x86, kvm: Remove cast obsoleted by set_64bit() prototype cleanup
[sfrench/cifs-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 PT64_LEVEL_BITS 9
96
97 #define PT64_LEVEL_SHIFT(level) \
98                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
99
100 #define PT64_LEVEL_MASK(level) \
101                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
102
103 #define PT64_INDEX(address, level)\
104         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
105
106
107 #define PT32_LEVEL_BITS 10
108
109 #define PT32_LEVEL_SHIFT(level) \
110                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
111
112 #define PT32_LEVEL_MASK(level) \
113                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
114 #define PT32_LVL_OFFSET_MASK(level) \
115         (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
116                                                 * PT32_LEVEL_BITS))) - 1))
117
118 #define PT32_INDEX(address, level)\
119         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
120
121
122 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
123 #define PT64_DIR_BASE_ADDR_MASK \
124         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
125 #define PT64_LVL_ADDR_MASK(level) \
126         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
127                                                 * PT64_LEVEL_BITS))) - 1))
128 #define PT64_LVL_OFFSET_MASK(level) \
129         (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
130                                                 * PT64_LEVEL_BITS))) - 1))
131
132 #define PT32_BASE_ADDR_MASK PAGE_MASK
133 #define PT32_DIR_BASE_ADDR_MASK \
134         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
135 #define PT32_LVL_ADDR_MASK(level) \
136         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
137                                             * PT32_LEVEL_BITS))) - 1))
138
139 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
140                         | PT64_NX_MASK)
141
142 #define RMAP_EXT 4
143
144 #define ACC_EXEC_MASK    1
145 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
146 #define ACC_USER_MASK    PT_USER_MASK
147 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
148
149 #include <trace/events/kvm.h>
150
151 #define CREATE_TRACE_POINTS
152 #include "mmutrace.h"
153
154 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
155
156 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
157
158 struct kvm_rmap_desc {
159         u64 *sptes[RMAP_EXT];
160         struct kvm_rmap_desc *more;
161 };
162
163 struct kvm_shadow_walk_iterator {
164         u64 addr;
165         hpa_t shadow_addr;
166         int level;
167         u64 *sptep;
168         unsigned index;
169 };
170
171 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
172         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
173              shadow_walk_okay(&(_walker));                      \
174              shadow_walk_next(&(_walker)))
175
176 typedef void (*mmu_parent_walk_fn) (struct kvm_mmu_page *sp, u64 *spte);
177
178 static struct kmem_cache *pte_chain_cache;
179 static struct kmem_cache *rmap_desc_cache;
180 static struct kmem_cache *mmu_page_header_cache;
181
182 static u64 __read_mostly shadow_trap_nonpresent_pte;
183 static u64 __read_mostly shadow_notrap_nonpresent_pte;
184 static u64 __read_mostly shadow_base_present_pte;
185 static u64 __read_mostly shadow_nx_mask;
186 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
187 static u64 __read_mostly shadow_user_mask;
188 static u64 __read_mostly shadow_accessed_mask;
189 static u64 __read_mostly shadow_dirty_mask;
190
191 static inline u64 rsvd_bits(int s, int e)
192 {
193         return ((1ULL << (e - s + 1)) - 1) << s;
194 }
195
196 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
197 {
198         shadow_trap_nonpresent_pte = trap_pte;
199         shadow_notrap_nonpresent_pte = notrap_pte;
200 }
201 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
202
203 void kvm_mmu_set_base_ptes(u64 base_pte)
204 {
205         shadow_base_present_pte = base_pte;
206 }
207 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
208
209 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
210                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
211 {
212         shadow_user_mask = user_mask;
213         shadow_accessed_mask = accessed_mask;
214         shadow_dirty_mask = dirty_mask;
215         shadow_nx_mask = nx_mask;
216         shadow_x_mask = x_mask;
217 }
218 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
219
220 static bool is_write_protection(struct kvm_vcpu *vcpu)
221 {
222         return kvm_read_cr0_bits(vcpu, X86_CR0_WP);
223 }
224
225 static int is_cpuid_PSE36(void)
226 {
227         return 1;
228 }
229
230 static int is_nx(struct kvm_vcpu *vcpu)
231 {
232         return vcpu->arch.efer & EFER_NX;
233 }
234
235 static int is_shadow_present_pte(u64 pte)
236 {
237         return pte != shadow_trap_nonpresent_pte
238                 && pte != shadow_notrap_nonpresent_pte;
239 }
240
241 static int is_large_pte(u64 pte)
242 {
243         return pte & PT_PAGE_SIZE_MASK;
244 }
245
246 static int is_writable_pte(unsigned long pte)
247 {
248         return pte & PT_WRITABLE_MASK;
249 }
250
251 static int is_dirty_gpte(unsigned long pte)
252 {
253         return pte & PT_DIRTY_MASK;
254 }
255
256 static int is_rmap_spte(u64 pte)
257 {
258         return is_shadow_present_pte(pte);
259 }
260
261 static int is_last_spte(u64 pte, int level)
262 {
263         if (level == PT_PAGE_TABLE_LEVEL)
264                 return 1;
265         if (is_large_pte(pte))
266                 return 1;
267         return 0;
268 }
269
270 static pfn_t spte_to_pfn(u64 pte)
271 {
272         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
273 }
274
275 static gfn_t pse36_gfn_delta(u32 gpte)
276 {
277         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
278
279         return (gpte & PT32_DIR_PSE36_MASK) << shift;
280 }
281
282 static void __set_spte(u64 *sptep, u64 spte)
283 {
284         set_64bit(sptep, spte);
285 }
286
287 static u64 __xchg_spte(u64 *sptep, u64 new_spte)
288 {
289 #ifdef CONFIG_X86_64
290         return xchg(sptep, new_spte);
291 #else
292         u64 old_spte;
293
294         do {
295                 old_spte = *sptep;
296         } while (cmpxchg64(sptep, old_spte, new_spte) != old_spte);
297
298         return old_spte;
299 #endif
300 }
301
302 static void update_spte(u64 *sptep, u64 new_spte)
303 {
304         u64 old_spte;
305
306         if (!shadow_accessed_mask || (new_spte & shadow_accessed_mask) ||
307               !is_rmap_spte(*sptep))
308                 __set_spte(sptep, new_spte);
309         else {
310                 old_spte = __xchg_spte(sptep, new_spte);
311                 if (old_spte & shadow_accessed_mask)
312                         mark_page_accessed(pfn_to_page(spte_to_pfn(old_spte)));
313         }
314 }
315
316 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
317                                   struct kmem_cache *base_cache, int min)
318 {
319         void *obj;
320
321         if (cache->nobjs >= min)
322                 return 0;
323         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
324                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
325                 if (!obj)
326                         return -ENOMEM;
327                 cache->objects[cache->nobjs++] = obj;
328         }
329         return 0;
330 }
331
332 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc,
333                                   struct kmem_cache *cache)
334 {
335         while (mc->nobjs)
336                 kmem_cache_free(cache, mc->objects[--mc->nobjs]);
337 }
338
339 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
340                                        int min)
341 {
342         struct page *page;
343
344         if (cache->nobjs >= min)
345                 return 0;
346         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
347                 page = alloc_page(GFP_KERNEL);
348                 if (!page)
349                         return -ENOMEM;
350                 cache->objects[cache->nobjs++] = page_address(page);
351         }
352         return 0;
353 }
354
355 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
356 {
357         while (mc->nobjs)
358                 free_page((unsigned long)mc->objects[--mc->nobjs]);
359 }
360
361 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
362 {
363         int r;
364
365         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
366                                    pte_chain_cache, 4);
367         if (r)
368                 goto out;
369         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
370                                    rmap_desc_cache, 4);
371         if (r)
372                 goto out;
373         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
374         if (r)
375                 goto out;
376         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
377                                    mmu_page_header_cache, 4);
378 out:
379         return r;
380 }
381
382 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
383 {
384         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache, pte_chain_cache);
385         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache, rmap_desc_cache);
386         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
387         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache,
388                                 mmu_page_header_cache);
389 }
390
391 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
392                                     size_t size)
393 {
394         void *p;
395
396         BUG_ON(!mc->nobjs);
397         p = mc->objects[--mc->nobjs];
398         return p;
399 }
400
401 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
402 {
403         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
404                                       sizeof(struct kvm_pte_chain));
405 }
406
407 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
408 {
409         kmem_cache_free(pte_chain_cache, pc);
410 }
411
412 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
413 {
414         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
415                                       sizeof(struct kvm_rmap_desc));
416 }
417
418 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
419 {
420         kmem_cache_free(rmap_desc_cache, rd);
421 }
422
423 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
424 {
425         if (!sp->role.direct)
426                 return sp->gfns[index];
427
428         return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
429 }
430
431 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
432 {
433         if (sp->role.direct)
434                 BUG_ON(gfn != kvm_mmu_page_get_gfn(sp, index));
435         else
436                 sp->gfns[index] = gfn;
437 }
438
439 /*
440  * Return the pointer to the largepage write count for a given
441  * gfn, handling slots that are not large page aligned.
442  */
443 static int *slot_largepage_idx(gfn_t gfn,
444                                struct kvm_memory_slot *slot,
445                                int level)
446 {
447         unsigned long idx;
448
449         idx = (gfn >> KVM_HPAGE_GFN_SHIFT(level)) -
450               (slot->base_gfn >> KVM_HPAGE_GFN_SHIFT(level));
451         return &slot->lpage_info[level - 2][idx].write_count;
452 }
453
454 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
455 {
456         struct kvm_memory_slot *slot;
457         int *write_count;
458         int i;
459
460         slot = gfn_to_memslot(kvm, gfn);
461         for (i = PT_DIRECTORY_LEVEL;
462              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
463                 write_count   = slot_largepage_idx(gfn, slot, i);
464                 *write_count += 1;
465         }
466 }
467
468 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
469 {
470         struct kvm_memory_slot *slot;
471         int *write_count;
472         int i;
473
474         slot = gfn_to_memslot(kvm, gfn);
475         for (i = PT_DIRECTORY_LEVEL;
476              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
477                 write_count   = slot_largepage_idx(gfn, slot, i);
478                 *write_count -= 1;
479                 WARN_ON(*write_count < 0);
480         }
481 }
482
483 static int has_wrprotected_page(struct kvm *kvm,
484                                 gfn_t gfn,
485                                 int level)
486 {
487         struct kvm_memory_slot *slot;
488         int *largepage_idx;
489
490         slot = gfn_to_memslot(kvm, gfn);
491         if (slot) {
492                 largepage_idx = slot_largepage_idx(gfn, slot, level);
493                 return *largepage_idx;
494         }
495
496         return 1;
497 }
498
499 static int host_mapping_level(struct kvm *kvm, gfn_t gfn)
500 {
501         unsigned long page_size;
502         int i, ret = 0;
503
504         page_size = kvm_host_page_size(kvm, gfn);
505
506         for (i = PT_PAGE_TABLE_LEVEL;
507              i < (PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES); ++i) {
508                 if (page_size >= KVM_HPAGE_SIZE(i))
509                         ret = i;
510                 else
511                         break;
512         }
513
514         return ret;
515 }
516
517 static int mapping_level(struct kvm_vcpu *vcpu, gfn_t large_gfn)
518 {
519         struct kvm_memory_slot *slot;
520         int host_level, level, max_level;
521
522         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
523         if (slot && slot->dirty_bitmap)
524                 return PT_PAGE_TABLE_LEVEL;
525
526         host_level = host_mapping_level(vcpu->kvm, large_gfn);
527
528         if (host_level == PT_PAGE_TABLE_LEVEL)
529                 return host_level;
530
531         max_level = kvm_x86_ops->get_lpage_level() < host_level ?
532                 kvm_x86_ops->get_lpage_level() : host_level;
533
534         for (level = PT_DIRECTORY_LEVEL; level <= max_level; ++level)
535                 if (has_wrprotected_page(vcpu->kvm, large_gfn, level))
536                         break;
537
538         return level - 1;
539 }
540
541 /*
542  * Take gfn and return the reverse mapping to it.
543  */
544
545 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int level)
546 {
547         struct kvm_memory_slot *slot;
548         unsigned long idx;
549
550         slot = gfn_to_memslot(kvm, gfn);
551         if (likely(level == PT_PAGE_TABLE_LEVEL))
552                 return &slot->rmap[gfn - slot->base_gfn];
553
554         idx = (gfn >> KVM_HPAGE_GFN_SHIFT(level)) -
555                 (slot->base_gfn >> KVM_HPAGE_GFN_SHIFT(level));
556
557         return &slot->lpage_info[level - 2][idx].rmap_pde;
558 }
559
560 /*
561  * Reverse mapping data structures:
562  *
563  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
564  * that points to page_address(page).
565  *
566  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
567  * containing more mappings.
568  *
569  * Returns the number of rmap entries before the spte was added or zero if
570  * the spte was not added.
571  *
572  */
573 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
574 {
575         struct kvm_mmu_page *sp;
576         struct kvm_rmap_desc *desc;
577         unsigned long *rmapp;
578         int i, count = 0;
579
580         if (!is_rmap_spte(*spte))
581                 return count;
582         sp = page_header(__pa(spte));
583         kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
584         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
585         if (!*rmapp) {
586                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
587                 *rmapp = (unsigned long)spte;
588         } else if (!(*rmapp & 1)) {
589                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
590                 desc = mmu_alloc_rmap_desc(vcpu);
591                 desc->sptes[0] = (u64 *)*rmapp;
592                 desc->sptes[1] = spte;
593                 *rmapp = (unsigned long)desc | 1;
594         } else {
595                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
596                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
597                 while (desc->sptes[RMAP_EXT-1] && desc->more) {
598                         desc = desc->more;
599                         count += RMAP_EXT;
600                 }
601                 if (desc->sptes[RMAP_EXT-1]) {
602                         desc->more = mmu_alloc_rmap_desc(vcpu);
603                         desc = desc->more;
604                 }
605                 for (i = 0; desc->sptes[i]; ++i)
606                         ;
607                 desc->sptes[i] = spte;
608         }
609         return count;
610 }
611
612 static void rmap_desc_remove_entry(unsigned long *rmapp,
613                                    struct kvm_rmap_desc *desc,
614                                    int i,
615                                    struct kvm_rmap_desc *prev_desc)
616 {
617         int j;
618
619         for (j = RMAP_EXT - 1; !desc->sptes[j] && j > i; --j)
620                 ;
621         desc->sptes[i] = desc->sptes[j];
622         desc->sptes[j] = NULL;
623         if (j != 0)
624                 return;
625         if (!prev_desc && !desc->more)
626                 *rmapp = (unsigned long)desc->sptes[0];
627         else
628                 if (prev_desc)
629                         prev_desc->more = desc->more;
630                 else
631                         *rmapp = (unsigned long)desc->more | 1;
632         mmu_free_rmap_desc(desc);
633 }
634
635 static void rmap_remove(struct kvm *kvm, u64 *spte)
636 {
637         struct kvm_rmap_desc *desc;
638         struct kvm_rmap_desc *prev_desc;
639         struct kvm_mmu_page *sp;
640         gfn_t gfn;
641         unsigned long *rmapp;
642         int i;
643
644         sp = page_header(__pa(spte));
645         gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
646         rmapp = gfn_to_rmap(kvm, gfn, sp->role.level);
647         if (!*rmapp) {
648                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
649                 BUG();
650         } else if (!(*rmapp & 1)) {
651                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
652                 if ((u64 *)*rmapp != spte) {
653                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
654                                spte, *spte);
655                         BUG();
656                 }
657                 *rmapp = 0;
658         } else {
659                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
660                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
661                 prev_desc = NULL;
662                 while (desc) {
663                         for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i)
664                                 if (desc->sptes[i] == spte) {
665                                         rmap_desc_remove_entry(rmapp,
666                                                                desc, i,
667                                                                prev_desc);
668                                         return;
669                                 }
670                         prev_desc = desc;
671                         desc = desc->more;
672                 }
673                 pr_err("rmap_remove: %p %llx many->many\n", spte, *spte);
674                 BUG();
675         }
676 }
677
678 static void set_spte_track_bits(u64 *sptep, u64 new_spte)
679 {
680         pfn_t pfn;
681         u64 old_spte = *sptep;
682
683         if (!shadow_accessed_mask || !is_shadow_present_pte(old_spte) ||
684               old_spte & shadow_accessed_mask) {
685                 __set_spte(sptep, new_spte);
686         } else
687                 old_spte = __xchg_spte(sptep, new_spte);
688
689         if (!is_rmap_spte(old_spte))
690                 return;
691         pfn = spte_to_pfn(old_spte);
692         if (!shadow_accessed_mask || old_spte & shadow_accessed_mask)
693                 kvm_set_pfn_accessed(pfn);
694         if (is_writable_pte(old_spte))
695                 kvm_set_pfn_dirty(pfn);
696 }
697
698 static void drop_spte(struct kvm *kvm, u64 *sptep, u64 new_spte)
699 {
700         set_spte_track_bits(sptep, new_spte);
701         rmap_remove(kvm, sptep);
702 }
703
704 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
705 {
706         struct kvm_rmap_desc *desc;
707         u64 *prev_spte;
708         int i;
709
710         if (!*rmapp)
711                 return NULL;
712         else if (!(*rmapp & 1)) {
713                 if (!spte)
714                         return (u64 *)*rmapp;
715                 return NULL;
716         }
717         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
718         prev_spte = NULL;
719         while (desc) {
720                 for (i = 0; i < RMAP_EXT && desc->sptes[i]; ++i) {
721                         if (prev_spte == spte)
722                                 return desc->sptes[i];
723                         prev_spte = desc->sptes[i];
724                 }
725                 desc = desc->more;
726         }
727         return NULL;
728 }
729
730 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
731 {
732         unsigned long *rmapp;
733         u64 *spte;
734         int i, write_protected = 0;
735
736         rmapp = gfn_to_rmap(kvm, gfn, PT_PAGE_TABLE_LEVEL);
737
738         spte = rmap_next(kvm, rmapp, NULL);
739         while (spte) {
740                 BUG_ON(!spte);
741                 BUG_ON(!(*spte & PT_PRESENT_MASK));
742                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
743                 if (is_writable_pte(*spte)) {
744                         update_spte(spte, *spte & ~PT_WRITABLE_MASK);
745                         write_protected = 1;
746                 }
747                 spte = rmap_next(kvm, rmapp, spte);
748         }
749         if (write_protected) {
750                 pfn_t pfn;
751
752                 spte = rmap_next(kvm, rmapp, NULL);
753                 pfn = spte_to_pfn(*spte);
754                 kvm_set_pfn_dirty(pfn);
755         }
756
757         /* check for huge page mappings */
758         for (i = PT_DIRECTORY_LEVEL;
759              i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
760                 rmapp = gfn_to_rmap(kvm, gfn, i);
761                 spte = rmap_next(kvm, rmapp, NULL);
762                 while (spte) {
763                         BUG_ON(!spte);
764                         BUG_ON(!(*spte & PT_PRESENT_MASK));
765                         BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
766                         pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
767                         if (is_writable_pte(*spte)) {
768                                 drop_spte(kvm, spte,
769                                           shadow_trap_nonpresent_pte);
770                                 --kvm->stat.lpages;
771                                 spte = NULL;
772                                 write_protected = 1;
773                         }
774                         spte = rmap_next(kvm, rmapp, spte);
775                 }
776         }
777
778         return write_protected;
779 }
780
781 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
782                            unsigned long data)
783 {
784         u64 *spte;
785         int need_tlb_flush = 0;
786
787         while ((spte = rmap_next(kvm, rmapp, NULL))) {
788                 BUG_ON(!(*spte & PT_PRESENT_MASK));
789                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
790                 drop_spte(kvm, spte, shadow_trap_nonpresent_pte);
791                 need_tlb_flush = 1;
792         }
793         return need_tlb_flush;
794 }
795
796 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
797                              unsigned long data)
798 {
799         int need_flush = 0;
800         u64 *spte, new_spte;
801         pte_t *ptep = (pte_t *)data;
802         pfn_t new_pfn;
803
804         WARN_ON(pte_huge(*ptep));
805         new_pfn = pte_pfn(*ptep);
806         spte = rmap_next(kvm, rmapp, NULL);
807         while (spte) {
808                 BUG_ON(!is_shadow_present_pte(*spte));
809                 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
810                 need_flush = 1;
811                 if (pte_write(*ptep)) {
812                         drop_spte(kvm, spte, shadow_trap_nonpresent_pte);
813                         spte = rmap_next(kvm, rmapp, NULL);
814                 } else {
815                         new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
816                         new_spte |= (u64)new_pfn << PAGE_SHIFT;
817
818                         new_spte &= ~PT_WRITABLE_MASK;
819                         new_spte &= ~SPTE_HOST_WRITEABLE;
820                         new_spte &= ~shadow_accessed_mask;
821                         set_spte_track_bits(spte, new_spte);
822                         spte = rmap_next(kvm, rmapp, spte);
823                 }
824         }
825         if (need_flush)
826                 kvm_flush_remote_tlbs(kvm);
827
828         return 0;
829 }
830
831 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
832                           unsigned long data,
833                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
834                                          unsigned long data))
835 {
836         int i, j;
837         int ret;
838         int retval = 0;
839         struct kvm_memslots *slots;
840
841         slots = kvm_memslots(kvm);
842
843         for (i = 0; i < slots->nmemslots; i++) {
844                 struct kvm_memory_slot *memslot = &slots->memslots[i];
845                 unsigned long start = memslot->userspace_addr;
846                 unsigned long end;
847
848                 end = start + (memslot->npages << PAGE_SHIFT);
849                 if (hva >= start && hva < end) {
850                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
851
852                         ret = handler(kvm, &memslot->rmap[gfn_offset], data);
853
854                         for (j = 0; j < KVM_NR_PAGE_SIZES - 1; ++j) {
855                                 unsigned long idx;
856                                 int sh;
857
858                                 sh = KVM_HPAGE_GFN_SHIFT(PT_DIRECTORY_LEVEL+j);
859                                 idx = ((memslot->base_gfn+gfn_offset) >> sh) -
860                                         (memslot->base_gfn >> sh);
861                                 ret |= handler(kvm,
862                                         &memslot->lpage_info[j][idx].rmap_pde,
863                                         data);
864                         }
865                         trace_kvm_age_page(hva, memslot, ret);
866                         retval |= ret;
867                 }
868         }
869
870         return retval;
871 }
872
873 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
874 {
875         return kvm_handle_hva(kvm, hva, 0, kvm_unmap_rmapp);
876 }
877
878 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
879 {
880         kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
881 }
882
883 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
884                          unsigned long data)
885 {
886         u64 *spte;
887         int young = 0;
888
889         /*
890          * Emulate the accessed bit for EPT, by checking if this page has
891          * an EPT mapping, and clearing it if it does. On the next access,
892          * a new EPT mapping will be established.
893          * This has some overhead, but not as much as the cost of swapping
894          * out actively used pages or breaking up actively used hugepages.
895          */
896         if (!shadow_accessed_mask)
897                 return kvm_unmap_rmapp(kvm, rmapp, data);
898
899         spte = rmap_next(kvm, rmapp, NULL);
900         while (spte) {
901                 int _young;
902                 u64 _spte = *spte;
903                 BUG_ON(!(_spte & PT_PRESENT_MASK));
904                 _young = _spte & PT_ACCESSED_MASK;
905                 if (_young) {
906                         young = 1;
907                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
908                 }
909                 spte = rmap_next(kvm, rmapp, spte);
910         }
911         return young;
912 }
913
914 #define RMAP_RECYCLE_THRESHOLD 1000
915
916 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
917 {
918         unsigned long *rmapp;
919         struct kvm_mmu_page *sp;
920
921         sp = page_header(__pa(spte));
922
923         rmapp = gfn_to_rmap(vcpu->kvm, gfn, sp->role.level);
924
925         kvm_unmap_rmapp(vcpu->kvm, rmapp, 0);
926         kvm_flush_remote_tlbs(vcpu->kvm);
927 }
928
929 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
930 {
931         return kvm_handle_hva(kvm, hva, 0, kvm_age_rmapp);
932 }
933
934 #ifdef MMU_DEBUG
935 static int is_empty_shadow_page(u64 *spt)
936 {
937         u64 *pos;
938         u64 *end;
939
940         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
941                 if (is_shadow_present_pte(*pos)) {
942                         printk(KERN_ERR "%s: %p %llx\n", __func__,
943                                pos, *pos);
944                         return 0;
945                 }
946         return 1;
947 }
948 #endif
949
950 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
951 {
952         ASSERT(is_empty_shadow_page(sp->spt));
953         hlist_del(&sp->hash_link);
954         list_del(&sp->link);
955         __free_page(virt_to_page(sp->spt));
956         if (!sp->role.direct)
957                 __free_page(virt_to_page(sp->gfns));
958         kmem_cache_free(mmu_page_header_cache, sp);
959         ++kvm->arch.n_free_mmu_pages;
960 }
961
962 static unsigned kvm_page_table_hashfn(gfn_t gfn)
963 {
964         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
965 }
966
967 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
968                                                u64 *parent_pte, int direct)
969 {
970         struct kvm_mmu_page *sp;
971
972         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
973         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
974         if (!direct)
975                 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache,
976                                                   PAGE_SIZE);
977         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
978         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
979         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
980         sp->multimapped = 0;
981         sp->parent_pte = parent_pte;
982         --vcpu->kvm->arch.n_free_mmu_pages;
983         return sp;
984 }
985
986 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
987                                     struct kvm_mmu_page *sp, u64 *parent_pte)
988 {
989         struct kvm_pte_chain *pte_chain;
990         struct hlist_node *node;
991         int i;
992
993         if (!parent_pte)
994                 return;
995         if (!sp->multimapped) {
996                 u64 *old = sp->parent_pte;
997
998                 if (!old) {
999                         sp->parent_pte = parent_pte;
1000                         return;
1001                 }
1002                 sp->multimapped = 1;
1003                 pte_chain = mmu_alloc_pte_chain(vcpu);
1004                 INIT_HLIST_HEAD(&sp->parent_ptes);
1005                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
1006                 pte_chain->parent_ptes[0] = old;
1007         }
1008         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
1009                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
1010                         continue;
1011                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
1012                         if (!pte_chain->parent_ptes[i]) {
1013                                 pte_chain->parent_ptes[i] = parent_pte;
1014                                 return;
1015                         }
1016         }
1017         pte_chain = mmu_alloc_pte_chain(vcpu);
1018         BUG_ON(!pte_chain);
1019         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
1020         pte_chain->parent_ptes[0] = parent_pte;
1021 }
1022
1023 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
1024                                        u64 *parent_pte)
1025 {
1026         struct kvm_pte_chain *pte_chain;
1027         struct hlist_node *node;
1028         int i;
1029
1030         if (!sp->multimapped) {
1031                 BUG_ON(sp->parent_pte != parent_pte);
1032                 sp->parent_pte = NULL;
1033                 return;
1034         }
1035         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1036                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1037                         if (!pte_chain->parent_ptes[i])
1038                                 break;
1039                         if (pte_chain->parent_ptes[i] != parent_pte)
1040                                 continue;
1041                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
1042                                 && pte_chain->parent_ptes[i + 1]) {
1043                                 pte_chain->parent_ptes[i]
1044                                         = pte_chain->parent_ptes[i + 1];
1045                                 ++i;
1046                         }
1047                         pte_chain->parent_ptes[i] = NULL;
1048                         if (i == 0) {
1049                                 hlist_del(&pte_chain->link);
1050                                 mmu_free_pte_chain(pte_chain);
1051                                 if (hlist_empty(&sp->parent_ptes)) {
1052                                         sp->multimapped = 0;
1053                                         sp->parent_pte = NULL;
1054                                 }
1055                         }
1056                         return;
1057                 }
1058         BUG();
1059 }
1060
1061 static void mmu_parent_walk(struct kvm_mmu_page *sp, mmu_parent_walk_fn fn)
1062 {
1063         struct kvm_pte_chain *pte_chain;
1064         struct hlist_node *node;
1065         struct kvm_mmu_page *parent_sp;
1066         int i;
1067
1068         if (!sp->multimapped && sp->parent_pte) {
1069                 parent_sp = page_header(__pa(sp->parent_pte));
1070                 fn(parent_sp, sp->parent_pte);
1071                 return;
1072         }
1073
1074         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
1075                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
1076                         u64 *spte = pte_chain->parent_ptes[i];
1077
1078                         if (!spte)
1079                                 break;
1080                         parent_sp = page_header(__pa(spte));
1081                         fn(parent_sp, spte);
1082                 }
1083 }
1084
1085 static void mark_unsync(struct kvm_mmu_page *sp, u64 *spte);
1086 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1087 {
1088         mmu_parent_walk(sp, mark_unsync);
1089 }
1090
1091 static void mark_unsync(struct kvm_mmu_page *sp, u64 *spte)
1092 {
1093         unsigned int index;
1094
1095         index = spte - sp->spt;
1096         if (__test_and_set_bit(index, sp->unsync_child_bitmap))
1097                 return;
1098         if (sp->unsync_children++)
1099                 return;
1100         kvm_mmu_mark_parents_unsync(sp);
1101 }
1102
1103 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1104                                     struct kvm_mmu_page *sp)
1105 {
1106         int i;
1107
1108         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1109                 sp->spt[i] = shadow_trap_nonpresent_pte;
1110 }
1111
1112 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1113                                struct kvm_mmu_page *sp, bool clear_unsync)
1114 {
1115         return 1;
1116 }
1117
1118 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
1119 {
1120 }
1121
1122 #define KVM_PAGE_ARRAY_NR 16
1123
1124 struct kvm_mmu_pages {
1125         struct mmu_page_and_offset {
1126                 struct kvm_mmu_page *sp;
1127                 unsigned int idx;
1128         } page[KVM_PAGE_ARRAY_NR];
1129         unsigned int nr;
1130 };
1131
1132 #define for_each_unsync_children(bitmap, idx)           \
1133         for (idx = find_first_bit(bitmap, 512);         \
1134              idx < 512;                                 \
1135              idx = find_next_bit(bitmap, 512, idx+1))
1136
1137 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1138                          int idx)
1139 {
1140         int i;
1141
1142         if (sp->unsync)
1143                 for (i=0; i < pvec->nr; i++)
1144                         if (pvec->page[i].sp == sp)
1145                                 return 0;
1146
1147         pvec->page[pvec->nr].sp = sp;
1148         pvec->page[pvec->nr].idx = idx;
1149         pvec->nr++;
1150         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1151 }
1152
1153 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1154                            struct kvm_mmu_pages *pvec)
1155 {
1156         int i, ret, nr_unsync_leaf = 0;
1157
1158         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1159                 struct kvm_mmu_page *child;
1160                 u64 ent = sp->spt[i];
1161
1162                 if (!is_shadow_present_pte(ent) || is_large_pte(ent))
1163                         goto clear_child_bitmap;
1164
1165                 child = page_header(ent & PT64_BASE_ADDR_MASK);
1166
1167                 if (child->unsync_children) {
1168                         if (mmu_pages_add(pvec, child, i))
1169                                 return -ENOSPC;
1170
1171                         ret = __mmu_unsync_walk(child, pvec);
1172                         if (!ret)
1173                                 goto clear_child_bitmap;
1174                         else if (ret > 0)
1175                                 nr_unsync_leaf += ret;
1176                         else
1177                                 return ret;
1178                 } else if (child->unsync) {
1179                         nr_unsync_leaf++;
1180                         if (mmu_pages_add(pvec, child, i))
1181                                 return -ENOSPC;
1182                 } else
1183                          goto clear_child_bitmap;
1184
1185                 continue;
1186
1187 clear_child_bitmap:
1188                 __clear_bit(i, sp->unsync_child_bitmap);
1189                 sp->unsync_children--;
1190                 WARN_ON((int)sp->unsync_children < 0);
1191         }
1192
1193
1194         return nr_unsync_leaf;
1195 }
1196
1197 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1198                            struct kvm_mmu_pages *pvec)
1199 {
1200         if (!sp->unsync_children)
1201                 return 0;
1202
1203         mmu_pages_add(pvec, sp, 0);
1204         return __mmu_unsync_walk(sp, pvec);
1205 }
1206
1207 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1208 {
1209         WARN_ON(!sp->unsync);
1210         trace_kvm_mmu_sync_page(sp);
1211         sp->unsync = 0;
1212         --kvm->stat.mmu_unsync;
1213 }
1214
1215 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1216                                     struct list_head *invalid_list);
1217 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1218                                     struct list_head *invalid_list);
1219
1220 #define for_each_gfn_sp(kvm, sp, gfn, pos)                              \
1221   hlist_for_each_entry(sp, pos,                                         \
1222    &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link)   \
1223         if ((sp)->gfn != (gfn)) {} else
1224
1225 #define for_each_gfn_indirect_valid_sp(kvm, sp, gfn, pos)               \
1226   hlist_for_each_entry(sp, pos,                                         \
1227    &(kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)], hash_link)   \
1228                 if ((sp)->gfn != (gfn) || (sp)->role.direct ||          \
1229                         (sp)->role.invalid) {} else
1230
1231 /* @sp->gfn should be write-protected at the call site */
1232 static int __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1233                            struct list_head *invalid_list, bool clear_unsync)
1234 {
1235         if (sp->role.cr4_pae != !!is_pae(vcpu)) {
1236                 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1237                 return 1;
1238         }
1239
1240         if (clear_unsync)
1241                 kvm_unlink_unsync_page(vcpu->kvm, sp);
1242
1243         if (vcpu->arch.mmu.sync_page(vcpu, sp, clear_unsync)) {
1244                 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1245                 return 1;
1246         }
1247
1248         kvm_mmu_flush_tlb(vcpu);
1249         return 0;
1250 }
1251
1252 static int kvm_sync_page_transient(struct kvm_vcpu *vcpu,
1253                                    struct kvm_mmu_page *sp)
1254 {
1255         LIST_HEAD(invalid_list);
1256         int ret;
1257
1258         ret = __kvm_sync_page(vcpu, sp, &invalid_list, false);
1259         if (ret)
1260                 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1261
1262         return ret;
1263 }
1264
1265 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1266                          struct list_head *invalid_list)
1267 {
1268         return __kvm_sync_page(vcpu, sp, invalid_list, true);
1269 }
1270
1271 /* @gfn should be write-protected at the call site */
1272 static void kvm_sync_pages(struct kvm_vcpu *vcpu,  gfn_t gfn)
1273 {
1274         struct kvm_mmu_page *s;
1275         struct hlist_node *node;
1276         LIST_HEAD(invalid_list);
1277         bool flush = false;
1278
1279         for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1280                 if (!s->unsync)
1281                         continue;
1282
1283                 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1284                 if ((s->role.cr4_pae != !!is_pae(vcpu)) ||
1285                         (vcpu->arch.mmu.sync_page(vcpu, s, true))) {
1286                         kvm_mmu_prepare_zap_page(vcpu->kvm, s, &invalid_list);
1287                         continue;
1288                 }
1289                 kvm_unlink_unsync_page(vcpu->kvm, s);
1290                 flush = true;
1291         }
1292
1293         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1294         if (flush)
1295                 kvm_mmu_flush_tlb(vcpu);
1296 }
1297
1298 struct mmu_page_path {
1299         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1300         unsigned int idx[PT64_ROOT_LEVEL-1];
1301 };
1302
1303 #define for_each_sp(pvec, sp, parents, i)                       \
1304                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1305                         sp = pvec.page[i].sp;                   \
1306                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1307                         i = mmu_pages_next(&pvec, &parents, i))
1308
1309 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1310                           struct mmu_page_path *parents,
1311                           int i)
1312 {
1313         int n;
1314
1315         for (n = i+1; n < pvec->nr; n++) {
1316                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1317
1318                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1319                         parents->idx[0] = pvec->page[n].idx;
1320                         return n;
1321                 }
1322
1323                 parents->parent[sp->role.level-2] = sp;
1324                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1325         }
1326
1327         return n;
1328 }
1329
1330 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1331 {
1332         struct kvm_mmu_page *sp;
1333         unsigned int level = 0;
1334
1335         do {
1336                 unsigned int idx = parents->idx[level];
1337
1338                 sp = parents->parent[level];
1339                 if (!sp)
1340                         return;
1341
1342                 --sp->unsync_children;
1343                 WARN_ON((int)sp->unsync_children < 0);
1344                 __clear_bit(idx, sp->unsync_child_bitmap);
1345                 level++;
1346         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1347 }
1348
1349 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1350                                struct mmu_page_path *parents,
1351                                struct kvm_mmu_pages *pvec)
1352 {
1353         parents->parent[parent->role.level-1] = NULL;
1354         pvec->nr = 0;
1355 }
1356
1357 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1358                               struct kvm_mmu_page *parent)
1359 {
1360         int i;
1361         struct kvm_mmu_page *sp;
1362         struct mmu_page_path parents;
1363         struct kvm_mmu_pages pages;
1364         LIST_HEAD(invalid_list);
1365
1366         kvm_mmu_pages_init(parent, &parents, &pages);
1367         while (mmu_unsync_walk(parent, &pages)) {
1368                 int protected = 0;
1369
1370                 for_each_sp(pages, sp, parents, i)
1371                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1372
1373                 if (protected)
1374                         kvm_flush_remote_tlbs(vcpu->kvm);
1375
1376                 for_each_sp(pages, sp, parents, i) {
1377                         kvm_sync_page(vcpu, sp, &invalid_list);
1378                         mmu_pages_clear_parents(&parents);
1379                 }
1380                 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
1381                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1382                 kvm_mmu_pages_init(parent, &parents, &pages);
1383         }
1384 }
1385
1386 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1387                                              gfn_t gfn,
1388                                              gva_t gaddr,
1389                                              unsigned level,
1390                                              int direct,
1391                                              unsigned access,
1392                                              u64 *parent_pte)
1393 {
1394         union kvm_mmu_page_role role;
1395         unsigned quadrant;
1396         struct kvm_mmu_page *sp;
1397         struct hlist_node *node;
1398         bool need_sync = false;
1399
1400         role = vcpu->arch.mmu.base_role;
1401         role.level = level;
1402         role.direct = direct;
1403         if (role.direct)
1404                 role.cr4_pae = 0;
1405         role.access = access;
1406         if (!tdp_enabled && vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1407                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1408                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1409                 role.quadrant = quadrant;
1410         }
1411         for_each_gfn_sp(vcpu->kvm, sp, gfn, node) {
1412                 if (!need_sync && sp->unsync)
1413                         need_sync = true;
1414
1415                 if (sp->role.word != role.word)
1416                         continue;
1417
1418                 if (sp->unsync && kvm_sync_page_transient(vcpu, sp))
1419                         break;
1420
1421                 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1422                 if (sp->unsync_children) {
1423                         kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
1424                         kvm_mmu_mark_parents_unsync(sp);
1425                 } else if (sp->unsync)
1426                         kvm_mmu_mark_parents_unsync(sp);
1427
1428                 trace_kvm_mmu_get_page(sp, false);
1429                 return sp;
1430         }
1431         ++vcpu->kvm->stat.mmu_cache_miss;
1432         sp = kvm_mmu_alloc_page(vcpu, parent_pte, direct);
1433         if (!sp)
1434                 return sp;
1435         sp->gfn = gfn;
1436         sp->role = role;
1437         hlist_add_head(&sp->hash_link,
1438                 &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
1439         if (!direct) {
1440                 if (rmap_write_protect(vcpu->kvm, gfn))
1441                         kvm_flush_remote_tlbs(vcpu->kvm);
1442                 if (level > PT_PAGE_TABLE_LEVEL && need_sync)
1443                         kvm_sync_pages(vcpu, gfn);
1444
1445                 account_shadowed(vcpu->kvm, gfn);
1446         }
1447         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1448                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1449         else
1450                 nonpaging_prefetch_page(vcpu, sp);
1451         trace_kvm_mmu_get_page(sp, true);
1452         return sp;
1453 }
1454
1455 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1456                              struct kvm_vcpu *vcpu, u64 addr)
1457 {
1458         iterator->addr = addr;
1459         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1460         iterator->level = vcpu->arch.mmu.shadow_root_level;
1461         if (iterator->level == PT32E_ROOT_LEVEL) {
1462                 iterator->shadow_addr
1463                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1464                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1465                 --iterator->level;
1466                 if (!iterator->shadow_addr)
1467                         iterator->level = 0;
1468         }
1469 }
1470
1471 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1472 {
1473         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1474                 return false;
1475
1476         if (iterator->level == PT_PAGE_TABLE_LEVEL)
1477                 if (is_large_pte(*iterator->sptep))
1478                         return false;
1479
1480         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1481         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1482         return true;
1483 }
1484
1485 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1486 {
1487         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1488         --iterator->level;
1489 }
1490
1491 static void link_shadow_page(u64 *sptep, struct kvm_mmu_page *sp)
1492 {
1493         u64 spte;
1494
1495         spte = __pa(sp->spt)
1496                 | PT_PRESENT_MASK | PT_ACCESSED_MASK
1497                 | PT_WRITABLE_MASK | PT_USER_MASK;
1498         __set_spte(sptep, spte);
1499 }
1500
1501 static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1502 {
1503         if (is_large_pte(*sptep)) {
1504                 drop_spte(vcpu->kvm, sptep, shadow_trap_nonpresent_pte);
1505                 kvm_flush_remote_tlbs(vcpu->kvm);
1506         }
1507 }
1508
1509 static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1510                                    unsigned direct_access)
1511 {
1512         if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
1513                 struct kvm_mmu_page *child;
1514
1515                 /*
1516                  * For the direct sp, if the guest pte's dirty bit
1517                  * changed form clean to dirty, it will corrupt the
1518                  * sp's access: allow writable in the read-only sp,
1519                  * so we should update the spte at this point to get
1520                  * a new sp with the correct access.
1521                  */
1522                 child = page_header(*sptep & PT64_BASE_ADDR_MASK);
1523                 if (child->role.access == direct_access)
1524                         return;
1525
1526                 mmu_page_remove_parent_pte(child, sptep);
1527                 __set_spte(sptep, shadow_trap_nonpresent_pte);
1528                 kvm_flush_remote_tlbs(vcpu->kvm);
1529         }
1530 }
1531
1532 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1533                                          struct kvm_mmu_page *sp)
1534 {
1535         unsigned i;
1536         u64 *pt;
1537         u64 ent;
1538
1539         pt = sp->spt;
1540
1541         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1542                 ent = pt[i];
1543
1544                 if (is_shadow_present_pte(ent)) {
1545                         if (!is_last_spte(ent, sp->role.level)) {
1546                                 ent &= PT64_BASE_ADDR_MASK;
1547                                 mmu_page_remove_parent_pte(page_header(ent),
1548                                                            &pt[i]);
1549                         } else {
1550                                 if (is_large_pte(ent))
1551                                         --kvm->stat.lpages;
1552                                 drop_spte(kvm, &pt[i],
1553                                           shadow_trap_nonpresent_pte);
1554                         }
1555                 }
1556                 pt[i] = shadow_trap_nonpresent_pte;
1557         }
1558 }
1559
1560 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1561 {
1562         mmu_page_remove_parent_pte(sp, parent_pte);
1563 }
1564
1565 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1566 {
1567         int i;
1568         struct kvm_vcpu *vcpu;
1569
1570         kvm_for_each_vcpu(i, vcpu, kvm)
1571                 vcpu->arch.last_pte_updated = NULL;
1572 }
1573
1574 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1575 {
1576         u64 *parent_pte;
1577
1578         while (sp->multimapped || sp->parent_pte) {
1579                 if (!sp->multimapped)
1580                         parent_pte = sp->parent_pte;
1581                 else {
1582                         struct kvm_pte_chain *chain;
1583
1584                         chain = container_of(sp->parent_ptes.first,
1585                                              struct kvm_pte_chain, link);
1586                         parent_pte = chain->parent_ptes[0];
1587                 }
1588                 BUG_ON(!parent_pte);
1589                 kvm_mmu_put_page(sp, parent_pte);
1590                 __set_spte(parent_pte, shadow_trap_nonpresent_pte);
1591         }
1592 }
1593
1594 static int mmu_zap_unsync_children(struct kvm *kvm,
1595                                    struct kvm_mmu_page *parent,
1596                                    struct list_head *invalid_list)
1597 {
1598         int i, zapped = 0;
1599         struct mmu_page_path parents;
1600         struct kvm_mmu_pages pages;
1601
1602         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1603                 return 0;
1604
1605         kvm_mmu_pages_init(parent, &parents, &pages);
1606         while (mmu_unsync_walk(parent, &pages)) {
1607                 struct kvm_mmu_page *sp;
1608
1609                 for_each_sp(pages, sp, parents, i) {
1610                         kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
1611                         mmu_pages_clear_parents(&parents);
1612                         zapped++;
1613                 }
1614                 kvm_mmu_pages_init(parent, &parents, &pages);
1615         }
1616
1617         return zapped;
1618 }
1619
1620 static int kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1621                                     struct list_head *invalid_list)
1622 {
1623         int ret;
1624
1625         trace_kvm_mmu_prepare_zap_page(sp);
1626         ++kvm->stat.mmu_shadow_zapped;
1627         ret = mmu_zap_unsync_children(kvm, sp, invalid_list);
1628         kvm_mmu_page_unlink_children(kvm, sp);
1629         kvm_mmu_unlink_parents(kvm, sp);
1630         if (!sp->role.invalid && !sp->role.direct)
1631                 unaccount_shadowed(kvm, sp->gfn);
1632         if (sp->unsync)
1633                 kvm_unlink_unsync_page(kvm, sp);
1634         if (!sp->root_count) {
1635                 /* Count self */
1636                 ret++;
1637                 list_move(&sp->link, invalid_list);
1638         } else {
1639                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1640                 kvm_reload_remote_mmus(kvm);
1641         }
1642
1643         sp->role.invalid = 1;
1644         kvm_mmu_reset_last_pte_updated(kvm);
1645         return ret;
1646 }
1647
1648 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1649                                     struct list_head *invalid_list)
1650 {
1651         struct kvm_mmu_page *sp;
1652
1653         if (list_empty(invalid_list))
1654                 return;
1655
1656         kvm_flush_remote_tlbs(kvm);
1657
1658         do {
1659                 sp = list_first_entry(invalid_list, struct kvm_mmu_page, link);
1660                 WARN_ON(!sp->role.invalid || sp->root_count);
1661                 kvm_mmu_free_page(kvm, sp);
1662         } while (!list_empty(invalid_list));
1663
1664 }
1665
1666 /*
1667  * Changing the number of mmu pages allocated to the vm
1668  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1669  */
1670 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1671 {
1672         int used_pages;
1673         LIST_HEAD(invalid_list);
1674
1675         used_pages = kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages;
1676         used_pages = max(0, used_pages);
1677
1678         /*
1679          * If we set the number of mmu pages to be smaller be than the
1680          * number of actived pages , we must to free some mmu pages before we
1681          * change the value
1682          */
1683
1684         if (used_pages > kvm_nr_mmu_pages) {
1685                 while (used_pages > kvm_nr_mmu_pages &&
1686                         !list_empty(&kvm->arch.active_mmu_pages)) {
1687                         struct kvm_mmu_page *page;
1688
1689                         page = container_of(kvm->arch.active_mmu_pages.prev,
1690                                             struct kvm_mmu_page, link);
1691                         used_pages -= kvm_mmu_prepare_zap_page(kvm, page,
1692                                                                &invalid_list);
1693                 }
1694                 kvm_mmu_commit_zap_page(kvm, &invalid_list);
1695                 kvm_nr_mmu_pages = used_pages;
1696                 kvm->arch.n_free_mmu_pages = 0;
1697         }
1698         else
1699                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1700                                          - kvm->arch.n_alloc_mmu_pages;
1701
1702         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1703 }
1704
1705 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1706 {
1707         struct kvm_mmu_page *sp;
1708         struct hlist_node *node;
1709         LIST_HEAD(invalid_list);
1710         int r;
1711
1712         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1713         r = 0;
1714
1715         for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
1716                 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1717                          sp->role.word);
1718                 r = 1;
1719                 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
1720         }
1721         kvm_mmu_commit_zap_page(kvm, &invalid_list);
1722         return r;
1723 }
1724
1725 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1726 {
1727         struct kvm_mmu_page *sp;
1728         struct hlist_node *node;
1729         LIST_HEAD(invalid_list);
1730
1731         for_each_gfn_indirect_valid_sp(kvm, sp, gfn, node) {
1732                 pgprintk("%s: zap %lx %x\n",
1733                          __func__, gfn, sp->role.word);
1734                 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
1735         }
1736         kvm_mmu_commit_zap_page(kvm, &invalid_list);
1737 }
1738
1739 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1740 {
1741         int slot = memslot_id(kvm, gfn);
1742         struct kvm_mmu_page *sp = page_header(__pa(pte));
1743
1744         __set_bit(slot, sp->slot_bitmap);
1745 }
1746
1747 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1748 {
1749         int i;
1750         u64 *pt = sp->spt;
1751
1752         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1753                 return;
1754
1755         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1756                 if (pt[i] == shadow_notrap_nonpresent_pte)
1757                         __set_spte(&pt[i], shadow_trap_nonpresent_pte);
1758         }
1759 }
1760
1761 /*
1762  * The function is based on mtrr_type_lookup() in
1763  * arch/x86/kernel/cpu/mtrr/generic.c
1764  */
1765 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1766                          u64 start, u64 end)
1767 {
1768         int i;
1769         u64 base, mask;
1770         u8 prev_match, curr_match;
1771         int num_var_ranges = KVM_NR_VAR_MTRR;
1772
1773         if (!mtrr_state->enabled)
1774                 return 0xFF;
1775
1776         /* Make end inclusive end, instead of exclusive */
1777         end--;
1778
1779         /* Look in fixed ranges. Just return the type as per start */
1780         if (mtrr_state->have_fixed && (start < 0x100000)) {
1781                 int idx;
1782
1783                 if (start < 0x80000) {
1784                         idx = 0;
1785                         idx += (start >> 16);
1786                         return mtrr_state->fixed_ranges[idx];
1787                 } else if (start < 0xC0000) {
1788                         idx = 1 * 8;
1789                         idx += ((start - 0x80000) >> 14);
1790                         return mtrr_state->fixed_ranges[idx];
1791                 } else if (start < 0x1000000) {
1792                         idx = 3 * 8;
1793                         idx += ((start - 0xC0000) >> 12);
1794                         return mtrr_state->fixed_ranges[idx];
1795                 }
1796         }
1797
1798         /*
1799          * Look in variable ranges
1800          * Look of multiple ranges matching this address and pick type
1801          * as per MTRR precedence
1802          */
1803         if (!(mtrr_state->enabled & 2))
1804                 return mtrr_state->def_type;
1805
1806         prev_match = 0xFF;
1807         for (i = 0; i < num_var_ranges; ++i) {
1808                 unsigned short start_state, end_state;
1809
1810                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1811                         continue;
1812
1813                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1814                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1815                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1816                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1817
1818                 start_state = ((start & mask) == (base & mask));
1819                 end_state = ((end & mask) == (base & mask));
1820                 if (start_state != end_state)
1821                         return 0xFE;
1822
1823                 if ((start & mask) != (base & mask))
1824                         continue;
1825
1826                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1827                 if (prev_match == 0xFF) {
1828                         prev_match = curr_match;
1829                         continue;
1830                 }
1831
1832                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1833                     curr_match == MTRR_TYPE_UNCACHABLE)
1834                         return MTRR_TYPE_UNCACHABLE;
1835
1836                 if ((prev_match == MTRR_TYPE_WRBACK &&
1837                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1838                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1839                      curr_match == MTRR_TYPE_WRBACK)) {
1840                         prev_match = MTRR_TYPE_WRTHROUGH;
1841                         curr_match = MTRR_TYPE_WRTHROUGH;
1842                 }
1843
1844                 if (prev_match != curr_match)
1845                         return MTRR_TYPE_UNCACHABLE;
1846         }
1847
1848         if (prev_match != 0xFF)
1849                 return prev_match;
1850
1851         return mtrr_state->def_type;
1852 }
1853
1854 u8 kvm_get_guest_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1855 {
1856         u8 mtrr;
1857
1858         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1859                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1860         if (mtrr == 0xfe || mtrr == 0xff)
1861                 mtrr = MTRR_TYPE_WRBACK;
1862         return mtrr;
1863 }
1864 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type);
1865
1866 static void __kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1867 {
1868         trace_kvm_mmu_unsync_page(sp);
1869         ++vcpu->kvm->stat.mmu_unsync;
1870         sp->unsync = 1;
1871
1872         kvm_mmu_mark_parents_unsync(sp);
1873         mmu_convert_notrap(sp);
1874 }
1875
1876 static void kvm_unsync_pages(struct kvm_vcpu *vcpu,  gfn_t gfn)
1877 {
1878         struct kvm_mmu_page *s;
1879         struct hlist_node *node;
1880
1881         for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1882                 if (s->unsync)
1883                         continue;
1884                 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
1885                 __kvm_unsync_page(vcpu, s);
1886         }
1887 }
1888
1889 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1890                                   bool can_unsync)
1891 {
1892         struct kvm_mmu_page *s;
1893         struct hlist_node *node;
1894         bool need_unsync = false;
1895
1896         for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn, node) {
1897                 if (!can_unsync)
1898                         return 1;
1899
1900                 if (s->role.level != PT_PAGE_TABLE_LEVEL)
1901                         return 1;
1902
1903                 if (!need_unsync && !s->unsync) {
1904                         if (!oos_shadow)
1905                                 return 1;
1906                         need_unsync = true;
1907                 }
1908         }
1909         if (need_unsync)
1910                 kvm_unsync_pages(vcpu, gfn);
1911         return 0;
1912 }
1913
1914 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1915                     unsigned pte_access, int user_fault,
1916                     int write_fault, int dirty, int level,
1917                     gfn_t gfn, pfn_t pfn, bool speculative,
1918                     bool can_unsync, bool reset_host_protection)
1919 {
1920         u64 spte;
1921         int ret = 0;
1922
1923         /*
1924          * We don't set the accessed bit, since we sometimes want to see
1925          * whether the guest actually used the pte (in order to detect
1926          * demand paging).
1927          */
1928         spte = shadow_base_present_pte | shadow_dirty_mask;
1929         if (!speculative)
1930                 spte |= shadow_accessed_mask;
1931         if (!dirty)
1932                 pte_access &= ~ACC_WRITE_MASK;
1933         if (pte_access & ACC_EXEC_MASK)
1934                 spte |= shadow_x_mask;
1935         else
1936                 spte |= shadow_nx_mask;
1937         if (pte_access & ACC_USER_MASK)
1938                 spte |= shadow_user_mask;
1939         if (level > PT_PAGE_TABLE_LEVEL)
1940                 spte |= PT_PAGE_SIZE_MASK;
1941         if (tdp_enabled)
1942                 spte |= kvm_x86_ops->get_mt_mask(vcpu, gfn,
1943                         kvm_is_mmio_pfn(pfn));
1944
1945         if (reset_host_protection)
1946                 spte |= SPTE_HOST_WRITEABLE;
1947
1948         spte |= (u64)pfn << PAGE_SHIFT;
1949
1950         if ((pte_access & ACC_WRITE_MASK)
1951             || (!tdp_enabled && write_fault && !is_write_protection(vcpu)
1952                 && !user_fault)) {
1953
1954                 if (level > PT_PAGE_TABLE_LEVEL &&
1955                     has_wrprotected_page(vcpu->kvm, gfn, level)) {
1956                         ret = 1;
1957                         drop_spte(vcpu->kvm, sptep, shadow_trap_nonpresent_pte);
1958                         goto done;
1959                 }
1960
1961                 spte |= PT_WRITABLE_MASK;
1962
1963                 if (!tdp_enabled && !(pte_access & ACC_WRITE_MASK))
1964                         spte &= ~PT_USER_MASK;
1965
1966                 /*
1967                  * Optimization: for pte sync, if spte was writable the hash
1968                  * lookup is unnecessary (and expensive). Write protection
1969                  * is responsibility of mmu_get_page / kvm_sync_page.
1970                  * Same reasoning can be applied to dirty page accounting.
1971                  */
1972                 if (!can_unsync && is_writable_pte(*sptep))
1973                         goto set_pte;
1974
1975                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1976                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1977                                  __func__, gfn);
1978                         ret = 1;
1979                         pte_access &= ~ACC_WRITE_MASK;
1980                         if (is_writable_pte(spte))
1981                                 spte &= ~PT_WRITABLE_MASK;
1982                 }
1983         }
1984
1985         if (pte_access & ACC_WRITE_MASK)
1986                 mark_page_dirty(vcpu->kvm, gfn);
1987
1988 set_pte:
1989         if (is_writable_pte(*sptep) && !is_writable_pte(spte))
1990                 kvm_set_pfn_dirty(pfn);
1991         update_spte(sptep, spte);
1992 done:
1993         return ret;
1994 }
1995
1996 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
1997                          unsigned pt_access, unsigned pte_access,
1998                          int user_fault, int write_fault, int dirty,
1999                          int *ptwrite, int level, gfn_t gfn,
2000                          pfn_t pfn, bool speculative,
2001                          bool reset_host_protection)
2002 {
2003         int was_rmapped = 0;
2004         int rmap_count;
2005
2006         pgprintk("%s: spte %llx access %x write_fault %d"
2007                  " user_fault %d gfn %lx\n",
2008                  __func__, *sptep, pt_access,
2009                  write_fault, user_fault, gfn);
2010
2011         if (is_rmap_spte(*sptep)) {
2012                 /*
2013                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
2014                  * the parent of the now unreachable PTE.
2015                  */
2016                 if (level > PT_PAGE_TABLE_LEVEL &&
2017                     !is_large_pte(*sptep)) {
2018                         struct kvm_mmu_page *child;
2019                         u64 pte = *sptep;
2020
2021                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2022                         mmu_page_remove_parent_pte(child, sptep);
2023                         __set_spte(sptep, shadow_trap_nonpresent_pte);
2024                         kvm_flush_remote_tlbs(vcpu->kvm);
2025                 } else if (pfn != spte_to_pfn(*sptep)) {
2026                         pgprintk("hfn old %lx new %lx\n",
2027                                  spte_to_pfn(*sptep), pfn);
2028                         drop_spte(vcpu->kvm, sptep, shadow_trap_nonpresent_pte);
2029                         kvm_flush_remote_tlbs(vcpu->kvm);
2030                 } else
2031                         was_rmapped = 1;
2032         }
2033
2034         if (set_spte(vcpu, sptep, pte_access, user_fault, write_fault,
2035                       dirty, level, gfn, pfn, speculative, true,
2036                       reset_host_protection)) {
2037                 if (write_fault)
2038                         *ptwrite = 1;
2039                 kvm_mmu_flush_tlb(vcpu);
2040         }
2041
2042         pgprintk("%s: setting spte %llx\n", __func__, *sptep);
2043         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
2044                  is_large_pte(*sptep)? "2MB" : "4kB",
2045                  *sptep & PT_PRESENT_MASK ?"RW":"R", gfn,
2046                  *sptep, sptep);
2047         if (!was_rmapped && is_large_pte(*sptep))
2048                 ++vcpu->kvm->stat.lpages;
2049
2050         page_header_update_slot(vcpu->kvm, sptep, gfn);
2051         if (!was_rmapped) {
2052                 rmap_count = rmap_add(vcpu, sptep, gfn);
2053                 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
2054                         rmap_recycle(vcpu, sptep, gfn);
2055         }
2056         kvm_release_pfn_clean(pfn);
2057         if (speculative) {
2058                 vcpu->arch.last_pte_updated = sptep;
2059                 vcpu->arch.last_pte_gfn = gfn;
2060         }
2061 }
2062
2063 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
2064 {
2065 }
2066
2067 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
2068                         int level, gfn_t gfn, pfn_t pfn)
2069 {
2070         struct kvm_shadow_walk_iterator iterator;
2071         struct kvm_mmu_page *sp;
2072         int pt_write = 0;
2073         gfn_t pseudo_gfn;
2074
2075         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
2076                 if (iterator.level == level) {
2077                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
2078                                      0, write, 1, &pt_write,
2079                                      level, gfn, pfn, false, true);
2080                         ++vcpu->stat.pf_fixed;
2081                         break;
2082                 }
2083
2084                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
2085                         u64 base_addr = iterator.addr;
2086
2087                         base_addr &= PT64_LVL_ADDR_MASK(iterator.level);
2088                         pseudo_gfn = base_addr >> PAGE_SHIFT;
2089                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
2090                                               iterator.level - 1,
2091                                               1, ACC_ALL, iterator.sptep);
2092                         if (!sp) {
2093                                 pgprintk("nonpaging_map: ENOMEM\n");
2094                                 kvm_release_pfn_clean(pfn);
2095                                 return -ENOMEM;
2096                         }
2097
2098                         __set_spte(iterator.sptep,
2099                                    __pa(sp->spt)
2100                                    | PT_PRESENT_MASK | PT_WRITABLE_MASK
2101                                    | shadow_user_mask | shadow_x_mask);
2102                 }
2103         }
2104         return pt_write;
2105 }
2106
2107 static void kvm_send_hwpoison_signal(struct kvm *kvm, gfn_t gfn)
2108 {
2109         char buf[1];
2110         void __user *hva;
2111         int r;
2112
2113         /* Touch the page, so send SIGBUS */
2114         hva = (void __user *)gfn_to_hva(kvm, gfn);
2115         r = copy_from_user(buf, hva, 1);
2116 }
2117
2118 static int kvm_handle_bad_page(struct kvm *kvm, gfn_t gfn, pfn_t pfn)
2119 {
2120         kvm_release_pfn_clean(pfn);
2121         if (is_hwpoison_pfn(pfn)) {
2122                 kvm_send_hwpoison_signal(kvm, gfn);
2123                 return 0;
2124         } else if (is_fault_pfn(pfn))
2125                 return -EFAULT;
2126
2127         return 1;
2128 }
2129
2130 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
2131 {
2132         int r;
2133         int level;
2134         pfn_t pfn;
2135         unsigned long mmu_seq;
2136
2137         level = mapping_level(vcpu, gfn);
2138
2139         /*
2140          * This path builds a PAE pagetable - so we can map 2mb pages at
2141          * maximum. Therefore check if the level is larger than that.
2142          */
2143         if (level > PT_DIRECTORY_LEVEL)
2144                 level = PT_DIRECTORY_LEVEL;
2145
2146         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2147
2148         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2149         smp_rmb();
2150         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2151
2152         /* mmio */
2153         if (is_error_pfn(pfn))
2154                 return kvm_handle_bad_page(vcpu->kvm, gfn, pfn);
2155
2156         spin_lock(&vcpu->kvm->mmu_lock);
2157         if (mmu_notifier_retry(vcpu, mmu_seq))
2158                 goto out_unlock;
2159         kvm_mmu_free_some_pages(vcpu);
2160         r = __direct_map(vcpu, v, write, level, gfn, pfn);
2161         spin_unlock(&vcpu->kvm->mmu_lock);
2162
2163
2164         return r;
2165
2166 out_unlock:
2167         spin_unlock(&vcpu->kvm->mmu_lock);
2168         kvm_release_pfn_clean(pfn);
2169         return 0;
2170 }
2171
2172
2173 static void mmu_free_roots(struct kvm_vcpu *vcpu)
2174 {
2175         int i;
2176         struct kvm_mmu_page *sp;
2177         LIST_HEAD(invalid_list);
2178
2179         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2180                 return;
2181         spin_lock(&vcpu->kvm->mmu_lock);
2182         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2183                 hpa_t root = vcpu->arch.mmu.root_hpa;
2184
2185                 sp = page_header(root);
2186                 --sp->root_count;
2187                 if (!sp->root_count && sp->role.invalid) {
2188                         kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
2189                         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2190                 }
2191                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2192                 spin_unlock(&vcpu->kvm->mmu_lock);
2193                 return;
2194         }
2195         for (i = 0; i < 4; ++i) {
2196                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2197
2198                 if (root) {
2199                         root &= PT64_BASE_ADDR_MASK;
2200                         sp = page_header(root);
2201                         --sp->root_count;
2202                         if (!sp->root_count && sp->role.invalid)
2203                                 kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2204                                                          &invalid_list);
2205                 }
2206                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2207         }
2208         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2209         spin_unlock(&vcpu->kvm->mmu_lock);
2210         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2211 }
2212
2213 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
2214 {
2215         int ret = 0;
2216
2217         if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
2218                 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
2219                 ret = 1;
2220         }
2221
2222         return ret;
2223 }
2224
2225 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
2226 {
2227         int i;
2228         gfn_t root_gfn;
2229         struct kvm_mmu_page *sp;
2230         int direct = 0;
2231         u64 pdptr;
2232
2233         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
2234
2235         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2236                 hpa_t root = vcpu->arch.mmu.root_hpa;
2237
2238                 ASSERT(!VALID_PAGE(root));
2239                 if (mmu_check_root(vcpu, root_gfn))
2240                         return 1;
2241                 if (tdp_enabled) {
2242                         direct = 1;
2243                         root_gfn = 0;
2244                 }
2245                 spin_lock(&vcpu->kvm->mmu_lock);
2246                 kvm_mmu_free_some_pages(vcpu);
2247                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
2248                                       PT64_ROOT_LEVEL, direct,
2249                                       ACC_ALL, NULL);
2250                 root = __pa(sp->spt);
2251                 ++sp->root_count;
2252                 spin_unlock(&vcpu->kvm->mmu_lock);
2253                 vcpu->arch.mmu.root_hpa = root;
2254                 return 0;
2255         }
2256         direct = !is_paging(vcpu);
2257         for (i = 0; i < 4; ++i) {
2258                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2259
2260                 ASSERT(!VALID_PAGE(root));
2261                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
2262                         pdptr = kvm_pdptr_read(vcpu, i);
2263                         if (!is_present_gpte(pdptr)) {
2264                                 vcpu->arch.mmu.pae_root[i] = 0;
2265                                 continue;
2266                         }
2267                         root_gfn = pdptr >> PAGE_SHIFT;
2268                 } else if (vcpu->arch.mmu.root_level == 0)
2269                         root_gfn = 0;
2270                 if (mmu_check_root(vcpu, root_gfn))
2271                         return 1;
2272                 if (tdp_enabled) {
2273                         direct = 1;
2274                         root_gfn = i << 30;
2275                 }
2276                 spin_lock(&vcpu->kvm->mmu_lock);
2277                 kvm_mmu_free_some_pages(vcpu);
2278                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
2279                                       PT32_ROOT_LEVEL, direct,
2280                                       ACC_ALL, NULL);
2281                 root = __pa(sp->spt);
2282                 ++sp->root_count;
2283                 spin_unlock(&vcpu->kvm->mmu_lock);
2284
2285                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
2286         }
2287         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
2288         return 0;
2289 }
2290
2291 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2292 {
2293         int i;
2294         struct kvm_mmu_page *sp;
2295
2296         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2297                 return;
2298         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2299                 hpa_t root = vcpu->arch.mmu.root_hpa;
2300                 sp = page_header(root);
2301                 mmu_sync_children(vcpu, sp);
2302                 return;
2303         }
2304         for (i = 0; i < 4; ++i) {
2305                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2306
2307                 if (root && VALID_PAGE(root)) {
2308                         root &= PT64_BASE_ADDR_MASK;
2309                         sp = page_header(root);
2310                         mmu_sync_children(vcpu, sp);
2311                 }
2312         }
2313 }
2314
2315 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2316 {
2317         spin_lock(&vcpu->kvm->mmu_lock);
2318         mmu_sync_roots(vcpu);
2319         spin_unlock(&vcpu->kvm->mmu_lock);
2320 }
2321
2322 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr,
2323                                   u32 access, u32 *error)
2324 {
2325         if (error)
2326                 *error = 0;
2327         return vaddr;
2328 }
2329
2330 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2331                                 u32 error_code)
2332 {
2333         gfn_t gfn;
2334         int r;
2335
2336         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2337         r = mmu_topup_memory_caches(vcpu);
2338         if (r)
2339                 return r;
2340
2341         ASSERT(vcpu);
2342         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2343
2344         gfn = gva >> PAGE_SHIFT;
2345
2346         return nonpaging_map(vcpu, gva & PAGE_MASK,
2347                              error_code & PFERR_WRITE_MASK, gfn);
2348 }
2349
2350 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2351                                 u32 error_code)
2352 {
2353         pfn_t pfn;
2354         int r;
2355         int level;
2356         gfn_t gfn = gpa >> PAGE_SHIFT;
2357         unsigned long mmu_seq;
2358
2359         ASSERT(vcpu);
2360         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2361
2362         r = mmu_topup_memory_caches(vcpu);
2363         if (r)
2364                 return r;
2365
2366         level = mapping_level(vcpu, gfn);
2367
2368         gfn &= ~(KVM_PAGES_PER_HPAGE(level) - 1);
2369
2370         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2371         smp_rmb();
2372         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2373         if (is_error_pfn(pfn))
2374                 return kvm_handle_bad_page(vcpu->kvm, gfn, pfn);
2375         spin_lock(&vcpu->kvm->mmu_lock);
2376         if (mmu_notifier_retry(vcpu, mmu_seq))
2377                 goto out_unlock;
2378         kvm_mmu_free_some_pages(vcpu);
2379         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2380                          level, gfn, pfn);
2381         spin_unlock(&vcpu->kvm->mmu_lock);
2382
2383         return r;
2384
2385 out_unlock:
2386         spin_unlock(&vcpu->kvm->mmu_lock);
2387         kvm_release_pfn_clean(pfn);
2388         return 0;
2389 }
2390
2391 static void nonpaging_free(struct kvm_vcpu *vcpu)
2392 {
2393         mmu_free_roots(vcpu);
2394 }
2395
2396 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2397 {
2398         struct kvm_mmu *context = &vcpu->arch.mmu;
2399
2400         context->new_cr3 = nonpaging_new_cr3;
2401         context->page_fault = nonpaging_page_fault;
2402         context->gva_to_gpa = nonpaging_gva_to_gpa;
2403         context->free = nonpaging_free;
2404         context->prefetch_page = nonpaging_prefetch_page;
2405         context->sync_page = nonpaging_sync_page;
2406         context->invlpg = nonpaging_invlpg;
2407         context->root_level = 0;
2408         context->shadow_root_level = PT32E_ROOT_LEVEL;
2409         context->root_hpa = INVALID_PAGE;
2410         return 0;
2411 }
2412
2413 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2414 {
2415         ++vcpu->stat.tlb_flush;
2416         kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
2417 }
2418
2419 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2420 {
2421         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2422         mmu_free_roots(vcpu);
2423 }
2424
2425 static void inject_page_fault(struct kvm_vcpu *vcpu,
2426                               u64 addr,
2427                               u32 err_code)
2428 {
2429         kvm_inject_page_fault(vcpu, addr, err_code);
2430 }
2431
2432 static void paging_free(struct kvm_vcpu *vcpu)
2433 {
2434         nonpaging_free(vcpu);
2435 }
2436
2437 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2438 {
2439         int bit7;
2440
2441         bit7 = (gpte >> 7) & 1;
2442         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2443 }
2444
2445 #define PTTYPE 64
2446 #include "paging_tmpl.h"
2447 #undef PTTYPE
2448
2449 #define PTTYPE 32
2450 #include "paging_tmpl.h"
2451 #undef PTTYPE
2452
2453 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2454 {
2455         struct kvm_mmu *context = &vcpu->arch.mmu;
2456         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2457         u64 exb_bit_rsvd = 0;
2458
2459         if (!is_nx(vcpu))
2460                 exb_bit_rsvd = rsvd_bits(63, 63);
2461         switch (level) {
2462         case PT32_ROOT_LEVEL:
2463                 /* no rsvd bits for 2 level 4K page table entries */
2464                 context->rsvd_bits_mask[0][1] = 0;
2465                 context->rsvd_bits_mask[0][0] = 0;
2466                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2467
2468                 if (!is_pse(vcpu)) {
2469                         context->rsvd_bits_mask[1][1] = 0;
2470                         break;
2471                 }
2472
2473                 if (is_cpuid_PSE36())
2474                         /* 36bits PSE 4MB page */
2475                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2476                 else
2477                         /* 32 bits PSE 4MB page */
2478                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2479                 break;
2480         case PT32E_ROOT_LEVEL:
2481                 context->rsvd_bits_mask[0][2] =
2482                         rsvd_bits(maxphyaddr, 63) |
2483                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2484                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2485                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2486                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2487                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2488                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2489                         rsvd_bits(maxphyaddr, 62) |
2490                         rsvd_bits(13, 20);              /* large page */
2491                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2492                 break;
2493         case PT64_ROOT_LEVEL:
2494                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2495                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2496                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2497                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2498                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2499                         rsvd_bits(maxphyaddr, 51);
2500                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2501                         rsvd_bits(maxphyaddr, 51);
2502                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2503                 context->rsvd_bits_mask[1][2] = exb_bit_rsvd |
2504                         rsvd_bits(maxphyaddr, 51) |
2505                         rsvd_bits(13, 29);
2506                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2507                         rsvd_bits(maxphyaddr, 51) |
2508                         rsvd_bits(13, 20);              /* large page */
2509                 context->rsvd_bits_mask[1][0] = context->rsvd_bits_mask[0][0];
2510                 break;
2511         }
2512 }
2513
2514 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2515 {
2516         struct kvm_mmu *context = &vcpu->arch.mmu;
2517
2518         ASSERT(is_pae(vcpu));
2519         context->new_cr3 = paging_new_cr3;
2520         context->page_fault = paging64_page_fault;
2521         context->gva_to_gpa = paging64_gva_to_gpa;
2522         context->prefetch_page = paging64_prefetch_page;
2523         context->sync_page = paging64_sync_page;
2524         context->invlpg = paging64_invlpg;
2525         context->free = paging_free;
2526         context->root_level = level;
2527         context->shadow_root_level = level;
2528         context->root_hpa = INVALID_PAGE;
2529         return 0;
2530 }
2531
2532 static int paging64_init_context(struct kvm_vcpu *vcpu)
2533 {
2534         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2535         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2536 }
2537
2538 static int paging32_init_context(struct kvm_vcpu *vcpu)
2539 {
2540         struct kvm_mmu *context = &vcpu->arch.mmu;
2541
2542         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2543         context->new_cr3 = paging_new_cr3;
2544         context->page_fault = paging32_page_fault;
2545         context->gva_to_gpa = paging32_gva_to_gpa;
2546         context->free = paging_free;
2547         context->prefetch_page = paging32_prefetch_page;
2548         context->sync_page = paging32_sync_page;
2549         context->invlpg = paging32_invlpg;
2550         context->root_level = PT32_ROOT_LEVEL;
2551         context->shadow_root_level = PT32E_ROOT_LEVEL;
2552         context->root_hpa = INVALID_PAGE;
2553         return 0;
2554 }
2555
2556 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2557 {
2558         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2559         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2560 }
2561
2562 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2563 {
2564         struct kvm_mmu *context = &vcpu->arch.mmu;
2565
2566         context->new_cr3 = nonpaging_new_cr3;
2567         context->page_fault = tdp_page_fault;
2568         context->free = nonpaging_free;
2569         context->prefetch_page = nonpaging_prefetch_page;
2570         context->sync_page = nonpaging_sync_page;
2571         context->invlpg = nonpaging_invlpg;
2572         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2573         context->root_hpa = INVALID_PAGE;
2574
2575         if (!is_paging(vcpu)) {
2576                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2577                 context->root_level = 0;
2578         } else if (is_long_mode(vcpu)) {
2579                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2580                 context->gva_to_gpa = paging64_gva_to_gpa;
2581                 context->root_level = PT64_ROOT_LEVEL;
2582         } else if (is_pae(vcpu)) {
2583                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2584                 context->gva_to_gpa = paging64_gva_to_gpa;
2585                 context->root_level = PT32E_ROOT_LEVEL;
2586         } else {
2587                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2588                 context->gva_to_gpa = paging32_gva_to_gpa;
2589                 context->root_level = PT32_ROOT_LEVEL;
2590         }
2591
2592         return 0;
2593 }
2594
2595 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2596 {
2597         int r;
2598
2599         ASSERT(vcpu);
2600         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2601
2602         if (!is_paging(vcpu))
2603                 r = nonpaging_init_context(vcpu);
2604         else if (is_long_mode(vcpu))
2605                 r = paging64_init_context(vcpu);
2606         else if (is_pae(vcpu))
2607                 r = paging32E_init_context(vcpu);
2608         else
2609                 r = paging32_init_context(vcpu);
2610
2611         vcpu->arch.mmu.base_role.cr4_pae = !!is_pae(vcpu);
2612         vcpu->arch.mmu.base_role.cr0_wp = is_write_protection(vcpu);
2613
2614         return r;
2615 }
2616
2617 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2618 {
2619         vcpu->arch.update_pte.pfn = bad_pfn;
2620
2621         if (tdp_enabled)
2622                 return init_kvm_tdp_mmu(vcpu);
2623         else
2624                 return init_kvm_softmmu(vcpu);
2625 }
2626
2627 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2628 {
2629         ASSERT(vcpu);
2630         if (VALID_PAGE(vcpu->arch.mmu.root_hpa))
2631                 /* mmu.free() should set root_hpa = INVALID_PAGE */
2632                 vcpu->arch.mmu.free(vcpu);
2633 }
2634
2635 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2636 {
2637         destroy_kvm_mmu(vcpu);
2638         return init_kvm_mmu(vcpu);
2639 }
2640 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2641
2642 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2643 {
2644         int r;
2645
2646         r = mmu_topup_memory_caches(vcpu);
2647         if (r)
2648                 goto out;
2649         r = mmu_alloc_roots(vcpu);
2650         spin_lock(&vcpu->kvm->mmu_lock);
2651         mmu_sync_roots(vcpu);
2652         spin_unlock(&vcpu->kvm->mmu_lock);
2653         if (r)
2654                 goto out;
2655         /* set_cr3() should ensure TLB has been flushed */
2656         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2657 out:
2658         return r;
2659 }
2660 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2661
2662 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2663 {
2664         mmu_free_roots(vcpu);
2665 }
2666
2667 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2668                                   struct kvm_mmu_page *sp,
2669                                   u64 *spte)
2670 {
2671         u64 pte;
2672         struct kvm_mmu_page *child;
2673
2674         pte = *spte;
2675         if (is_shadow_present_pte(pte)) {
2676                 if (is_last_spte(pte, sp->role.level))
2677                         drop_spte(vcpu->kvm, spte, shadow_trap_nonpresent_pte);
2678                 else {
2679                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2680                         mmu_page_remove_parent_pte(child, spte);
2681                 }
2682         }
2683         __set_spte(spte, shadow_trap_nonpresent_pte);
2684         if (is_large_pte(pte))
2685                 --vcpu->kvm->stat.lpages;
2686 }
2687
2688 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2689                                   struct kvm_mmu_page *sp,
2690                                   u64 *spte,
2691                                   const void *new)
2692 {
2693         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2694                 ++vcpu->kvm->stat.mmu_pde_zapped;
2695                 return;
2696         }
2697
2698         if (is_rsvd_bits_set(vcpu, *(u64 *)new, PT_PAGE_TABLE_LEVEL))
2699                 return;
2700
2701         ++vcpu->kvm->stat.mmu_pte_updated;
2702         if (!sp->role.cr4_pae)
2703                 paging32_update_pte(vcpu, sp, spte, new);
2704         else
2705                 paging64_update_pte(vcpu, sp, spte, new);
2706 }
2707
2708 static bool need_remote_flush(u64 old, u64 new)
2709 {
2710         if (!is_shadow_present_pte(old))
2711                 return false;
2712         if (!is_shadow_present_pte(new))
2713                 return true;
2714         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2715                 return true;
2716         old ^= PT64_NX_MASK;
2717         new ^= PT64_NX_MASK;
2718         return (old & ~new & PT64_PERM_MASK) != 0;
2719 }
2720
2721 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, bool zap_page,
2722                                     bool remote_flush, bool local_flush)
2723 {
2724         if (zap_page)
2725                 return;
2726
2727         if (remote_flush)
2728                 kvm_flush_remote_tlbs(vcpu->kvm);
2729         else if (local_flush)
2730                 kvm_mmu_flush_tlb(vcpu);
2731 }
2732
2733 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2734 {
2735         u64 *spte = vcpu->arch.last_pte_updated;
2736
2737         return !!(spte && (*spte & shadow_accessed_mask));
2738 }
2739
2740 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2741                                           u64 gpte)
2742 {
2743         gfn_t gfn;
2744         pfn_t pfn;
2745
2746         if (!is_present_gpte(gpte))
2747                 return;
2748         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2749
2750         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2751         smp_rmb();
2752         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2753
2754         if (is_error_pfn(pfn)) {
2755                 kvm_release_pfn_clean(pfn);
2756                 return;
2757         }
2758         vcpu->arch.update_pte.gfn = gfn;
2759         vcpu->arch.update_pte.pfn = pfn;
2760 }
2761
2762 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2763 {
2764         u64 *spte = vcpu->arch.last_pte_updated;
2765
2766         if (spte
2767             && vcpu->arch.last_pte_gfn == gfn
2768             && shadow_accessed_mask
2769             && !(*spte & shadow_accessed_mask)
2770             && is_shadow_present_pte(*spte))
2771                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2772 }
2773
2774 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2775                        const u8 *new, int bytes,
2776                        bool guest_initiated)
2777 {
2778         gfn_t gfn = gpa >> PAGE_SHIFT;
2779         union kvm_mmu_page_role mask = { .word = 0 };
2780         struct kvm_mmu_page *sp;
2781         struct hlist_node *node;
2782         LIST_HEAD(invalid_list);
2783         u64 entry, gentry;
2784         u64 *spte;
2785         unsigned offset = offset_in_page(gpa);
2786         unsigned pte_size;
2787         unsigned page_offset;
2788         unsigned misaligned;
2789         unsigned quadrant;
2790         int level;
2791         int flooded = 0;
2792         int npte;
2793         int r;
2794         int invlpg_counter;
2795         bool remote_flush, local_flush, zap_page;
2796
2797         zap_page = remote_flush = local_flush = false;
2798
2799         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2800
2801         invlpg_counter = atomic_read(&vcpu->kvm->arch.invlpg_counter);
2802
2803         /*
2804          * Assume that the pte write on a page table of the same type
2805          * as the current vcpu paging mode.  This is nearly always true
2806          * (might be false while changing modes).  Note it is verified later
2807          * by update_pte().
2808          */
2809         if ((is_pae(vcpu) && bytes == 4) || !new) {
2810                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2811                 if (is_pae(vcpu)) {
2812                         gpa &= ~(gpa_t)7;
2813                         bytes = 8;
2814                 }
2815                 r = kvm_read_guest(vcpu->kvm, gpa, &gentry, min(bytes, 8));
2816                 if (r)
2817                         gentry = 0;
2818                 new = (const u8 *)&gentry;
2819         }
2820
2821         switch (bytes) {
2822         case 4:
2823                 gentry = *(const u32 *)new;
2824                 break;
2825         case 8:
2826                 gentry = *(const u64 *)new;
2827                 break;
2828         default:
2829                 gentry = 0;
2830                 break;
2831         }
2832
2833         mmu_guess_page_from_pte_write(vcpu, gpa, gentry);
2834         spin_lock(&vcpu->kvm->mmu_lock);
2835         if (atomic_read(&vcpu->kvm->arch.invlpg_counter) != invlpg_counter)
2836                 gentry = 0;
2837         kvm_mmu_access_page(vcpu, gfn);
2838         kvm_mmu_free_some_pages(vcpu);
2839         ++vcpu->kvm->stat.mmu_pte_write;
2840         kvm_mmu_audit(vcpu, "pre pte write");
2841         if (guest_initiated) {
2842                 if (gfn == vcpu->arch.last_pt_write_gfn
2843                     && !last_updated_pte_accessed(vcpu)) {
2844                         ++vcpu->arch.last_pt_write_count;
2845                         if (vcpu->arch.last_pt_write_count >= 3)
2846                                 flooded = 1;
2847                 } else {
2848                         vcpu->arch.last_pt_write_gfn = gfn;
2849                         vcpu->arch.last_pt_write_count = 1;
2850                         vcpu->arch.last_pte_updated = NULL;
2851                 }
2852         }
2853
2854         mask.cr0_wp = mask.cr4_pae = mask.nxe = 1;
2855         for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn, node) {
2856                 pte_size = sp->role.cr4_pae ? 8 : 4;
2857                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2858                 misaligned |= bytes < 4;
2859                 if (misaligned || flooded) {
2860                         /*
2861                          * Misaligned accesses are too much trouble to fix
2862                          * up; also, they usually indicate a page is not used
2863                          * as a page table.
2864                          *
2865                          * If we're seeing too many writes to a page,
2866                          * it may no longer be a page table, or we may be
2867                          * forking, in which case it is better to unmap the
2868                          * page.
2869                          */
2870                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2871                                  gpa, bytes, sp->role.word);
2872                         zap_page |= !!kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2873                                                      &invalid_list);
2874                         ++vcpu->kvm->stat.mmu_flooded;
2875                         continue;
2876                 }
2877                 page_offset = offset;
2878                 level = sp->role.level;
2879                 npte = 1;
2880                 if (!sp->role.cr4_pae) {
2881                         page_offset <<= 1;      /* 32->64 */
2882                         /*
2883                          * A 32-bit pde maps 4MB while the shadow pdes map
2884                          * only 2MB.  So we need to double the offset again
2885                          * and zap two pdes instead of one.
2886                          */
2887                         if (level == PT32_ROOT_LEVEL) {
2888                                 page_offset &= ~7; /* kill rounding error */
2889                                 page_offset <<= 1;
2890                                 npte = 2;
2891                         }
2892                         quadrant = page_offset >> PAGE_SHIFT;
2893                         page_offset &= ~PAGE_MASK;
2894                         if (quadrant != sp->role.quadrant)
2895                                 continue;
2896                 }
2897                 local_flush = true;
2898                 spte = &sp->spt[page_offset / sizeof(*spte)];
2899                 while (npte--) {
2900                         entry = *spte;
2901                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2902                         if (gentry &&
2903                               !((sp->role.word ^ vcpu->arch.mmu.base_role.word)
2904                               & mask.word))
2905                                 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
2906                         if (!remote_flush && need_remote_flush(entry, *spte))
2907                                 remote_flush = true;
2908                         ++spte;
2909                 }
2910         }
2911         mmu_pte_write_flush_tlb(vcpu, zap_page, remote_flush, local_flush);
2912         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2913         kvm_mmu_audit(vcpu, "post pte write");
2914         spin_unlock(&vcpu->kvm->mmu_lock);
2915         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2916                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2917                 vcpu->arch.update_pte.pfn = bad_pfn;
2918         }
2919 }
2920
2921 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2922 {
2923         gpa_t gpa;
2924         int r;
2925
2926         if (tdp_enabled)
2927                 return 0;
2928
2929         gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2930
2931         spin_lock(&vcpu->kvm->mmu_lock);
2932         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2933         spin_unlock(&vcpu->kvm->mmu_lock);
2934         return r;
2935 }
2936 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2937
2938 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2939 {
2940         int free_pages;
2941         LIST_HEAD(invalid_list);
2942
2943         free_pages = vcpu->kvm->arch.n_free_mmu_pages;
2944         while (free_pages < KVM_REFILL_PAGES &&
2945                !list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2946                 struct kvm_mmu_page *sp;
2947
2948                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2949                                   struct kvm_mmu_page, link);
2950                 free_pages += kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2951                                                        &invalid_list);
2952                 ++vcpu->kvm->stat.mmu_recycled;
2953         }
2954         kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2955 }
2956
2957 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2958 {
2959         int r;
2960         enum emulation_result er;
2961
2962         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2963         if (r < 0)
2964                 goto out;
2965
2966         if (!r) {
2967                 r = 1;
2968                 goto out;
2969         }
2970
2971         r = mmu_topup_memory_caches(vcpu);
2972         if (r)
2973                 goto out;
2974
2975         er = emulate_instruction(vcpu, cr2, error_code, 0);
2976
2977         switch (er) {
2978         case EMULATE_DONE:
2979                 return 1;
2980         case EMULATE_DO_MMIO:
2981                 ++vcpu->stat.mmio_exits;
2982                 /* fall through */
2983         case EMULATE_FAIL:
2984                 return 0;
2985         default:
2986                 BUG();
2987         }
2988 out:
2989         return r;
2990 }
2991 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2992
2993 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2994 {
2995         vcpu->arch.mmu.invlpg(vcpu, gva);
2996         kvm_mmu_flush_tlb(vcpu);
2997         ++vcpu->stat.invlpg;
2998 }
2999 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
3000
3001 void kvm_enable_tdp(void)
3002 {
3003         tdp_enabled = true;
3004 }
3005 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
3006
3007 void kvm_disable_tdp(void)
3008 {
3009         tdp_enabled = false;
3010 }
3011 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
3012
3013 static void free_mmu_pages(struct kvm_vcpu *vcpu)
3014 {
3015         free_page((unsigned long)vcpu->arch.mmu.pae_root);
3016 }
3017
3018 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
3019 {
3020         struct page *page;
3021         int i;
3022
3023         ASSERT(vcpu);
3024
3025         /*
3026          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
3027          * Therefore we need to allocate shadow page tables in the first
3028          * 4GB of memory, which happens to fit the DMA32 zone.
3029          */
3030         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
3031         if (!page)
3032                 return -ENOMEM;
3033
3034         vcpu->arch.mmu.pae_root = page_address(page);
3035         for (i = 0; i < 4; ++i)
3036                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
3037
3038         return 0;
3039 }
3040
3041 int kvm_mmu_create(struct kvm_vcpu *vcpu)
3042 {
3043         ASSERT(vcpu);
3044         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3045
3046         return alloc_mmu_pages(vcpu);
3047 }
3048
3049 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
3050 {
3051         ASSERT(vcpu);
3052         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
3053
3054         return init_kvm_mmu(vcpu);
3055 }
3056
3057 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
3058 {
3059         ASSERT(vcpu);
3060
3061         destroy_kvm_mmu(vcpu);
3062         free_mmu_pages(vcpu);
3063         mmu_free_memory_caches(vcpu);
3064 }
3065
3066 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
3067 {
3068         struct kvm_mmu_page *sp;
3069
3070         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
3071                 int i;
3072                 u64 *pt;
3073
3074                 if (!test_bit(slot, sp->slot_bitmap))
3075                         continue;
3076
3077                 pt = sp->spt;
3078                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
3079                         /* avoid RMW */
3080                         if (is_writable_pte(pt[i]))
3081                                 pt[i] &= ~PT_WRITABLE_MASK;
3082         }
3083         kvm_flush_remote_tlbs(kvm);
3084 }
3085
3086 void kvm_mmu_zap_all(struct kvm *kvm)
3087 {
3088         struct kvm_mmu_page *sp, *node;
3089         LIST_HEAD(invalid_list);
3090
3091         spin_lock(&kvm->mmu_lock);
3092 restart:
3093         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
3094                 if (kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list))
3095                         goto restart;
3096
3097         kvm_mmu_commit_zap_page(kvm, &invalid_list);
3098         spin_unlock(&kvm->mmu_lock);
3099 }
3100
3101 static int kvm_mmu_remove_some_alloc_mmu_pages(struct kvm *kvm,
3102                                                struct list_head *invalid_list)
3103 {
3104         struct kvm_mmu_page *page;
3105
3106         page = container_of(kvm->arch.active_mmu_pages.prev,
3107                             struct kvm_mmu_page, link);
3108         return kvm_mmu_prepare_zap_page(kvm, page, invalid_list);
3109 }
3110
3111 static int mmu_shrink(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
3112 {
3113         struct kvm *kvm;
3114         struct kvm *kvm_freed = NULL;
3115         int cache_count = 0;
3116
3117         spin_lock(&kvm_lock);
3118
3119         list_for_each_entry(kvm, &vm_list, vm_list) {
3120                 int npages, idx, freed_pages;
3121                 LIST_HEAD(invalid_list);
3122
3123                 idx = srcu_read_lock(&kvm->srcu);
3124                 spin_lock(&kvm->mmu_lock);
3125                 npages = kvm->arch.n_alloc_mmu_pages -
3126                          kvm->arch.n_free_mmu_pages;
3127                 cache_count += npages;
3128                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
3129                         freed_pages = kvm_mmu_remove_some_alloc_mmu_pages(kvm,
3130                                                           &invalid_list);
3131                         cache_count -= freed_pages;
3132                         kvm_freed = kvm;
3133                 }
3134                 nr_to_scan--;
3135
3136                 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3137                 spin_unlock(&kvm->mmu_lock);
3138                 srcu_read_unlock(&kvm->srcu, idx);
3139         }
3140         if (kvm_freed)
3141                 list_move_tail(&kvm_freed->vm_list, &vm_list);
3142
3143         spin_unlock(&kvm_lock);
3144
3145         return cache_count;
3146 }
3147
3148 static struct shrinker mmu_shrinker = {
3149         .shrink = mmu_shrink,
3150         .seeks = DEFAULT_SEEKS * 10,
3151 };
3152
3153 static void mmu_destroy_caches(void)
3154 {
3155         if (pte_chain_cache)
3156                 kmem_cache_destroy(pte_chain_cache);
3157         if (rmap_desc_cache)
3158                 kmem_cache_destroy(rmap_desc_cache);
3159         if (mmu_page_header_cache)
3160                 kmem_cache_destroy(mmu_page_header_cache);
3161 }
3162
3163 void kvm_mmu_module_exit(void)
3164 {
3165         mmu_destroy_caches();
3166         unregister_shrinker(&mmu_shrinker);
3167 }
3168
3169 int kvm_mmu_module_init(void)
3170 {
3171         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
3172                                             sizeof(struct kvm_pte_chain),
3173                                             0, 0, NULL);
3174         if (!pte_chain_cache)
3175                 goto nomem;
3176         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
3177                                             sizeof(struct kvm_rmap_desc),
3178                                             0, 0, NULL);
3179         if (!rmap_desc_cache)
3180                 goto nomem;
3181
3182         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
3183                                                   sizeof(struct kvm_mmu_page),
3184                                                   0, 0, NULL);
3185         if (!mmu_page_header_cache)
3186                 goto nomem;
3187
3188         register_shrinker(&mmu_shrinker);
3189
3190         return 0;
3191
3192 nomem:
3193         mmu_destroy_caches();
3194         return -ENOMEM;
3195 }
3196
3197 /*
3198  * Caculate mmu pages needed for kvm.
3199  */
3200 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
3201 {
3202         int i;
3203         unsigned int nr_mmu_pages;
3204         unsigned int  nr_pages = 0;
3205         struct kvm_memslots *slots;
3206
3207         slots = kvm_memslots(kvm);
3208
3209         for (i = 0; i < slots->nmemslots; i++)
3210                 nr_pages += slots->memslots[i].npages;
3211
3212         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
3213         nr_mmu_pages = max(nr_mmu_pages,
3214                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
3215
3216         return nr_mmu_pages;
3217 }
3218
3219 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3220                                 unsigned len)
3221 {
3222         if (len > buffer->len)
3223                 return NULL;
3224         return buffer->ptr;
3225 }
3226
3227 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
3228                                 unsigned len)
3229 {
3230         void *ret;
3231
3232         ret = pv_mmu_peek_buffer(buffer, len);
3233         if (!ret)
3234                 return ret;
3235         buffer->ptr += len;
3236         buffer->len -= len;
3237         buffer->processed += len;
3238         return ret;
3239 }
3240
3241 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
3242                              gpa_t addr, gpa_t value)
3243 {
3244         int bytes = 8;
3245         int r;
3246
3247         if (!is_long_mode(vcpu) && !is_pae(vcpu))
3248                 bytes = 4;
3249
3250         r = mmu_topup_memory_caches(vcpu);
3251         if (r)
3252                 return r;
3253
3254         if (!emulator_write_phys(vcpu, addr, &value, bytes))
3255                 return -EFAULT;
3256
3257         return 1;
3258 }
3259
3260 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
3261 {
3262         (void)kvm_set_cr3(vcpu, vcpu->arch.cr3);
3263         return 1;
3264 }
3265
3266 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
3267 {
3268         spin_lock(&vcpu->kvm->mmu_lock);
3269         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
3270         spin_unlock(&vcpu->kvm->mmu_lock);
3271         return 1;
3272 }
3273
3274 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
3275                              struct kvm_pv_mmu_op_buffer *buffer)
3276 {
3277         struct kvm_mmu_op_header *header;
3278
3279         header = pv_mmu_peek_buffer(buffer, sizeof *header);
3280         if (!header)
3281                 return 0;
3282         switch (header->op) {
3283         case KVM_MMU_OP_WRITE_PTE: {
3284                 struct kvm_mmu_op_write_pte *wpte;
3285
3286                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
3287                 if (!wpte)
3288                         return 0;
3289                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
3290                                         wpte->pte_val);
3291         }
3292         case KVM_MMU_OP_FLUSH_TLB: {
3293                 struct kvm_mmu_op_flush_tlb *ftlb;
3294
3295                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
3296                 if (!ftlb)
3297                         return 0;
3298                 return kvm_pv_mmu_flush_tlb(vcpu);
3299         }
3300         case KVM_MMU_OP_RELEASE_PT: {
3301                 struct kvm_mmu_op_release_pt *rpt;
3302
3303                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3304                 if (!rpt)
3305                         return 0;
3306                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3307         }
3308         default: return 0;
3309         }
3310 }
3311
3312 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3313                   gpa_t addr, unsigned long *ret)
3314 {
3315         int r;
3316         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3317
3318         buffer->ptr = buffer->buf;
3319         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3320         buffer->processed = 0;
3321
3322         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3323         if (r)
3324                 goto out;