arch/x86/kvm/mmu.c

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