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