 0, NULL);
3091         if (!pte_chain_cache)
3092                 goto nomem;
3093         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
3094                                             sizeof(struct kvm_rmap_desc),
3095                                             0, 0, NULL);
3096         if (!rmap_desc_cache)
3097                 goto nomem;
3098
3099         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
3100                                                   sizeof(struct kvm_mmu_page),
3101                                                   0, 0, NULL);
3102         if (!mmu_page_header_cache)
3103                 goto nomem;
3104
3105         register_shrinker(&mmu_shrinker);
3106
3107         return 0;
3108
3109 nomem:
3110         mmu_destroy_caches();
3111         return -ENOMEM;
3112 }
3113
3114 /*
3115  * Caculate mmu pages needed for kvm.
3116  */
3117 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3118 {
3119         int i;
3120         unsigned int nr_mmu_pages;
3121         unsigned int  nr_pages = 0;
3122         struct kvm_memslots *slots;
3123
3124         slots = kvm_memslots(kvm);
3125
3126         for (i = 0; i < slots->nmemslots; i++)
3127                 nr_pages += slots->memslots[i].npages;
3128
3129         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3130         nr_mmu_pages = max(nr_mmu_pages,
3131                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3132
3133         return nr_mmu_pages;
3134 }
3135
3136 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3137                                 unsigned len)
3138 {
3139         if (len > buffer->len)
3140                 return NULL;
3141         return buffer->ptr;
3142 }
3143
3144 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3145                                 unsigned len)
3146 {
3147         void *ret;
3148
3149         ret = pv_mmu_peek_buffer(buffer, len);
3150         if (!ret)
3151                 return ret;
3152         buffer->ptr += len;
3153         buffer->len -= len;
3154         buffer->processed += len;
3155         return ret;
3156 }
3157
3158 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3159                              gpa_t addr, gpa_t value)
3160 {
3161         int bytes = 8;
3162         int r;
3163
3164         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3165                 bytes = 4;
3166
3167         r = mmu_topup_memory_caches(vcpu);
3168         if (r)
3169                 return r;
3170
3171         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3172                 return -EFAULT;
3173
3174         return 1;
3175 }
3176
3177 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3178 {
3179         (void)kvm_set_cr3(vcpu, vcpu->arch.cr3);
3180         return 1;
3181 }
3182
3183 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3184 {
3185         spin_lock(&vcpu->kvm->mmu_lock);
3186         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3187         spin_unlock(&vcpu->kvm->mmu_lock);
3188         return 1;
3189 }
3190
3191 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3192                              struct kvm_pv_mmu_op_buffer *buffer)
3193 {
3194         struct kvm_mmu_op_header *header;
3195
3196         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3197         if (!header)
3198                 return 0;
3199         switch (header->op) {
3200         case KVM_MMU_OP_WRITE_PTE: {
3201                 struct kvm_mmu_op_write_pte *wpte;
3202
3203                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3204                 if (!wpte)
3205                         return 0;
3206                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3207                                         wpte->pte_val);
3208         }
3209         case KVM_MMU_OP_FLUSH_TLB: {
3210                 struct kvm_mmu_op_flush_tlb *ftlb;
3211
3212                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3213                 if (!ftlb)
3214                         return 0;
3215                 return kvm_pv_mmu_flush_tlb(vcpu);
3216         }
3217         case KVM_MMU_OP_RELEASE_PT: {
3218                 struct kvm_mmu_op_release_pt *rpt;
3219
3220                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3221                 if (!rpt)
3222                         return 0;
3223                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3224         }
3225         default: return 0;
3226         }
3227 }
3228
3229 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3230                   gpa_t addr, unsigned long *ret)
3231 {
3232         int r;
3233         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3234
3235         buffer->ptr = buffer->buf;
3236         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3237         buffer->processed = 0;
3238
3239         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3240         if (r)
3241                 goto out;
3242
3243         while (buffer->len) {
3244                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3245                 if (r < 0)
3246                         goto out;
3247                 if (r == 0)
3248                         break;
3249         }
3250
3251         r = 1;
3252 out:
3253         *ret = buffer->processed;
3254         return r;
3255 }
3256
3257 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu *vcpu, u64 addr, u64 sptes[4])
3258 {
3259         struct kvm_shadow_walk_iterator iterator;
3260         int nr_sptes = 0;
3261
3262         spin_lock(&vcpu->kvm->mmu_lock);
3263         for_each_shadow_entry(vcpu, addr, iterator) {
3264                 sptes[iterator.level-1] = *iterator.sptep;
3265                 nr_sptes++;
3266                 if (!is_shadow_present_pte(*iterator.sptep))
3267                         break;
3268         }
3269         spin_unlock(&vcpu->kvm->mmu_lock);
3270
3271         return nr_sptes;
3272 }
3273 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
3274
3275 #ifdef AUDIT
3276
3277 static const char *audit_msg;
3278
3279 static gva_t canonicalize(gva_t gva)
3280 {
3281 #ifdef CONFIG_X86_64
3282         gva = (long long)(gva << 16) >> 16;
3283 #endif
3284         return gva;
3285 }
3286
3287
3288 typedef void (*inspect_spte_fn) (struct kvm *kvm, u64 *sptep);
3289
3290 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
3291                             inspect_spte_fn fn)
3292 {
3293         int i;
3294
3295         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3296                 u64 ent = sp->spt[i];
3297
3298                 if (is_shadow_present_pte(ent)) {
3299                         if (!is_last_spte(ent, sp->role.level)) {
3300                                 struct kvm_mmu_page *child;
3301                                 child = page_header(ent & PT64_BASE_ADDR_MASK);
3302                                 __mmu_spte_walk(kvm, child, fn);
3303                         } else
3304                                 fn(kvm, &sp->spt[i]);
3305                 }
3306         }
3307 }
3308
3309 static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
3310 {
3311         int i;
3312         struct kvm_mmu_page *sp;
3313
3314         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
3315                 return;
3316         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
3317                 hpa_t root = vcpu->arch.mmu.root_hpa;
3318                 sp = page_header(root);
3319                 __mmu_spte_walk(vcpu->kvm, sp, fn);
3320                 return;
3321         }
3322         for (i = 0; i < 4; ++i) {
3323                 hpa_t root = vcpu->arch.mmu.pae_root[i];
3324
3325                 if (root && VALID_PAGE(root)) {
3326                         root &= PT64_BASE_ADDR_MASK;
3327                         sp = page_header(root);
3328                         __mmu_spte_walk(vcpu->kvm, sp, fn);
3329                 }
3330         }
3331         return;
3332 }
3333
3334 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3335                                 gva_t va, int level)
3336 {
3337         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3338         int i;
3339         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3340
3341         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3342                 u64 ent = pt[i];
3343
3344                 if (ent == shadow_trap_nonpresent_pte)
3345                         continue;
3346
3347                 va = canonicalize(va);
3348                 if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
3349                         audit_mappings_page(vcpu, ent, va, level - 1);
3350                 else {
3351                         gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, va, NULL);
3352                         gfn_t gfn = gpa >> PAGE_SHIFT;
3353                         pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
3354                         hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
3355
3356                         if (is_error_pfn(pfn)) {
3357                                 kvm_release_pfn_clean(pfn);
3358                                 continue;
3359                         }
3360
3361                         if (is_shadow_present_pte(ent)
3362                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3363                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3364                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3365                                        audit_msg, vcpu->arch.mmu.root_level,
3366                                        va, gpa, hpa, ent,
3367                                        is_shadow_present_pte(ent));
3368                         else if (ent == shadow_notrap_nonpresent_pte
3369                                  && !is_error_hpa(hpa))
3370                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3371                                        " valid guest gva %lx\n", audit_msg, va);
3372                         kvm_release_pfn_clean(pfn);
3373
3374                 }
3375         }
3376 }
3377
3378 static void audit_mappings(struct kvm_vcpu *vcpu)
3379 {
3380         unsigned i;
3381
3382         if (vcpu->arch.mmu.root_level == 4)
3383                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3384         else
3385                 for (i = 0; i < 4; ++i)
3386                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3387                                 audit_mappings_page(vcpu,
3388                                                     vcpu->arch.mmu.pae_root[i],
3389                                                     i << 30,
3390                                                     2);
3391 }
3392
3393 static int count_rmaps(struct kvm_vcpu *vcpu)
3394 {
3395         struct kvm *kvm = vcpu->kvm;
3396         struct kvm_memslots *slots;
3397         int nmaps = 0;
3398         int i, j, k, idx;
3399
3400         idx = srcu_read_lock(&kvm->srcu);
3401         slots = kvm_memslots(kvm);
3402         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3403                 struct kvm_memory_slot *m = &slots->memslots[i];
3404                 struct kvm_rmap_desc *d;
3405
3406                 for (j = 0; j < m->npages; ++j) {
3407                         unsigned long *rmapp = &m->rmap[j];
3408
3409                         if (!*rmapp)
3410                                 continue;
3411                         if (!(*rmapp & 1)) {
3412                                 ++nmaps;
3413                                 continue;
3414                         }
3415                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3416                         while (d) {
3417                                 for (k = 0; k < RMAP_EXT; ++k)
3418                                         if (d->sptes[k])
3419                                                 ++nmaps;
3420                                         else
3421                                                 break;
3422                                 d = d->more;
3423                         }
3424                 }
3425         }
3426         srcu_read_unlock(&kvm->srcu, idx);
3427         return nmaps;
3428 }
3429
3430 void inspect_spte_has_rmap(struct kvm *kvm, u64 *sptep)
3431 {
3432         unsigned long *rmapp;
3433         struct kvm_mmu_page *rev_sp;
3434         gfn_t gfn;
3435
3436         if (is_writable_pte(*sptep)) {
3437                 rev_sp = page_header(__pa(sptep));
3438                 gfn = kvm_mmu_page_get_gfn(rev_sp, sptep - rev_sp->spt);
3439
3440                 if (!gfn_to_memslot(kvm, gfn)) {
3441                         if (!printk_ratelimit())
3442                                 return;
3443                         printk(KERN_ERR "%s: no memslot for gfn %ld\n",
3444                                          audit_msg, gfn);
3445                         printk(KERN_ERR "%s: index %ld of sp (gfn=%lx)\n",
3446                                audit_msg, (long int)(sptep - rev_sp->spt),
3447                                         rev_sp->gfn);
3448                         dump_stack();
3449                         return;
3450                 }
3451
3452                 rmapp = gfn_to_rmap(kvm, gfn, rev_sp->role.level);
3453                 if (!*rmapp) {
3454                         if (!printk_ratelimit())
3455                                 return;
3456                         printk(KERN_ERR "%s: no rmap for writable spte %llx\n",
3457                                          audit_msg, *sptep);
3458                         dump_stack();
3459                 }
3460         }
3461
3462 }
3463
3464 void audit_writable_sptes_have_rmaps(struct kvm_vcpu *vcpu)
3465 {
3466         mmu_spte_walk(vcpu, inspect_spte_has_rmap);
3467 }
3468
3469 static void check_writable_mappings_rmap(struct kvm_vcpu *vcpu)
3470 {
3471         struct kvm_mmu_page *sp;
3472         int i;
3473
3474         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3475                 u64 *pt = sp->spt;
3476
3477                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3478                         continue;
3479
3480                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3481                         u64 ent = pt[i];
3482
3483                         if (!(ent & PT_PRESENT_MASK))
3484                                 continue;
3485                         if (!is_writable_pte(ent))
3486                                 continue;
3487                         inspect_spte_has_rmap(vcpu->kvm, &pt[i]);
3488                 }
3489         }
3490         return;
3491 }
3492
3493 static void audit_rmap(struct kvm_vcpu *vcpu)
3494 {
3495         check_writable_mappings_rmap(vcpu);
3496         count_rmaps(vcpu);
3497 }
3498
3499 static void audit_write_protection(struct kvm_vcpu *vcpu)
3500 {
3501         struct kvm_mmu_page *sp;
3502         struct kvm_memory_slot *slot;
3503         unsigned long *rmapp;
3504         u64 *spte;
3505         gfn_t gfn;
3506
3507         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3508                 if (sp->role.direct)
3509                         continue;
3510                 if (sp->unsync)
3511                         continue;
3512
3513                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3514                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3515                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3516
3517                 spte = rmap_next(vcpu->kvm, rmapp, NULL);
3518                 while (spte) {
3519                         if (is_writable_pte(*spte))
3520                                 printk(KERN_ERR "%s: (%s) shadow page has "
3521                                 "writable mappings: gfn %lx role %x\n",
3522                                __func__, audit_msg, sp->gfn,
3523                                sp->role.word);
3524                         spte = rmap_next(vcpu->kvm, rmapp, spte);
3525                 }
3526         }
3527 }
3528
3529 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3530 {
3531         int olddbg = dbg;
3532
3533         dbg = 0;
3534         audit_msg = msg;
3535         audit_rmap(vcpu);
3536         audit_write_protection(vcpu);
3537         if (strcmp("pre pte write", audit_msg) != 0)
3538                 audit_mappings(vcpu);
3539         audit_writable_sptes_have_rmaps(vcpu);
3540         dbg = olddbg;
3541 }
3542
3543 #endif