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
  20 #include "vmx.h"
  21 #include "kvm.h"
  22
  23 #include <linux/types.h>
  24 #include <linux/string.h>
  25 #include <linux/mm.h>
  26 #include <linux/highmem.h>
  27 #include <linux/module.h>
  28
  29 #include <asm/page.h>
  30 #include <asm/cmpxchg.h>
  31
  32 #undef MMU_DEBUG
  33
  34 #undef AUDIT
  35
  36 #ifdef AUDIT
  37 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  38 #else
  39 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  40 #endif
  41
  42 #ifdef MMU_DEBUG
  43
  44 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  45 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  46
  47 #else
  48
  49 #define pgprintk(x...) do { } while (0)
  50 #define rmap_printk(x...) do { } while (0)
  51
  52 #endif
  53
  54 #if defined(MMU_DEBUG) || defined(AUDIT)
  55 static int dbg = 1;
  56 #endif
  57
  58 #ifndef MMU_DEBUG
  59 #define ASSERT(x) do { } while (0)
  60 #else
  61 #define ASSERT(x)                                                       \
  62         if (!(x)) {                                                     \
  63                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  64                        __FILE__, __LINE__, #x);                         \
  65         }
  66 #endif
  67
  68 #define PT64_PT_BITS 9
  69 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  70 #define PT32_PT_BITS 10
  71 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  72
  73 #define PT_WRITABLE_SHIFT 1
  74
  75 #define PT_PRESENT_MASK (1ULL << 0)
  76 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  77 #define PT_USER_MASK (1ULL << 2)
  78 #define PT_PWT_MASK (1ULL << 3)
  79 #define PT_PCD_MASK (1ULL << 4)
  80 #define PT_ACCESSED_MASK (1ULL << 5)
  81 #define PT_DIRTY_MASK (1ULL << 6)
  82 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  83 #define PT_PAT_MASK (1ULL << 7)
  84 #define PT_GLOBAL_MASK (1ULL << 8)
  85 #define PT64_NX_MASK (1ULL << 63)
  86
  87 #define PT_PAT_SHIFT 7
  88 #define PT_DIR_PAT_SHIFT 12
  89 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  90
  91 #define PT32_DIR_PSE36_SIZE 4
  92 #define PT32_DIR_PSE36_SHIFT 13
  93 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  94
  95
  96 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  97 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  98
  99 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 100
 101 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 102
 103 #define PT64_LEVEL_BITS 9
 104
 105 #define PT64_LEVEL_SHIFT(level) \
 106                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
 107
 108 #define PT64_LEVEL_MASK(level) \
 109                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 110
 111 #define PT64_INDEX(address, level)\
 112         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 113
 114
 115 #define PT32_LEVEL_BITS 10
 116
 117 #define PT32_LEVEL_SHIFT(level) \
 118                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
 119
 120 #define PT32_LEVEL_MASK(level) \
 121                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 122
 123 #define PT32_INDEX(address, level)\
 124         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 125
 126
 127 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 128 #define PT64_DIR_BASE_ADDR_MASK \
 129         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 130
 131 #define PT32_BASE_ADDR_MASK PAGE_MASK
 132 #define PT32_DIR_BASE_ADDR_MASK \
 133         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 134
 135
 136 #define PFERR_PRESENT_MASK (1U << 0)
 137 #define PFERR_WRITE_MASK (1U << 1)
 138 #define PFERR_USER_MASK (1U << 2)
 139 #define PFERR_FETCH_MASK (1U << 4)
 140
 141 #define PT64_ROOT_LEVEL 4
 142 #define PT32_ROOT_LEVEL 2
 143 #define PT32E_ROOT_LEVEL 3
 144
 145 #define PT_DIRECTORY_LEVEL 2
 146 #define PT_PAGE_TABLE_LEVEL 1
 147
 148 #define RMAP_EXT 4
 149
 150 struct kvm_rmap_desc {
 151         u64 *shadow_ptes[RMAP_EXT];
 152         struct kvm_rmap_desc *more;
 153 };
 154
 155 static struct kmem_cache *pte_chain_cache;
 156 static struct kmem_cache *rmap_desc_cache;
 157 static struct kmem_cache *mmu_page_header_cache;
 158
 159 static int is_write_protection(struct kvm_vcpu *vcpu)
 160 {
 161         return vcpu->cr0 & X86_CR0_WP;
 162 }
 163
 164 static int is_cpuid_PSE36(void)
 165 {
 166         return 1;
 167 }
 168
 169 static int is_nx(struct kvm_vcpu *vcpu)
 170 {
 171         return vcpu->shadow_efer & EFER_NX;
 172 }
 173
 174 static int is_present_pte(unsigned long pte)
 175 {
 176         return pte & PT_PRESENT_MASK;
 177 }
 178
 179 static int is_writeble_pte(unsigned long pte)
 180 {
 181         return pte & PT_WRITABLE_MASK;
 182 }
 183
 184 static int is_io_pte(unsigned long pte)
 185 {
 186         return pte & PT_SHADOW_IO_MARK;
 187 }
 188
 189 static int is_rmap_pte(u64 pte)
 190 {
 191         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 192                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 193 }
 194
 195 static void set_shadow_pte(u64 *sptep, u64 spte)
 196 {
 197 #ifdef CONFIG_X86_64
 198         set_64bit((unsigned long *)sptep, spte);
 199 #else
 200         set_64bit((unsigned long long *)sptep, spte);
 201 #endif
 202 }
 203
 204 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 205                                   struct kmem_cache *base_cache, int min)
 206 {
 207         void *obj;
 208
 209         if (cache->nobjs >= min)
 210                 return 0;
 211         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 212                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 213                 if (!obj)
 214                         return -ENOMEM;
 215                 cache->objects[cache->nobjs++] = obj;
 216         }
 217         return 0;
 218 }
 219
 220 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 221 {
 222         while (mc->nobjs)
 223                 kfree(mc->objects[--mc->nobjs]);
 224 }
 225
 226 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 227                                        int min)
 228 {
 229         struct page *page;
 230
 231         if (cache->nobjs >= min)
 232                 return 0;
 233         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 234                 page = alloc_page(GFP_KERNEL);
 235                 if (!page)
 236                         return -ENOMEM;
 237                 set_page_private(page, 0);
 238                 cache->objects[cache->nobjs++] = page_address(page);
 239         }
 240         return 0;
 241 }
 242
 243 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 244 {
 245         while (mc->nobjs)
 246                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 247 }
 248
 249 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 250 {
 251         int r;
 252
 253         kvm_mmu_free_some_pages(vcpu);
 254         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 255                                    pte_chain_cache, 4);
 256         if (r)
 257                 goto out;
 258         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 259                                    rmap_desc_cache, 1);
 260         if (r)
 261                 goto out;
 262         r = mmu_topup_memory_cache_page(&vcpu->mmu_page_cache, 4);
 263         if (r)
 264                 goto out;
 265         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
 266                                    mmu_page_header_cache, 4);
 267 out:
 268         return r;
 269 }
 270
 271 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 272 {
 273         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 274         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 275         mmu_free_memory_cache_page(&vcpu->mmu_page_cache);
 276         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
 277 }
 278
 279 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 280                                     size_t size)
 281 {
 282         void *p;
 283
 284         BUG_ON(!mc->nobjs);
 285         p = mc->objects[--mc->nobjs];
 286         memset(p, 0, size);
 287         return p;
 288 }
 289
 290 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 291 {
 292         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 293                                       sizeof(struct kvm_pte_chain));
 294 }
 295
 296 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 297 {
 298         kfree(pc);
 299 }
 300
 301 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 302 {
 303         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 304                                       sizeof(struct kvm_rmap_desc));
 305 }
 306
 307 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 308 {
 309         kfree(rd);
 310 }
 311
 312 /*
 313  * Reverse mapping data structures:
 314  *
 315  * If page->private bit zero is zero, then page->private points to the
 316  * shadow page table entry that points to page_address(page).
 317  *
 318  * If page->private bit zero is one, (then page->private & ~1) points
 319  * to a struct kvm_rmap_desc containing more mappings.
 320  */
 321 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 322 {
 323         struct page *page;
 324         struct kvm_rmap_desc *desc;
 325         int i;
 326
 327         if (!is_rmap_pte(*spte))
 328                 return;
 329         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 330         if (!page_private(page)) {
 331                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 332                 set_page_private(page,(unsigned long)spte);
 333         } else if (!(page_private(page) & 1)) {
 334                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 335                 desc = mmu_alloc_rmap_desc(vcpu);
 336                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 337                 desc->shadow_ptes[1] = spte;
 338                 set_page_private(page,(unsigned long)desc | 1);
 339         } else {
 340                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 341                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 342                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 343                         desc = desc->more;
 344                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 345                         desc->more = mmu_alloc_rmap_desc(vcpu);
 346                         desc = desc->more;
 347                 }
 348                 for (i = 0; desc->shadow_ptes[i]; ++i)
 349                         ;
 350                 desc->shadow_ptes[i] = spte;
 351         }
 352 }
 353
 354 static void rmap_desc_remove_entry(struct page *page,
 355                                    struct kvm_rmap_desc *desc,
 356                                    int i,
 357                                    struct kvm_rmap_desc *prev_desc)
 358 {
 359         int j;
 360
 361         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 362                 ;
 363         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 364         desc->shadow_ptes[j] = NULL;
 365         if (j != 0)
 366                 return;
 367         if (!prev_desc && !desc->more)
 368                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 369         else
 370                 if (prev_desc)
 371                         prev_desc->more = desc->more;
 372                 else
 373                         set_page_private(page,(unsigned long)desc->more | 1);
 374         mmu_free_rmap_desc(desc);
 375 }
 376
 377 static void rmap_remove(u64 *spte)
 378 {
 379         struct page *page;
 380         struct kvm_rmap_desc *desc;
 381         struct kvm_rmap_desc *prev_desc;
 382         int i;
 383
 384         if (!is_rmap_pte(*spte))
 385                 return;
 386         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 387         if (!page_private(page)) {
 388                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 389                 BUG();
 390         } else if (!(page_private(page) & 1)) {
 391                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 392                 if ((u64 *)page_private(page) != spte) {
 393                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 394                                spte, *spte);
 395                         BUG();
 396                 }
 397                 set_page_private(page,0);
 398         } else {
 399                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 400                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 401                 prev_desc = NULL;
 402                 while (desc) {
 403                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 404                                 if (desc->shadow_ptes[i] == spte) {
 405                                         rmap_desc_remove_entry(page,
 406                                                                desc, i,
 407                                                                prev_desc);
 408                                         return;
 409                                 }
 410                         prev_desc = desc;
 411                         desc = desc->more;
 412                 }
 413                 BUG();
 414         }
 415 }
 416
 417 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 418 {
 419         struct kvm *kvm = vcpu->kvm;
 420         struct page *page;
 421         struct kvm_rmap_desc *desc;
 422         u64 *spte;
 423
 424         page = gfn_to_page(kvm, gfn);
 425         BUG_ON(!page);
 426
 427         while (page_private(page)) {
 428                 if (!(page_private(page) & 1))
 429                         spte = (u64 *)page_private(page);
 430                 else {
 431                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 432                         spte = desc->shadow_ptes[0];
 433                 }
 434                 BUG_ON(!spte);
 435                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 436                        != page_to_pfn(page));
 437                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 438                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 439                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 440                 rmap_remove(spte);
 441                 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 442                 kvm_flush_remote_tlbs(vcpu->kvm);
 443         }
 444 }
 445
 446 #ifdef MMU_DEBUG
 447 static int is_empty_shadow_page(u64 *spt)
 448 {
 449         u64 *pos;
 450         u64 *end;
 451
 452         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 453                 if (*pos != 0) {
 454                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 455                                pos, *pos);
 456                         return 0;
 457                 }
 458         return 1;
 459 }
 460 #endif
 461
 462 static void kvm_mmu_free_page(struct kvm *kvm,
 463                               struct kvm_mmu_page *page_head)
 464 {
 465         ASSERT(is_empty_shadow_page(page_head->spt));
 466         list_del(&page_head->link);
 467         __free_page(virt_to_page(page_head->spt));
 468         kfree(page_head);
 469         ++kvm->n_free_mmu_pages;
 470 }
 471
 472 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 473 {
 474         return gfn;
 475 }
 476
 477 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 478                                                u64 *parent_pte)
 479 {
 480         struct kvm_mmu_page *page;
 481
 482         if (!vcpu->kvm->n_free_mmu_pages)
 483                 return NULL;
 484
 485         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 486                                       sizeof *page);
 487         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 488         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 489         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 490         ASSERT(is_empty_shadow_page(page->spt));
 491         page->slot_bitmap = 0;
 492         page->multimapped = 0;
 493         page->parent_pte = parent_pte;
 494         --vcpu->kvm->n_free_mmu_pages;
 495         return page;
 496 }
 497
 498 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 499                                     struct kvm_mmu_page *page, u64 *parent_pte)
 500 {
 501         struct kvm_pte_chain *pte_chain;
 502         struct hlist_node *node;
 503         int i;
 504
 505         if (!parent_pte)
 506                 return;
 507         if (!page->multimapped) {
 508                 u64 *old = page->parent_pte;
 509
 510                 if (!old) {
 511                         page->parent_pte = parent_pte;
 512                         return;
 513                 }
 514                 page->multimapped = 1;
 515                 pte_chain = mmu_alloc_pte_chain(vcpu);
 516                 INIT_HLIST_HEAD(&page->parent_ptes);
 517                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 518                 pte_chain->parent_ptes[0] = old;
 519         }
 520         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 521                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 522                         continue;
 523                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 524                         if (!pte_chain->parent_ptes[i]) {
 525                                 pte_chain->parent_ptes[i] = parent_pte;
 526                                 return;
 527                         }
 528         }
 529         pte_chain = mmu_alloc_pte_chain(vcpu);
 530         BUG_ON(!pte_chain);
 531         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 532         pte_chain->parent_ptes[0] = parent_pte;
 533 }
 534
 535 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
 536                                        u64 *parent_pte)
 537 {
 538         struct kvm_pte_chain *pte_chain;
 539         struct hlist_node *node;
 540         int i;
 541
 542         if (!page->multimapped) {
 543                 BUG_ON(page->parent_pte != parent_pte);
 544                 page->parent_pte = NULL;
 545                 return;
 546         }
 547         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 548                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 549                         if (!pte_chain->parent_ptes[i])
 550                                 break;
 551                         if (pte_chain->parent_ptes[i] != parent_pte)
 552                                 continue;
 553                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 554                                 && pte_chain->parent_ptes[i + 1]) {
 555                                 pte_chain->parent_ptes[i]
 556                                         = pte_chain->parent_ptes[i + 1];
 557                                 ++i;
 558                         }
 559                         pte_chain->parent_ptes[i] = NULL;
 560                         if (i == 0) {
 561                                 hlist_del(&pte_chain->link);
 562                                 mmu_free_pte_chain(pte_chain);
 563                                 if (hlist_empty(&page->parent_ptes)) {
 564                                         page->multimapped = 0;
 565                                         page->parent_pte = NULL;
 566                                 }
 567                         }
 568                         return;
 569                 }
 570         BUG();
 571 }
 572
 573 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 574                                                 gfn_t gfn)
 575 {
 576         unsigned index;
 577         struct hlist_head *bucket;
 578         struct kvm_mmu_page *page;
 579         struct hlist_node *node;
 580
 581         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 582         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 583         bucket = &vcpu->kvm->mmu_page_hash[index];
 584         hlist_for_each_entry(page, node, bucket, hash_link)
 585                 if (page->gfn == gfn && !page->role.metaphysical) {
 586                         pgprintk("%s: found role %x\n",
 587                                  __FUNCTION__, page->role.word);
 588                         return page;
 589                 }
 590         return NULL;
 591 }
 592
 593 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 594                                              gfn_t gfn,
 595                                              gva_t gaddr,
 596                                              unsigned level,
 597                                              int metaphysical,
 598                                              unsigned hugepage_access,
 599                                              u64 *parent_pte)
 600 {
 601         union kvm_mmu_page_role role;
 602         unsigned index;
 603         unsigned quadrant;
 604         struct hlist_head *bucket;
 605         struct kvm_mmu_page *page;
 606         struct hlist_node *node;
 607
 608         role.word = 0;
 609         role.glevels = vcpu->mmu.root_level;
 610         role.level = level;
 611         role.metaphysical = metaphysical;
 612         role.hugepage_access = hugepage_access;
 613         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 614                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 615                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 616                 role.quadrant = quadrant;
 617         }
 618         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 619                  gfn, role.word);
 620         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 621         bucket = &vcpu->kvm->mmu_page_hash[index];
 622         hlist_for_each_entry(page, node, bucket, hash_link)
 623                 if (page->gfn == gfn && page->role.word == role.word) {
 624                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 625                         pgprintk("%s: found\n", __FUNCTION__);
 626                         return page;
 627                 }
 628         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 629         if (!page)
 630                 return page;
 631         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 632         page->gfn = gfn;
 633         page->role = role;
 634         hlist_add_head(&page->hash_link, bucket);
 635         if (!metaphysical)
 636                 rmap_write_protect(vcpu, gfn);
 637         return page;
 638 }
 639
 640 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 641                                          struct kvm_mmu_page *page)
 642 {
 643         unsigned i;
 644         u64 *pt;
 645         u64 ent;
 646
 647         pt = page->spt;
 648
 649         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 650                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 651                         if (pt[i] & PT_PRESENT_MASK)
 652                                 rmap_remove(&pt[i]);
 653                         pt[i] = 0;
 654                 }
 655                 kvm_flush_remote_tlbs(kvm);
 656                 return;
 657         }
 658
 659         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 660                 ent = pt[i];
 661
 662                 pt[i] = 0;
 663                 if (!(ent & PT_PRESENT_MASK))
 664                         continue;
 665                 ent &= PT64_BASE_ADDR_MASK;
 666                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 667         }
 668         kvm_flush_remote_tlbs(kvm);
 669 }
 670
 671 static void kvm_mmu_put_page(struct kvm_mmu_page *page,
 672                              u64 *parent_pte)
 673 {
 674         mmu_page_remove_parent_pte(page, parent_pte);
 675 }
 676
 677 static void kvm_mmu_zap_page(struct kvm *kvm,
 678                              struct kvm_mmu_page *page)
 679 {
 680         u64 *parent_pte;
 681
 682         while (page->multimapped || page->parent_pte) {
 683                 if (!page->multimapped)
 684                         parent_pte = page->parent_pte;
 685                 else {
 686                         struct kvm_pte_chain *chain;
 687
 688                         chain = container_of(page->parent_ptes.first,
 689                                              struct kvm_pte_chain, link);
 690                         parent_pte = chain->parent_ptes[0];
 691                 }
 692                 BUG_ON(!parent_pte);
 693                 kvm_mmu_put_page(page, parent_pte);
 694                 set_shadow_pte(parent_pte, 0);
 695         }
 696         kvm_mmu_page_unlink_children(kvm, page);
 697         if (!page->root_count) {
 698                 hlist_del(&page->hash_link);
 699                 kvm_mmu_free_page(kvm, page);
 700         } else
 701                 list_move(&page->link, &kvm->active_mmu_pages);
 702 }
 703
 704 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 705 {
 706         unsigned index;
 707         struct hlist_head *bucket;
 708         struct kvm_mmu_page *page;
 709         struct hlist_node *node, *n;
 710         int r;
 711
 712         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 713         r = 0;
 714         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 715         bucket = &vcpu->kvm->mmu_page_hash[index];
 716         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 717                 if (page->gfn == gfn && !page->role.metaphysical) {
 718                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 719                                  page->role.word);
 720                         kvm_mmu_zap_page(vcpu->kvm, page);
 721                         r = 1;
 722                 }
 723         return r;
 724 }
 725
 726 static void mmu_unshadow(struct kvm_vcpu *vcpu, gfn_t gfn)
 727 {
 728         struct kvm_mmu_page *page;
 729
 730         while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
 731                 pgprintk("%s: zap %lx %x\n",
 732                          __FUNCTION__, gfn, page->role.word);
 733                 kvm_mmu_zap_page(vcpu->kvm, page);
 734         }
 735 }
 736
 737 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 738 {
 739         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 740         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 741
 742         __set_bit(slot, &page_head->slot_bitmap);
 743 }
 744
 745 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 746 {
 747         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 748
 749         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 750 }
 751
 752 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 753 {
 754         struct page *page;
 755
 756         ASSERT((gpa & HPA_ERR_MASK) == 0);
 757         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 758         if (!page)
 759                 return gpa | HPA_ERR_MASK;
 760         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 761                 | (gpa & (PAGE_SIZE-1));
 762 }
 763
 764 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 765 {
 766         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 767
 768         if (gpa == UNMAPPED_GVA)
 769                 return UNMAPPED_GVA;
 770         return gpa_to_hpa(vcpu, gpa);
 771 }
 772
 773 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 774 {
 775         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 776
 777         if (gpa == UNMAPPED_GVA)
 778                 return NULL;
 779         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 780 }
 781
 782 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 783 {
 784 }
 785
 786 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 787 {
 788         int level = PT32E_ROOT_LEVEL;
 789         hpa_t table_addr = vcpu->mmu.root_hpa;
 790
 791         for (; ; level--) {
 792                 u32 index = PT64_INDEX(v, level);
 793                 u64 *table;
 794                 u64 pte;
 795
 796                 ASSERT(VALID_PAGE(table_addr));
 797                 table = __va(table_addr);
 798
 799                 if (level == 1) {
 800                         pte = table[index];
 801                         if (is_present_pte(pte) && is_writeble_pte(pte))
 802                                 return 0;
 803                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 804                         page_header_update_slot(vcpu->kvm, table, v);
 805                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 806                                                                 PT_USER_MASK;
 807                         rmap_add(vcpu, &table[index]);
 808                         return 0;
 809                 }
 810
 811                 if (table[index] == 0) {
 812                         struct kvm_mmu_page *new_table;
 813                         gfn_t pseudo_gfn;
 814
 815                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 816                                 >> PAGE_SHIFT;
 817                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 818                                                      v, level - 1,
 819                                                      1, 0, &table[index]);
 820                         if (!new_table) {
 821                                 pgprintk("nonpaging_map: ENOMEM\n");
 822                                 return -ENOMEM;
 823                         }
 824
 825                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 826                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 827                 }
 828                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 829         }
 830 }
 831
 832 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 833 {
 834         int i;
 835         struct kvm_mmu_page *page;
 836
 837         if (!VALID_PAGE(vcpu->mmu.root_hpa))
 838                 return;
 839 #ifdef CONFIG_X86_64
 840         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 841                 hpa_t root = vcpu->mmu.root_hpa;
 842
 843                 page = page_header(root);
 844                 --page->root_count;
 845                 vcpu->mmu.root_hpa = INVALID_PAGE;
 846                 return;
 847         }
 848 #endif
 849         for (i = 0; i < 4; ++i) {
 850                 hpa_t root = vcpu->mmu.pae_root[i];
 851
 852                 if (root) {
 853                         root &= PT64_BASE_ADDR_MASK;
 854                         page = page_header(root);
 855                         --page->root_count;
 856                 }
 857                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 858         }
 859         vcpu->mmu.root_hpa = INVALID_PAGE;
 860 }
 861
 862 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 863 {
 864         int i;
 865         gfn_t root_gfn;
 866         struct kvm_mmu_page *page;
 867
 868         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 869
 870 #ifdef CONFIG_X86_64
 871         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 872                 hpa_t root = vcpu->mmu.root_hpa;
 873
 874                 ASSERT(!VALID_PAGE(root));
 875                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 876                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 877                 root = __pa(page->spt);
 878                 ++page->root_count;
 879                 vcpu->mmu.root_hpa = root;
 880                 return;
 881         }
 882 #endif
 883         for (i = 0; i < 4; ++i) {
 884                 hpa_t root = vcpu->mmu.pae_root[i];
 885
 886                 ASSERT(!VALID_PAGE(root));
 887                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 888                         if (!is_present_pte(vcpu->pdptrs[i])) {
 889                                 vcpu->mmu.pae_root[i] = 0;
 890                                 continue;
 891                         }
 892                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 893                 } else if (vcpu->mmu.root_level == 0)
 894                         root_gfn = 0;
 895                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 896                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 897                                         0, NULL);
 898                 root = __pa(page->spt);
 899                 ++page->root_count;
 900                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 901         }
 902         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 903 }
 904
 905 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 906 {
 907         return vaddr;
 908 }
 909
 910 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 911                                u32 error_code)
 912 {
 913         gpa_t addr = gva;
 914         hpa_t paddr;
 915         int r;
 916
 917         r = mmu_topup_memory_caches(vcpu);
 918         if (r)
 919                 return r;
 920
 921         ASSERT(vcpu);
 922         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 923
 924
 925         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 926
 927         if (is_error_hpa(paddr))
 928                 return 1;
 929
 930         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 931 }
 932
 933 static void nonpaging_free(struct kvm_vcpu *vcpu)
 934 {
 935         mmu_free_roots(vcpu);
 936 }
 937
 938 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 939 {
 940         struct kvm_mmu *context = &vcpu->mmu;
 941
 942         context->new_cr3 = nonpaging_new_cr3;
 943         context->page_fault = nonpaging_page_fault;
 944         context->gva_to_gpa = nonpaging_gva_to_gpa;
 945         context->free = nonpaging_free;
 946         context->root_level = 0;
 947         context->shadow_root_level = PT32E_ROOT_LEVEL;
 948         context->root_hpa = INVALID_PAGE;
 949         return 0;
 950 }
 951
 952 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 953 {
 954         ++vcpu->stat.tlb_flush;
 955         kvm_x86_ops->tlb_flush(vcpu);
 956 }
 957
 958 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 959 {
 960         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 961         mmu_free_roots(vcpu);
 962 }
 963
 964 static void inject_page_fault(struct kvm_vcpu *vcpu,
 965                               u64 addr,
 966                               u32 err_code)
 967 {
 968         kvm_x86_ops->inject_page_fault(vcpu, addr, err_code);
 969 }
 970
 971 static void paging_free(struct kvm_vcpu *vcpu)
 972 {
 973         nonpaging_free(vcpu);
 974 }
 975
 976 #define PTTYPE 64
 977 #include "paging_tmpl.h"
 978 #undef PTTYPE
 979
 980 #define PTTYPE 32
 981 #include "paging_tmpl.h"
 982 #undef PTTYPE
 983
 984 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
 985 {
 986         struct kvm_mmu *context = &vcpu->mmu;
 987
 988         ASSERT(is_pae(vcpu));
 989         context->new_cr3 = paging_new_cr3;
 990         context->page_fault = paging64_page_fault;
 991         context->gva_to_gpa = paging64_gva_to_gpa;
 992         context->free = paging_free;
 993         context->root_level = level;
 994         context->shadow_root_level = level;
 995         context->root_hpa = INVALID_PAGE;
 996         return 0;
 997 }
 998
 999 static int paging64_init_context(struct kvm_vcpu *vcpu)
1000 {
1001         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1002 }
1003
1004 static int paging32_init_context(struct kvm_vcpu *vcpu)
1005 {
1006         struct kvm_mmu *context = &vcpu->mmu;
1007
1008         context->new_cr3 = paging_new_cr3;
1009         context->page_fault = paging32_page_fault;
1010         context->gva_to_gpa = paging32_gva_to_gpa;
1011         context->free = paging_free;
1012         context->root_level = PT32_ROOT_LEVEL;
1013         context->shadow_root_level = PT32E_ROOT_LEVEL;
1014         context->root_hpa = INVALID_PAGE;
1015         return 0;
1016 }
1017
1018 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1019 {
1020         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1021 }
1022
1023 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1024 {
1025         ASSERT(vcpu);
1026         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1027
1028         if (!is_paging(vcpu))
1029                 return nonpaging_init_context(vcpu);
1030         else if (is_long_mode(vcpu))
1031                 return paging64_init_context(vcpu);
1032         else if (is_pae(vcpu))
1033                 return paging32E_init_context(vcpu);
1034         else
1035                 return paging32_init_context(vcpu);
1036 }
1037
1038 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1039 {
1040         ASSERT(vcpu);
1041         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1042                 vcpu->mmu.free(vcpu);
1043                 vcpu->mmu.root_hpa = INVALID_PAGE;
1044         }
1045 }
1046
1047 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1048 {
1049         destroy_kvm_mmu(vcpu);
1050         return init_kvm_mmu(vcpu);
1051 }
1052
1053 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1054 {
1055         int r;
1056
1057         mutex_lock(&vcpu->kvm->lock);
1058         r = mmu_topup_memory_caches(vcpu);
1059         if (r)
1060                 goto out;
1061         mmu_alloc_roots(vcpu);
1062         kvm_x86_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1063         kvm_mmu_flush_tlb(vcpu);
1064 out:
1065         mutex_unlock(&vcpu->kvm->lock);
1066         return r;
1067 }
1068 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1069
1070 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1071 {
1072         mmu_free_roots(vcpu);
1073 }
1074
1075 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1076                                   struct kvm_mmu_page *page,
1077                                   u64 *spte)
1078 {
1079         u64 pte;
1080         struct kvm_mmu_page *child;
1081
1082         pte = *spte;
1083         if (is_present_pte(pte)) {
1084                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1085                         rmap_remove(spte);
1086                 else {
1087                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1088                         mmu_page_remove_parent_pte(child, spte);
1089                 }
1090         }
1091         *spte = 0;
1092         kvm_flush_remote_tlbs(vcpu->kvm);
1093 }
1094
1095 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1096                                   struct kvm_mmu_page *page,
1097                                   u64 *spte,
1098                                   const void *new, int bytes)
1099 {
1100         if (page->role.level != PT_PAGE_TABLE_LEVEL)
1101                 return;
1102
1103         if (page->role.glevels == PT32_ROOT_LEVEL)
1104                 paging32_update_pte(vcpu, page, spte, new, bytes);
1105         else
1106                 paging64_update_pte(vcpu, page, spte, new, bytes);
1107 }
1108
1109 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1110                        const u8 *new, int bytes)
1111 {
1112         gfn_t gfn = gpa >> PAGE_SHIFT;
1113         struct kvm_mmu_page *page;
1114         struct hlist_node *node, *n;
1115         struct hlist_head *bucket;
1116         unsigned index;
1117         u64 *spte;
1118         unsigned offset = offset_in_page(gpa);
1119         unsigned pte_size;
1120         unsigned page_offset;
1121         unsigned misaligned;
1122         unsigned quadrant;
1123         int level;
1124         int flooded = 0;
1125         int npte;
1126
1127         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1128         if (gfn == vcpu->last_pt_write_gfn) {
1129                 ++vcpu->last_pt_write_count;
1130                 if (vcpu->last_pt_write_count >= 3)
1131                         flooded = 1;
1132         } else {
1133                 vcpu->last_pt_write_gfn = gfn;
1134                 vcpu->last_pt_write_count = 1;
1135         }
1136         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1137         bucket = &vcpu->kvm->mmu_page_hash[index];
1138         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1139                 if (page->gfn != gfn || page->role.metaphysical)
1140                         continue;
1141                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1142                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1143                 misaligned |= bytes < 4;
1144                 if (misaligned || flooded) {
1145                         /*
1146                          * Misaligned accesses are too much trouble to fix
1147                          * up; also, they usually indicate a page is not used
1148                          * as a page table.
1149                          *
1150                          * If we're seeing too many writes to a page,
1151                          * it may no longer be a page table, or we may be
1152                          * forking, in which case it is better to unmap the
1153                          * page.
1154                          */
1155                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1156                                  gpa, bytes, page->role.word);
1157                         kvm_mmu_zap_page(vcpu->kvm, page);
1158                         continue;
1159                 }
1160                 page_offset = offset;
1161                 level = page->role.level;
1162                 npte = 1;
1163                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1164                         page_offset <<= 1;      /* 32->64 */
1165                         /*
1166                          * A 32-bit pde maps 4MB while the shadow pdes map
1167                          * only 2MB.  So we need to double the offset again
1168                          * and zap two pdes instead of one.
1169                          */
1170                         if (level == PT32_ROOT_LEVEL) {
1171                                 page_offset &= ~7; /* kill rounding error */
1172                                 page_offset <<= 1;
1173                                 npte = 2;
1174                         }
1175                         quadrant = page_offset >> PAGE_SHIFT;
1176                         page_offset &= ~PAGE_MASK;
1177                         if (quadrant != page->role.quadrant)
1178                                 continue;
1179                 }
1180                 spte = &page->spt[page_offset / sizeof(*spte)];
1181                 while (npte--) {
1182                         mmu_pte_write_zap_pte(vcpu, page, spte);
1183                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes);
1184                         ++spte;
1185                 }
1186         }
1187 }
1188
1189 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1190 {
1191         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1192
1193         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1194 }
1195
1196 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1197 {
1198         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1199                 struct kvm_mmu_page *page;
1200
1201                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1202                                     struct kvm_mmu_page, link);
1203                 kvm_mmu_zap_page(vcpu->kvm, page);
1204         }
1205 }
1206
1207 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1208 {
1209         struct kvm_mmu_page *page;
1210
1211         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1212                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1213                                     struct kvm_mmu_page, link);
1214                 kvm_mmu_zap_page(vcpu->kvm, page);
1215         }
1216         free_page((unsigned long)vcpu->mmu.pae_root);
1217 }
1218
1219 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1220 {
1221         struct page *page;
1222         int i;
1223
1224         ASSERT(vcpu);
1225
1226         vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1227
1228         /*
1229          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1230          * Therefore we need to allocate shadow page tables in the first
1231          * 4GB of memory, which happens to fit the DMA32 zone.
1232          */
1233         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1234         if (!page)
1235                 goto error_1;
1236         vcpu->mmu.pae_root = page_address(page);
1237         for (i = 0; i < 4; ++i)
1238                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1239
1240         return 0;
1241
1242 error_1:
1243         free_mmu_pages(vcpu);
1244         return -ENOMEM;
1245 }
1246
1247 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1248 {
1249         ASSERT(vcpu);
1250         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1251
1252         return alloc_mmu_pages(vcpu);
1253 }
1254
1255 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1256 {
1257         ASSERT(vcpu);
1258         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1259
1260         return init_kvm_mmu(vcpu);
1261 }
1262
1263 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1264 {
1265         ASSERT(vcpu);
1266
1267         destroy_kvm_mmu(vcpu);
1268         free_mmu_pages(vcpu);
1269         mmu_free_memory_caches(vcpu);
1270 }
1271
1272 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1273 {
1274         struct kvm_mmu_page *page;
1275
1276         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1277                 int i;
1278                 u64 *pt;
1279
1280                 if (!test_bit(slot, &page->slot_bitmap))
1281                         continue;
1282
1283                 pt = page->spt;
1284                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1285                         /* avoid RMW */
1286                         if (pt[i] & PT_WRITABLE_MASK) {
1287                                 rmap_remove(&pt[i]);
1288                                 pt[i] &= ~PT_WRITABLE_MASK;
1289                         }
1290         }
1291 }
1292
1293 void kvm_mmu_zap_all(struct kvm *kvm)
1294 {
1295         struct kvm_mmu_page *page, *node;
1296
1297         list_for_each_entry_safe(page, node, &kvm->active_mmu_pages, link)
1298                 kvm_mmu_zap_page(kvm, page);
1299
1300         kvm_flush_remote_tlbs(kvm);
1301 }
1302
1303 void kvm_mmu_module_exit(void)
1304 {
1305         if (pte_chain_cache)
1306                 kmem_cache_destroy(pte_chain_cache);
1307         if (rmap_desc_cache)
1308                 kmem_cache_destroy(rmap_desc_cache);
1309         if (mmu_page_header_cache)
1310                 kmem_cache_destroy(mmu_page_header_cache);
1311 }
1312
1313 int kvm_mmu_module_init(void)
1314 {
1315         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1316                                             sizeof(struct kvm_pte_chain),
1317                                             0, 0, NULL);
1318         if (!pte_chain_cache)
1319                 goto nomem;
1320         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1321                                             sizeof(struct kvm_rmap_desc),
1322                                             0, 0, NULL);
1323         if (!rmap_desc_cache)
1324                 goto nomem;
1325
1326         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1327                                                   sizeof(struct kvm_mmu_page),
1328                                                   0, 0, NULL);
1329         if (!mmu_page_header_cache)
1330                 goto nomem;
1331
1332         return 0;
1333
1334 nomem:
1335         kvm_mmu_module_exit();
1336         return -ENOMEM;
1337 }
1338
1339 #ifdef AUDIT
1340
1341 static const char *audit_msg;
1342
1343 static gva_t canonicalize(gva_t gva)
1344 {
1345 #ifdef CONFIG_X86_64
1346         gva = (long long)(gva << 16) >> 16;
1347 #endif
1348         return gva;
1349 }
1350
1351 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1352                                 gva_t va, int level)
1353 {
1354         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1355         int i;
1356         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1357
1358         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1359                 u64 ent = pt[i];
1360
1361                 if (!(ent & PT_PRESENT_MASK))
1362                         continue;
1363
1364                 va = canonicalize(va);
1365                 if (level > 1)
1366                         audit_mappings_page(vcpu, ent, va, level - 1);
1367                 else {
1368                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1369                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1370
1371                         if ((ent & PT_PRESENT_MASK)
1372                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1373                                 printk(KERN_ERR "audit error: (%s) levels %d"
1374                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1375                                        audit_msg, vcpu->mmu.root_level,
1376                                        va, gpa, hpa, ent);
1377                 }
1378         }
1379 }
1380
1381 static void audit_mappings(struct kvm_vcpu *vcpu)
1382 {
1383         unsigned i;
1384
1385         if (vcpu->mmu.root_level == 4)
1386                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1387         else
1388                 for (i = 0; i < 4; ++i)
1389                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1390                                 audit_mappings_page(vcpu,
1391                                                     vcpu->mmu.pae_root[i],
1392                                                     i << 30,
1393                                                     2);
1394 }
1395
1396 static int count_rmaps(struct kvm_vcpu *vcpu)
1397 {
1398         int nmaps = 0;
1399         int i, j, k;
1400
1401         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1402                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1403                 struct kvm_rmap_desc *d;
1404
1405                 for (j = 0; j < m->npages; ++j) {
1406                         struct page *page = m->phys_mem[j];
1407
1408                         if (!page->private)
1409                                 continue;
1410                         if (!(page->private & 1)) {
1411                                 ++nmaps;
1412                                 continue;
1413                         }
1414                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1415                         while (d) {
1416                                 for (k = 0; k < RMAP_EXT; ++k)
1417                                         if (d->shadow_ptes[k])
1418                                                 ++nmaps;
1419                                         else
1420                                                 break;
1421                                 d = d->more;
1422                         }
1423                 }
1424         }
1425         return nmaps;
1426 }
1427
1428 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1429 {
1430         int nmaps = 0;
1431         struct kvm_mmu_page *page;
1432         int i;
1433
1434         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1435                 u64 *pt = page->spt;
1436
1437                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1438                         continue;
1439
1440                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1441                         u64 ent = pt[i];
1442
1443                         if (!(ent & PT_PRESENT_MASK))
1444                                 continue;
1445                         if (!(ent & PT_WRITABLE_MASK))
1446                                 continue;
1447                         ++nmaps;
1448                 }
1449         }
1450         return nmaps;
1451 }
1452
1453 static void audit_rmap(struct kvm_vcpu *vcpu)
1454 {
1455         int n_rmap = count_rmaps(vcpu);
1456         int n_actual = count_writable_mappings(vcpu);
1457
1458         if (n_rmap != n_actual)
1459                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1460                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1461 }
1462
1463 static void audit_write_protection(struct kvm_vcpu *vcpu)
1464 {
1465         struct kvm_mmu_page *page;
1466
1467         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1468                 hfn_t hfn;
1469                 struct page *pg;
1470
1471                 if (page->role.metaphysical)
1472                         continue;
1473
1474                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1475                         >> PAGE_SHIFT;
1476                 pg = pfn_to_page(hfn);
1477                 if (pg->private)
1478                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1479                                " mappings: gfn %lx role %x\n",
1480                                __FUNCTION__, audit_msg, page->gfn,
1481                                page->role.word);
1482         }
1483 }
1484
1485 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1486 {
1487         int olddbg = dbg;
1488
1489         dbg = 0;
1490         audit_msg = msg;
1491         audit_rmap(vcpu);
1492         audit_write_protection(vcpu);
1493         audit_mappings(vcpu);
1494         dbg = olddbg;
1495 }
1496
1497 #endif