arch/x86/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "mmu.h"
  21
  22 #include <linux/kvm_host.h>
  23 #include <linux/types.h>
  24 #include <linux/string.h>
  25 #include <linux/mm.h>
  26 #include <linux/highmem.h>
  27 #include <linux/module.h>
  28 #include <linux/swap.h>
  29 #include <linux/hugetlb.h>
  30 #include <linux/compiler.h>
  31
  32 #include <asm/page.h>
  33 #include <asm/cmpxchg.h>
  34 #include <asm/io.h>
  35 #include <asm/vmx.h>
  36
  37 /*
  38  * When setting this variable to true it enables Two-Dimensional-Paging
  39  * where the hardware walks 2 page tables:
  40  * 1. the guest-virtual to guest-physical
  41  * 2. while doing 1. it walks guest-physical to host-physical
  42  * If the hardware supports that we don't need to do shadow paging.
  43  */
  44 bool tdp_enabled = false;
  45
  46 #undef MMU_DEBUG
  47
  48 #undef AUDIT
  49
  50 #ifdef AUDIT
  51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  52 #else
  53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  54 #endif
  55
  56 #ifdef MMU_DEBUG
  57
  58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  60
  61 #else
  62
  63 #define pgprintk(x...) do { } while (0)
  64 #define rmap_printk(x...) do { } while (0)
  65
  66 #endif
  67
  68 #if defined(MMU_DEBUG) || defined(AUDIT)
  69 static int dbg = 0;
  70 module_param(dbg, bool, 0644);
  71 #endif
  72
  73 static int oos_shadow = 1;
  74 module_param(oos_shadow, bool, 0644);
  75
  76 #ifndef MMU_DEBUG
  77 #define ASSERT(x) do { } while (0)
  78 #else
  79 #define ASSERT(x)                                                       \
  80         if (!(x)) {                                                     \
  81                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  82                        __FILE__, __LINE__, #x);                         \
  83         }
  84 #endif
  85
  86 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  87 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  88
  89 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  90
  91 #define PT64_LEVEL_BITS 9
  92
  93 #define PT64_LEVEL_SHIFT(level) \
  94                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  95
  96 #define PT64_LEVEL_MASK(level) \
  97                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  98
  99 #define PT64_INDEX(address, level)\
 100         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 101
 102
 103 #define PT32_LEVEL_BITS 10
 104
 105 #define PT32_LEVEL_SHIFT(level) \
 106                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 107
 108 #define PT32_LEVEL_MASK(level) \
 109                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 110
 111 #define PT32_INDEX(address, level)\
 112         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 113
 114
 115 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 116 #define PT64_DIR_BASE_ADDR_MASK \
 117         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 118
 119 #define PT32_BASE_ADDR_MASK PAGE_MASK
 120 #define PT32_DIR_BASE_ADDR_MASK \
 121         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 122
 123 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 124                         | PT64_NX_MASK)
 125
 126 #define PFERR_PRESENT_MASK (1U << 0)
 127 #define PFERR_WRITE_MASK (1U << 1)
 128 #define PFERR_USER_MASK (1U << 2)
 129 #define PFERR_FETCH_MASK (1U << 4)
 130
 131 #define PT_DIRECTORY_LEVEL 2
 132 #define PT_PAGE_TABLE_LEVEL 1
 133
 134 #define RMAP_EXT 4
 135
 136 #define ACC_EXEC_MASK    1
 137 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 138 #define ACC_USER_MASK    PT_USER_MASK
 139 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 140
 141 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 142
 143 struct kvm_rmap_desc {
 144         u64 *shadow_ptes[RMAP_EXT];
 145         struct kvm_rmap_desc *more;
 146 };
 147
 148 struct kvm_shadow_walk {
 149         int (*entry)(struct kvm_shadow_walk *walk, struct kvm_vcpu *vcpu,
 150                      u64 addr, u64 *spte, int level);
 151 };
 152
 153 struct kvm_unsync_walk {
 154         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 155 };
 156
 157 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 158
 159 static struct kmem_cache *pte_chain_cache;
 160 static struct kmem_cache *rmap_desc_cache;
 161 static struct kmem_cache *mmu_page_header_cache;
 162
 163 static u64 __read_mostly shadow_trap_nonpresent_pte;
 164 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 165 static u64 __read_mostly shadow_base_present_pte;
 166 static u64 __read_mostly shadow_nx_mask;
 167 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 168 static u64 __read_mostly shadow_user_mask;
 169 static u64 __read_mostly shadow_accessed_mask;
 170 static u64 __read_mostly shadow_dirty_mask;
 171 static u64 __read_mostly shadow_mt_mask;
 172
 173 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 174 {
 175         shadow_trap_nonpresent_pte = trap_pte;
 176         shadow_notrap_nonpresent_pte = notrap_pte;
 177 }
 178 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 179
 180 void kvm_mmu_set_base_ptes(u64 base_pte)
 181 {
 182         shadow_base_present_pte = base_pte;
 183 }
 184 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 185
 186 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 187                 u64 dirty_mask, u64 nx_mask, u64 x_mask, u64 mt_mask)
 188 {
 189         shadow_user_mask = user_mask;
 190         shadow_accessed_mask = accessed_mask;
 191         shadow_dirty_mask = dirty_mask;
 192         shadow_nx_mask = nx_mask;
 193         shadow_x_mask = x_mask;
 194         shadow_mt_mask = mt_mask;
 195 }
 196 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 197
 198 static int is_write_protection(struct kvm_vcpu *vcpu)
 199 {
 200         return vcpu->arch.cr0 & X86_CR0_WP;
 201 }
 202
 203 static int is_cpuid_PSE36(void)
 204 {
 205         return 1;
 206 }
 207
 208 static int is_nx(struct kvm_vcpu *vcpu)
 209 {
 210         return vcpu->arch.shadow_efer & EFER_NX;
 211 }
 212
 213 static int is_present_pte(unsigned long pte)
 214 {
 215         return pte & PT_PRESENT_MASK;
 216 }
 217
 218 static int is_shadow_present_pte(u64 pte)
 219 {
 220         return pte != shadow_trap_nonpresent_pte
 221                 && pte != shadow_notrap_nonpresent_pte;
 222 }
 223
 224 static int is_large_pte(u64 pte)
 225 {
 226         return pte & PT_PAGE_SIZE_MASK;
 227 }
 228
 229 static int is_writeble_pte(unsigned long pte)
 230 {
 231         return pte & PT_WRITABLE_MASK;
 232 }
 233
 234 static int is_dirty_pte(unsigned long pte)
 235 {
 236         return pte & shadow_dirty_mask;
 237 }
 238
 239 static int is_rmap_pte(u64 pte)
 240 {
 241         return is_shadow_present_pte(pte);
 242 }
 243
 244 static pfn_t spte_to_pfn(u64 pte)
 245 {
 246         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 247 }
 248
 249 static gfn_t pse36_gfn_delta(u32 gpte)
 250 {
 251         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 252
 253         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 254 }
 255
 256 static void set_shadow_pte(u64 *sptep, u64 spte)
 257 {
 258 #ifdef CONFIG_X86_64
 259         set_64bit((unsigned long *)sptep, spte);
 260 #else
 261         set_64bit((unsigned long long *)sptep, spte);
 262 #endif
 263 }
 264
 265 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 266                                   struct kmem_cache *base_cache, int min)
 267 {
 268         void *obj;
 269
 270         if (cache->nobjs >= min)
 271                 return 0;
 272         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 273                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 274                 if (!obj)
 275                         return -ENOMEM;
 276                 cache->objects[cache->nobjs++] = obj;
 277         }
 278         return 0;
 279 }
 280
 281 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 282 {
 283         while (mc->nobjs)
 284                 kfree(mc->objects[--mc->nobjs]);
 285 }
 286
 287 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 288                                        int min)
 289 {
 290         struct page *page;
 291
 292         if (cache->nobjs >= min)
 293                 return 0;
 294         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 295                 page = alloc_page(GFP_KERNEL);
 296                 if (!page)
 297                         return -ENOMEM;
 298                 set_page_private(page, 0);
 299                 cache->objects[cache->nobjs++] = page_address(page);
 300         }
 301         return 0;
 302 }
 303
 304 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 305 {
 306         while (mc->nobjs)
 307                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 308 }
 309
 310 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 311 {
 312         int r;
 313
 314         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 315                                    pte_chain_cache, 4);
 316         if (r)
 317                 goto out;
 318         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 319                                    rmap_desc_cache, 4);
 320         if (r)
 321                 goto out;
 322         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 323         if (r)
 324                 goto out;
 325         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 326                                    mmu_page_header_cache, 4);
 327 out:
 328         return r;
 329 }
 330
 331 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 332 {
 333         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 334         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 335         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 336         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 337 }
 338
 339 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 340                                     size_t size)
 341 {
 342         void *p;
 343
 344         BUG_ON(!mc->nobjs);
 345         p = mc->objects[--mc->nobjs];
 346         memset(p, 0, size);
 347         return p;
 348 }
 349
 350 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 351 {
 352         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 353                                       sizeof(struct kvm_pte_chain));
 354 }
 355
 356 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 357 {
 358         kfree(pc);
 359 }
 360
 361 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 362 {
 363         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 364                                       sizeof(struct kvm_rmap_desc));
 365 }
 366
 367 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 368 {
 369         kfree(rd);
 370 }
 371
 372 /*
 373  * Return the pointer to the largepage write count for a given
 374  * gfn, handling slots that are not large page aligned.
 375  */
 376 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 377 {
 378         unsigned long idx;
 379
 380         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 381               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 382         return &slot->lpage_info[idx].write_count;
 383 }
 384
 385 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 386 {
 387         int *write_count;
 388
 389         gfn = unalias_gfn(kvm, gfn);
 390         write_count = slot_largepage_idx(gfn,
 391                                          gfn_to_memslot_unaliased(kvm, gfn));
 392         *write_count += 1;
 393 }
 394
 395 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 396 {
 397         int *write_count;
 398
 399         gfn = unalias_gfn(kvm, gfn);
 400         write_count = slot_largepage_idx(gfn,
 401                                          gfn_to_memslot_unaliased(kvm, gfn));
 402         *write_count -= 1;
 403         WARN_ON(*write_count < 0);
 404 }
 405
 406 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 407 {
 408         struct kvm_memory_slot *slot;
 409         int *largepage_idx;
 410
 411         gfn = unalias_gfn(kvm, gfn);
 412         slot = gfn_to_memslot_unaliased(kvm, gfn);
 413         if (slot) {
 414                 largepage_idx = slot_largepage_idx(gfn, slot);
 415                 return *largepage_idx;
 416         }
 417
 418         return 1;
 419 }
 420
 421 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 422 {
 423         struct vm_area_struct *vma;
 424         unsigned long addr;
 425         int ret = 0;
 426
 427         addr = gfn_to_hva(kvm, gfn);
 428         if (kvm_is_error_hva(addr))
 429                 return ret;
 430
 431         down_read(&current->mm->mmap_sem);
 432         vma = find_vma(current->mm, addr);
 433         if (vma && is_vm_hugetlb_page(vma))
 434                 ret = 1;
 435         up_read(&current->mm->mmap_sem);
 436
 437         return ret;
 438 }
 439
 440 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 441 {
 442         struct kvm_memory_slot *slot;
 443
 444         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 445                 return 0;
 446
 447         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 448                 return 0;
 449
 450         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 451         if (slot && slot->dirty_bitmap)
 452                 return 0;
 453
 454         return 1;
 455 }
 456
 457 /*
 458  * Take gfn and return the reverse mapping to it.
 459  * Note: gfn must be unaliased before this function get called
 460  */
 461
 462 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 463 {
 464         struct kvm_memory_slot *slot;
 465         unsigned long idx;
 466
 467         slot = gfn_to_memslot(kvm, gfn);
 468         if (!lpage)
 469                 return &slot->rmap[gfn - slot->base_gfn];
 470
 471         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 472               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 473
 474         return &slot->lpage_info[idx].rmap_pde;
 475 }
 476
 477 /*
 478  * Reverse mapping data structures:
 479  *
 480  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 481  * that points to page_address(page).
 482  *
 483  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 484  * containing more mappings.
 485  */
 486 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 487 {
 488         struct kvm_mmu_page *sp;
 489         struct kvm_rmap_desc *desc;
 490         unsigned long *rmapp;
 491         int i;
 492
 493         if (!is_rmap_pte(*spte))
 494                 return;
 495         gfn = unalias_gfn(vcpu->kvm, gfn);
 496         sp = page_header(__pa(spte));
 497         sp->gfns[spte - sp->spt] = gfn;
 498         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 499         if (!*rmapp) {
 500                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 501                 *rmapp = (unsigned long)spte;
 502         } else if (!(*rmapp & 1)) {
 503                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 504                 desc = mmu_alloc_rmap_desc(vcpu);
 505                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 506                 desc->shadow_ptes[1] = spte;
 507                 *rmapp = (unsigned long)desc | 1;
 508         } else {
 509                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 510                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 511                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 512                         desc = desc->more;
 513                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 514                         desc->more = mmu_alloc_rmap_desc(vcpu);
 515                         desc = desc->more;
 516                 }
 517                 for (i = 0; desc->shadow_ptes[i]; ++i)
 518                         ;
 519                 desc->shadow_ptes[i] = spte;
 520         }
 521 }
 522
 523 static void rmap_desc_remove_entry(unsigned long *rmapp,
 524                                    struct kvm_rmap_desc *desc,
 525                                    int i,
 526                                    struct kvm_rmap_desc *prev_desc)
 527 {
 528         int j;
 529
 530         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 531                 ;
 532         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 533         desc->shadow_ptes[j] = NULL;
 534         if (j != 0)
 535                 return;
 536         if (!prev_desc && !desc->more)
 537                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 538         else
 539                 if (prev_desc)
 540                         prev_desc->more = desc->more;
 541                 else
 542                         *rmapp = (unsigned long)desc->more | 1;
 543         mmu_free_rmap_desc(desc);
 544 }
 545
 546 static void rmap_remove(struct kvm *kvm, u64 *spte)
 547 {
 548         struct kvm_rmap_desc *desc;
 549         struct kvm_rmap_desc *prev_desc;
 550         struct kvm_mmu_page *sp;
 551         pfn_t pfn;
 552         unsigned long *rmapp;
 553         int i;
 554
 555         if (!is_rmap_pte(*spte))
 556                 return;
 557         sp = page_header(__pa(spte));
 558         pfn = spte_to_pfn(*spte);
 559         if (*spte & shadow_accessed_mask)
 560                 kvm_set_pfn_accessed(pfn);
 561         if (is_writeble_pte(*spte))
 562                 kvm_release_pfn_dirty(pfn);
 563         else
 564                 kvm_release_pfn_clean(pfn);
 565         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 566         if (!*rmapp) {
 567                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 568                 BUG();
 569         } else if (!(*rmapp & 1)) {
 570                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 571                 if ((u64 *)*rmapp != spte) {
 572                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 573                                spte, *spte);
 574                         BUG();
 575                 }
 576                 *rmapp = 0;
 577         } else {
 578                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 579                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 580                 prev_desc = NULL;
 581                 while (desc) {
 582                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 583                                 if (desc->shadow_ptes[i] == spte) {
 584                                         rmap_desc_remove_entry(rmapp,
 585                                                                desc, i,
 586                                                                prev_desc);
 587                                         return;
 588                                 }
 589                         prev_desc = desc;
 590                         desc = desc->more;
 591                 }
 592                 BUG();
 593         }
 594 }
 595
 596 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 597 {
 598         struct kvm_rmap_desc *desc;
 599         struct kvm_rmap_desc *prev_desc;
 600         u64 *prev_spte;
 601         int i;
 602
 603         if (!*rmapp)
 604                 return NULL;
 605         else if (!(*rmapp & 1)) {
 606                 if (!spte)
 607                         return (u64 *)*rmapp;
 608                 return NULL;
 609         }
 610         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 611         prev_desc = NULL;
 612         prev_spte = NULL;
 613         while (desc) {
 614                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 615                         if (prev_spte == spte)
 616                                 return desc->shadow_ptes[i];
 617                         prev_spte = desc->shadow_ptes[i];
 618                 }
 619                 desc = desc->more;
 620         }
 621         return NULL;
 622 }
 623
 624 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 625 {
 626         unsigned long *rmapp;
 627         u64 *spte;
 628         int write_protected = 0;
 629
 630         gfn = unalias_gfn(kvm, gfn);
 631         rmapp = gfn_to_rmap(kvm, gfn, 0);
 632
 633         spte = rmap_next(kvm, rmapp, NULL);
 634         while (spte) {
 635                 BUG_ON(!spte);
 636                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 637                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 638                 if (is_writeble_pte(*spte)) {
 639                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 640                         write_protected = 1;
 641                 }
 642                 spte = rmap_next(kvm, rmapp, spte);
 643         }
 644         if (write_protected) {
 645                 pfn_t pfn;
 646
 647                 spte = rmap_next(kvm, rmapp, NULL);
 648                 pfn = spte_to_pfn(*spte);
 649                 kvm_set_pfn_dirty(pfn);
 650         }
 651
 652         /* check for huge page mappings */
 653         rmapp = gfn_to_rmap(kvm, gfn, 1);
 654         spte = rmap_next(kvm, rmapp, NULL);
 655         while (spte) {
 656                 BUG_ON(!spte);
 657                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 658                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 659                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 660                 if (is_writeble_pte(*spte)) {
 661                         rmap_remove(kvm, spte);
 662                         --kvm->stat.lpages;
 663                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 664                         spte = NULL;
 665                         write_protected = 1;
 666                 }
 667                 spte = rmap_next(kvm, rmapp, spte);
 668         }
 669
 670         if (write_protected)
 671                 kvm_flush_remote_tlbs(kvm);
 672 }
 673
 674 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 675 {
 676         u64 *spte;
 677         int need_tlb_flush = 0;
 678
 679         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 680                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 681                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 682                 rmap_remove(kvm, spte);
 683                 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 684                 need_tlb_flush = 1;
 685         }
 686         return need_tlb_flush;
 687 }
 688
 689 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 690                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 691 {
 692         int i;
 693         int retval = 0;
 694
 695         /*
 696          * If mmap_sem isn't taken, we can look the memslots with only
 697          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 698          */
 699         for (i = 0; i < kvm->nmemslots; i++) {
 700                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 701                 unsigned long start = memslot->userspace_addr;
 702                 unsigned long end;
 703
 704                 /* mmu_lock protects userspace_addr */
 705                 if (!start)
 706                         continue;
 707
 708                 end = start + (memslot->npages << PAGE_SHIFT);
 709                 if (hva >= start && hva < end) {
 710                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 711                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 712                         retval |= handler(kvm,
 713                                           &memslot->lpage_info[
 714                                                   gfn_offset /
 715                                                   KVM_PAGES_PER_HPAGE].rmap_pde);
 716                 }
 717         }
 718
 719         return retval;
 720 }
 721
 722 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 723 {
 724         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 725 }
 726
 727 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 728 {
 729         u64 *spte;
 730         int young = 0;
 731
 732         /* always return old for EPT */
 733         if (!shadow_accessed_mask)
 734                 return 0;
 735
 736         spte = rmap_next(kvm, rmapp, NULL);
 737         while (spte) {
 738                 int _young;
 739                 u64 _spte = *spte;
 740                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 741                 _young = _spte & PT_ACCESSED_MASK;
 742                 if (_young) {
 743                         young = 1;
 744                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 745                 }
 746                 spte = rmap_next(kvm, rmapp, spte);
 747         }
 748         return young;
 749 }
 750
 751 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 752 {
 753         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 754 }
 755
 756 #ifdef MMU_DEBUG
 757 static int is_empty_shadow_page(u64 *spt)
 758 {
 759         u64 *pos;
 760         u64 *end;
 761
 762         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 763                 if (is_shadow_present_pte(*pos)) {
 764                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 765                                pos, *pos);
 766                         return 0;
 767                 }
 768         return 1;
 769 }
 770 #endif
 771
 772 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 773 {
 774         ASSERT(is_empty_shadow_page(sp->spt));
 775         list_del(&sp->link);
 776         __free_page(virt_to_page(sp->spt));
 777         __free_page(virt_to_page(sp->gfns));
 778         kfree(sp);
 779         ++kvm->arch.n_free_mmu_pages;
 780 }
 781
 782 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 783 {
 784         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 785 }
 786
 787 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 788                                                u64 *parent_pte)
 789 {
 790         struct kvm_mmu_page *sp;
 791
 792         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 793         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 794         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 795         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 796         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 797         ASSERT(is_empty_shadow_page(sp->spt));
 798         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 799         sp->multimapped = 0;
 800         sp->parent_pte = parent_pte;
 801         --vcpu->kvm->arch.n_free_mmu_pages;
 802         return sp;
 803 }
 804
 805 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 806                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 807 {
 808         struct kvm_pte_chain *pte_chain;
 809         struct hlist_node *node;
 810         int i;
 811
 812         if (!parent_pte)
 813                 return;
 814         if (!sp->multimapped) {
 815                 u64 *old = sp->parent_pte;
 816
 817                 if (!old) {
 818                         sp->parent_pte = parent_pte;
 819                         return;
 820                 }
 821                 sp->multimapped = 1;
 822                 pte_chain = mmu_alloc_pte_chain(vcpu);
 823                 INIT_HLIST_HEAD(&sp->parent_ptes);
 824                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 825                 pte_chain->parent_ptes[0] = old;
 826         }
 827         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 828                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 829                         continue;
 830                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 831                         if (!pte_chain->parent_ptes[i]) {
 832                                 pte_chain->parent_ptes[i] = parent_pte;
 833                                 return;
 834                         }
 835         }
 836         pte_chain = mmu_alloc_pte_chain(vcpu);
 837         BUG_ON(!pte_chain);
 838         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 839         pte_chain->parent_ptes[0] = parent_pte;
 840 }
 841
 842 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 843                                        u64 *parent_pte)
 844 {
 845         struct kvm_pte_chain *pte_chain;
 846         struct hlist_node *node;
 847         int i;
 848
 849         if (!sp->multimapped) {
 850                 BUG_ON(sp->parent_pte != parent_pte);
 851                 sp->parent_pte = NULL;
 852                 return;
 853         }
 854         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 855                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 856                         if (!pte_chain->parent_ptes[i])
 857                                 break;
 858                         if (pte_chain->parent_ptes[i] != parent_pte)
 859                                 continue;
 860                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 861                                 && pte_chain->parent_ptes[i + 1]) {
 862                                 pte_chain->parent_ptes[i]
 863                                         = pte_chain->parent_ptes[i + 1];
 864                                 ++i;
 865                         }
 866                         pte_chain->parent_ptes[i] = NULL;
 867                         if (i == 0) {
 868                                 hlist_del(&pte_chain->link);
 869                                 mmu_free_pte_chain(pte_chain);
 870                                 if (hlist_empty(&sp->parent_ptes)) {
 871                                         sp->multimapped = 0;
 872                                         sp->parent_pte = NULL;
 873                                 }
 874                         }
 875                         return;
 876                 }
 877         BUG();
 878 }
 879
 880
 881 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 882                             mmu_parent_walk_fn fn)
 883 {
 884         struct kvm_pte_chain *pte_chain;
 885         struct hlist_node *node;
 886         struct kvm_mmu_page *parent_sp;
 887         int i;
 888
 889         if (!sp->multimapped && sp->parent_pte) {
 890                 parent_sp = page_header(__pa(sp->parent_pte));
 891                 fn(vcpu, parent_sp);
 892                 mmu_parent_walk(vcpu, parent_sp, fn);
 893                 return;
 894         }
 895         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 896                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 897                         if (!pte_chain->parent_ptes[i])
 898                                 break;
 899                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 900                         fn(vcpu, parent_sp);
 901                         mmu_parent_walk(vcpu, parent_sp, fn);
 902                 }
 903 }
 904
 905 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
 906 {
 907         unsigned int index;
 908         struct kvm_mmu_page *sp = page_header(__pa(spte));
 909
 910         index = spte - sp->spt;
 911         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
 912                 sp->unsync_children++;
 913         WARN_ON(!sp->unsync_children);
 914 }
 915
 916 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
 917 {
 918         struct kvm_pte_chain *pte_chain;
 919         struct hlist_node *node;
 920         int i;
 921
 922         if (!sp->parent_pte)
 923                 return;
 924
 925         if (!sp->multimapped) {
 926                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
 927                 return;
 928         }
 929
 930         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 931                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 932                         if (!pte_chain->parent_ptes[i])
 933                                 break;
 934                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
 935                 }
 936 }
 937
 938 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
 939 {
 940         kvm_mmu_update_parents_unsync(sp);
 941         return 1;
 942 }
 943
 944 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
 945                                         struct kvm_mmu_page *sp)
 946 {
 947         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
 948         kvm_mmu_update_parents_unsync(sp);
 949 }
 950
 951 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 952                                     struct kvm_mmu_page *sp)
 953 {
 954         int i;
 955
 956         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 957                 sp->spt[i] = shadow_trap_nonpresent_pte;
 958 }
 959
 960 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
 961                                struct kvm_mmu_page *sp)
 962 {
 963         return 1;
 964 }
 965
 966 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
 967 {
 968 }
 969
 970 #define KVM_PAGE_ARRAY_NR 16
 971
 972 struct kvm_mmu_pages {
 973         struct mmu_page_and_offset {
 974                 struct kvm_mmu_page *sp;
 975                 unsigned int idx;
 976         } page[KVM_PAGE_ARRAY_NR];
 977         unsigned int nr;
 978 };
 979
 980 #define for_each_unsync_children(bitmap, idx)           \
 981         for (idx = find_first_bit(bitmap, 512);         \
 982              idx < 512;                                 \
 983              idx = find_next_bit(bitmap, 512, idx+1))
 984
 985 int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
 986                    int idx)
 987 {
 988         int i;
 989
 990         if (sp->unsync)
 991                 for (i=0; i < pvec->nr; i++)
 992                         if (pvec->page[i].sp == sp)
 993                                 return 0;
 994
 995         pvec->page[pvec->nr].sp = sp;
 996         pvec->page[pvec->nr].idx = idx;
 997         pvec->nr++;
 998         return (pvec->nr == KVM_PAGE_ARRAY_NR);
 999 }
1000
1001 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1002                            struct kvm_mmu_pages *pvec)
1003 {
1004         int i, ret, nr_unsync_leaf = 0;
1005
1006         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1007                 u64 ent = sp->spt[i];
1008
1009                 if (is_shadow_present_pte(ent)) {
1010                         struct kvm_mmu_page *child;
1011                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1012
1013                         if (child->unsync_children) {
1014                                 if (mmu_pages_add(pvec, child, i))
1015                                         return -ENOSPC;
1016
1017                                 ret = __mmu_unsync_walk(child, pvec);
1018                                 if (!ret)
1019                                         __clear_bit(i, sp->unsync_child_bitmap);
1020                                 else if (ret > 0)
1021                                         nr_unsync_leaf += ret;
1022                                 else
1023                                         return ret;
1024                         }
1025
1026                         if (child->unsync) {
1027                                 nr_unsync_leaf++;
1028                                 if (mmu_pages_add(pvec, child, i))
1029                                         return -ENOSPC;
1030                         }
1031                 }
1032         }
1033
1034         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1035                 sp->unsync_children = 0;
1036
1037         return nr_unsync_leaf;
1038 }
1039
1040 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1041                            struct kvm_mmu_pages *pvec)
1042 {
1043         if (!sp->unsync_children)
1044                 return 0;
1045
1046         mmu_pages_add(pvec, sp, 0);
1047         return __mmu_unsync_walk(sp, pvec);
1048 }
1049
1050 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1051 {
1052         unsigned index;
1053         struct hlist_head *bucket;
1054         struct kvm_mmu_page *sp;
1055         struct hlist_node *node;
1056
1057         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1058         index = kvm_page_table_hashfn(gfn);
1059         bucket = &kvm->arch.mmu_page_hash[index];
1060         hlist_for_each_entry(sp, node, bucket, hash_link)
1061                 if (sp->gfn == gfn && !sp->role.metaphysical
1062                     && !sp->role.invalid) {
1063                         pgprintk("%s: found role %x\n",
1064                                  __func__, sp->role.word);
1065                         return sp;
1066                 }
1067         return NULL;
1068 }
1069
1070 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1071 {
1072         WARN_ON(!sp->unsync);
1073         sp->unsync = 0;
1074         --kvm->stat.mmu_unsync;
1075 }
1076
1077 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1078
1079 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1080 {
1081         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1082                 kvm_mmu_zap_page(vcpu->kvm, sp);
1083                 return 1;
1084         }
1085
1086         rmap_write_protect(vcpu->kvm, sp->gfn);
1087         kvm_unlink_unsync_page(vcpu->kvm, sp);
1088         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1089                 kvm_mmu_zap_page(vcpu->kvm, sp);
1090                 return 1;
1091         }
1092
1093         kvm_mmu_flush_tlb(vcpu);
1094         return 0;
1095 }
1096
1097 struct mmu_page_path {
1098         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1099         unsigned int idx[PT64_ROOT_LEVEL-1];
1100 };
1101
1102 #define for_each_sp(pvec, sp, parents, i)                       \
1103                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1104                         sp = pvec.page[i].sp;                   \
1105                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1106                         i = mmu_pages_next(&pvec, &parents, i))
1107
1108 int mmu_pages_next(struct kvm_mmu_pages *pvec, struct mmu_page_path *parents,
1109                    int i)
1110 {
1111         int n;
1112
1113         for (n = i+1; n < pvec->nr; n++) {
1114                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1115
1116                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1117                         parents->idx[0] = pvec->page[n].idx;
1118                         return n;
1119                 }
1120
1121                 parents->parent[sp->role.level-2] = sp;
1122                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1123         }
1124
1125         return n;
1126 }
1127
1128 void mmu_pages_clear_parents(struct mmu_page_path *parents)
1129 {
1130         struct kvm_mmu_page *sp;
1131         unsigned int level = 0;
1132
1133         do {
1134                 unsigned int idx = parents->idx[level];
1135
1136                 sp = parents->parent[level];
1137                 if (!sp)
1138                         return;
1139
1140                 --sp->unsync_children;
1141                 WARN_ON((int)sp->unsync_children < 0);
1142                 __clear_bit(idx, sp->unsync_child_bitmap);
1143                 level++;
1144         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1145 }
1146
1147 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1148                                struct mmu_page_path *parents,
1149                                struct kvm_mmu_pages *pvec)
1150 {
1151         parents->parent[parent->role.level-1] = NULL;
1152         pvec->nr = 0;
1153 }
1154
1155 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1156                               struct kvm_mmu_page *parent)
1157 {
1158         int i;
1159         struct kvm_mmu_page *sp;
1160         struct mmu_page_path parents;
1161         struct kvm_mmu_pages pages;
1162
1163         kvm_mmu_pages_init(parent, &parents, &pages);
1164         while (mmu_unsync_walk(parent, &pages)) {
1165                 for_each_sp(pages, sp, parents, i) {
1166                         kvm_sync_page(vcpu, sp);
1167                         mmu_pages_clear_parents(&parents);
1168                 }
1169                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1170                 kvm_mmu_pages_init(parent, &parents, &pages);
1171         }
1172 }
1173
1174 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1175                                              gfn_t gfn,
1176                                              gva_t gaddr,
1177                                              unsigned level,
1178                                              int metaphysical,
1179                                              unsigned access,
1180                                              u64 *parent_pte)
1181 {
1182         union kvm_mmu_page_role role;
1183         unsigned index;
1184         unsigned quadrant;
1185         struct hlist_head *bucket;
1186         struct kvm_mmu_page *sp;
1187         struct hlist_node *node, *tmp;
1188
1189         role.word = 0;
1190         role.glevels = vcpu->arch.mmu.root_level;
1191         role.level = level;
1192         role.metaphysical = metaphysical;
1193         role.access = access;
1194         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1195                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1196                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1197                 role.quadrant = quadrant;
1198         }
1199         pgprintk("%s: looking gfn %lx role %x\n", __func__,
1200                  gfn, role.word);
1201         index = kvm_page_table_hashfn(gfn);
1202         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1203         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1204                 if (sp->gfn == gfn) {
1205                         if (sp->unsync)
1206                                 if (kvm_sync_page(vcpu, sp))
1207                                         continue;
1208
1209                         if (sp->role.word != role.word)
1210                                 continue;
1211
1212                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1213                         if (sp->unsync_children) {
1214                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1215                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1216                         }
1217                         pgprintk("%s: found\n", __func__);
1218                         return sp;
1219                 }
1220         ++vcpu->kvm->stat.mmu_cache_miss;
1221         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1222         if (!sp)
1223                 return sp;
1224         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1225         sp->gfn = gfn;
1226         sp->role = role;
1227         hlist_add_head(&sp->hash_link, bucket);
1228         if (!metaphysical) {
1229                 rmap_write_protect(vcpu->kvm, gfn);
1230                 account_shadowed(vcpu->kvm, gfn);
1231         }
1232         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1233                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1234         else
1235                 nonpaging_prefetch_page(vcpu, sp);
1236         return sp;
1237 }
1238
1239 static int walk_shadow(struct kvm_shadow_walk *walker,
1240                        struct kvm_vcpu *vcpu, u64 addr)
1241 {
1242         hpa_t shadow_addr;
1243         int level;
1244         int r;
1245         u64 *sptep;
1246         unsigned index;
1247
1248         shadow_addr = vcpu->arch.mmu.root_hpa;
1249         level = vcpu->arch.mmu.shadow_root_level;
1250         if (level == PT32E_ROOT_LEVEL) {
1251                 shadow_addr = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1252                 shadow_addr &= PT64_BASE_ADDR_MASK;
1253                 --level;
1254         }
1255
1256         while (level >= PT_PAGE_TABLE_LEVEL) {
1257                 index = SHADOW_PT_INDEX(addr, level);
1258                 sptep = ((u64 *)__va(shadow_addr)) + index;
1259                 r = walker->entry(walker, vcpu, addr, sptep, level);
1260                 if (r)
1261                         return r;
1262                 shadow_addr = *sptep & PT64_BASE_ADDR_MASK;
1263                 --level;
1264         }
1265         return 0;
1266 }
1267
1268 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1269                                          struct kvm_mmu_page *sp)
1270 {
1271         unsigned i;
1272         u64 *pt;
1273         u64 ent;
1274
1275         pt = sp->spt;
1276
1277         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1278                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1279                         if (is_shadow_present_pte(pt[i]))
1280                                 rmap_remove(kvm, &pt[i]);
1281                         pt[i] = shadow_trap_nonpresent_pte;
1282                 }
1283                 return;
1284         }
1285
1286         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1287                 ent = pt[i];
1288
1289                 if (is_shadow_present_pte(ent)) {
1290                         if (!is_large_pte(ent)) {
1291                                 ent &= PT64_BASE_ADDR_MASK;
1292                                 mmu_page_remove_parent_pte(page_header(ent),
1293                                                            &pt[i]);
1294                         } else {
1295                                 --kvm->stat.lpages;
1296                                 rmap_remove(kvm, &pt[i]);
1297                         }
1298                 }
1299                 pt[i] = shadow_trap_nonpresent_pte;
1300         }
1301 }
1302
1303 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1304 {
1305         mmu_page_remove_parent_pte(sp, parent_pte);
1306 }
1307
1308 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1309 {
1310         int i;
1311
1312         for (i = 0; i < KVM_MAX_VCPUS; ++i)
1313                 if (kvm->vcpus[i])
1314                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
1315 }
1316
1317 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1318 {
1319         u64 *parent_pte;
1320
1321         while (sp->multimapped || sp->parent_pte) {
1322                 if (!sp->multimapped)
1323                         parent_pte = sp->parent_pte;
1324                 else {
1325                         struct kvm_pte_chain *chain;
1326
1327                         chain = container_of(sp->parent_ptes.first,
1328                                              struct kvm_pte_chain, link);
1329                         parent_pte = chain->parent_ptes[0];
1330                 }
1331                 BUG_ON(!parent_pte);
1332                 kvm_mmu_put_page(sp, parent_pte);
1333                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1334         }
1335 }
1336
1337 static int mmu_zap_unsync_children(struct kvm *kvm,
1338                                    struct kvm_mmu_page *parent)
1339 {
1340         int i, zapped = 0;
1341         struct mmu_page_path parents;
1342         struct kvm_mmu_pages pages;
1343
1344         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1345                 return 0;
1346
1347         kvm_mmu_pages_init(parent, &parents, &pages);
1348         while (mmu_unsync_walk(parent, &pages)) {
1349                 struct kvm_mmu_page *sp;
1350
1351                 for_each_sp(pages, sp, parents, i) {
1352                         kvm_mmu_zap_page(kvm, sp);
1353                         mmu_pages_clear_parents(&parents);
1354                 }
1355                 zapped += pages.nr;
1356                 kvm_mmu_pages_init(parent, &parents, &pages);
1357         }
1358
1359         return zapped;
1360 }
1361
1362 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1363 {
1364         int ret;
1365         ++kvm->stat.mmu_shadow_zapped;
1366         ret = mmu_zap_unsync_children(kvm, sp);
1367         kvm_mmu_page_unlink_children(kvm, sp);
1368         kvm_mmu_unlink_parents(kvm, sp);
1369         kvm_flush_remote_tlbs(kvm);
1370         if (!sp->role.invalid && !sp->role.metaphysical)
1371                 unaccount_shadowed(kvm, sp->gfn);
1372         if (sp->unsync)
1373                 kvm_unlink_unsync_page(kvm, sp);
1374         if (!sp->root_count) {
1375                 hlist_del(&sp->hash_link);
1376                 kvm_mmu_free_page(kvm, sp);
1377         } else {
1378                 sp->role.invalid = 1;
1379                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1380                 kvm_reload_remote_mmus(kvm);
1381         }
1382         kvm_mmu_reset_last_pte_updated(kvm);
1383         return ret;
1384 }
1385
1386 /*
1387  * Changing the number of mmu pages allocated to the vm
1388  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1389  */
1390 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1391 {
1392         /*
1393          * If we set the number of mmu pages to be smaller be than the
1394          * number of actived pages , we must to free some mmu pages before we
1395          * change the value
1396          */
1397
1398         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1399             kvm_nr_mmu_pages) {
1400                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1401                                        - kvm->arch.n_free_mmu_pages;
1402
1403                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1404                         struct kvm_mmu_page *page;
1405
1406                         page = container_of(kvm->arch.active_mmu_pages.prev,
1407                                             struct kvm_mmu_page, link);
1408                         kvm_mmu_zap_page(kvm, page);
1409                         n_used_mmu_pages--;
1410                 }
1411                 kvm->arch.n_free_mmu_pages = 0;
1412         }
1413         else
1414                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1415                                          - kvm->arch.n_alloc_mmu_pages;
1416
1417         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1418 }
1419
1420 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1421 {
1422         unsigned index;
1423         struct hlist_head *bucket;
1424         struct kvm_mmu_page *sp;
1425         struct hlist_node *node, *n;
1426         int r;
1427
1428         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1429         r = 0;
1430         index = kvm_page_table_hashfn(gfn);
1431         bucket = &kvm->arch.mmu_page_hash[index];
1432         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1433                 if (sp->gfn == gfn && !sp->role.metaphysical) {
1434                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1435                                  sp->role.word);
1436                         r = 1;
1437                         if (kvm_mmu_zap_page(kvm, sp))
1438                                 n = bucket->first;
1439                 }
1440         return r;
1441 }
1442
1443 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1444 {
1445         struct kvm_mmu_page *sp;
1446
1447         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1448                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1449                 kvm_mmu_zap_page(kvm, sp);
1450         }
1451 }
1452
1453 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1454 {
1455         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1456         struct kvm_mmu_page *sp = page_header(__pa(pte));
1457
1458         __set_bit(slot, sp->slot_bitmap);
1459 }
1460
1461 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1462 {
1463         int i;
1464         u64 *pt = sp->spt;
1465
1466         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1467                 return;
1468
1469         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1470                 if (pt[i] == shadow_notrap_nonpresent_pte)
1471                         set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1472         }
1473 }
1474
1475 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1476 {
1477         struct page *page;
1478
1479         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1480
1481         if (gpa == UNMAPPED_GVA)
1482                 return NULL;
1483
1484         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1485
1486         return page;
1487 }
1488
1489 /*
1490  * The function is based on mtrr_type_lookup() in
1491  * arch/x86/kernel/cpu/mtrr/generic.c
1492  */
1493 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1494                          u64 start, u64 end)
1495 {
1496         int i;
1497         u64 base, mask;
1498         u8 prev_match, curr_match;
1499         int num_var_ranges = KVM_NR_VAR_MTRR;
1500
1501         if (!mtrr_state->enabled)
1502                 return 0xFF;
1503
1504         /* Make end inclusive end, instead of exclusive */
1505         end--;
1506
1507         /* Look in fixed ranges. Just return the type as per start */
1508         if (mtrr_state->have_fixed && (start < 0x100000)) {
1509                 int idx;
1510
1511                 if (start < 0x80000) {
1512                         idx = 0;
1513                         idx += (start >> 16);
1514                         return mtrr_state->fixed_ranges[idx];
1515                 } else if (start < 0xC0000) {
1516                         idx = 1 * 8;
1517                         idx += ((start - 0x80000) >> 14);
1518                         return mtrr_state->fixed_ranges[idx];
1519                 } else if (start < 0x1000000) {
1520                         idx = 3 * 8;
1521                         idx += ((start - 0xC0000) >> 12);
1522                         return mtrr_state->fixed_ranges[idx];
1523                 }
1524         }
1525
1526         /*
1527          * Look in variable ranges
1528          * Look of multiple ranges matching this address and pick type
1529          * as per MTRR precedence
1530          */
1531         if (!(mtrr_state->enabled & 2))
1532                 return mtrr_state->def_type;
1533
1534         prev_match = 0xFF;
1535         for (i = 0; i < num_var_ranges; ++i) {
1536                 unsigned short start_state, end_state;
1537
1538                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1539                         continue;
1540
1541                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1542                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1543                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1544                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1545
1546                 start_state = ((start & mask) == (base & mask));
1547                 end_state = ((end & mask) == (base & mask));
1548                 if (start_state != end_state)
1549                         return 0xFE;
1550
1551                 if ((start & mask) != (base & mask))
1552                         continue;
1553
1554                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1555                 if (prev_match == 0xFF) {
1556                         prev_match = curr_match;
1557                         continue;
1558                 }
1559
1560                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1561                     curr_match == MTRR_TYPE_UNCACHABLE)
1562                         return MTRR_TYPE_UNCACHABLE;
1563
1564                 if ((prev_match == MTRR_TYPE_WRBACK &&
1565                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1566                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1567                      curr_match == MTRR_TYPE_WRBACK)) {
1568                         prev_match = MTRR_TYPE_WRTHROUGH;
1569                         curr_match = MTRR_TYPE_WRTHROUGH;
1570                 }
1571
1572                 if (prev_match != curr_match)
1573                         return MTRR_TYPE_UNCACHABLE;
1574         }
1575
1576         if (prev_match != 0xFF)
1577                 return prev_match;
1578
1579         return mtrr_state->def_type;
1580 }
1581
1582 static u8 get_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1583 {
1584         u8 mtrr;
1585
1586         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1587                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1588         if (mtrr == 0xfe || mtrr == 0xff)
1589                 mtrr = MTRR_TYPE_WRBACK;
1590         return mtrr;
1591 }
1592
1593 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1594 {
1595         unsigned index;
1596         struct hlist_head *bucket;
1597         struct kvm_mmu_page *s;
1598         struct hlist_node *node, *n;
1599
1600         index = kvm_page_table_hashfn(sp->gfn);
1601         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1602         /* don't unsync if pagetable is shadowed with multiple roles */
1603         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1604                 if (s->gfn != sp->gfn || s->role.metaphysical)
1605                         continue;
1606                 if (s->role.word != sp->role.word)
1607                         return 1;
1608         }
1609         kvm_mmu_mark_parents_unsync(vcpu, sp);
1610         ++vcpu->kvm->stat.mmu_unsync;
1611         sp->unsync = 1;
1612         mmu_convert_notrap(sp);
1613         return 0;
1614 }
1615
1616 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1617                                   bool can_unsync)
1618 {
1619         struct kvm_mmu_page *shadow;
1620
1621         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1622         if (shadow) {
1623                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1624                         return 1;
1625                 if (shadow->unsync)
1626                         return 0;
1627                 if (can_unsync && oos_shadow)
1628                         return kvm_unsync_page(vcpu, shadow);
1629                 return 1;
1630         }
1631         return 0;
1632 }
1633
1634 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1635                     unsigned pte_access, int user_fault,
1636                     int write_fault, int dirty, int largepage,
1637                     gfn_t gfn, pfn_t pfn, bool speculative,
1638                     bool can_unsync)
1639 {
1640         u64 spte;
1641         int ret = 0;
1642         u64 mt_mask = shadow_mt_mask;
1643
1644         /*
1645          * We don't set the accessed bit, since we sometimes want to see
1646          * whether the guest actually used the pte (in order to detect
1647          * demand paging).
1648          */
1649         spte = shadow_base_present_pte | shadow_dirty_mask;
1650         if (!speculative)
1651                 spte |= shadow_accessed_mask;
1652         if (!dirty)
1653                 pte_access &= ~ACC_WRITE_MASK;
1654         if (pte_access & ACC_EXEC_MASK)
1655                 spte |= shadow_x_mask;
1656         else
1657                 spte |= shadow_nx_mask;
1658         if (pte_access & ACC_USER_MASK)
1659                 spte |= shadow_user_mask;
1660         if (largepage)
1661                 spte |= PT_PAGE_SIZE_MASK;
1662         if (mt_mask) {
1663                 mt_mask = get_memory_type(vcpu, gfn) <<
1664                           kvm_x86_ops->get_mt_mask_shift();
1665                 spte |= mt_mask;
1666         }
1667
1668         spte |= (u64)pfn << PAGE_SHIFT;
1669
1670         if ((pte_access & ACC_WRITE_MASK)
1671             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1672
1673                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1674                         ret = 1;
1675                         spte = shadow_trap_nonpresent_pte;
1676                         goto set_pte;
1677                 }
1678
1679                 spte |= PT_WRITABLE_MASK;
1680
1681                 /*
1682                  * Optimization: for pte sync, if spte was writable the hash
1683                  * lookup is unnecessary (and expensive). Write protection
1684                  * is responsibility of mmu_get_page / kvm_sync_page.
1685                  * Same reasoning can be applied to dirty page accounting.
1686                  */
1687                 if (!can_unsync && is_writeble_pte(*shadow_pte))
1688                         goto set_pte;
1689
1690                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1691                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1692                                  __func__, gfn);
1693                         ret = 1;
1694                         pte_access &= ~ACC_WRITE_MASK;
1695                         if (is_writeble_pte(spte))
1696                                 spte &= ~PT_WRITABLE_MASK;
1697                 }
1698         }
1699
1700         if (pte_access & ACC_WRITE_MASK)
1701                 mark_page_dirty(vcpu->kvm, gfn);
1702
1703 set_pte:
1704         set_shadow_pte(shadow_pte, spte);
1705         return ret;
1706 }
1707
1708 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1709                          unsigned pt_access, unsigned pte_access,
1710                          int user_fault, int write_fault, int dirty,
1711                          int *ptwrite, int largepage, gfn_t gfn,
1712                          pfn_t pfn, bool speculative)
1713 {
1714         int was_rmapped = 0;
1715         int was_writeble = is_writeble_pte(*shadow_pte);
1716
1717         pgprintk("%s: spte %llx access %x write_fault %d"
1718                  " user_fault %d gfn %lx\n",
1719                  __func__, *shadow_pte, pt_access,
1720                  write_fault, user_fault, gfn);
1721
1722         if (is_rmap_pte(*shadow_pte)) {
1723                 /*
1724                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1725                  * the parent of the now unreachable PTE.
1726                  */
1727                 if (largepage && !is_large_pte(*shadow_pte)) {
1728                         struct kvm_mmu_page *child;
1729                         u64 pte = *shadow_pte;
1730
1731                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1732                         mmu_page_remove_parent_pte(child, shadow_pte);
1733                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1734                         pgprintk("hfn old %lx new %lx\n",
1735                                  spte_to_pfn(*shadow_pte), pfn);
1736                         rmap_remove(vcpu->kvm, shadow_pte);
1737                 } else {
1738                         if (largepage)
1739                                 was_rmapped = is_large_pte(*shadow_pte);
1740                         else
1741                                 was_rmapped = 1;
1742                 }
1743         }
1744         if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1745                       dirty, largepage, gfn, pfn, speculative, true)) {
1746                 if (write_fault)
1747                         *ptwrite = 1;
1748                 kvm_x86_ops->tlb_flush(vcpu);
1749         }
1750
1751         pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1752         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1753                  is_large_pte(*shadow_pte)? "2MB" : "4kB",
1754                  is_present_pte(*shadow_pte)?"RW":"R", gfn,
1755                  *shadow_pte, shadow_pte);
1756         if (!was_rmapped && is_large_pte(*shadow_pte))
1757                 ++vcpu->kvm->stat.lpages;
1758
1759         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1760         if (!was_rmapped) {
1761                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1762                 if (!is_rmap_pte(*shadow_pte))
1763                         kvm_release_pfn_clean(pfn);
1764         } else {
1765                 if (was_writeble)
1766                         kvm_release_pfn_dirty(pfn);
1767                 else
1768                         kvm_release_pfn_clean(pfn);
1769         }
1770         if (speculative) {
1771                 vcpu->arch.last_pte_updated = shadow_pte;
1772                 vcpu->arch.last_pte_gfn = gfn;
1773         }
1774 }
1775
1776 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1777 {
1778 }
1779
1780 struct direct_shadow_walk {
1781         struct kvm_shadow_walk walker;
1782         pfn_t pfn;
1783         int write;
1784         int largepage;
1785         int pt_write;
1786 };
1787
1788 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1789                             struct kvm_vcpu *vcpu,
1790                             u64 addr, u64 *sptep, int level)
1791 {
1792         struct direct_shadow_walk *walk =
1793                 container_of(_walk, struct direct_shadow_walk, walker);
1794         struct kvm_mmu_page *sp;
1795         gfn_t pseudo_gfn;
1796         gfn_t gfn = addr >> PAGE_SHIFT;
1797
1798         if (level == PT_PAGE_TABLE_LEVEL
1799             || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1800                 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1801                              0, walk->write, 1, &walk->pt_write,
1802                              walk->largepage, gfn, walk->pfn, false);
1803                 ++vcpu->stat.pf_fixed;
1804                 return 1;
1805         }
1806
1807         if (*sptep == shadow_trap_nonpresent_pte) {
1808                 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1809                 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1810                                       1, ACC_ALL, sptep);
1811                 if (!sp) {
1812                         pgprintk("nonpaging_map: ENOMEM\n");
1813                         kvm_release_pfn_clean(walk->pfn);
1814                         return -ENOMEM;
1815                 }
1816
1817                 set_shadow_pte(sptep,
1818                                __pa(sp->spt)
1819                                | PT_PRESENT_MASK | PT_WRITABLE_MASK
1820                                | shadow_user_mask | shadow_x_mask);
1821         }
1822         return 0;
1823 }
1824
1825 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1826                         int largepage, gfn_t gfn, pfn_t pfn)
1827 {
1828         int r;
1829         struct direct_shadow_walk walker = {
1830                 .walker = { .entry = direct_map_entry, },
1831                 .pfn = pfn,
1832                 .largepage = largepage,
1833                 .write = write,
1834                 .pt_write = 0,
1835         };
1836
1837         r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1838         if (r < 0)
1839                 return r;
1840         return walker.pt_write;
1841 }
1842
1843 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1844 {
1845         int r;
1846         int largepage = 0;
1847         pfn_t pfn;
1848         unsigned long mmu_seq;
1849
1850         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1851                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1852                 largepage = 1;
1853         }
1854
1855         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1856         smp_rmb();
1857         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1858
1859         /* mmio */
1860         if (is_error_pfn(pfn)) {
1861                 kvm_release_pfn_clean(pfn);
1862                 return 1;
1863         }
1864
1865         spin_lock(&vcpu->kvm->mmu_lock);
1866         if (mmu_notifier_retry(vcpu, mmu_seq))
1867                 goto out_unlock;
1868         kvm_mmu_free_some_pages(vcpu);
1869         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1870         spin_unlock(&vcpu->kvm->mmu_lock);
1871
1872
1873         return r;
1874
1875 out_unlock:
1876         spin_unlock(&vcpu->kvm->mmu_lock);
1877         kvm_release_pfn_clean(pfn);
1878         return 0;
1879 }
1880
1881
1882 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1883 {
1884         int i;
1885         struct kvm_mmu_page *sp;
1886
1887         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1888                 return;
1889         spin_lock(&vcpu->kvm->mmu_lock);
1890         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1891                 hpa_t root = vcpu->arch.mmu.root_hpa;
1892
1893                 sp = page_header(root);
1894                 --sp->root_count;
1895                 if (!sp->root_count && sp->role.invalid)
1896                         kvm_mmu_zap_page(vcpu->kvm, sp);
1897                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1898                 spin_unlock(&vcpu->kvm->mmu_lock);
1899                 return;
1900         }
1901         for (i = 0; i < 4; ++i) {
1902                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1903
1904                 if (root) {
1905                         root &= PT64_BASE_ADDR_MASK;
1906                         sp = page_header(root);
1907                         --sp->root_count;
1908                         if (!sp->root_count && sp->role.invalid)
1909                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1910                 }
1911                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1912         }
1913         spin_unlock(&vcpu->kvm->mmu_lock);
1914         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1915 }
1916
1917 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1918 {
1919         int i;
1920         gfn_t root_gfn;
1921         struct kvm_mmu_page *sp;
1922         int metaphysical = 0;
1923
1924         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1925
1926         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1927                 hpa_t root = vcpu->arch.mmu.root_hpa;
1928
1929                 ASSERT(!VALID_PAGE(root));
1930                 if (tdp_enabled)
1931                         metaphysical = 1;
1932                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1933                                       PT64_ROOT_LEVEL, metaphysical,
1934                                       ACC_ALL, NULL);
1935                 root = __pa(sp->spt);
1936                 ++sp->root_count;
1937                 vcpu->arch.mmu.root_hpa = root;
1938                 return;
1939         }
1940         metaphysical = !is_paging(vcpu);
1941         if (tdp_enabled)
1942                 metaphysical = 1;
1943         for (i = 0; i < 4; ++i) {
1944                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1945
1946                 ASSERT(!VALID_PAGE(root));
1947                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1948                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1949                                 vcpu->arch.mmu.pae_root[i] = 0;
1950                                 continue;
1951                         }
1952                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1953                 } else if (vcpu->arch.mmu.root_level == 0)
1954                         root_gfn = 0;
1955                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1956                                       PT32_ROOT_LEVEL, metaphysical,
1957                                       ACC_ALL, NULL);
1958                 root = __pa(sp->spt);
1959                 ++sp->root_count;
1960                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1961         }
1962         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1963 }
1964
1965 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
1966 {
1967         int i;
1968         struct kvm_mmu_page *sp;
1969
1970         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1971                 return;
1972         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1973                 hpa_t root = vcpu->arch.mmu.root_hpa;
1974                 sp = page_header(root);
1975                 mmu_sync_children(vcpu, sp);
1976                 return;
1977         }
1978         for (i = 0; i < 4; ++i) {
1979                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1980
1981                 if (root) {
1982                         root &= PT64_BASE_ADDR_MASK;
1983                         sp = page_header(root);
1984                         mmu_sync_children(vcpu, sp);
1985                 }
1986         }
1987 }
1988
1989 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
1990 {
1991         spin_lock(&vcpu->kvm->mmu_lock);
1992         mmu_sync_roots(vcpu);
1993         spin_unlock(&vcpu->kvm->mmu_lock);
1994 }
1995
1996 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1997 {
1998         return vaddr;
1999 }
2000
2001 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2002                                 u32 error_code)
2003 {
2004         gfn_t gfn;
2005         int r;
2006
2007         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2008         r = mmu_topup_memory_caches(vcpu);
2009         if (r)
2010                 return r;
2011
2012         ASSERT(vcpu);
2013         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2014
2015         gfn = gva >> PAGE_SHIFT;
2016
2017         return nonpaging_map(vcpu, gva & PAGE_MASK,
2018                              error_code & PFERR_WRITE_MASK, gfn);
2019 }
2020
2021 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2022                                 u32 error_code)
2023 {
2024         pfn_t pfn;
2025         int r;
2026         int largepage = 0;
2027         gfn_t gfn = gpa >> PAGE_SHIFT;
2028         unsigned long mmu_seq;
2029
2030         ASSERT(vcpu);
2031         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2032
2033         r = mmu_topup_memory_caches(vcpu);
2034         if (r)
2035                 return r;
2036
2037         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
2038                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2039                 largepage = 1;
2040         }
2041         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2042         smp_rmb();
2043         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2044         if (is_error_pfn(pfn)) {
2045                 kvm_release_pfn_clean(pfn);
2046                 return 1;
2047         }
2048         spin_lock(&vcpu->kvm->mmu_lock);
2049         if (mmu_notifier_retry(vcpu, mmu_seq))
2050                 goto out_unlock;
2051         kvm_mmu_free_some_pages(vcpu);
2052         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2053                          largepage, gfn, pfn);
2054         spin_unlock(&vcpu->kvm->mmu_lock);
2055
2056         return r;
2057
2058 out_unlock:
2059         spin_unlock(&vcpu->kvm->mmu_lock);
2060         kvm_release_pfn_clean(pfn);
2061         return 0;
2062 }
2063
2064 static void nonpaging_free(struct kvm_vcpu *vcpu)
2065 {
2066         mmu_free_roots(vcpu);
2067 }
2068
2069 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2070 {
2071         struct kvm_mmu *context = &vcpu->arch.mmu;
2072
2073         context->new_cr3 = nonpaging_new_cr3;
2074         context->page_fault = nonpaging_page_fault;
2075         context->gva_to_gpa = nonpaging_gva_to_gpa;
2076         context->free = nonpaging_free;
2077         context->prefetch_page = nonpaging_prefetch_page;
2078         context->sync_page = nonpaging_sync_page;
2079         context->invlpg = nonpaging_invlpg;
2080         context->root_level = 0;
2081         context->shadow_root_level = PT32E_ROOT_LEVEL;
2082         context->root_hpa = INVALID_PAGE;
2083         return 0;
2084 }
2085
2086 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2087 {
2088         ++vcpu->stat.tlb_flush;
2089         kvm_x86_ops->tlb_flush(vcpu);
2090 }
2091
2092 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2093 {
2094         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2095         mmu_free_roots(vcpu);
2096 }
2097
2098 static void inject_page_fault(struct kvm_vcpu *vcpu,
2099                               u64 addr,
2100                               u32 err_code)
2101 {
2102         kvm_inject_page_fault(vcpu, addr, err_code);
2103 }
2104
2105 static void paging_free(struct kvm_vcpu *vcpu)
2106 {
2107         nonpaging_free(vcpu);
2108 }
2109
2110 #define PTTYPE 64
2111 #include "paging_tmpl.h"
2112 #undef PTTYPE
2113
2114 #define PTTYPE 32
2115 #include "paging_tmpl.h"
2116 #undef PTTYPE
2117
2118 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2119 {
2120         struct kvm_mmu *context = &vcpu->arch.mmu;
2121
2122         ASSERT(is_pae(vcpu));
2123         context->new_cr3 = paging_new_cr3;
2124         context->page_fault = paging64_page_fault;
2125         context->gva_to_gpa = paging64_gva_to_gpa;
2126         context->prefetch_page = paging64_prefetch_page;
2127         context->sync_page = paging64_sync_page;
2128         context->invlpg = paging64_invlpg;
2129         context->free = paging_free;
2130         context->root_level = level;
2131         context->shadow_root_level = level;
2132         context->root_hpa = INVALID_PAGE;
2133         return 0;
2134 }
2135
2136 static int paging64_init_context(struct kvm_vcpu *vcpu)
2137 {
2138         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2139 }
2140
2141 static int paging32_init_context(struct kvm_vcpu *vcpu)
2142 {
2143         struct kvm_mmu *context = &vcpu->arch.mmu;
2144
2145         context->new_cr3 = paging_new_cr3;
2146         context->page_fault = paging32_page_fault;
2147         context->gva_to_gpa = paging32_gva_to_gpa;
2148         context->free = paging_free;
2149         context->prefetch_page = paging32_prefetch_page;
2150         context->sync_page = paging32_sync_page;
2151         context->invlpg = paging32_invlpg;
2152         context->root_level = PT32_ROOT_LEVEL;
2153         context->shadow_root_level = PT32E_ROOT_LEVEL;
2154         context->root_hpa = INVALID_PAGE;
2155         return 0;
2156 }
2157
2158 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2159 {
2160         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2161 }
2162
2163 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2164 {
2165         struct kvm_mmu *context = &vcpu->arch.mmu;
2166
2167         context->new_cr3 = nonpaging_new_cr3;
2168         context->page_fault = tdp_page_fault;
2169         context->free = nonpaging_free;
2170         context->prefetch_page = nonpaging_prefetch_page;
2171         context->sync_page = nonpaging_sync_page;
2172         context->invlpg = nonpaging_invlpg;
2173         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2174         context->root_hpa = INVALID_PAGE;
2175
2176         if (!is_paging(vcpu)) {
2177                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2178                 context->root_level = 0;
2179         } else if (is_long_mode(vcpu)) {
2180                 context->gva_to_gpa = paging64_gva_to_gpa;
2181                 context->root_level = PT64_ROOT_LEVEL;
2182         } else if (is_pae(vcpu)) {
2183                 context->gva_to_gpa = paging64_gva_to_gpa;
2184                 context->root_level = PT32E_ROOT_LEVEL;
2185         } else {
2186                 context->gva_to_gpa = paging32_gva_to_gpa;
2187                 context->root_level = PT32_ROOT_LEVEL;
2188         }
2189
2190         return 0;
2191 }
2192
2193 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2194 {
2195         ASSERT(vcpu);
2196         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2197
2198         if (!is_paging(vcpu))
2199                 return nonpaging_init_context(vcpu);
2200         else if (is_long_mode(vcpu))
2201                 return paging64_init_context(vcpu);
2202         else if (is_pae(vcpu))
2203                 return paging32E_init_context(vcpu);
2204         else
2205                 return paging32_init_context(vcpu);
2206 }
2207
2208 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2209 {
2210         vcpu->arch.update_pte.pfn = bad_pfn;
2211
2212         if (tdp_enabled)
2213                 return init_kvm_tdp_mmu(vcpu);
2214         else
2215                 return init_kvm_softmmu(vcpu);
2216 }
2217
2218 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2219 {
2220         ASSERT(vcpu);
2221         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2222                 vcpu->arch.mmu.free(vcpu);
2223                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2224         }
2225 }
2226
2227 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2228 {
2229         destroy_kvm_mmu(vcpu);
2230         return init_kvm_mmu(vcpu);
2231 }
2232 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2233
2234 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2235 {
2236         int r;
2237
2238         r = mmu_topup_memory_caches(vcpu);
2239         if (r)
2240                 goto out;
2241         spin_lock(&vcpu->kvm->mmu_lock);
2242         kvm_mmu_free_some_pages(vcpu);
2243         mmu_alloc_roots(vcpu);
2244         mmu_sync_roots(vcpu);
2245         spin_unlock(&vcpu->kvm->mmu_lock);
2246         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2247         kvm_mmu_flush_tlb(vcpu);
2248 out:
2249         return r;
2250 }
2251 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2252
2253 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2254 {
2255         mmu_free_roots(vcpu);
2256 }
2257
2258 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2259                                   struct kvm_mmu_page *sp,
2260                                   u64 *spte)
2261 {
2262         u64 pte;
2263         struct kvm_mmu_page *child;
2264
2265         pte = *spte;
2266         if (is_shadow_present_pte(pte)) {
2267                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2268                     is_large_pte(pte))
2269                         rmap_remove(vcpu->kvm, spte);
2270                 else {
2271                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2272                         mmu_page_remove_parent_pte(child, spte);
2273                 }
2274         }
2275         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2276         if (is_large_pte(pte))
2277                 --vcpu->kvm->stat.lpages;
2278 }
2279
2280 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2281                                   struct kvm_mmu_page *sp,
2282                                   u64 *spte,
2283                                   const void *new)
2284 {
2285         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2286                 if (!vcpu->arch.update_pte.largepage ||
2287                     sp->role.glevels == PT32_ROOT_LEVEL) {
2288                         ++vcpu->kvm->stat.mmu_pde_zapped;
2289                         return;
2290                 }
2291         }
2292
2293         ++vcpu->kvm->stat.mmu_pte_updated;
2294         if (sp->role.glevels == PT32_ROOT_LEVEL)
2295                 paging32_update_pte(vcpu, sp, spte, new);
2296         else
2297                 paging64_update_pte(vcpu, sp, spte, new);
2298 }
2299
2300 static bool need_remote_flush(u64 old, u64 new)
2301 {
2302         if (!is_shadow_present_pte(old))
2303                 return false;
2304         if (!is_shadow_present_pte(new))
2305                 return true;
2306         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2307                 return true;
2308         old ^= PT64_NX_MASK;
2309         new ^= PT64_NX_MASK;
2310         return (old & ~new & PT64_PERM_MASK) != 0;
2311 }
2312
2313 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2314 {
2315         if (need_remote_flush(old, new))
2316                 kvm_flush_remote_tlbs(vcpu->kvm);
2317         else
2318                 kvm_mmu_flush_tlb(vcpu);
2319 }
2320
2321 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2322 {
2323         u64 *spte = vcpu->arch.last_pte_updated;
2324
2325         return !!(spte && (*spte & shadow_accessed_mask));
2326 }
2327
2328 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2329                                           const u8 *new, int bytes)
2330 {
2331         gfn_t gfn;
2332         int r;
2333         u64 gpte = 0;
2334         pfn_t pfn;
2335
2336         vcpu->arch.update_pte.largepage = 0;
2337
2338         if (bytes != 4 && bytes != 8)
2339                 return;
2340
2341         /*
2342          * Assume that the pte write on a page table of the same type
2343          * as the current vcpu paging mode.  This is nearly always true
2344          * (might be false while changing modes).  Note it is verified later
2345          * by update_pte().
2346          */
2347         if (is_pae(vcpu)) {
2348                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2349                 if ((bytes == 4) && (gpa % 4 == 0)) {
2350                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2351                         if (r)
2352                                 return;
2353                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2354                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2355                         memcpy((void *)&gpte, new, 8);
2356                 }
2357         } else {
2358                 if ((bytes == 4) && (gpa % 4 == 0))
2359                         memcpy((void *)&gpte, new, 4);
2360         }
2361         if (!is_present_pte(gpte))
2362                 return;
2363         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2364
2365         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2366                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2367                 vcpu->arch.update_pte.largepage = 1;
2368         }
2369         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2370         smp_rmb();
2371         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2372
2373         if (is_error_pfn(pfn)) {
2374                 kvm_release_pfn_clean(pfn);
2375                 return;
2376         }
2377         vcpu->arch.update_pte.gfn = gfn;
2378         vcpu->arch.update_pte.pfn = pfn;
2379 }
2380
2381 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2382 {
2383         u64 *spte = vcpu->arch.last_pte_updated;
2384
2385         if (spte
2386             && vcpu->arch.last_pte_gfn == gfn
2387             && shadow_accessed_mask
2388             && !(*spte & shadow_accessed_mask)
2389             && is_shadow_present_pte(*spte))
2390                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2391 }
2392
2393 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2394                        const u8 *new, int bytes)
2395 {
2396         gfn_t gfn = gpa >> PAGE_SHIFT;
2397         struct kvm_mmu_page *sp;
2398         struct hlist_node *node, *n;
2399         struct hlist_head *bucket;
2400         unsigned index;
2401         u64 entry, gentry;
2402         u64 *spte;
2403         unsigned offset = offset_in_page(gpa);
2404         unsigned pte_size;
2405         unsigned page_offset;
2406         unsigned misaligned;
2407         unsigned quadrant;
2408         int level;
2409         int flooded = 0;
2410         int npte;
2411         int r;
2412
2413         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2414         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2415         spin_lock(&vcpu->kvm->mmu_lock);
2416         kvm_mmu_access_page(vcpu, gfn);
2417         kvm_mmu_free_some_pages(vcpu);
2418         ++vcpu->kvm->stat.mmu_pte_write;
2419         kvm_mmu_audit(vcpu, "pre pte write");
2420         if (gfn == vcpu->arch.last_pt_write_gfn
2421             && !last_updated_pte_accessed(vcpu)) {
2422                 ++vcpu->arch.last_pt_write_count;
2423                 if (vcpu->arch.last_pt_write_count >= 3)
2424                         flooded = 1;
2425         } else {
2426                 vcpu->arch.last_pt_write_gfn = gfn;
2427                 vcpu->arch.last_pt_write_count = 1;
2428                 vcpu->arch.last_pte_updated = NULL;
2429         }
2430         index = kvm_page_table_hashfn(gfn);
2431         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2432         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2433                 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
2434                         continue;
2435                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2436                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2437                 misaligned |= bytes < 4;
2438                 if (misaligned || flooded) {
2439                         /*
2440                          * Misaligned accesses are too much trouble to fix
2441                          * up; also, they usually indicate a page is not used
2442                          * as a page table.
2443                          *
2444                          * If we're seeing too many writes to a page,
2445                          * it may no longer be a page table, or we may be
2446                          * forking, in which case it is better to unmap the
2447                          * page.
2448                          */
2449                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2450                                  gpa, bytes, sp->role.word);
2451                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2452                                 n = bucket->first;
2453                         ++vcpu->kvm->stat.mmu_flooded;
2454                         continue;
2455                 }
2456                 page_offset = offset;
2457                 level = sp->role.level;
2458                 npte = 1;
2459                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2460                         page_offset <<= 1;      /* 32->64 */
2461                         /*
2462                          * A 32-bit pde maps 4MB while the shadow pdes map
2463                          * only 2MB.  So we need to double the offset again
2464                          * and zap two pdes instead of one.
2465                          */
2466                         if (level == PT32_ROOT_LEVEL) {
2467                                 page_offset &= ~7; /* kill rounding error */
2468                                 page_offset <<= 1;
2469                                 npte = 2;
2470                         }
2471                         quadrant = page_offset >> PAGE_SHIFT;
2472                         page_offset &= ~PAGE_MASK;
2473                         if (quadrant != sp->role.quadrant)
2474                                 continue;
2475                 }
2476                 spte = &sp->spt[page_offset / sizeof(*spte)];
2477                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2478                         gentry = 0;
2479                         r = kvm_read_guest_atomic(vcpu->kvm,
2480                                                   gpa & ~(u64)(pte_size - 1),
2481                                                   &gentry, pte_size);
2482                         new = (const void *)&gentry;
2483                         if (r < 0)
2484                                 new = NULL;
2485                 }
2486                 while (npte--) {
2487                         entry = *spte;
2488                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2489                         if (new)
2490                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2491                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2492                         ++spte;
2493                 }
2494         }
2495         kvm_mmu_audit(vcpu, "post pte write");
2496         spin_unlock(&vcpu->kvm->mmu_lock);
2497         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2498                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2499                 vcpu->arch.update_pte.pfn = bad_pfn;
2500         }
2501 }
2502
2503 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2504 {
2505         gpa_t gpa;
2506         int r;
2507
2508         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2509
2510         spin_lock(&vcpu->kvm->mmu_lock);
2511         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2512         spin_unlock(&vcpu->kvm->mmu_lock);
2513         return r;
2514 }
2515 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2516
2517 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2518 {
2519         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2520                 struct kvm_mmu_page *sp;
2521
2522                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2523                                   struct kvm_mmu_page, link);
2524                 kvm_mmu_zap_page(vcpu->kvm, sp);
2525                 ++vcpu->kvm->stat.mmu_recycled;
2526         }
2527 }
2528
2529 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2530 {
2531         int r;
2532         enum emulation_result er;
2533
2534         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2535         if (r < 0)
2536                 goto out;
2537
2538         if (!r) {
2539                 r = 1;
2540                 goto out;
2541         }
2542
2543         r = mmu_topup_memory_caches(vcpu);
2544         if (r)
2545                 goto out;
2546
2547         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2548
2549         switch (er) {
2550         case EMULATE_DONE:
2551                 return 1;
2552         case EMULATE_DO_MMIO:
2553                 ++vcpu->stat.mmio_exits;
2554                 return 0;
2555         case EMULATE_FAIL:
2556                 kvm_report_emulation_failure(vcpu, "pagetable");
2557                 return 1;
2558         default:
2559                 BUG();
2560         }
2561 out:
2562         return r;
2563 }
2564 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2565
2566 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2567 {
2568         spin_lock(&vcpu->kvm->mmu_lock);
2569         vcpu->arch.mmu.invlpg(vcpu, gva);
2570         spin_unlock(&vcpu->kvm->mmu_lock);
2571         kvm_mmu_flush_tlb(vcpu);
2572         ++vcpu->stat.invlpg;
2573 }
2574 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2575
2576 void kvm_enable_tdp(void)
2577 {
2578         tdp_enabled = true;
2579 }
2580 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2581
2582 void kvm_disable_tdp(void)
2583 {
2584         tdp_enabled = false;
2585 }
2586 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2587
2588 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2589 {
2590         struct kvm_mmu_page *sp;
2591
2592         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2593                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2594                                   struct kvm_mmu_page, link);
2595                 kvm_mmu_zap_page(vcpu->kvm, sp);
2596                 cond_resched();
2597         }
2598         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2599 }
2600
2601 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2602 {
2603         struct page *page;
2604         int i;
2605
2606         ASSERT(vcpu);
2607
2608         if (vcpu->kvm->arch.n_requested_mmu_pages)
2609                 vcpu->kvm->arch.n_free_mmu_pages =
2610                                         vcpu->kvm->arch.n_requested_mmu_pages;
2611         else
2612                 vcpu->kvm->arch.n_free_mmu_pages =
2613                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2614         /*
2615          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2616          * Therefore we need to allocate shadow page tables in the first
2617          * 4GB of memory, which happens to fit the DMA32 zone.
2618          */
2619         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2620         if (!page)
2621                 goto error_1;
2622         vcpu->arch.mmu.pae_root = page_address(page);
2623         for (i = 0; i < 4; ++i)
2624                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2625
2626         return 0;
2627
2628 error_1:
2629         free_mmu_pages(vcpu);
2630         return -ENOMEM;
2631 }
2632
2633 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2634 {
2635         ASSERT(vcpu);
2636         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2637
2638         return alloc_mmu_pages(vcpu);
2639 }
2640
2641 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2642 {
2643         ASSERT(vcpu);
2644         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2645
2646         return init_kvm_mmu(vcpu);
2647 }
2648
2649 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2650 {
2651         ASSERT(vcpu);
2652
2653         destroy_kvm_mmu(vcpu);
2654         free_mmu_pages(vcpu);
2655         mmu_free_memory_caches(vcpu);
2656 }
2657
2658 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2659 {
2660         struct kvm_mmu_page *sp;
2661
2662         spin_lock(&kvm->mmu_lock);
2663         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2664                 int i;
2665                 u64 *pt;
2666
2667                 if (!test_bit(slot, sp->slot_bitmap))
2668                         continue;
2669
2670                 pt = sp->spt;
2671                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2672                         /* avoid RMW */
2673                         if (pt[i] & PT_WRITABLE_MASK)
2674                                 pt[i] &= ~PT_WRITABLE_MASK;
2675         }
2676         kvm_flush_remote_tlbs(kvm);
2677         spin_unlock(&kvm->mmu_lock);
2678 }
2679
2680 void kvm_mmu_zap_all(struct kvm *kvm)
2681 {
2682         struct kvm_mmu_page *sp, *node;
2683
2684         spin_lock(&kvm->mmu_lock);
2685         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2686                 if (kvm_mmu_zap_page(kvm, sp))
2687                         node = container_of(kvm->arch.active_mmu_pages.next,
2688                                             struct kvm_mmu_page, link);
2689         spin_unlock(&kvm->mmu_lock);
2690
2691         kvm_flush_remote_tlbs(kvm);
2692 }
2693
2694 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2695 {
2696         struct kvm_mmu_page *page;
2697
2698         page = container_of(kvm->arch.active_mmu_pages.prev,
2699                             struct kvm_mmu_page, link);
2700         kvm_mmu_zap_page(kvm, page);
2701 }
2702
2703 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2704 {
2705         struct kvm *kvm;
2706         struct kvm *kvm_freed = NULL;
2707         int cache_count = 0;
2708
2709         spin_lock(&kvm_lock);
2710
2711         list_for_each_entry(kvm, &vm_list, vm_list) {
2712                 int npages;
2713
2714                 if (!down_read_trylock(&kvm->slots_lock))
2715                         continue;
2716                 spin_lock(&kvm->mmu_lock);
2717                 npages = kvm->arch.n_alloc_mmu_pages -
2718                          kvm->arch.n_free_mmu_pages;
2719                 cache_count += npages;
2720                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2721                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2722                         cache_count--;
2723                         kvm_freed = kvm;
2724                 }
2725                 nr_to_scan--;
2726
2727                 spin_unlock(&kvm->mmu_lock);
2728                 up_read(&kvm->slots_lock);
2729         }
2730         if (kvm_freed)
2731                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2732
2733         spin_unlock(&kvm_lock);
2734
2735         return cache_count;
2736 }
2737
2738 static struct shrinker mmu_shrinker = {
2739         .shrink = mmu_shrink,
2740         .seeks = DEFAULT_SEEKS * 10,
2741 };
2742
2743 static void mmu_destroy_caches(void)
2744 {
2745         if (pte_chain_cache)
2746                 kmem_cache_destroy(pte_chain_cache);
2747         if (rmap_desc_cache)
2748                 kmem_cache_destroy(rmap_desc_cache);
2749         if (mmu_page_header_cache)
2750                 kmem_cache_destroy(mmu_page_header_cache);
2751 }
2752
2753 void kvm_mmu_module_exit(void)
2754 {
2755         mmu_destroy_caches();
2756         unregister_shrinker(&mmu_shrinker);
2757 }
2758
2759 int kvm_mmu_module_init(void)
2760 {
2761         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2762                                             sizeof(struct kvm_pte_chain),
2763                                             0, 0, NULL);
2764         if (!pte_chain_cache)
2765                 goto nomem;
2766         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2767                                             sizeof(struct kvm_rmap_desc),
2768                                             0, 0, NULL);
2769         if (!rmap_desc_cache)
2770                 goto nomem;
2771
2772         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2773                                                   sizeof(struct kvm_mmu_page),
2774                                                   0, 0, NULL);
2775         if (!mmu_page_header_cache)
2776                 goto nomem;
2777
2778         register_shrinker(&mmu_shrinker);
2779
2780         return 0;
2781
2782 nomem:
2783         mmu_destroy_caches();
2784         return -ENOMEM;
2785 }
2786
2787 /*
2788  * Caculate mmu pages needed for kvm.
2789  */
2790 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2791 {
2792         int i;
2793         unsigned int nr_mmu_pages;
2794         unsigned int  nr_pages = 0;
2795
2796         for (i = 0; i < kvm->nmemslots; i++)
2797                 nr_pages += kvm->memslots[i].npages;
2798
2799         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2800         nr_mmu_pages = max(nr_mmu_pages,
2801                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2802
2803         return nr_mmu_pages;
2804 }
2805
2806 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2807                                 unsigned len)
2808 {
2809         if (len > buffer->len)
2810                 return NULL;
2811         return buffer->ptr;
2812 }
2813
2814 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2815                                 unsigned len)
2816 {
2817         void *ret;
2818
2819         ret = pv_mmu_peek_buffer(buffer, len);
2820         if (!ret)
2821                 return ret;
2822         buffer->ptr += len;
2823         buffer->len -= len;
2824         buffer->processed += len;
2825         return ret;
2826 }
2827
2828 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2829                              gpa_t addr, gpa_t value)
2830 {
2831         int bytes = 8;
2832         int r;
2833
2834         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2835                 bytes = 4;
2836
2837         r = mmu_topup_memory_caches(vcpu);
2838         if (r)
2839                 return r;
2840
2841         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2842                 return -EFAULT;
2843
2844         return 1;
2845 }
2846
2847 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2848 {
2849         kvm_x86_ops->tlb_flush(vcpu);
2850         set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
2851         return 1;
2852 }
2853
2854 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2855 {
2856         spin_lock(&vcpu->kvm->mmu_lock);
2857         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2858         spin_unlock(&vcpu->kvm->mmu_lock);
2859         return 1;
2860 }
2861
2862 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2863                              struct kvm_pv_mmu_op_buffer *buffer)
2864 {
2865         struct kvm_mmu_op_header *header;
2866
2867         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2868         if (!header)
2869                 return 0;
2870         switch (header->op) {
2871         case KVM_MMU_OP_WRITE_PTE: {
2872                 struct kvm_mmu_op_write_pte *wpte;
2873
2874                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2875                 if (!wpte)
2876                         return 0;
2877                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2878                                         wpte->pte_val);
2879         }
2880         case KVM_MMU_OP_FLUSH_TLB: {
2881                 struct kvm_mmu_op_flush_tlb *ftlb;
2882
2883                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2884                 if (!ftlb)
2885                         return 0;
2886                 return kvm_pv_mmu_flush_tlb(vcpu);
2887         }
2888         case KVM_MMU_OP_RELEASE_PT: {
2889                 struct kvm_mmu_op_release_pt *rpt;
2890
2891                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2892                 if (!rpt)
2893                         return 0;
2894                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2895         }
2896         default: return 0;
2897         }
2898 }
2899
2900 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2901                   gpa_t addr, unsigned long *ret)
2902 {
2903         int r;
2904         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2905
2906         buffer->ptr = buffer->buf;
2907         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
2908         buffer->processed = 0;
2909
2910         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
2911         if (r)
2912                 goto out;
2913
2914         while (buffer->len) {
2915                 r = kvm_pv_mmu_op_one(vcpu, buffer);
2916                 if (r < 0)
2917                         goto out;
2918                 if (r == 0)
2919                         break;
2920         }
2921
2922         r = 1;
2923 out:
2924         *ret = buffer->processed;
2925         return r;
2926 }
2927
2928 #ifdef AUDIT
2929
2930 static const char *audit_msg;
2931
2932 static gva_t canonicalize(gva_t gva)
2933 {
2934 #ifdef CONFIG_X86_64
2935         gva = (long long)(gva << 16) >> 16;
2936 #endif
2937         return gva;
2938 }
2939
2940 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2941                                 gva_t va, int level)
2942 {
2943         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2944         int i;
2945         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2946
2947         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2948                 u64 ent = pt[i];
2949
2950                 if (ent == shadow_trap_nonpresent_pte)
2951                         continue;
2952
2953                 va = canonicalize(va);
2954                 if (level > 1) {
2955                         if (ent == shadow_notrap_nonpresent_pte)
2956                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2957                                        " in nonleaf level: levels %d gva %lx"
2958                                        " level %d pte %llx\n", audit_msg,
2959                                        vcpu->arch.mmu.root_level, va, level, ent);
2960
2961                         audit_mappings_page(vcpu, ent, va, level - 1);
2962                 } else {
2963                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2964                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2965
2966                         if (is_shadow_present_pte(ent)
2967                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2968                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2969                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2970                                        audit_msg, vcpu->arch.mmu.root_level,
2971                                        va, gpa, hpa, ent,
2972                                        is_shadow_present_pte(ent));
2973                         else if (ent == shadow_notrap_nonpresent_pte
2974                                  && !is_error_hpa(hpa))
2975                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2976                                        " valid guest gva %lx\n", audit_msg, va);
2977                         kvm_release_pfn_clean(pfn);
2978
2979                 }
2980         }
2981 }
2982
2983 static void audit_mappings(struct kvm_vcpu *vcpu)
2984 {
2985         unsigned i;
2986
2987         if (vcpu->arch.mmu.root_level == 4)
2988                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2989         else
2990                 for (i = 0; i < 4; ++i)
2991                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2992                                 audit_mappings_page(vcpu,
2993                                                     vcpu->arch.mmu.pae_root[i],
2994                                                     i << 30,
2995                                                     2);
2996 }
2997
2998 static int count_rmaps(struct kvm_vcpu *vcpu)
2999 {
3000         int nmaps = 0;
3001         int i, j, k;
3002
3003         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3004                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3005                 struct kvm_rmap_desc *d;
3006
3007                 for (j = 0; j < m->npages; ++j) {
3008                         unsigned long *rmapp = &m->rmap[j];
3009
3010                         if (!*rmapp)
3011                                 continue;
3012                         if (!(*rmapp & 1)) {
3013                                 ++nmaps;
3014                                 continue;
3015                         }
3016                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3017                         while (d) {
3018                                 for (k = 0; k < RMAP_EXT; ++k)
3019                                         if (d->shadow_ptes[k])
3020                                                 ++nmaps;
3021                                         else
3022                                                 break;
3023                                 d = d->more;
3024                         }
3025                 }
3026         }
3027         return nmaps;
3028 }
3029
3030 static int count_writable_mappings(struct kvm_vcpu *vcpu)
3031 {
3032         int nmaps = 0;
3033         struct kvm_mmu_page *sp;
3034         int i;
3035
3036         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3037                 u64 *pt = sp->spt;
3038
3039                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3040                         continue;
3041
3042                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3043                         u64 ent = pt[i];
3044
3045                         if (!(ent & PT_PRESENT_MASK))
3046                                 continue;
3047                         if (!(ent & PT_WRITABLE_MASK))
3048                                 continue;
3049                         ++nmaps;
3050                 }
3051         }
3052         return nmaps;
3053 }
3054
3055 static void audit_rmap(struct kvm_vcpu *vcpu)
3056 {
3057         int n_rmap = count_rmaps(vcpu);
3058         int n_actual = count_writable_mappings(vcpu);
3059
3060         if (n_rmap != n_actual)
3061                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
3062                        __func__, audit_msg, n_rmap, n_actual);
3063 }
3064
3065 static void audit_write_protection(struct kvm_vcpu *vcpu)
3066 {
3067         struct kvm_mmu_page *sp;
3068         struct kvm_memory_slot *slot;
3069         unsigned long *rmapp;
3070         gfn_t gfn;
3071
3072         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3073                 if (sp->role.metaphysical)
3074                         continue;
3075
3076                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3077                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3078                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3079                 if (*rmapp)
3080                         printk(KERN_ERR "%s: (%s) shadow page has writable"
3081                                " mappings: gfn %lx role %x\n",
3082                                __func__, audit_msg, sp->gfn,
3083                                sp->role.word);
3084         }
3085 }
3086
3087 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3088 {
3089         int olddbg = dbg;
3090
3091         dbg = 0;
3092         audit_msg = msg;
3093         audit_rmap(vcpu);
3094         audit_write_protection(vcpu);
3095         audit_mappings(vcpu);
3096         dbg = olddbg;
3097 }
3098
3099 #endif