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 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 139
 140 struct kvm_rmap_desc {
 141         u64 *shadow_ptes[RMAP_EXT];
 142         struct kvm_rmap_desc *more;
 143 };
 144
 145 struct kvm_shadow_walk {
 146         int (*entry)(struct kvm_shadow_walk *walk, struct kvm_vcpu *vcpu,
 147                      u64 addr, u64 *spte, int level);
 148 };
 149
 150 static struct kmem_cache *pte_chain_cache;
 151 static struct kmem_cache *rmap_desc_cache;
 152 static struct kmem_cache *mmu_page_header_cache;
 153
 154 static u64 __read_mostly shadow_trap_nonpresent_pte;
 155 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 156 static u64 __read_mostly shadow_base_present_pte;
 157 static u64 __read_mostly shadow_nx_mask;
 158 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 159 static u64 __read_mostly shadow_user_mask;
 160 static u64 __read_mostly shadow_accessed_mask;
 161 static u64 __read_mostly shadow_dirty_mask;
 162
 163 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 164 {
 165         shadow_trap_nonpresent_pte = trap_pte;
 166         shadow_notrap_nonpresent_pte = notrap_pte;
 167 }
 168 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 169
 170 void kvm_mmu_set_base_ptes(u64 base_pte)
 171 {
 172         shadow_base_present_pte = base_pte;
 173 }
 174 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 175
 176 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 177                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 178 {
 179         shadow_user_mask = user_mask;
 180         shadow_accessed_mask = accessed_mask;
 181         shadow_dirty_mask = dirty_mask;
 182         shadow_nx_mask = nx_mask;
 183         shadow_x_mask = x_mask;
 184 }
 185 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 186
 187 static int is_write_protection(struct kvm_vcpu *vcpu)
 188 {
 189         return vcpu->arch.cr0 & X86_CR0_WP;
 190 }
 191
 192 static int is_cpuid_PSE36(void)
 193 {
 194         return 1;
 195 }
 196
 197 static int is_nx(struct kvm_vcpu *vcpu)
 198 {
 199         return vcpu->arch.shadow_efer & EFER_NX;
 200 }
 201
 202 static int is_present_pte(unsigned long pte)
 203 {
 204         return pte & PT_PRESENT_MASK;
 205 }
 206
 207 static int is_shadow_present_pte(u64 pte)
 208 {
 209         return pte != shadow_trap_nonpresent_pte
 210                 && pte != shadow_notrap_nonpresent_pte;
 211 }
 212
 213 static int is_large_pte(u64 pte)
 214 {
 215         return pte & PT_PAGE_SIZE_MASK;
 216 }
 217
 218 static int is_writeble_pte(unsigned long pte)
 219 {
 220         return pte & PT_WRITABLE_MASK;
 221 }
 222
 223 static int is_dirty_pte(unsigned long pte)
 224 {
 225         return pte & shadow_dirty_mask;
 226 }
 227
 228 static int is_rmap_pte(u64 pte)
 229 {
 230         return is_shadow_present_pte(pte);
 231 }
 232
 233 static pfn_t spte_to_pfn(u64 pte)
 234 {
 235         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 236 }
 237
 238 static gfn_t pse36_gfn_delta(u32 gpte)
 239 {
 240         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 241
 242         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 243 }
 244
 245 static void set_shadow_pte(u64 *sptep, u64 spte)
 246 {
 247 #ifdef CONFIG_X86_64
 248         set_64bit((unsigned long *)sptep, spte);
 249 #else
 250         set_64bit((unsigned long long *)sptep, spte);
 251 #endif
 252 }
 253
 254 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 255                                   struct kmem_cache *base_cache, int min)
 256 {
 257         void *obj;
 258
 259         if (cache->nobjs >= min)
 260                 return 0;
 261         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 262                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 263                 if (!obj)
 264                         return -ENOMEM;
 265                 cache->objects[cache->nobjs++] = obj;
 266         }
 267         return 0;
 268 }
 269
 270 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 271 {
 272         while (mc->nobjs)
 273                 kfree(mc->objects[--mc->nobjs]);
 274 }
 275
 276 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 277                                        int min)
 278 {
 279         struct page *page;
 280
 281         if (cache->nobjs >= min)
 282                 return 0;
 283         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 284                 page = alloc_page(GFP_KERNEL);
 285                 if (!page)
 286                         return -ENOMEM;
 287                 set_page_private(page, 0);
 288                 cache->objects[cache->nobjs++] = page_address(page);
 289         }
 290         return 0;
 291 }
 292
 293 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 294 {
 295         while (mc->nobjs)
 296                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 297 }
 298
 299 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 300 {
 301         int r;
 302
 303         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 304                                    pte_chain_cache, 4);
 305         if (r)
 306                 goto out;
 307         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 308                                    rmap_desc_cache, 1);
 309         if (r)
 310                 goto out;
 311         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 312         if (r)
 313                 goto out;
 314         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 315                                    mmu_page_header_cache, 4);
 316 out:
 317         return r;
 318 }
 319
 320 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 321 {
 322         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 323         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 324         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 325         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 326 }
 327
 328 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 329                                     size_t size)
 330 {
 331         void *p;
 332
 333         BUG_ON(!mc->nobjs);
 334         p = mc->objects[--mc->nobjs];
 335         memset(p, 0, size);
 336         return p;
 337 }
 338
 339 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 340 {
 341         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 342                                       sizeof(struct kvm_pte_chain));
 343 }
 344
 345 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 346 {
 347         kfree(pc);
 348 }
 349
 350 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 351 {
 352         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 353                                       sizeof(struct kvm_rmap_desc));
 354 }
 355
 356 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 357 {
 358         kfree(rd);
 359 }
 360
 361 /*
 362  * Return the pointer to the largepage write count for a given
 363  * gfn, handling slots that are not large page aligned.
 364  */
 365 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 366 {
 367         unsigned long idx;
 368
 369         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 370               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 371         return &slot->lpage_info[idx].write_count;
 372 }
 373
 374 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 375 {
 376         int *write_count;
 377
 378         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 379         *write_count += 1;
 380 }
 381
 382 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 383 {
 384         int *write_count;
 385
 386         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 387         *write_count -= 1;
 388         WARN_ON(*write_count < 0);
 389 }
 390
 391 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 392 {
 393         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
 394         int *largepage_idx;
 395
 396         if (slot) {
 397                 largepage_idx = slot_largepage_idx(gfn, slot);
 398                 return *largepage_idx;
 399         }
 400
 401         return 1;
 402 }
 403
 404 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 405 {
 406         struct vm_area_struct *vma;
 407         unsigned long addr;
 408
 409         addr = gfn_to_hva(kvm, gfn);
 410         if (kvm_is_error_hva(addr))
 411                 return 0;
 412
 413         vma = find_vma(current->mm, addr);
 414         if (vma && is_vm_hugetlb_page(vma))
 415                 return 1;
 416
 417         return 0;
 418 }
 419
 420 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 421 {
 422         struct kvm_memory_slot *slot;
 423
 424         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 425                 return 0;
 426
 427         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 428                 return 0;
 429
 430         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 431         if (slot && slot->dirty_bitmap)
 432                 return 0;
 433
 434         return 1;
 435 }
 436
 437 /*
 438  * Take gfn and return the reverse mapping to it.
 439  * Note: gfn must be unaliased before this function get called
 440  */
 441
 442 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 443 {
 444         struct kvm_memory_slot *slot;
 445         unsigned long idx;
 446
 447         slot = gfn_to_memslot(kvm, gfn);
 448         if (!lpage)
 449                 return &slot->rmap[gfn - slot->base_gfn];
 450
 451         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 452               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 453
 454         return &slot->lpage_info[idx].rmap_pde;
 455 }
 456
 457 /*
 458  * Reverse mapping data structures:
 459  *
 460  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 461  * that points to page_address(page).
 462  *
 463  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 464  * containing more mappings.
 465  */
 466 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 467 {
 468         struct kvm_mmu_page *sp;
 469         struct kvm_rmap_desc *desc;
 470         unsigned long *rmapp;
 471         int i;
 472
 473         if (!is_rmap_pte(*spte))
 474                 return;
 475         gfn = unalias_gfn(vcpu->kvm, gfn);
 476         sp = page_header(__pa(spte));
 477         sp->gfns[spte - sp->spt] = gfn;
 478         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 479         if (!*rmapp) {
 480                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 481                 *rmapp = (unsigned long)spte;
 482         } else if (!(*rmapp & 1)) {
 483                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 484                 desc = mmu_alloc_rmap_desc(vcpu);
 485                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 486                 desc->shadow_ptes[1] = spte;
 487                 *rmapp = (unsigned long)desc | 1;
 488         } else {
 489                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 490                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 491                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 492                         desc = desc->more;
 493                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 494                         desc->more = mmu_alloc_rmap_desc(vcpu);
 495                         desc = desc->more;
 496                 }
 497                 for (i = 0; desc->shadow_ptes[i]; ++i)
 498                         ;
 499                 desc->shadow_ptes[i] = spte;
 500         }
 501 }
 502
 503 static void rmap_desc_remove_entry(unsigned long *rmapp,
 504                                    struct kvm_rmap_desc *desc,
 505                                    int i,
 506                                    struct kvm_rmap_desc *prev_desc)
 507 {
 508         int j;
 509
 510         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 511                 ;
 512         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 513         desc->shadow_ptes[j] = NULL;
 514         if (j != 0)
 515                 return;
 516         if (!prev_desc && !desc->more)
 517                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 518         else
 519                 if (prev_desc)
 520                         prev_desc->more = desc->more;
 521                 else
 522                         *rmapp = (unsigned long)desc->more | 1;
 523         mmu_free_rmap_desc(desc);
 524 }
 525
 526 static void rmap_remove(struct kvm *kvm, u64 *spte)
 527 {
 528         struct kvm_rmap_desc *desc;
 529         struct kvm_rmap_desc *prev_desc;
 530         struct kvm_mmu_page *sp;
 531         pfn_t pfn;
 532         unsigned long *rmapp;
 533         int i;
 534
 535         if (!is_rmap_pte(*spte))
 536                 return;
 537         sp = page_header(__pa(spte));
 538         pfn = spte_to_pfn(*spte);
 539         if (*spte & shadow_accessed_mask)
 540                 kvm_set_pfn_accessed(pfn);
 541         if (is_writeble_pte(*spte))
 542                 kvm_release_pfn_dirty(pfn);
 543         else
 544                 kvm_release_pfn_clean(pfn);
 545         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 546         if (!*rmapp) {
 547                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 548                 BUG();
 549         } else if (!(*rmapp & 1)) {
 550                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 551                 if ((u64 *)*rmapp != spte) {
 552                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 553                                spte, *spte);
 554                         BUG();
 555                 }
 556                 *rmapp = 0;
 557         } else {
 558                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 559                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 560                 prev_desc = NULL;
 561                 while (desc) {
 562                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 563                                 if (desc->shadow_ptes[i] == spte) {
 564                                         rmap_desc_remove_entry(rmapp,
 565                                                                desc, i,
 566                                                                prev_desc);
 567                                         return;
 568                                 }
 569                         prev_desc = desc;
 570                         desc = desc->more;
 571                 }
 572                 BUG();
 573         }
 574 }
 575
 576 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 577 {
 578         struct kvm_rmap_desc *desc;
 579         struct kvm_rmap_desc *prev_desc;
 580         u64 *prev_spte;
 581         int i;
 582
 583         if (!*rmapp)
 584                 return NULL;
 585         else if (!(*rmapp & 1)) {
 586                 if (!spte)
 587                         return (u64 *)*rmapp;
 588                 return NULL;
 589         }
 590         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 591         prev_desc = NULL;
 592         prev_spte = NULL;
 593         while (desc) {
 594                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 595                         if (prev_spte == spte)
 596                                 return desc->shadow_ptes[i];
 597                         prev_spte = desc->shadow_ptes[i];
 598                 }
 599                 desc = desc->more;
 600         }
 601         return NULL;
 602 }
 603
 604 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 605 {
 606         unsigned long *rmapp;
 607         u64 *spte;
 608         int write_protected = 0;
 609
 610         gfn = unalias_gfn(kvm, gfn);
 611         rmapp = gfn_to_rmap(kvm, gfn, 0);
 612
 613         spte = rmap_next(kvm, rmapp, NULL);
 614         while (spte) {
 615                 BUG_ON(!spte);
 616                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 617                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 618                 if (is_writeble_pte(*spte)) {
 619                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 620                         write_protected = 1;
 621                 }
 622                 spte = rmap_next(kvm, rmapp, spte);
 623         }
 624         if (write_protected) {
 625                 pfn_t pfn;
 626
 627                 spte = rmap_next(kvm, rmapp, NULL);
 628                 pfn = spte_to_pfn(*spte);
 629                 kvm_set_pfn_dirty(pfn);
 630         }
 631
 632         /* check for huge page mappings */
 633         rmapp = gfn_to_rmap(kvm, gfn, 1);
 634         spte = rmap_next(kvm, rmapp, NULL);
 635         while (spte) {
 636                 BUG_ON(!spte);
 637                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 638                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 639                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 640                 if (is_writeble_pte(*spte)) {
 641                         rmap_remove(kvm, spte);
 642                         --kvm->stat.lpages;
 643                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 644                         spte = NULL;
 645                         write_protected = 1;
 646                 }
 647                 spte = rmap_next(kvm, rmapp, spte);
 648         }
 649
 650         if (write_protected)
 651                 kvm_flush_remote_tlbs(kvm);
 652
 653         account_shadowed(kvm, gfn);
 654 }
 655
 656 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 657 {
 658         u64 *spte;
 659         int need_tlb_flush = 0;
 660
 661         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 662                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 663                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 664                 rmap_remove(kvm, spte);
 665                 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 666                 need_tlb_flush = 1;
 667         }
 668         return need_tlb_flush;
 669 }
 670
 671 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 672                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 673 {
 674         int i;
 675         int retval = 0;
 676
 677         /*
 678          * If mmap_sem isn't taken, we can look the memslots with only
 679          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 680          */
 681         for (i = 0; i < kvm->nmemslots; i++) {
 682                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 683                 unsigned long start = memslot->userspace_addr;
 684                 unsigned long end;
 685
 686                 /* mmu_lock protects userspace_addr */
 687                 if (!start)
 688                         continue;
 689
 690                 end = start + (memslot->npages << PAGE_SHIFT);
 691                 if (hva >= start && hva < end) {
 692                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 693                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 694                         retval |= handler(kvm,
 695                                           &memslot->lpage_info[
 696                                                   gfn_offset /
 697                                                   KVM_PAGES_PER_HPAGE].rmap_pde);
 698                 }
 699         }
 700
 701         return retval;
 702 }
 703
 704 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 705 {
 706         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 707 }
 708
 709 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 710 {
 711         u64 *spte;
 712         int young = 0;
 713
 714         /* always return old for EPT */
 715         if (!shadow_accessed_mask)
 716                 return 0;
 717
 718         spte = rmap_next(kvm, rmapp, NULL);
 719         while (spte) {
 720                 int _young;
 721                 u64 _spte = *spte;
 722                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 723                 _young = _spte & PT_ACCESSED_MASK;
 724                 if (_young) {
 725                         young = 1;
 726                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 727                 }
 728                 spte = rmap_next(kvm, rmapp, spte);
 729         }
 730         return young;
 731 }
 732
 733 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 734 {
 735         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 736 }
 737
 738 #ifdef MMU_DEBUG
 739 static int is_empty_shadow_page(u64 *spt)
 740 {
 741         u64 *pos;
 742         u64 *end;
 743
 744         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 745                 if (is_shadow_present_pte(*pos)) {
 746                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 747                                pos, *pos);
 748                         return 0;
 749                 }
 750         return 1;
 751 }
 752 #endif
 753
 754 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 755 {
 756         ASSERT(is_empty_shadow_page(sp->spt));
 757         list_del(&sp->link);
 758         __free_page(virt_to_page(sp->spt));
 759         __free_page(virt_to_page(sp->gfns));
 760         kfree(sp);
 761         ++kvm->arch.n_free_mmu_pages;
 762 }
 763
 764 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 765 {
 766         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 767 }
 768
 769 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 770                                                u64 *parent_pte)
 771 {
 772         struct kvm_mmu_page *sp;
 773
 774         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 775         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 776         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 777         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 778         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 779         ASSERT(is_empty_shadow_page(sp->spt));
 780         sp->slot_bitmap = 0;
 781         sp->multimapped = 0;
 782         sp->parent_pte = parent_pte;
 783         --vcpu->kvm->arch.n_free_mmu_pages;
 784         return sp;
 785 }
 786
 787 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 788                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 789 {
 790         struct kvm_pte_chain *pte_chain;
 791         struct hlist_node *node;
 792         int i;
 793
 794         if (!parent_pte)
 795                 return;
 796         if (!sp->multimapped) {
 797                 u64 *old = sp->parent_pte;
 798
 799                 if (!old) {
 800                         sp->parent_pte = parent_pte;
 801                         return;
 802                 }
 803                 sp->multimapped = 1;
 804                 pte_chain = mmu_alloc_pte_chain(vcpu);
 805                 INIT_HLIST_HEAD(&sp->parent_ptes);
 806                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 807                 pte_chain->parent_ptes[0] = old;
 808         }
 809         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 810                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 811                         continue;
 812                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 813                         if (!pte_chain->parent_ptes[i]) {
 814                                 pte_chain->parent_ptes[i] = parent_pte;
 815                                 return;
 816                         }
 817         }
 818         pte_chain = mmu_alloc_pte_chain(vcpu);
 819         BUG_ON(!pte_chain);
 820         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 821         pte_chain->parent_ptes[0] = parent_pte;
 822 }
 823
 824 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 825                                        u64 *parent_pte)
 826 {
 827         struct kvm_pte_chain *pte_chain;
 828         struct hlist_node *node;
 829         int i;
 830
 831         if (!sp->multimapped) {
 832                 BUG_ON(sp->parent_pte != parent_pte);
 833                 sp->parent_pte = NULL;
 834                 return;
 835         }
 836         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 837                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 838                         if (!pte_chain->parent_ptes[i])
 839                                 break;
 840                         if (pte_chain->parent_ptes[i] != parent_pte)
 841                                 continue;
 842                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 843                                 && pte_chain->parent_ptes[i + 1]) {
 844                                 pte_chain->parent_ptes[i]
 845                                         = pte_chain->parent_ptes[i + 1];
 846                                 ++i;
 847                         }
 848                         pte_chain->parent_ptes[i] = NULL;
 849                         if (i == 0) {
 850                                 hlist_del(&pte_chain->link);
 851                                 mmu_free_pte_chain(pte_chain);
 852                                 if (hlist_empty(&sp->parent_ptes)) {
 853                                         sp->multimapped = 0;
 854                                         sp->parent_pte = NULL;
 855                                 }
 856                         }
 857                         return;
 858                 }
 859         BUG();
 860 }
 861
 862 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 863                                     struct kvm_mmu_page *sp)
 864 {
 865         int i;
 866
 867         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 868                 sp->spt[i] = shadow_trap_nonpresent_pte;
 869 }
 870
 871 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
 872 {
 873         unsigned index;
 874         struct hlist_head *bucket;
 875         struct kvm_mmu_page *sp;
 876         struct hlist_node *node;
 877
 878         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 879         index = kvm_page_table_hashfn(gfn);
 880         bucket = &kvm->arch.mmu_page_hash[index];
 881         hlist_for_each_entry(sp, node, bucket, hash_link)
 882                 if (sp->gfn == gfn && !sp->role.metaphysical
 883                     && !sp->role.invalid) {
 884                         pgprintk("%s: found role %x\n",
 885                                  __func__, sp->role.word);
 886                         return sp;
 887                 }
 888         return NULL;
 889 }
 890
 891 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 892                                              gfn_t gfn,
 893                                              gva_t gaddr,
 894                                              unsigned level,
 895                                              int metaphysical,
 896                                              unsigned access,
 897                                              u64 *parent_pte)
 898 {
 899         union kvm_mmu_page_role role;
 900         unsigned index;
 901         unsigned quadrant;
 902         struct hlist_head *bucket;
 903         struct kvm_mmu_page *sp;
 904         struct hlist_node *node;
 905
 906         role.word = 0;
 907         role.glevels = vcpu->arch.mmu.root_level;
 908         role.level = level;
 909         role.metaphysical = metaphysical;
 910         role.access = access;
 911         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
 912                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 913                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 914                 role.quadrant = quadrant;
 915         }
 916         pgprintk("%s: looking gfn %lx role %x\n", __func__,
 917                  gfn, role.word);
 918         index = kvm_page_table_hashfn(gfn);
 919         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
 920         hlist_for_each_entry(sp, node, bucket, hash_link)
 921                 if (sp->gfn == gfn && sp->role.word == role.word) {
 922                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
 923                         pgprintk("%s: found\n", __func__);
 924                         return sp;
 925                 }
 926         ++vcpu->kvm->stat.mmu_cache_miss;
 927         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
 928         if (!sp)
 929                 return sp;
 930         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
 931         sp->gfn = gfn;
 932         sp->role = role;
 933         hlist_add_head(&sp->hash_link, bucket);
 934         if (!metaphysical)
 935                 rmap_write_protect(vcpu->kvm, gfn);
 936         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
 937                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
 938         else
 939                 nonpaging_prefetch_page(vcpu, sp);
 940         return sp;
 941 }
 942
 943 static int walk_shadow(struct kvm_shadow_walk *walker,
 944                        struct kvm_vcpu *vcpu, u64 addr)
 945 {
 946         hpa_t shadow_addr;
 947         int level;
 948         int r;
 949         u64 *sptep;
 950         unsigned index;
 951
 952         shadow_addr = vcpu->arch.mmu.root_hpa;
 953         level = vcpu->arch.mmu.shadow_root_level;
 954         if (level == PT32E_ROOT_LEVEL) {
 955                 shadow_addr = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
 956                 shadow_addr &= PT64_BASE_ADDR_MASK;
 957                 --level;
 958         }
 959
 960         while (level >= PT_PAGE_TABLE_LEVEL) {
 961                 index = SHADOW_PT_INDEX(addr, level);
 962                 sptep = ((u64 *)__va(shadow_addr)) + index;
 963                 r = walker->entry(walker, vcpu, addr, sptep, level);
 964                 if (r)
 965                         return r;
 966                 shadow_addr = *sptep & PT64_BASE_ADDR_MASK;
 967                 --level;
 968         }
 969         return 0;
 970 }
 971
 972 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 973                                          struct kvm_mmu_page *sp)
 974 {
 975         unsigned i;
 976         u64 *pt;
 977         u64 ent;
 978
 979         pt = sp->spt;
 980
 981         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
 982                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 983                         if (is_shadow_present_pte(pt[i]))
 984                                 rmap_remove(kvm, &pt[i]);
 985                         pt[i] = shadow_trap_nonpresent_pte;
 986                 }
 987                 return;
 988         }
 989
 990         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 991                 ent = pt[i];
 992
 993                 if (is_shadow_present_pte(ent)) {
 994                         if (!is_large_pte(ent)) {
 995                                 ent &= PT64_BASE_ADDR_MASK;
 996                                 mmu_page_remove_parent_pte(page_header(ent),
 997                                                            &pt[i]);
 998                         } else {
 999                                 --kvm->stat.lpages;
1000                                 rmap_remove(kvm, &pt[i]);
1001                         }
1002                 }
1003                 pt[i] = shadow_trap_nonpresent_pte;
1004         }
1005 }
1006
1007 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1008 {
1009         mmu_page_remove_parent_pte(sp, parent_pte);
1010 }
1011
1012 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1013 {
1014         int i;
1015
1016         for (i = 0; i < KVM_MAX_VCPUS; ++i)
1017                 if (kvm->vcpus[i])
1018                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
1019 }
1020
1021 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1022 {
1023         u64 *parent_pte;
1024
1025         while (sp->multimapped || sp->parent_pte) {
1026                 if (!sp->multimapped)
1027                         parent_pte = sp->parent_pte;
1028                 else {
1029                         struct kvm_pte_chain *chain;
1030
1031                         chain = container_of(sp->parent_ptes.first,
1032                                              struct kvm_pte_chain, link);
1033                         parent_pte = chain->parent_ptes[0];
1034                 }
1035                 BUG_ON(!parent_pte);
1036                 kvm_mmu_put_page(sp, parent_pte);
1037                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1038         }
1039 }
1040
1041 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1042 {
1043         ++kvm->stat.mmu_shadow_zapped;
1044         kvm_mmu_page_unlink_children(kvm, sp);
1045         kvm_mmu_unlink_parents(kvm, sp);
1046         kvm_flush_remote_tlbs(kvm);
1047         if (!sp->role.invalid && !sp->role.metaphysical)
1048                 unaccount_shadowed(kvm, sp->gfn);
1049         if (!sp->root_count) {
1050                 hlist_del(&sp->hash_link);
1051                 kvm_mmu_free_page(kvm, sp);
1052         } else {
1053                 sp->role.invalid = 1;
1054                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1055                 kvm_reload_remote_mmus(kvm);
1056         }
1057         kvm_mmu_reset_last_pte_updated(kvm);
1058 }
1059
1060 /*
1061  * Changing the number of mmu pages allocated to the vm
1062  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1063  */
1064 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1065 {
1066         /*
1067          * If we set the number of mmu pages to be smaller be than the
1068          * number of actived pages , we must to free some mmu pages before we
1069          * change the value
1070          */
1071
1072         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1073             kvm_nr_mmu_pages) {
1074                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1075                                        - kvm->arch.n_free_mmu_pages;
1076
1077                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1078                         struct kvm_mmu_page *page;
1079
1080                         page = container_of(kvm->arch.active_mmu_pages.prev,
1081                                             struct kvm_mmu_page, link);
1082                         kvm_mmu_zap_page(kvm, page);
1083                         n_used_mmu_pages--;
1084                 }
1085                 kvm->arch.n_free_mmu_pages = 0;
1086         }
1087         else
1088                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1089                                          - kvm->arch.n_alloc_mmu_pages;
1090
1091         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1092 }
1093
1094 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1095 {
1096         unsigned index;
1097         struct hlist_head *bucket;
1098         struct kvm_mmu_page *sp;
1099         struct hlist_node *node, *n;
1100         int r;
1101
1102         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1103         r = 0;
1104         index = kvm_page_table_hashfn(gfn);
1105         bucket = &kvm->arch.mmu_page_hash[index];
1106         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1107                 if (sp->gfn == gfn && !sp->role.metaphysical) {
1108                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1109                                  sp->role.word);
1110                         kvm_mmu_zap_page(kvm, sp);
1111                         r = 1;
1112                 }
1113         return r;
1114 }
1115
1116 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1117 {
1118         struct kvm_mmu_page *sp;
1119
1120         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1121                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1122                 kvm_mmu_zap_page(kvm, sp);
1123         }
1124 }
1125
1126 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1127 {
1128         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1129         struct kvm_mmu_page *sp = page_header(__pa(pte));
1130
1131         __set_bit(slot, &sp->slot_bitmap);
1132 }
1133
1134 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1135 {
1136         struct page *page;
1137
1138         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1139
1140         if (gpa == UNMAPPED_GVA)
1141                 return NULL;
1142
1143         down_read(&current->mm->mmap_sem);
1144         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1145         up_read(&current->mm->mmap_sem);
1146
1147         return page;
1148 }
1149
1150 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1151                          unsigned pt_access, unsigned pte_access,
1152                          int user_fault, int write_fault, int dirty,
1153                          int *ptwrite, int largepage, gfn_t gfn,
1154                          pfn_t pfn, bool speculative)
1155 {
1156         u64 spte;
1157         int was_rmapped = 0;
1158         int was_writeble = is_writeble_pte(*shadow_pte);
1159
1160         pgprintk("%s: spte %llx access %x write_fault %d"
1161                  " user_fault %d gfn %lx\n",
1162                  __func__, *shadow_pte, pt_access,
1163                  write_fault, user_fault, gfn);
1164
1165         if (is_rmap_pte(*shadow_pte)) {
1166                 /*
1167                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1168                  * the parent of the now unreachable PTE.
1169                  */
1170                 if (largepage && !is_large_pte(*shadow_pte)) {
1171                         struct kvm_mmu_page *child;
1172                         u64 pte = *shadow_pte;
1173
1174                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1175                         mmu_page_remove_parent_pte(child, shadow_pte);
1176                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1177                         pgprintk("hfn old %lx new %lx\n",
1178                                  spte_to_pfn(*shadow_pte), pfn);
1179                         rmap_remove(vcpu->kvm, shadow_pte);
1180                 } else {
1181                         if (largepage)
1182                                 was_rmapped = is_large_pte(*shadow_pte);
1183                         else
1184                                 was_rmapped = 1;
1185                 }
1186         }
1187
1188         /*
1189          * We don't set the accessed bit, since we sometimes want to see
1190          * whether the guest actually used the pte (in order to detect
1191          * demand paging).
1192          */
1193         spte = shadow_base_present_pte | shadow_dirty_mask;
1194         if (!speculative)
1195                 spte |= shadow_accessed_mask;
1196         if (!dirty)
1197                 pte_access &= ~ACC_WRITE_MASK;
1198         if (pte_access & ACC_EXEC_MASK)
1199                 spte |= shadow_x_mask;
1200         else
1201                 spte |= shadow_nx_mask;
1202         if (pte_access & ACC_USER_MASK)
1203                 spte |= shadow_user_mask;
1204         if (largepage)
1205                 spte |= PT_PAGE_SIZE_MASK;
1206
1207         spte |= (u64)pfn << PAGE_SHIFT;
1208
1209         if ((pte_access & ACC_WRITE_MASK)
1210             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1211                 struct kvm_mmu_page *shadow;
1212
1213                 spte |= PT_WRITABLE_MASK;
1214
1215                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1216                 if (shadow ||
1217                    (largepage && has_wrprotected_page(vcpu->kvm, gfn))) {
1218                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1219                                  __func__, gfn);
1220                         pte_access &= ~ACC_WRITE_MASK;
1221                         if (is_writeble_pte(spte)) {
1222                                 spte &= ~PT_WRITABLE_MASK;
1223                                 kvm_x86_ops->tlb_flush(vcpu);
1224                         }
1225                         if (write_fault)
1226                                 *ptwrite = 1;
1227                 }
1228         }
1229
1230         if (pte_access & ACC_WRITE_MASK)
1231                 mark_page_dirty(vcpu->kvm, gfn);
1232
1233         pgprintk("%s: setting spte %llx\n", __func__, spte);
1234         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1235                  (spte&PT_PAGE_SIZE_MASK)? "2MB" : "4kB",
1236                  (spte&PT_WRITABLE_MASK)?"RW":"R", gfn, spte, shadow_pte);
1237         set_shadow_pte(shadow_pte, spte);
1238         if (!was_rmapped && (spte & PT_PAGE_SIZE_MASK)
1239             && (spte & PT_PRESENT_MASK))
1240                 ++vcpu->kvm->stat.lpages;
1241
1242         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1243         if (!was_rmapped) {
1244                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1245                 if (!is_rmap_pte(*shadow_pte))
1246                         kvm_release_pfn_clean(pfn);
1247         } else {
1248                 if (was_writeble)
1249                         kvm_release_pfn_dirty(pfn);
1250                 else
1251                         kvm_release_pfn_clean(pfn);
1252         }
1253         if (speculative) {
1254                 vcpu->arch.last_pte_updated = shadow_pte;
1255                 vcpu->arch.last_pte_gfn = gfn;
1256         }
1257 }
1258
1259 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1260 {
1261 }
1262
1263 struct direct_shadow_walk {
1264         struct kvm_shadow_walk walker;
1265         pfn_t pfn;
1266         int write;
1267         int largepage;
1268         int pt_write;
1269 };
1270
1271 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1272                             struct kvm_vcpu *vcpu,
1273                             u64 addr, u64 *sptep, int level)
1274 {
1275         struct direct_shadow_walk *walk =
1276                 container_of(_walk, struct direct_shadow_walk, walker);
1277         struct kvm_mmu_page *sp;
1278         gfn_t pseudo_gfn;
1279         gfn_t gfn = addr >> PAGE_SHIFT;
1280
1281         if (level == PT_PAGE_TABLE_LEVEL
1282             || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1283                 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1284                              0, walk->write, 1, &walk->pt_write,
1285                              walk->largepage, gfn, walk->pfn, false);
1286                 ++vcpu->stat.pf_fixed;
1287                 return 1;
1288         }
1289
1290         if (*sptep == shadow_trap_nonpresent_pte) {
1291                 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1292                 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1293                                       1, ACC_ALL, sptep);
1294                 if (!sp) {
1295                         pgprintk("nonpaging_map: ENOMEM\n");
1296                         kvm_release_pfn_clean(walk->pfn);
1297                         return -ENOMEM;
1298                 }
1299
1300                 set_shadow_pte(sptep,
1301                                __pa(sp->spt)
1302                                | PT_PRESENT_MASK | PT_WRITABLE_MASK
1303                                | shadow_user_mask | shadow_x_mask);
1304         }
1305         return 0;
1306 }
1307
1308 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1309                         int largepage, gfn_t gfn, pfn_t pfn)
1310 {
1311         int r;
1312         struct direct_shadow_walk walker = {
1313                 .walker = { .entry = direct_map_entry, },
1314                 .pfn = pfn,
1315                 .largepage = largepage,
1316                 .write = write,
1317                 .pt_write = 0,
1318         };
1319
1320         r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1321         if (r < 0)
1322                 return r;
1323         return walker.pt_write;
1324 }
1325
1326 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1327 {
1328         int r;
1329         int largepage = 0;
1330         pfn_t pfn;
1331         unsigned long mmu_seq;
1332
1333         down_read(&current->mm->mmap_sem);
1334         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1335                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1336                 largepage = 1;
1337         }
1338
1339         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1340         /* implicit mb(), we'll read before PT lock is unlocked */
1341         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1342         up_read(&current->mm->mmap_sem);
1343
1344         /* mmio */
1345         if (is_error_pfn(pfn)) {
1346                 kvm_release_pfn_clean(pfn);
1347                 return 1;
1348         }
1349
1350         spin_lock(&vcpu->kvm->mmu_lock);
1351         if (mmu_notifier_retry(vcpu, mmu_seq))
1352                 goto out_unlock;
1353         kvm_mmu_free_some_pages(vcpu);
1354         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1355         spin_unlock(&vcpu->kvm->mmu_lock);
1356
1357
1358         return r;
1359
1360 out_unlock:
1361         spin_unlock(&vcpu->kvm->mmu_lock);
1362         kvm_release_pfn_clean(pfn);
1363         return 0;
1364 }
1365
1366
1367 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1368 {
1369         int i;
1370         struct kvm_mmu_page *sp;
1371
1372         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1373                 return;
1374         spin_lock(&vcpu->kvm->mmu_lock);
1375         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1376                 hpa_t root = vcpu->arch.mmu.root_hpa;
1377
1378                 sp = page_header(root);
1379                 --sp->root_count;
1380                 if (!sp->root_count && sp->role.invalid)
1381                         kvm_mmu_zap_page(vcpu->kvm, sp);
1382                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1383                 spin_unlock(&vcpu->kvm->mmu_lock);
1384                 return;
1385         }
1386         for (i = 0; i < 4; ++i) {
1387                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1388
1389                 if (root) {
1390                         root &= PT64_BASE_ADDR_MASK;
1391                         sp = page_header(root);
1392                         --sp->root_count;
1393                         if (!sp->root_count && sp->role.invalid)
1394                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1395                 }
1396                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1397         }
1398         spin_unlock(&vcpu->kvm->mmu_lock);
1399         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1400 }
1401
1402 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1403 {
1404         int i;
1405         gfn_t root_gfn;
1406         struct kvm_mmu_page *sp;
1407         int metaphysical = 0;
1408
1409         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1410
1411         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1412                 hpa_t root = vcpu->arch.mmu.root_hpa;
1413
1414                 ASSERT(!VALID_PAGE(root));
1415                 if (tdp_enabled)
1416                         metaphysical = 1;
1417                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1418                                       PT64_ROOT_LEVEL, metaphysical,
1419                                       ACC_ALL, NULL);
1420                 root = __pa(sp->spt);
1421                 ++sp->root_count;
1422                 vcpu->arch.mmu.root_hpa = root;
1423                 return;
1424         }
1425         metaphysical = !is_paging(vcpu);
1426         if (tdp_enabled)
1427                 metaphysical = 1;
1428         for (i = 0; i < 4; ++i) {
1429                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1430
1431                 ASSERT(!VALID_PAGE(root));
1432                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1433                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1434                                 vcpu->arch.mmu.pae_root[i] = 0;
1435                                 continue;
1436                         }
1437                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1438                 } else if (vcpu->arch.mmu.root_level == 0)
1439                         root_gfn = 0;
1440                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1441                                       PT32_ROOT_LEVEL, metaphysical,
1442                                       ACC_ALL, NULL);
1443                 root = __pa(sp->spt);
1444                 ++sp->root_count;
1445                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1446         }
1447         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1448 }
1449
1450 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1451 {
1452         return vaddr;
1453 }
1454
1455 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1456                                 u32 error_code)
1457 {
1458         gfn_t gfn;
1459         int r;
1460
1461         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1462         r = mmu_topup_memory_caches(vcpu);
1463         if (r)
1464                 return r;
1465
1466         ASSERT(vcpu);
1467         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1468
1469         gfn = gva >> PAGE_SHIFT;
1470
1471         return nonpaging_map(vcpu, gva & PAGE_MASK,
1472                              error_code & PFERR_WRITE_MASK, gfn);
1473 }
1474
1475 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1476                                 u32 error_code)
1477 {
1478         pfn_t pfn;
1479         int r;
1480         int largepage = 0;
1481         gfn_t gfn = gpa >> PAGE_SHIFT;
1482         unsigned long mmu_seq;
1483
1484         ASSERT(vcpu);
1485         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1486
1487         r = mmu_topup_memory_caches(vcpu);
1488         if (r)
1489                 return r;
1490
1491         down_read(&current->mm->mmap_sem);
1492         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1493                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1494                 largepage = 1;
1495         }
1496         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1497         /* implicit mb(), we'll read before PT lock is unlocked */
1498         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1499         up_read(&current->mm->mmap_sem);
1500         if (is_error_pfn(pfn)) {
1501                 kvm_release_pfn_clean(pfn);
1502                 return 1;
1503         }
1504         spin_lock(&vcpu->kvm->mmu_lock);
1505         if (mmu_notifier_retry(vcpu, mmu_seq))
1506                 goto out_unlock;
1507         kvm_mmu_free_some_pages(vcpu);
1508         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1509                          largepage, gfn, pfn);
1510         spin_unlock(&vcpu->kvm->mmu_lock);
1511
1512         return r;
1513
1514 out_unlock:
1515         spin_unlock(&vcpu->kvm->mmu_lock);
1516         kvm_release_pfn_clean(pfn);
1517         return 0;
1518 }
1519
1520 static void nonpaging_free(struct kvm_vcpu *vcpu)
1521 {
1522         mmu_free_roots(vcpu);
1523 }
1524
1525 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1526 {
1527         struct kvm_mmu *context = &vcpu->arch.mmu;
1528
1529         context->new_cr3 = nonpaging_new_cr3;
1530         context->page_fault = nonpaging_page_fault;
1531         context->gva_to_gpa = nonpaging_gva_to_gpa;
1532         context->free = nonpaging_free;
1533         context->prefetch_page = nonpaging_prefetch_page;
1534         context->root_level = 0;
1535         context->shadow_root_level = PT32E_ROOT_LEVEL;
1536         context->root_hpa = INVALID_PAGE;
1537         return 0;
1538 }
1539
1540 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1541 {
1542         ++vcpu->stat.tlb_flush;
1543         kvm_x86_ops->tlb_flush(vcpu);
1544 }
1545
1546 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1547 {
1548         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1549         mmu_free_roots(vcpu);
1550 }
1551
1552 static void inject_page_fault(struct kvm_vcpu *vcpu,
1553                               u64 addr,
1554                               u32 err_code)
1555 {
1556         kvm_inject_page_fault(vcpu, addr, err_code);
1557 }
1558
1559 static void paging_free(struct kvm_vcpu *vcpu)
1560 {
1561         nonpaging_free(vcpu);
1562 }
1563
1564 #define PTTYPE 64
1565 #include "paging_tmpl.h"
1566 #undef PTTYPE
1567
1568 #define PTTYPE 32
1569 #include "paging_tmpl.h"
1570 #undef PTTYPE
1571
1572 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1573 {
1574         struct kvm_mmu *context = &vcpu->arch.mmu;
1575
1576         ASSERT(is_pae(vcpu));
1577         context->new_cr3 = paging_new_cr3;
1578         context->page_fault = paging64_page_fault;
1579         context->gva_to_gpa = paging64_gva_to_gpa;
1580         context->prefetch_page = paging64_prefetch_page;
1581         context->free = paging_free;
1582         context->root_level = level;
1583         context->shadow_root_level = level;
1584         context->root_hpa = INVALID_PAGE;
1585         return 0;
1586 }
1587
1588 static int paging64_init_context(struct kvm_vcpu *vcpu)
1589 {
1590         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1591 }
1592
1593 static int paging32_init_context(struct kvm_vcpu *vcpu)
1594 {
1595         struct kvm_mmu *context = &vcpu->arch.mmu;
1596
1597         context->new_cr3 = paging_new_cr3;
1598         context->page_fault = paging32_page_fault;
1599         context->gva_to_gpa = paging32_gva_to_gpa;
1600         context->free = paging_free;
1601         context->prefetch_page = paging32_prefetch_page;
1602         context->root_level = PT32_ROOT_LEVEL;
1603         context->shadow_root_level = PT32E_ROOT_LEVEL;
1604         context->root_hpa = INVALID_PAGE;
1605         return 0;
1606 }
1607
1608 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1609 {
1610         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1611 }
1612
1613 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1614 {
1615         struct kvm_mmu *context = &vcpu->arch.mmu;
1616
1617         context->new_cr3 = nonpaging_new_cr3;
1618         context->page_fault = tdp_page_fault;
1619         context->free = nonpaging_free;
1620         context->prefetch_page = nonpaging_prefetch_page;
1621         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
1622         context->root_hpa = INVALID_PAGE;
1623
1624         if (!is_paging(vcpu)) {
1625                 context->gva_to_gpa = nonpaging_gva_to_gpa;
1626                 context->root_level = 0;
1627         } else if (is_long_mode(vcpu)) {
1628                 context->gva_to_gpa = paging64_gva_to_gpa;
1629                 context->root_level = PT64_ROOT_LEVEL;
1630         } else if (is_pae(vcpu)) {
1631                 context->gva_to_gpa = paging64_gva_to_gpa;
1632                 context->root_level = PT32E_ROOT_LEVEL;
1633         } else {
1634                 context->gva_to_gpa = paging32_gva_to_gpa;
1635                 context->root_level = PT32_ROOT_LEVEL;
1636         }
1637
1638         return 0;
1639 }
1640
1641 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1642 {
1643         ASSERT(vcpu);
1644         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1645
1646         if (!is_paging(vcpu))
1647                 return nonpaging_init_context(vcpu);
1648         else if (is_long_mode(vcpu))
1649                 return paging64_init_context(vcpu);
1650         else if (is_pae(vcpu))
1651                 return paging32E_init_context(vcpu);
1652         else
1653                 return paging32_init_context(vcpu);
1654 }
1655
1656 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1657 {
1658         vcpu->arch.update_pte.pfn = bad_pfn;
1659
1660         if (tdp_enabled)
1661                 return init_kvm_tdp_mmu(vcpu);
1662         else
1663                 return init_kvm_softmmu(vcpu);
1664 }
1665
1666 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1667 {
1668         ASSERT(vcpu);
1669         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1670                 vcpu->arch.mmu.free(vcpu);
1671                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1672         }
1673 }
1674
1675 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1676 {
1677         destroy_kvm_mmu(vcpu);
1678         return init_kvm_mmu(vcpu);
1679 }
1680 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1681
1682 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1683 {
1684         int r;
1685
1686         r = mmu_topup_memory_caches(vcpu);
1687         if (r)
1688                 goto out;
1689         spin_lock(&vcpu->kvm->mmu_lock);
1690         kvm_mmu_free_some_pages(vcpu);
1691         mmu_alloc_roots(vcpu);
1692         spin_unlock(&vcpu->kvm->mmu_lock);
1693         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1694         kvm_mmu_flush_tlb(vcpu);
1695 out:
1696         return r;
1697 }
1698 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1699
1700 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1701 {
1702         mmu_free_roots(vcpu);
1703 }
1704
1705 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1706                                   struct kvm_mmu_page *sp,
1707                                   u64 *spte)
1708 {
1709         u64 pte;
1710         struct kvm_mmu_page *child;
1711
1712         pte = *spte;
1713         if (is_shadow_present_pte(pte)) {
1714                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
1715                     is_large_pte(pte))
1716                         rmap_remove(vcpu->kvm, spte);
1717                 else {
1718                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1719                         mmu_page_remove_parent_pte(child, spte);
1720                 }
1721         }
1722         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1723         if (is_large_pte(pte))
1724                 --vcpu->kvm->stat.lpages;
1725 }
1726
1727 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1728                                   struct kvm_mmu_page *sp,
1729                                   u64 *spte,
1730                                   const void *new)
1731 {
1732         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1733                 if (!vcpu->arch.update_pte.largepage ||
1734                     sp->role.glevels == PT32_ROOT_LEVEL) {
1735                         ++vcpu->kvm->stat.mmu_pde_zapped;
1736                         return;
1737                 }
1738         }
1739
1740         ++vcpu->kvm->stat.mmu_pte_updated;
1741         if (sp->role.glevels == PT32_ROOT_LEVEL)
1742                 paging32_update_pte(vcpu, sp, spte, new);
1743         else
1744                 paging64_update_pte(vcpu, sp, spte, new);
1745 }
1746
1747 static bool need_remote_flush(u64 old, u64 new)
1748 {
1749         if (!is_shadow_present_pte(old))
1750                 return false;
1751         if (!is_shadow_present_pte(new))
1752                 return true;
1753         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1754                 return true;
1755         old ^= PT64_NX_MASK;
1756         new ^= PT64_NX_MASK;
1757         return (old & ~new & PT64_PERM_MASK) != 0;
1758 }
1759
1760 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1761 {
1762         if (need_remote_flush(old, new))
1763                 kvm_flush_remote_tlbs(vcpu->kvm);
1764         else
1765                 kvm_mmu_flush_tlb(vcpu);
1766 }
1767
1768 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1769 {
1770         u64 *spte = vcpu->arch.last_pte_updated;
1771
1772         return !!(spte && (*spte & shadow_accessed_mask));
1773 }
1774
1775 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1776                                           const u8 *new, int bytes)
1777 {
1778         gfn_t gfn;
1779         int r;
1780         u64 gpte = 0;
1781         pfn_t pfn;
1782
1783         vcpu->arch.update_pte.largepage = 0;
1784
1785         if (bytes != 4 && bytes != 8)
1786                 return;
1787
1788         /*
1789          * Assume that the pte write on a page table of the same type
1790          * as the current vcpu paging mode.  This is nearly always true
1791          * (might be false while changing modes).  Note it is verified later
1792          * by update_pte().
1793          */
1794         if (is_pae(vcpu)) {
1795                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1796                 if ((bytes == 4) && (gpa % 4 == 0)) {
1797                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1798                         if (r)
1799                                 return;
1800                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1801                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1802                         memcpy((void *)&gpte, new, 8);
1803                 }
1804         } else {
1805                 if ((bytes == 4) && (gpa % 4 == 0))
1806                         memcpy((void *)&gpte, new, 4);
1807         }
1808         if (!is_present_pte(gpte))
1809                 return;
1810         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1811
1812         down_read(&current->mm->mmap_sem);
1813         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
1814                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1815                 vcpu->arch.update_pte.largepage = 1;
1816         }
1817         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
1818         /* implicit mb(), we'll read before PT lock is unlocked */
1819         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1820         up_read(&current->mm->mmap_sem);
1821
1822         if (is_error_pfn(pfn)) {
1823                 kvm_release_pfn_clean(pfn);
1824                 return;
1825         }
1826         vcpu->arch.update_pte.gfn = gfn;
1827         vcpu->arch.update_pte.pfn = pfn;
1828 }
1829
1830 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
1831 {
1832         u64 *spte = vcpu->arch.last_pte_updated;
1833
1834         if (spte
1835             && vcpu->arch.last_pte_gfn == gfn
1836             && shadow_accessed_mask
1837             && !(*spte & shadow_accessed_mask)
1838             && is_shadow_present_pte(*spte))
1839                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
1840 }
1841
1842 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1843                        const u8 *new, int bytes)
1844 {
1845         gfn_t gfn = gpa >> PAGE_SHIFT;
1846         struct kvm_mmu_page *sp;
1847         struct hlist_node *node, *n;
1848         struct hlist_head *bucket;
1849         unsigned index;
1850         u64 entry, gentry;
1851         u64 *spte;
1852         unsigned offset = offset_in_page(gpa);
1853         unsigned pte_size;
1854         unsigned page_offset;
1855         unsigned misaligned;
1856         unsigned quadrant;
1857         int level;
1858         int flooded = 0;
1859         int npte;
1860         int r;
1861
1862         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
1863         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1864         spin_lock(&vcpu->kvm->mmu_lock);
1865         kvm_mmu_access_page(vcpu, gfn);
1866         kvm_mmu_free_some_pages(vcpu);
1867         ++vcpu->kvm->stat.mmu_pte_write;
1868         kvm_mmu_audit(vcpu, "pre pte write");
1869         if (gfn == vcpu->arch.last_pt_write_gfn
1870             && !last_updated_pte_accessed(vcpu)) {
1871                 ++vcpu->arch.last_pt_write_count;
1872                 if (vcpu->arch.last_pt_write_count >= 3)
1873                         flooded = 1;
1874         } else {
1875                 vcpu->arch.last_pt_write_gfn = gfn;
1876                 vcpu->arch.last_pt_write_count = 1;
1877                 vcpu->arch.last_pte_updated = NULL;
1878         }
1879         index = kvm_page_table_hashfn(gfn);
1880         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1881         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1882                 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
1883                         continue;
1884                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1885                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1886                 misaligned |= bytes < 4;
1887                 if (misaligned || flooded) {
1888                         /*
1889                          * Misaligned accesses are too much trouble to fix
1890                          * up; also, they usually indicate a page is not used
1891                          * as a page table.
1892                          *
1893                          * If we're seeing too many writes to a page,
1894                          * it may no longer be a page table, or we may be
1895                          * forking, in which case it is better to unmap the
1896                          * page.
1897                          */
1898                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1899                                  gpa, bytes, sp->role.word);
1900                         kvm_mmu_zap_page(vcpu->kvm, sp);
1901                         ++vcpu->kvm->stat.mmu_flooded;
1902                         continue;
1903                 }
1904                 page_offset = offset;
1905                 level = sp->role.level;
1906                 npte = 1;
1907                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1908                         page_offset <<= 1;      /* 32->64 */
1909                         /*
1910                          * A 32-bit pde maps 4MB while the shadow pdes map
1911                          * only 2MB.  So we need to double the offset again
1912                          * and zap two pdes instead of one.
1913                          */
1914                         if (level == PT32_ROOT_LEVEL) {
1915                                 page_offset &= ~7; /* kill rounding error */
1916                                 page_offset <<= 1;
1917                                 npte = 2;
1918                         }
1919                         quadrant = page_offset >> PAGE_SHIFT;
1920                         page_offset &= ~PAGE_MASK;
1921                         if (quadrant != sp->role.quadrant)
1922                                 continue;
1923                 }
1924                 spte = &sp->spt[page_offset / sizeof(*spte)];
1925                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1926                         gentry = 0;
1927                         r = kvm_read_guest_atomic(vcpu->kvm,
1928                                                   gpa & ~(u64)(pte_size - 1),
1929                                                   &gentry, pte_size);
1930                         new = (const void *)&gentry;
1931                         if (r < 0)
1932                                 new = NULL;
1933                 }
1934                 while (npte--) {
1935                         entry = *spte;
1936                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1937                         if (new)
1938                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1939                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1940                         ++spte;
1941                 }
1942         }
1943         kvm_mmu_audit(vcpu, "post pte write");
1944         spin_unlock(&vcpu->kvm->mmu_lock);
1945         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
1946                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
1947                 vcpu->arch.update_pte.pfn = bad_pfn;
1948         }
1949 }
1950
1951 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1952 {
1953         gpa_t gpa;
1954         int r;
1955
1956         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1957
1958         spin_lock(&vcpu->kvm->mmu_lock);
1959         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1960         spin_unlock(&vcpu->kvm->mmu_lock);
1961         return r;
1962 }
1963 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
1964
1965 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1966 {
1967         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1968                 struct kvm_mmu_page *sp;
1969
1970                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1971                                   struct kvm_mmu_page, link);
1972                 kvm_mmu_zap_page(vcpu->kvm, sp);
1973                 ++vcpu->kvm->stat.mmu_recycled;
1974         }
1975 }
1976
1977 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1978 {
1979         int r;
1980         enum emulation_result er;
1981
1982         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1983         if (r < 0)
1984                 goto out;
1985
1986         if (!r) {
1987                 r = 1;
1988                 goto out;
1989         }
1990
1991         r = mmu_topup_memory_caches(vcpu);
1992         if (r)
1993                 goto out;
1994
1995         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1996
1997         switch (er) {
1998         case EMULATE_DONE:
1999                 return 1;
2000         case EMULATE_DO_MMIO:
2001                 ++vcpu->stat.mmio_exits;
2002                 return 0;
2003         case EMULATE_FAIL:
2004                 kvm_report_emulation_failure(vcpu, "pagetable");
2005                 return 1;
2006         default:
2007                 BUG();
2008         }
2009 out:
2010         return r;
2011 }
2012 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2013
2014 void kvm_enable_tdp(void)
2015 {
2016         tdp_enabled = true;
2017 }
2018 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2019
2020 void kvm_disable_tdp(void)
2021 {
2022         tdp_enabled = false;
2023 }
2024 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2025
2026 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2027 {
2028         struct kvm_mmu_page *sp;
2029
2030         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2031                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2032                                   struct kvm_mmu_page, link);
2033                 kvm_mmu_zap_page(vcpu->kvm, sp);
2034                 cond_resched();
2035         }
2036         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2037 }
2038
2039 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2040 {
2041         struct page *page;
2042         int i;
2043
2044         ASSERT(vcpu);
2045
2046         if (vcpu->kvm->arch.n_requested_mmu_pages)
2047                 vcpu->kvm->arch.n_free_mmu_pages =
2048                                         vcpu->kvm->arch.n_requested_mmu_pages;
2049         else
2050                 vcpu->kvm->arch.n_free_mmu_pages =
2051                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2052         /*
2053          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2054          * Therefore we need to allocate shadow page tables in the first
2055          * 4GB of memory, which happens to fit the DMA32 zone.
2056          */
2057         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2058         if (!page)
2059                 goto error_1;
2060         vcpu->arch.mmu.pae_root = page_address(page);
2061         for (i = 0; i < 4; ++i)
2062                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2063
2064         return 0;
2065
2066 error_1:
2067         free_mmu_pages(vcpu);
2068         return -ENOMEM;
2069 }
2070
2071 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2072 {
2073         ASSERT(vcpu);
2074         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2075
2076         return alloc_mmu_pages(vcpu);
2077 }
2078
2079 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2080 {
2081         ASSERT(vcpu);
2082         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2083
2084         return init_kvm_mmu(vcpu);
2085 }
2086
2087 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2088 {
2089         ASSERT(vcpu);
2090
2091         destroy_kvm_mmu(vcpu);
2092         free_mmu_pages(vcpu);
2093         mmu_free_memory_caches(vcpu);
2094 }
2095
2096 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2097 {
2098         struct kvm_mmu_page *sp;
2099
2100         spin_lock(&kvm->mmu_lock);
2101         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2102                 int i;
2103                 u64 *pt;
2104
2105                 if (!test_bit(slot, &sp->slot_bitmap))
2106                         continue;
2107
2108                 pt = sp->spt;
2109                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2110                         /* avoid RMW */
2111                         if (pt[i] & PT_WRITABLE_MASK)
2112                                 pt[i] &= ~PT_WRITABLE_MASK;
2113         }
2114         kvm_flush_remote_tlbs(kvm);
2115         spin_unlock(&kvm->mmu_lock);
2116 }
2117
2118 void kvm_mmu_zap_all(struct kvm *kvm)
2119 {
2120         struct kvm_mmu_page *sp, *node;
2121
2122         spin_lock(&kvm->mmu_lock);
2123         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2124                 kvm_mmu_zap_page(kvm, sp);
2125         spin_unlock(&kvm->mmu_lock);
2126
2127         kvm_flush_remote_tlbs(kvm);
2128 }
2129
2130 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2131 {
2132         struct kvm_mmu_page *page;
2133
2134         page = container_of(kvm->arch.active_mmu_pages.prev,
2135                             struct kvm_mmu_page, link);
2136         kvm_mmu_zap_page(kvm, page);
2137 }
2138
2139 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2140 {
2141         struct kvm *kvm;
2142         struct kvm *kvm_freed = NULL;
2143         int cache_count = 0;
2144
2145         spin_lock(&kvm_lock);
2146
2147         list_for_each_entry(kvm, &vm_list, vm_list) {
2148                 int npages;
2149
2150                 if (!down_read_trylock(&kvm->slots_lock))
2151                         continue;
2152                 spin_lock(&kvm->mmu_lock);
2153                 npages = kvm->arch.n_alloc_mmu_pages -
2154                          kvm->arch.n_free_mmu_pages;
2155                 cache_count += npages;
2156                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2157                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2158                         cache_count--;
2159                         kvm_freed = kvm;
2160                 }
2161                 nr_to_scan--;
2162
2163                 spin_unlock(&kvm->mmu_lock);
2164                 up_read(&kvm->slots_lock);
2165         }
2166         if (kvm_freed)
2167                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2168
2169         spin_unlock(&kvm_lock);
2170
2171         return cache_count;
2172 }
2173
2174 static struct shrinker mmu_shrinker = {
2175         .shrink = mmu_shrink,
2176         .seeks = DEFAULT_SEEKS * 10,
2177 };
2178
2179 static void mmu_destroy_caches(void)
2180 {
2181         if (pte_chain_cache)
2182                 kmem_cache_destroy(pte_chain_cache);
2183         if (rmap_desc_cache)
2184                 kmem_cache_destroy(rmap_desc_cache);
2185         if (mmu_page_header_cache)
2186                 kmem_cache_destroy(mmu_page_header_cache);
2187 }
2188
2189 void kvm_mmu_module_exit(void)
2190 {
2191         mmu_destroy_caches();
2192         unregister_shrinker(&mmu_shrinker);
2193 }
2194
2195 int kvm_mmu_module_init(void)
2196 {
2197         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2198                                             sizeof(struct kvm_pte_chain),
2199                                             0, 0, NULL);
2200         if (!pte_chain_cache)
2201                 goto nomem;
2202         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2203                                             sizeof(struct kvm_rmap_desc),
2204                                             0, 0, NULL);
2205         if (!rmap_desc_cache)
2206                 goto nomem;
2207
2208         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2209                                                   sizeof(struct kvm_mmu_page),
2210                                                   0, 0, NULL);
2211         if (!mmu_page_header_cache)
2212                 goto nomem;
2213
2214         register_shrinker(&mmu_shrinker);
2215
2216         return 0;
2217
2218 nomem:
2219         mmu_destroy_caches();
2220         return -ENOMEM;
2221 }
2222
2223 /*
2224  * Caculate mmu pages needed for kvm.
2225  */
2226 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2227 {
2228         int i;
2229         unsigned int nr_mmu_pages;
2230         unsigned int  nr_pages = 0;
2231
2232         for (i = 0; i < kvm->nmemslots; i++)
2233                 nr_pages += kvm->memslots[i].npages;
2234
2235         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2236         nr_mmu_pages = max(nr_mmu_pages,
2237                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2238
2239         return nr_mmu_pages;
2240 }
2241
2242 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2243                                 unsigned len)
2244 {
2245         if (len > buffer->len)
2246                 return NULL;
2247         return buffer->ptr;
2248 }
2249
2250 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2251                                 unsigned len)
2252 {
2253         void *ret;
2254
2255         ret = pv_mmu_peek_buffer(buffer, len);
2256         if (!ret)
2257                 return ret;
2258         buffer->ptr += len;
2259         buffer->len -= len;
2260         buffer->processed += len;
2261         return ret;
2262 }
2263
2264 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2265                              gpa_t addr, gpa_t value)
2266 {
2267         int bytes = 8;
2268         int r;
2269
2270         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2271                 bytes = 4;
2272
2273         r = mmu_topup_memory_caches(vcpu);
2274         if (r)
2275                 return r;
2276
2277         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2278                 return -EFAULT;
2279
2280         return 1;
2281 }
2282
2283 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2284 {
2285         kvm_x86_ops->tlb_flush(vcpu);
2286         return 1;
2287 }
2288
2289 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2290 {
2291         spin_lock(&vcpu->kvm->mmu_lock);
2292         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2293         spin_unlock(&vcpu->kvm->mmu_lock);
2294         return 1;
2295 }
2296
2297 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2298                              struct kvm_pv_mmu_op_buffer *buffer)
2299 {
2300         struct kvm_mmu_op_header *header;
2301
2302         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2303         if (!header)
2304                 return 0;
2305         switch (header->op) {
2306         case KVM_MMU_OP_WRITE_PTE: {
2307                 struct kvm_mmu_op_write_pte *wpte;
2308
2309                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2310                 if (!wpte)
2311                         return 0;
2312                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2313                                         wpte->pte_val);
2314         }
2315         case KVM_MMU_OP_FLUSH_TLB: {
2316                 struct kvm_mmu_op_flush_tlb *ftlb;
2317
2318                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2319                 if (!ftlb)
2320                         return 0;
2321                 return kvm_pv_mmu_flush_tlb(vcpu);
2322         }
2323         case KVM_MMU_OP_RELEASE_PT: {
2324                 struct kvm_mmu_op_release_pt *rpt;
2325
2326                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2327                 if (!rpt)
2328                         return 0;
2329                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2330         }
2331         default: return 0;
2332         }
2333 }
2334
2335 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2336                   gpa_t addr, unsigned long *ret)
2337 {
2338         int r;
2339         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2340
2341         buffer->ptr = buffer->buf;
2342         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
2343         buffer->processed = 0;
2344
2345         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
2346         if (r)
2347                 goto out;
2348
2349         while (buffer->len) {
2350                 r = kvm_pv_mmu_op_one(vcpu, buffer);
2351                 if (r < 0)
2352                         goto out;
2353                 if (r == 0)
2354                         break;
2355         }
2356
2357         r = 1;
2358 out:
2359         *ret = buffer->processed;
2360         return r;
2361 }
2362
2363 #ifdef AUDIT
2364
2365 static const char *audit_msg;
2366
2367 static gva_t canonicalize(gva_t gva)
2368 {
2369 #ifdef CONFIG_X86_64
2370         gva = (long long)(gva << 16) >> 16;
2371 #endif
2372         return gva;
2373 }
2374
2375 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2376                                 gva_t va, int level)
2377 {
2378         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2379         int i;
2380         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2381
2382         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2383                 u64 ent = pt[i];
2384
2385                 if (ent == shadow_trap_nonpresent_pte)
2386                         continue;
2387
2388                 va = canonicalize(va);
2389                 if (level > 1) {
2390                         if (ent == shadow_notrap_nonpresent_pte)
2391                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2392                                        " in nonleaf level: levels %d gva %lx"
2393                                        " level %d pte %llx\n", audit_msg,
2394                                        vcpu->arch.mmu.root_level, va, level, ent);
2395
2396                         audit_mappings_page(vcpu, ent, va, level - 1);
2397                 } else {
2398                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2399                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2400
2401                         if (is_shadow_present_pte(ent)
2402                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2403                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2404                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2405                                        audit_msg, vcpu->arch.mmu.root_level,
2406                                        va, gpa, hpa, ent,
2407                                        is_shadow_present_pte(ent));
2408                         else if (ent == shadow_notrap_nonpresent_pte
2409                                  && !is_error_hpa(hpa))
2410                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2411                                        " valid guest gva %lx\n", audit_msg, va);
2412                         kvm_release_pfn_clean(pfn);
2413
2414                 }
2415         }
2416 }
2417
2418 static void audit_mappings(struct kvm_vcpu *vcpu)
2419 {
2420         unsigned i;
2421
2422         if (vcpu->arch.mmu.root_level == 4)
2423                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2424         else
2425                 for (i = 0; i < 4; ++i)
2426                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2427                                 audit_mappings_page(vcpu,
2428                                                     vcpu->arch.mmu.pae_root[i],
2429                                                     i << 30,
2430                                                     2);
2431 }
2432
2433 static int count_rmaps(struct kvm_vcpu *vcpu)
2434 {
2435         int nmaps = 0;
2436         int i, j, k;
2437
2438         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2439                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2440                 struct kvm_rmap_desc *d;
2441
2442                 for (j = 0; j < m->npages; ++j) {
2443                         unsigned long *rmapp = &m->rmap[j];
2444
2445                         if (!*rmapp)
2446                                 continue;
2447                         if (!(*rmapp & 1)) {
2448                                 ++nmaps;
2449                                 continue;
2450                         }
2451                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2452                         while (d) {
2453                                 for (k = 0; k < RMAP_EXT; ++k)
2454                                         if (d->shadow_ptes[k])
2455                                                 ++nmaps;
2456                                         else
2457                                                 break;
2458                                 d = d->more;
2459                         }
2460                 }
2461         }
2462         return nmaps;
2463 }
2464
2465 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2466 {
2467         int nmaps = 0;
2468         struct kvm_mmu_page *sp;
2469         int i;
2470
2471         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2472                 u64 *pt = sp->spt;
2473
2474                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2475                         continue;
2476
2477                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2478                         u64 ent = pt[i];
2479
2480                         if (!(ent & PT_PRESENT_MASK))
2481                                 continue;
2482                         if (!(ent & PT_WRITABLE_MASK))
2483                                 continue;
2484                         ++nmaps;
2485                 }
2486         }
2487         return nmaps;
2488 }
2489
2490 static void audit_rmap(struct kvm_vcpu *vcpu)
2491 {
2492         int n_rmap = count_rmaps(vcpu);
2493         int n_actual = count_writable_mappings(vcpu);
2494
2495         if (n_rmap != n_actual)
2496                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2497                        __func__, audit_msg, n_rmap, n_actual);
2498 }
2499
2500 static void audit_write_protection(struct kvm_vcpu *vcpu)
2501 {
2502         struct kvm_mmu_page *sp;
2503         struct kvm_memory_slot *slot;
2504         unsigned long *rmapp;
2505         gfn_t gfn;
2506
2507         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2508                 if (sp->role.metaphysical)
2509                         continue;
2510
2511                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2512                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2513                 rmapp = &slot->rmap[gfn - slot->base_gfn];
2514                 if (*rmapp)
2515                         printk(KERN_ERR "%s: (%s) shadow page has writable"
2516                                " mappings: gfn %lx role %x\n",
2517                                __func__, audit_msg, sp->gfn,
2518                                sp->role.word);
2519         }
2520 }
2521
2522 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2523 {
2524         int olddbg = dbg;
2525
2526         dbg = 0;
2527         audit_msg = msg;
2528         audit_rmap(vcpu);
2529         audit_write_protection(vcpu);
2530         audit_mappings(vcpu);
2531         dbg = olddbg;
2532 }
2533
2534 #endif