drivers/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 #include <linux/types.h>
  20 #include <linux/string.h>
  21 #include <asm/page.h>
  22 #include <linux/mm.h>
  23 #include <linux/highmem.h>
  24 #include <linux/module.h>
  25
  26 #include "vmx.h"
  27 #include "kvm.h"
  28
  29 #undef MMU_DEBUG
  30
  31 #undef AUDIT
  32
  33 #ifdef AUDIT
  34 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  35 #else
  36 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  37 #endif
  38
  39 #ifdef MMU_DEBUG
  40
  41 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  42 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  43
  44 #else
  45
  46 #define pgprintk(x...) do { } while (0)
  47 #define rmap_printk(x...) do { } while (0)
  48
  49 #endif
  50
  51 #if defined(MMU_DEBUG) || defined(AUDIT)
  52 static int dbg = 1;
  53 #endif
  54
  55 #ifndef MMU_DEBUG
  56 #define ASSERT(x) do { } while (0)
  57 #else
  58 #define ASSERT(x)                                                       \
  59         if (!(x)) {                                                     \
  60                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  61                        __FILE__, __LINE__, #x);                         \
  62         }
  63 #endif
  64
  65 #define PT64_PT_BITS 9
  66 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  67 #define PT32_PT_BITS 10
  68 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  69
  70 #define PT_WRITABLE_SHIFT 1
  71
  72 #define PT_PRESENT_MASK (1ULL << 0)
  73 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  74 #define PT_USER_MASK (1ULL << 2)
  75 #define PT_PWT_MASK (1ULL << 3)
  76 #define PT_PCD_MASK (1ULL << 4)
  77 #define PT_ACCESSED_MASK (1ULL << 5)
  78 #define PT_DIRTY_MASK (1ULL << 6)
  79 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  80 #define PT_PAT_MASK (1ULL << 7)
  81 #define PT_GLOBAL_MASK (1ULL << 8)
  82 #define PT64_NX_MASK (1ULL << 63)
  83
  84 #define PT_PAT_SHIFT 7
  85 #define PT_DIR_PAT_SHIFT 12
  86 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  87
  88 #define PT32_DIR_PSE36_SIZE 4
  89 #define PT32_DIR_PSE36_SHIFT 13
  90 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  91
  92
  93 #define PT32_PTE_COPY_MASK \
  94         (PT_PRESENT_MASK | PT_ACCESSED_MASK | PT_DIRTY_MASK | PT_GLOBAL_MASK)
  95
  96 #define PT64_PTE_COPY_MASK (PT64_NX_MASK | PT32_PTE_COPY_MASK)
  97
  98 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  99 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
 100
 101 #define PT_SHADOW_PS_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 102 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 103
 104 #define PT_SHADOW_WRITABLE_SHIFT (PT_FIRST_AVAIL_BITS_SHIFT + 1)
 105 #define PT_SHADOW_WRITABLE_MASK (1ULL << PT_SHADOW_WRITABLE_SHIFT)
 106
 107 #define PT_SHADOW_USER_SHIFT (PT_SHADOW_WRITABLE_SHIFT + 1)
 108 #define PT_SHADOW_USER_MASK (1ULL << (PT_SHADOW_USER_SHIFT))
 109
 110 #define PT_SHADOW_BITS_OFFSET (PT_SHADOW_WRITABLE_SHIFT - PT_WRITABLE_SHIFT)
 111
 112 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 113
 114 #define PT64_LEVEL_BITS 9
 115
 116 #define PT64_LEVEL_SHIFT(level) \
 117                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
 118
 119 #define PT64_LEVEL_MASK(level) \
 120                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 121
 122 #define PT64_INDEX(address, level)\
 123         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 124
 125
 126 #define PT32_LEVEL_BITS 10
 127
 128 #define PT32_LEVEL_SHIFT(level) \
 129                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
 130
 131 #define PT32_LEVEL_MASK(level) \
 132                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 133
 134 #define PT32_INDEX(address, level)\
 135         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 136
 137
 138 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 139 #define PT64_DIR_BASE_ADDR_MASK \
 140         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 141
 142 #define PT32_BASE_ADDR_MASK PAGE_MASK
 143 #define PT32_DIR_BASE_ADDR_MASK \
 144         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 145
 146
 147 #define PFERR_PRESENT_MASK (1U << 0)
 148 #define PFERR_WRITE_MASK (1U << 1)
 149 #define PFERR_USER_MASK (1U << 2)
 150 #define PFERR_FETCH_MASK (1U << 4)
 151
 152 #define PT64_ROOT_LEVEL 4
 153 #define PT32_ROOT_LEVEL 2
 154 #define PT32E_ROOT_LEVEL 3
 155
 156 #define PT_DIRECTORY_LEVEL 2
 157 #define PT_PAGE_TABLE_LEVEL 1
 158
 159 #define RMAP_EXT 4
 160
 161 struct kvm_rmap_desc {
 162         u64 *shadow_ptes[RMAP_EXT];
 163         struct kvm_rmap_desc *more;
 164 };
 165
 166 static struct kmem_cache *pte_chain_cache;
 167 static struct kmem_cache *rmap_desc_cache;
 168 static struct kmem_cache *mmu_page_cache;
 169 static struct kmem_cache *mmu_page_header_cache;
 170
 171 static int is_write_protection(struct kvm_vcpu *vcpu)
 172 {
 173         return vcpu->cr0 & CR0_WP_MASK;
 174 }
 175
 176 static int is_cpuid_PSE36(void)
 177 {
 178         return 1;
 179 }
 180
 181 static int is_nx(struct kvm_vcpu *vcpu)
 182 {
 183         return vcpu->shadow_efer & EFER_NX;
 184 }
 185
 186 static int is_present_pte(unsigned long pte)
 187 {
 188         return pte & PT_PRESENT_MASK;
 189 }
 190
 191 static int is_writeble_pte(unsigned long pte)
 192 {
 193         return pte & PT_WRITABLE_MASK;
 194 }
 195
 196 static int is_io_pte(unsigned long pte)
 197 {
 198         return pte & PT_SHADOW_IO_MARK;
 199 }
 200
 201 static int is_rmap_pte(u64 pte)
 202 {
 203         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 204                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 205 }
 206
 207 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 208                                   struct kmem_cache *base_cache, int min,
 209                                   gfp_t gfp_flags)
 210 {
 211         void *obj;
 212
 213         if (cache->nobjs >= min)
 214                 return 0;
 215         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 216                 obj = kmem_cache_zalloc(base_cache, gfp_flags);
 217                 if (!obj)
 218                         return -ENOMEM;
 219                 cache->objects[cache->nobjs++] = obj;
 220         }
 221         return 0;
 222 }
 223
 224 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 225 {
 226         while (mc->nobjs)
 227                 kfree(mc->objects[--mc->nobjs]);
 228 }
 229
 230 static int __mmu_topup_memory_caches(struct kvm_vcpu *vcpu, gfp_t gfp_flags)
 231 {
 232         int r;
 233
 234         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 235                                    pte_chain_cache, 4, gfp_flags);
 236         if (r)
 237                 goto out;
 238         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 239                                    rmap_desc_cache, 1, gfp_flags);
 240         if (r)
 241                 goto out;
 242         r = mmu_topup_memory_cache(&vcpu->mmu_page_cache,
 243                                    mmu_page_cache, 4, gfp_flags);
 244         if (r)
 245                 goto out;
 246         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
 247                                    mmu_page_header_cache, 4, gfp_flags);
 248 out:
 249         return r;
 250 }
 251
 252 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 253 {
 254         int r;
 255
 256         r = __mmu_topup_memory_caches(vcpu, GFP_NOWAIT);
 257         if (r < 0) {
 258                 spin_unlock(&vcpu->kvm->lock);
 259                 kvm_arch_ops->vcpu_put(vcpu);
 260                 r = __mmu_topup_memory_caches(vcpu, GFP_KERNEL);
 261                 kvm_arch_ops->vcpu_load(vcpu);
 262                 spin_lock(&vcpu->kvm->lock);
 263         }
 264         return r;
 265 }
 266
 267 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 268 {
 269         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 270         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 271         mmu_free_memory_cache(&vcpu->mmu_page_cache);
 272         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
 273 }
 274
 275 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 276                                     size_t size)
 277 {
 278         void *p;
 279
 280         BUG_ON(!mc->nobjs);
 281         p = mc->objects[--mc->nobjs];
 282         memset(p, 0, size);
 283         return p;
 284 }
 285
 286 static void mmu_memory_cache_free(struct kvm_mmu_memory_cache *mc, void *obj)
 287 {
 288         if (mc->nobjs < KVM_NR_MEM_OBJS)
 289                 mc->objects[mc->nobjs++] = obj;
 290         else
 291                 kfree(obj);
 292 }
 293
 294 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 295 {
 296         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 297                                       sizeof(struct kvm_pte_chain));
 298 }
 299
 300 static void mmu_free_pte_chain(struct kvm_vcpu *vcpu,
 301                                struct kvm_pte_chain *pc)
 302 {
 303         mmu_memory_cache_free(&vcpu->mmu_pte_chain_cache, pc);
 304 }
 305
 306 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 307 {
 308         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 309                                       sizeof(struct kvm_rmap_desc));
 310 }
 311
 312 static void mmu_free_rmap_desc(struct kvm_vcpu *vcpu,
 313                                struct kvm_rmap_desc *rd)
 314 {
 315         mmu_memory_cache_free(&vcpu->mmu_rmap_desc_cache, rd);
 316 }
 317
 318 /*
 319  * Reverse mapping data structures:
 320  *
 321  * If page->private bit zero is zero, then page->private points to the
 322  * shadow page table entry that points to page_address(page).
 323  *
 324  * If page->private bit zero is one, (then page->private & ~1) points
 325  * to a struct kvm_rmap_desc containing more mappings.
 326  */
 327 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 328 {
 329         struct page *page;
 330         struct kvm_rmap_desc *desc;
 331         int i;
 332
 333         if (!is_rmap_pte(*spte))
 334                 return;
 335         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 336         if (!page_private(page)) {
 337                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 338                 set_page_private(page,(unsigned long)spte);
 339         } else if (!(page_private(page) & 1)) {
 340                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 341                 desc = mmu_alloc_rmap_desc(vcpu);
 342                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 343                 desc->shadow_ptes[1] = spte;
 344                 set_page_private(page,(unsigned long)desc | 1);
 345         } else {
 346                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 347                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 348                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 349                         desc = desc->more;
 350                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 351                         desc->more = mmu_alloc_rmap_desc(vcpu);
 352                         desc = desc->more;
 353                 }
 354                 for (i = 0; desc->shadow_ptes[i]; ++i)
 355                         ;
 356                 desc->shadow_ptes[i] = spte;
 357         }
 358 }
 359
 360 static void rmap_desc_remove_entry(struct kvm_vcpu *vcpu,
 361                                    struct page *page,
 362                                    struct kvm_rmap_desc *desc,
 363                                    int i,
 364                                    struct kvm_rmap_desc *prev_desc)
 365 {
 366         int j;
 367
 368         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 369                 ;
 370         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 371         desc->shadow_ptes[j] = NULL;
 372         if (j != 0)
 373                 return;
 374         if (!prev_desc && !desc->more)
 375                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 376         else
 377                 if (prev_desc)
 378                         prev_desc->more = desc->more;
 379                 else
 380                         set_page_private(page,(unsigned long)desc->more | 1);
 381         mmu_free_rmap_desc(vcpu, desc);
 382 }
 383
 384 static void rmap_remove(struct kvm_vcpu *vcpu, u64 *spte)
 385 {
 386         struct page *page;
 387         struct kvm_rmap_desc *desc;
 388         struct kvm_rmap_desc *prev_desc;
 389         int i;
 390
 391         if (!is_rmap_pte(*spte))
 392                 return;
 393         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 394         if (!page_private(page)) {
 395                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 396                 BUG();
 397         } else if (!(page_private(page) & 1)) {
 398                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 399                 if ((u64 *)page_private(page) != spte) {
 400                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 401                                spte, *spte);
 402                         BUG();
 403                 }
 404                 set_page_private(page,0);
 405         } else {
 406                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 407                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 408                 prev_desc = NULL;
 409                 while (desc) {
 410                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 411                                 if (desc->shadow_ptes[i] == spte) {
 412                                         rmap_desc_remove_entry(vcpu, page,
 413                                                                desc, i,
 414                                                                prev_desc);
 415                                         return;
 416                                 }
 417                         prev_desc = desc;
 418                         desc = desc->more;
 419                 }
 420                 BUG();
 421         }
 422 }
 423
 424 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 425 {
 426         struct kvm *kvm = vcpu->kvm;
 427         struct page *page;
 428         struct kvm_rmap_desc *desc;
 429         u64 *spte;
 430
 431         page = gfn_to_page(kvm, gfn);
 432         BUG_ON(!page);
 433
 434         while (page_private(page)) {
 435                 if (!(page_private(page) & 1))
 436                         spte = (u64 *)page_private(page);
 437                 else {
 438                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 439                         spte = desc->shadow_ptes[0];
 440                 }
 441                 BUG_ON(!spte);
 442                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 443                        != page_to_pfn(page));
 444                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 445                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 446                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 447                 rmap_remove(vcpu, spte);
 448                 kvm_arch_ops->tlb_flush(vcpu);
 449                 *spte &= ~(u64)PT_WRITABLE_MASK;
 450         }
 451 }
 452
 453 #ifdef MMU_DEBUG
 454 static int is_empty_shadow_page(u64 *spt)
 455 {
 456         u64 *pos;
 457         u64 *end;
 458
 459         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 460                 if (*pos != 0) {
 461                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 462                                pos, *pos);
 463                         return 0;
 464                 }
 465         return 1;
 466 }
 467 #endif
 468
 469 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu,
 470                               struct kvm_mmu_page *page_head)
 471 {
 472         ASSERT(is_empty_shadow_page(page_head->spt));
 473         list_del(&page_head->link);
 474         mmu_memory_cache_free(&vcpu->mmu_page_cache, page_head->spt);
 475         mmu_memory_cache_free(&vcpu->mmu_page_header_cache, page_head);
 476         ++vcpu->kvm->n_free_mmu_pages;
 477 }
 478
 479 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 480 {
 481         return gfn;
 482 }
 483
 484 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 485                                                u64 *parent_pte)
 486 {
 487         struct kvm_mmu_page *page;
 488
 489         if (!vcpu->kvm->n_free_mmu_pages)
 490                 return NULL;
 491
 492         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 493                                       sizeof *page);
 494         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 495         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 496         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 497         ASSERT(is_empty_shadow_page(page->spt));
 498         page->slot_bitmap = 0;
 499         page->multimapped = 0;
 500         page->parent_pte = parent_pte;
 501         --vcpu->kvm->n_free_mmu_pages;
 502         return page;
 503 }
 504
 505 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 506                                     struct kvm_mmu_page *page, u64 *parent_pte)
 507 {
 508         struct kvm_pte_chain *pte_chain;
 509         struct hlist_node *node;
 510         int i;
 511
 512         if (!parent_pte)
 513                 return;
 514         if (!page->multimapped) {
 515                 u64 *old = page->parent_pte;
 516
 517                 if (!old) {
 518                         page->parent_pte = parent_pte;
 519                         return;
 520                 }
 521                 page->multimapped = 1;
 522                 pte_chain = mmu_alloc_pte_chain(vcpu);
 523                 INIT_HLIST_HEAD(&page->parent_ptes);
 524                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 525                 pte_chain->parent_ptes[0] = old;
 526         }
 527         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 528                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 529                         continue;
 530                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 531                         if (!pte_chain->parent_ptes[i]) {
 532                                 pte_chain->parent_ptes[i] = parent_pte;
 533                                 return;
 534                         }
 535         }
 536         pte_chain = mmu_alloc_pte_chain(vcpu);
 537         BUG_ON(!pte_chain);
 538         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 539         pte_chain->parent_ptes[0] = parent_pte;
 540 }
 541
 542 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
 543                                        struct kvm_mmu_page *page,
 544                                        u64 *parent_pte)
 545 {
 546         struct kvm_pte_chain *pte_chain;
 547         struct hlist_node *node;
 548         int i;
 549
 550         if (!page->multimapped) {
 551                 BUG_ON(page->parent_pte != parent_pte);
 552                 page->parent_pte = NULL;
 553                 return;
 554         }
 555         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 556                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 557                         if (!pte_chain->parent_ptes[i])
 558                                 break;
 559                         if (pte_chain->parent_ptes[i] != parent_pte)
 560                                 continue;
 561                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 562                                 && pte_chain->parent_ptes[i + 1]) {
 563                                 pte_chain->parent_ptes[i]
 564                                         = pte_chain->parent_ptes[i + 1];
 565                                 ++i;
 566                         }
 567                         pte_chain->parent_ptes[i] = NULL;
 568                         if (i == 0) {
 569                                 hlist_del(&pte_chain->link);
 570                                 mmu_free_pte_chain(vcpu, pte_chain);
 571                                 if (hlist_empty(&page->parent_ptes)) {
 572                                         page->multimapped = 0;
 573                                         page->parent_pte = NULL;
 574                                 }
 575                         }
 576                         return;
 577                 }
 578         BUG();
 579 }
 580
 581 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 582                                                 gfn_t gfn)
 583 {
 584         unsigned index;
 585         struct hlist_head *bucket;
 586         struct kvm_mmu_page *page;
 587         struct hlist_node *node;
 588
 589         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 590         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 591         bucket = &vcpu->kvm->mmu_page_hash[index];
 592         hlist_for_each_entry(page, node, bucket, hash_link)
 593                 if (page->gfn == gfn && !page->role.metaphysical) {
 594                         pgprintk("%s: found role %x\n",
 595                                  __FUNCTION__, page->role.word);
 596                         return page;
 597                 }
 598         return NULL;
 599 }
 600
 601 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 602                                              gfn_t gfn,
 603                                              gva_t gaddr,
 604                                              unsigned level,
 605                                              int metaphysical,
 606                                              unsigned hugepage_access,
 607                                              u64 *parent_pte)
 608 {
 609         union kvm_mmu_page_role role;
 610         unsigned index;
 611         unsigned quadrant;
 612         struct hlist_head *bucket;
 613         struct kvm_mmu_page *page;
 614         struct hlist_node *node;
 615
 616         role.word = 0;
 617         role.glevels = vcpu->mmu.root_level;
 618         role.level = level;
 619         role.metaphysical = metaphysical;
 620         role.hugepage_access = hugepage_access;
 621         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 622                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 623                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 624                 role.quadrant = quadrant;
 625         }
 626         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 627                  gfn, role.word);
 628         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 629         bucket = &vcpu->kvm->mmu_page_hash[index];
 630         hlist_for_each_entry(page, node, bucket, hash_link)
 631                 if (page->gfn == gfn && page->role.word == role.word) {
 632                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 633                         pgprintk("%s: found\n", __FUNCTION__);
 634                         return page;
 635                 }
 636         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 637         if (!page)
 638                 return page;
 639         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 640         page->gfn = gfn;
 641         page->role = role;
 642         hlist_add_head(&page->hash_link, bucket);
 643         if (!metaphysical)
 644                 rmap_write_protect(vcpu, gfn);
 645         return page;
 646 }
 647
 648 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 649                                          struct kvm_mmu_page *page)
 650 {
 651         unsigned i;
 652         u64 *pt;
 653         u64 ent;
 654
 655         pt = page->spt;
 656
 657         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 658                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 659                         if (pt[i] & PT_PRESENT_MASK)
 660                                 rmap_remove(vcpu, &pt[i]);
 661                         pt[i] = 0;
 662                 }
 663                 kvm_arch_ops->tlb_flush(vcpu);
 664                 return;
 665         }
 666
 667         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 668                 ent = pt[i];
 669
 670                 pt[i] = 0;
 671                 if (!(ent & PT_PRESENT_MASK))
 672                         continue;
 673                 ent &= PT64_BASE_ADDR_MASK;
 674                 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
 675         }
 676 }
 677
 678 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 679                              struct kvm_mmu_page *page,
 680                              u64 *parent_pte)
 681 {
 682         mmu_page_remove_parent_pte(vcpu, page, parent_pte);
 683 }
 684
 685 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 686                              struct kvm_mmu_page *page)
 687 {
 688         u64 *parent_pte;
 689
 690         while (page->multimapped || page->parent_pte) {
 691                 if (!page->multimapped)
 692                         parent_pte = page->parent_pte;
 693                 else {
 694                         struct kvm_pte_chain *chain;
 695
 696                         chain = container_of(page->parent_ptes.first,
 697                                              struct kvm_pte_chain, link);
 698                         parent_pte = chain->parent_ptes[0];
 699                 }
 700                 BUG_ON(!parent_pte);
 701                 kvm_mmu_put_page(vcpu, page, parent_pte);
 702                 *parent_pte = 0;
 703         }
 704         kvm_mmu_page_unlink_children(vcpu, page);
 705         if (!page->root_count) {
 706                 hlist_del(&page->hash_link);
 707                 kvm_mmu_free_page(vcpu, page);
 708         } else
 709                 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 710 }
 711
 712 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 713 {
 714         unsigned index;
 715         struct hlist_head *bucket;
 716         struct kvm_mmu_page *page;
 717         struct hlist_node *node, *n;
 718         int r;
 719
 720         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 721         r = 0;
 722         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 723         bucket = &vcpu->kvm->mmu_page_hash[index];
 724         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 725                 if (page->gfn == gfn && !page->role.metaphysical) {
 726                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 727                                  page->role.word);
 728                         kvm_mmu_zap_page(vcpu, page);
 729                         r = 1;
 730                 }
 731         return r;
 732 }
 733
 734 static void mmu_unshadow(struct kvm_vcpu *vcpu, gfn_t gfn)
 735 {
 736         struct kvm_mmu_page *page;
 737
 738         while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
 739                 pgprintk("%s: zap %lx %x\n",
 740                          __FUNCTION__, gfn, page->role.word);
 741                 kvm_mmu_zap_page(vcpu, page);
 742         }
 743 }
 744
 745 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 746 {
 747         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 748         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 749
 750         __set_bit(slot, &page_head->slot_bitmap);
 751 }
 752
 753 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 754 {
 755         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 756
 757         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 758 }
 759
 760 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 761 {
 762         struct page *page;
 763
 764         ASSERT((gpa & HPA_ERR_MASK) == 0);
 765         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 766         if (!page)
 767                 return gpa | HPA_ERR_MASK;
 768         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 769                 | (gpa & (PAGE_SIZE-1));
 770 }
 771
 772 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 773 {
 774         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 775
 776         if (gpa == UNMAPPED_GVA)
 777                 return UNMAPPED_GVA;
 778         return gpa_to_hpa(vcpu, gpa);
 779 }
 780
 781 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 782 {
 783         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 784
 785         if (gpa == UNMAPPED_GVA)
 786                 return NULL;
 787         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 788 }
 789
 790 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 791 {
 792 }
 793
 794 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 795 {
 796         int level = PT32E_ROOT_LEVEL;
 797         hpa_t table_addr = vcpu->mmu.root_hpa;
 798
 799         for (; ; level--) {
 800                 u32 index = PT64_INDEX(v, level);
 801                 u64 *table;
 802                 u64 pte;
 803
 804                 ASSERT(VALID_PAGE(table_addr));
 805                 table = __va(table_addr);
 806
 807                 if (level == 1) {
 808                         pte = table[index];
 809                         if (is_present_pte(pte) && is_writeble_pte(pte))
 810                                 return 0;
 811                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 812                         page_header_update_slot(vcpu->kvm, table, v);
 813                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 814                                                                 PT_USER_MASK;
 815                         rmap_add(vcpu, &table[index]);
 816                         return 0;
 817                 }
 818
 819                 if (table[index] == 0) {
 820                         struct kvm_mmu_page *new_table;
 821                         gfn_t pseudo_gfn;
 822
 823                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 824                                 >> PAGE_SHIFT;
 825                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 826                                                      v, level - 1,
 827                                                      1, 0, &table[index]);
 828                         if (!new_table) {
 829                                 pgprintk("nonpaging_map: ENOMEM\n");
 830                                 return -ENOMEM;
 831                         }
 832
 833                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 834                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 835                 }
 836                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 837         }
 838 }
 839
 840 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 841 {
 842         int i;
 843         struct kvm_mmu_page *page;
 844
 845 #ifdef CONFIG_X86_64
 846         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 847                 hpa_t root = vcpu->mmu.root_hpa;
 848
 849                 ASSERT(VALID_PAGE(root));
 850                 page = page_header(root);
 851                 --page->root_count;
 852                 vcpu->mmu.root_hpa = INVALID_PAGE;
 853                 return;
 854         }
 855 #endif
 856         for (i = 0; i < 4; ++i) {
 857                 hpa_t root = vcpu->mmu.pae_root[i];
 858
 859                 if (root) {
 860                         ASSERT(VALID_PAGE(root));
 861                         root &= PT64_BASE_ADDR_MASK;
 862                         page = page_header(root);
 863                         --page->root_count;
 864                 }
 865                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 866         }
 867         vcpu->mmu.root_hpa = INVALID_PAGE;
 868 }
 869
 870 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 871 {
 872         int i;
 873         gfn_t root_gfn;
 874         struct kvm_mmu_page *page;
 875
 876         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 877
 878 #ifdef CONFIG_X86_64
 879         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 880                 hpa_t root = vcpu->mmu.root_hpa;
 881
 882                 ASSERT(!VALID_PAGE(root));
 883                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 884                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 885                 root = __pa(page->spt);
 886                 ++page->root_count;
 887                 vcpu->mmu.root_hpa = root;
 888                 return;
 889         }
 890 #endif
 891         for (i = 0; i < 4; ++i) {
 892                 hpa_t root = vcpu->mmu.pae_root[i];
 893
 894                 ASSERT(!VALID_PAGE(root));
 895                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 896                         if (!is_present_pte(vcpu->pdptrs[i])) {
 897                                 vcpu->mmu.pae_root[i] = 0;
 898                                 continue;
 899                         }
 900                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 901                 } else if (vcpu->mmu.root_level == 0)
 902                         root_gfn = 0;
 903                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 904                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 905                                         0, NULL);
 906                 root = __pa(page->spt);
 907                 ++page->root_count;
 908                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 909         }
 910         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 911 }
 912
 913 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 914 {
 915         return vaddr;
 916 }
 917
 918 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 919                                u32 error_code)
 920 {
 921         gpa_t addr = gva;
 922         hpa_t paddr;
 923         int r;
 924
 925         r = mmu_topup_memory_caches(vcpu);
 926         if (r)
 927                 return r;
 928
 929         ASSERT(vcpu);
 930         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 931
 932
 933         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 934
 935         if (is_error_hpa(paddr))
 936                 return 1;
 937
 938         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 939 }
 940
 941 static void nonpaging_free(struct kvm_vcpu *vcpu)
 942 {
 943         mmu_free_roots(vcpu);
 944 }
 945
 946 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 947 {
 948         struct kvm_mmu *context = &vcpu->mmu;
 949
 950         context->new_cr3 = nonpaging_new_cr3;
 951         context->page_fault = nonpaging_page_fault;
 952         context->gva_to_gpa = nonpaging_gva_to_gpa;
 953         context->free = nonpaging_free;
 954         context->root_level = 0;
 955         context->shadow_root_level = PT32E_ROOT_LEVEL;
 956         mmu_alloc_roots(vcpu);
 957         ASSERT(VALID_PAGE(context->root_hpa));
 958         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 959         return 0;
 960 }
 961
 962 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 963 {
 964         ++vcpu->stat.tlb_flush;
 965         kvm_arch_ops->tlb_flush(vcpu);
 966 }
 967
 968 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 969 {
 970         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 971         mmu_free_roots(vcpu);
 972         if (unlikely(vcpu->kvm->n_free_mmu_pages < KVM_MIN_FREE_MMU_PAGES))
 973                 kvm_mmu_free_some_pages(vcpu);
 974         mmu_alloc_roots(vcpu);
 975         kvm_mmu_flush_tlb(vcpu);
 976         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 977 }
 978
 979 static void inject_page_fault(struct kvm_vcpu *vcpu,
 980                               u64 addr,
 981                               u32 err_code)
 982 {
 983         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
 984 }
 985
 986 static void paging_free(struct kvm_vcpu *vcpu)
 987 {
 988         nonpaging_free(vcpu);
 989 }
 990
 991 #define PTTYPE 64
 992 #include "paging_tmpl.h"
 993 #undef PTTYPE
 994
 995 #define PTTYPE 32
 996 #include "paging_tmpl.h"
 997 #undef PTTYPE
 998
 999 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1000 {
1001         struct kvm_mmu *context = &vcpu->mmu;
1002
1003         ASSERT(is_pae(vcpu));
1004         context->new_cr3 = paging_new_cr3;
1005         context->page_fault = paging64_page_fault;
1006         context->gva_to_gpa = paging64_gva_to_gpa;
1007         context->free = paging_free;
1008         context->root_level = level;
1009         context->shadow_root_level = level;
1010         mmu_alloc_roots(vcpu);
1011         ASSERT(VALID_PAGE(context->root_hpa));
1012         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1013                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1014         return 0;
1015 }
1016
1017 static int paging64_init_context(struct kvm_vcpu *vcpu)
1018 {
1019         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1020 }
1021
1022 static int paging32_init_context(struct kvm_vcpu *vcpu)
1023 {
1024         struct kvm_mmu *context = &vcpu->mmu;
1025
1026         context->new_cr3 = paging_new_cr3;
1027         context->page_fault = paging32_page_fault;
1028         context->gva_to_gpa = paging32_gva_to_gpa;
1029         context->free = paging_free;
1030         context->root_level = PT32_ROOT_LEVEL;
1031         context->shadow_root_level = PT32E_ROOT_LEVEL;
1032         mmu_alloc_roots(vcpu);
1033         ASSERT(VALID_PAGE(context->root_hpa));
1034         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1035                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1036         return 0;
1037 }
1038
1039 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1040 {
1041         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1042 }
1043
1044 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1045 {
1046         ASSERT(vcpu);
1047         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1048
1049         mmu_topup_memory_caches(vcpu);
1050         if (!is_paging(vcpu))
1051                 return nonpaging_init_context(vcpu);
1052         else if (is_long_mode(vcpu))
1053                 return paging64_init_context(vcpu);
1054         else if (is_pae(vcpu))
1055                 return paging32E_init_context(vcpu);
1056         else
1057                 return paging32_init_context(vcpu);
1058 }
1059
1060 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1061 {
1062         ASSERT(vcpu);
1063         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1064                 vcpu->mmu.free(vcpu);
1065                 vcpu->mmu.root_hpa = INVALID_PAGE;
1066         }
1067 }
1068
1069 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1070 {
1071         int r;
1072
1073         destroy_kvm_mmu(vcpu);
1074         r = init_kvm_mmu(vcpu);
1075         if (r < 0)
1076                 goto out;
1077         r = mmu_topup_memory_caches(vcpu);
1078 out:
1079         return r;
1080 }
1081
1082 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1083                                   struct kvm_mmu_page *page,
1084                                   u64 *spte)
1085 {
1086         u64 pte;
1087         struct kvm_mmu_page *child;
1088
1089         pte = *spte;
1090         if (is_present_pte(pte)) {
1091                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1092                         rmap_remove(vcpu, spte);
1093                 else {
1094                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1095                         mmu_page_remove_parent_pte(vcpu, child, spte);
1096                 }
1097         }
1098         *spte = 0;
1099 }
1100
1101 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1102                                   struct kvm_mmu_page *page,
1103                                   u64 *spte,
1104                                   const void *new, int bytes)
1105 {
1106         if (page->role.level != PT_PAGE_TABLE_LEVEL)
1107                 return;
1108
1109         if (page->role.glevels == PT32_ROOT_LEVEL)
1110                 paging32_update_pte(vcpu, page, spte, new, bytes);
1111         else
1112                 paging64_update_pte(vcpu, page, spte, new, bytes);
1113 }
1114
1115 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1116                        const u8 *old, const u8 *new, int bytes)
1117 {
1118         gfn_t gfn = gpa >> PAGE_SHIFT;
1119         struct kvm_mmu_page *page;
1120         struct hlist_node *node, *n;
1121         struct hlist_head *bucket;
1122         unsigned index;
1123         u64 *spte;
1124         unsigned offset = offset_in_page(gpa);
1125         unsigned pte_size;
1126         unsigned page_offset;
1127         unsigned misaligned;
1128         unsigned quadrant;
1129         int level;
1130         int flooded = 0;
1131         int npte;
1132
1133         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1134         if (gfn == vcpu->last_pt_write_gfn) {
1135                 ++vcpu->last_pt_write_count;
1136                 if (vcpu->last_pt_write_count >= 3)
1137                         flooded = 1;
1138         } else {
1139                 vcpu->last_pt_write_gfn = gfn;
1140                 vcpu->last_pt_write_count = 1;
1141         }
1142         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1143         bucket = &vcpu->kvm->mmu_page_hash[index];
1144         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1145                 if (page->gfn != gfn || page->role.metaphysical)
1146                         continue;
1147                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1148                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1149                 misaligned |= bytes < 4;
1150                 if (misaligned || flooded) {
1151                         /*
1152                          * Misaligned accesses are too much trouble to fix
1153                          * up; also, they usually indicate a page is not used
1154                          * as a page table.
1155                          *
1156                          * If we're seeing too many writes to a page,
1157                          * it may no longer be a page table, or we may be
1158                          * forking, in which case it is better to unmap the
1159                          * page.
1160                          */
1161                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1162                                  gpa, bytes, page->role.word);
1163                         kvm_mmu_zap_page(vcpu, page);
1164                         continue;
1165                 }
1166                 page_offset = offset;
1167                 level = page->role.level;
1168                 npte = 1;
1169                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1170                         page_offset <<= 1;      /* 32->64 */
1171                         /*
1172                          * A 32-bit pde maps 4MB while the shadow pdes map
1173                          * only 2MB.  So we need to double the offset again
1174                          * and zap two pdes instead of one.
1175                          */
1176                         if (level == PT32_ROOT_LEVEL) {
1177                                 page_offset &= ~7; /* kill rounding error */
1178                                 page_offset <<= 1;
1179                                 npte = 2;
1180                         }
1181                         quadrant = page_offset >> PAGE_SHIFT;
1182                         page_offset &= ~PAGE_MASK;
1183                         if (quadrant != page->role.quadrant)
1184                                 continue;
1185                 }
1186                 spte = &page->spt[page_offset / sizeof(*spte)];
1187                 while (npte--) {
1188                         mmu_pte_write_zap_pte(vcpu, page, spte);
1189                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes);
1190                         ++spte;
1191                 }
1192         }
1193 }
1194
1195 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1196 {
1197         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1198
1199         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1200 }
1201
1202 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1203 {
1204         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1205                 struct kvm_mmu_page *page;
1206
1207                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1208                                     struct kvm_mmu_page, link);
1209                 kvm_mmu_zap_page(vcpu, page);
1210         }
1211 }
1212 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1213
1214 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1215 {
1216         struct kvm_mmu_page *page;
1217
1218         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1219                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1220                                     struct kvm_mmu_page, link);
1221                 kvm_mmu_zap_page(vcpu, page);
1222         }
1223         free_page((unsigned long)vcpu->mmu.pae_root);
1224 }
1225
1226 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1227 {
1228         struct page *page;
1229         int i;
1230
1231         ASSERT(vcpu);
1232
1233         vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1234
1235         /*
1236          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1237          * Therefore we need to allocate shadow page tables in the first
1238          * 4GB of memory, which happens to fit the DMA32 zone.
1239          */
1240         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1241         if (!page)
1242                 goto error_1;
1243         vcpu->mmu.pae_root = page_address(page);
1244         for (i = 0; i < 4; ++i)
1245                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1246
1247         return 0;
1248
1249 error_1:
1250         free_mmu_pages(vcpu);
1251         return -ENOMEM;
1252 }
1253
1254 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1255 {
1256         ASSERT(vcpu);
1257         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1258
1259         return alloc_mmu_pages(vcpu);
1260 }
1261
1262 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1263 {
1264         ASSERT(vcpu);
1265         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1266
1267         return init_kvm_mmu(vcpu);
1268 }
1269
1270 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1271 {
1272         ASSERT(vcpu);
1273
1274         destroy_kvm_mmu(vcpu);
1275         free_mmu_pages(vcpu);
1276         mmu_free_memory_caches(vcpu);
1277 }
1278
1279 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1280 {
1281         struct kvm *kvm = vcpu->kvm;
1282         struct kvm_mmu_page *page;
1283
1284         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1285                 int i;
1286                 u64 *pt;
1287
1288                 if (!test_bit(slot, &page->slot_bitmap))
1289                         continue;
1290
1291                 pt = page->spt;
1292                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1293                         /* avoid RMW */
1294                         if (pt[i] & PT_WRITABLE_MASK) {
1295                                 rmap_remove(vcpu, &pt[i]);
1296                                 pt[i] &= ~PT_WRITABLE_MASK;
1297                         }
1298         }
1299 }
1300
1301 void kvm_mmu_zap_all(struct kvm_vcpu *vcpu)
1302 {
1303         destroy_kvm_mmu(vcpu);
1304
1305         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1306                 struct kvm_mmu_page *page;
1307
1308                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1309                                     struct kvm_mmu_page, link);
1310                 kvm_mmu_zap_page(vcpu, page);
1311         }
1312
1313         mmu_free_memory_caches(vcpu);
1314         kvm_arch_ops->tlb_flush(vcpu);
1315         init_kvm_mmu(vcpu);
1316 }
1317
1318 void kvm_mmu_module_exit(void)
1319 {
1320         if (pte_chain_cache)
1321                 kmem_cache_destroy(pte_chain_cache);
1322         if (rmap_desc_cache)
1323                 kmem_cache_destroy(rmap_desc_cache);
1324         if (mmu_page_cache)
1325                 kmem_cache_destroy(mmu_page_cache);
1326         if (mmu_page_header_cache)
1327                 kmem_cache_destroy(mmu_page_header_cache);
1328 }
1329
1330 int kvm_mmu_module_init(void)
1331 {
1332         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1333                                             sizeof(struct kvm_pte_chain),
1334                                             0, 0, NULL, NULL);
1335         if (!pte_chain_cache)
1336                 goto nomem;
1337         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1338                                             sizeof(struct kvm_rmap_desc),
1339                                             0, 0, NULL, NULL);
1340         if (!rmap_desc_cache)
1341                 goto nomem;
1342
1343         mmu_page_cache = kmem_cache_create("kvm_mmu_page",
1344                                            PAGE_SIZE,
1345                                            PAGE_SIZE, 0, NULL, NULL);
1346         if (!mmu_page_cache)
1347                 goto nomem;
1348
1349         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1350                                                   sizeof(struct kvm_mmu_page),
1351                                                   0, 0, NULL, NULL);
1352         if (!mmu_page_header_cache)
1353                 goto nomem;
1354
1355         return 0;
1356
1357 nomem:
1358         kvm_mmu_module_exit();
1359         return -ENOMEM;
1360 }
1361
1362 #ifdef AUDIT
1363
1364 static const char *audit_msg;
1365
1366 static gva_t canonicalize(gva_t gva)
1367 {
1368 #ifdef CONFIG_X86_64
1369         gva = (long long)(gva << 16) >> 16;
1370 #endif
1371         return gva;
1372 }
1373
1374 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1375                                 gva_t va, int level)
1376 {
1377         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1378         int i;
1379         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1380
1381         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1382                 u64 ent = pt[i];
1383
1384                 if (!(ent & PT_PRESENT_MASK))
1385                         continue;
1386
1387                 va = canonicalize(va);
1388                 if (level > 1)
1389                         audit_mappings_page(vcpu, ent, va, level - 1);
1390                 else {
1391                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1392                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1393
1394                         if ((ent & PT_PRESENT_MASK)
1395                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1396                                 printk(KERN_ERR "audit error: (%s) levels %d"
1397                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1398                                        audit_msg, vcpu->mmu.root_level,
1399                                        va, gpa, hpa, ent);
1400                 }
1401         }
1402 }
1403
1404 static void audit_mappings(struct kvm_vcpu *vcpu)
1405 {
1406         unsigned i;
1407
1408         if (vcpu->mmu.root_level == 4)
1409                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1410         else
1411                 for (i = 0; i < 4; ++i)
1412                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1413                                 audit_mappings_page(vcpu,
1414                                                     vcpu->mmu.pae_root[i],
1415                                                     i << 30,
1416                                                     2);
1417 }
1418
1419 static int count_rmaps(struct kvm_vcpu *vcpu)
1420 {
1421         int nmaps = 0;
1422         int i, j, k;
1423
1424         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1425                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1426                 struct kvm_rmap_desc *d;
1427
1428                 for (j = 0; j < m->npages; ++j) {
1429                         struct page *page = m->phys_mem[j];
1430
1431                         if (!page->private)
1432                                 continue;
1433                         if (!(page->private & 1)) {
1434                                 ++nmaps;
1435                                 continue;
1436                         }
1437                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1438                         while (d) {
1439                                 for (k = 0; k < RMAP_EXT; ++k)
1440                                         if (d->shadow_ptes[k])
1441                                                 ++nmaps;
1442                                         else
1443                                                 break;
1444                                 d = d->more;
1445                         }
1446                 }
1447         }
1448         return nmaps;
1449 }
1450
1451 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1452 {
1453         int nmaps = 0;
1454         struct kvm_mmu_page *page;
1455         int i;
1456
1457         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1458                 u64 *pt = page->spt;
1459
1460                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1461                         continue;
1462
1463                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1464                         u64 ent = pt[i];
1465
1466                         if (!(ent & PT_PRESENT_MASK))
1467                                 continue;
1468                         if (!(ent & PT_WRITABLE_MASK))
1469                                 continue;
1470                         ++nmaps;
1471                 }
1472         }
1473         return nmaps;
1474 }
1475
1476 static void audit_rmap(struct kvm_vcpu *vcpu)
1477 {
1478         int n_rmap = count_rmaps(vcpu);
1479         int n_actual = count_writable_mappings(vcpu);
1480
1481         if (n_rmap != n_actual)
1482                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1483                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1484 }
1485
1486 static void audit_write_protection(struct kvm_vcpu *vcpu)
1487 {
1488         struct kvm_mmu_page *page;
1489
1490         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1491                 hfn_t hfn;
1492                 struct page *pg;
1493
1494                 if (page->role.metaphysical)
1495                         continue;
1496
1497                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1498                         >> PAGE_SHIFT;
1499                 pg = pfn_to_page(hfn);
1500                 if (pg->private)
1501                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1502                                " mappings: gfn %lx role %x\n",
1503                                __FUNCTION__, audit_msg, page->gfn,
1504                                page->role.word);
1505         }
1506 }
1507
1508 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1509 {
1510         int olddbg = dbg;
1511
1512         dbg = 0;
1513         audit_msg = msg;
1514         audit_rmap(vcpu);
1515         audit_write_protection(vcpu);
1516         audit_mappings(vcpu);
1517         dbg = olddbg;
1518 }
1519
1520 #endif