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