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 u64 __read_mostly shadow_trap_nonpresent_pte;
 160 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 161
 162 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 163 {
 164         shadow_trap_nonpresent_pte = trap_pte;
 165         shadow_notrap_nonpresent_pte = notrap_pte;
 166 }
 167 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 168
 169 static int is_write_protection(struct kvm_vcpu *vcpu)
 170 {
 171         return vcpu->cr0 & X86_CR0_WP;
 172 }
 173
 174 static int is_cpuid_PSE36(void)
 175 {
 176         return 1;
 177 }
 178
 179 static int is_nx(struct kvm_vcpu *vcpu)
 180 {
 181         return vcpu->shadow_efer & EFER_NX;
 182 }
 183
 184 static int is_present_pte(unsigned long pte)
 185 {
 186         return pte & PT_PRESENT_MASK;
 187 }
 188
 189 static int is_shadow_present_pte(u64 pte)
 190 {
 191         pte &= ~PT_SHADOW_IO_MARK;
 192         return pte != shadow_trap_nonpresent_pte
 193                 && pte != shadow_notrap_nonpresent_pte;
 194 }
 195
 196 static int is_writeble_pte(unsigned long pte)
 197 {
 198         return pte & PT_WRITABLE_MASK;
 199 }
 200
 201 static int is_io_pte(unsigned long pte)
 202 {
 203         return pte & PT_SHADOW_IO_MARK;
 204 }
 205
 206 static int is_rmap_pte(u64 pte)
 207 {
 208         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 209                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 210 }
 211
 212 static void set_shadow_pte(u64 *sptep, u64 spte)
 213 {
 214 #ifdef CONFIG_X86_64
 215         set_64bit((unsigned long *)sptep, spte);
 216 #else
 217         set_64bit((unsigned long long *)sptep, spte);
 218 #endif
 219 }
 220
 221 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 222                                   struct kmem_cache *base_cache, int min)
 223 {
 224         void *obj;
 225
 226         if (cache->nobjs >= min)
 227                 return 0;
 228         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 229                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 230                 if (!obj)
 231                         return -ENOMEM;
 232                 cache->objects[cache->nobjs++] = obj;
 233         }
 234         return 0;
 235 }
 236
 237 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 238 {
 239         while (mc->nobjs)
 240                 kfree(mc->objects[--mc->nobjs]);
 241 }
 242
 243 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 244                                        int min)
 245 {
 246         struct page *page;
 247
 248         if (cache->nobjs >= min)
 249                 return 0;
 250         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 251                 page = alloc_page(GFP_KERNEL);
 252                 if (!page)
 253                         return -ENOMEM;
 254                 set_page_private(page, 0);
 255                 cache->objects[cache->nobjs++] = page_address(page);
 256         }
 257         return 0;
 258 }
 259
 260 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 261 {
 262         while (mc->nobjs)
 263                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 264 }
 265
 266 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 267 {
 268         int r;
 269
 270         kvm_mmu_free_some_pages(vcpu);
 271         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 272                                    pte_chain_cache, 4);
 273         if (r)
 274                 goto out;
 275         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 276                                    rmap_desc_cache, 1);
 277         if (r)
 278                 goto out;
 279         r = mmu_topup_memory_cache_page(&vcpu->mmu_page_cache, 4);
 280         if (r)
 281                 goto out;
 282         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
 283                                    mmu_page_header_cache, 4);
 284 out:
 285         return r;
 286 }
 287
 288 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 289 {
 290         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 291         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 292         mmu_free_memory_cache_page(&vcpu->mmu_page_cache);
 293         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
 294 }
 295
 296 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 297                                     size_t size)
 298 {
 299         void *p;
 300
 301         BUG_ON(!mc->nobjs);
 302         p = mc->objects[--mc->nobjs];
 303         memset(p, 0, size);
 304         return p;
 305 }
 306
 307 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 308 {
 309         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 310                                       sizeof(struct kvm_pte_chain));
 311 }
 312
 313 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 314 {
 315         kfree(pc);
 316 }
 317
 318 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 319 {
 320         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 321                                       sizeof(struct kvm_rmap_desc));
 322 }
 323
 324 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 325 {
 326         kfree(rd);
 327 }
 328
 329 /*
 330  * Reverse mapping data structures:
 331  *
 332  * If page->private bit zero is zero, then page->private points to the
 333  * shadow page table entry that points to page_address(page).
 334  *
 335  * If page->private bit zero is one, (then page->private & ~1) points
 336  * to a struct kvm_rmap_desc containing more mappings.
 337  */
 338 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 339 {
 340         struct page *page;
 341         struct kvm_rmap_desc *desc;
 342         int i;
 343
 344         if (!is_rmap_pte(*spte))
 345                 return;
 346         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 347         if (!page_private(page)) {
 348                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 349                 set_page_private(page,(unsigned long)spte);
 350         } else if (!(page_private(page) & 1)) {
 351                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 352                 desc = mmu_alloc_rmap_desc(vcpu);
 353                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 354                 desc->shadow_ptes[1] = spte;
 355                 set_page_private(page,(unsigned long)desc | 1);
 356         } else {
 357                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 358                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 359                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 360                         desc = desc->more;
 361                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 362                         desc->more = mmu_alloc_rmap_desc(vcpu);
 363                         desc = desc->more;
 364                 }
 365                 for (i = 0; desc->shadow_ptes[i]; ++i)
 366                         ;
 367                 desc->shadow_ptes[i] = spte;
 368         }
 369 }
 370
 371 static void rmap_desc_remove_entry(struct page *page,
 372                                    struct kvm_rmap_desc *desc,
 373                                    int i,
 374                                    struct kvm_rmap_desc *prev_desc)
 375 {
 376         int j;
 377
 378         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 379                 ;
 380         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 381         desc->shadow_ptes[j] = NULL;
 382         if (j != 0)
 383                 return;
 384         if (!prev_desc && !desc->more)
 385                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 386         else
 387                 if (prev_desc)
 388                         prev_desc->more = desc->more;
 389                 else
 390                         set_page_private(page,(unsigned long)desc->more | 1);
 391         mmu_free_rmap_desc(desc);
 392 }
 393
 394 static void rmap_remove(u64 *spte)
 395 {
 396         struct page *page;
 397         struct kvm_rmap_desc *desc;
 398         struct kvm_rmap_desc *prev_desc;
 399         int i;
 400
 401         if (!is_rmap_pte(*spte))
 402                 return;
 403         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 404         if (!page_private(page)) {
 405                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 406                 BUG();
 407         } else if (!(page_private(page) & 1)) {
 408                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 409                 if ((u64 *)page_private(page) != spte) {
 410                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 411                                spte, *spte);
 412                         BUG();
 413                 }
 414                 set_page_private(page,0);
 415         } else {
 416                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 417                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 418                 prev_desc = NULL;
 419                 while (desc) {
 420                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 421                                 if (desc->shadow_ptes[i] == spte) {
 422                                         rmap_desc_remove_entry(page,
 423                                                                desc, i,
 424                                                                prev_desc);
 425                                         return;
 426                                 }
 427                         prev_desc = desc;
 428                         desc = desc->more;
 429                 }
 430                 BUG();
 431         }
 432 }
 433
 434 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 435 {
 436         struct kvm *kvm = vcpu->kvm;
 437         struct page *page;
 438         struct kvm_rmap_desc *desc;
 439         u64 *spte;
 440
 441         page = gfn_to_page(kvm, gfn);
 442         BUG_ON(!page);
 443
 444         while (page_private(page)) {
 445                 if (!(page_private(page) & 1))
 446                         spte = (u64 *)page_private(page);
 447                 else {
 448                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 449                         spte = desc->shadow_ptes[0];
 450                 }
 451                 BUG_ON(!spte);
 452                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 453                        != page_to_pfn(page));
 454                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 455                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 456                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 457                 rmap_remove(spte);
 458                 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 459                 kvm_flush_remote_tlbs(vcpu->kvm);
 460         }
 461 }
 462
 463 #ifdef MMU_DEBUG
 464 static int is_empty_shadow_page(u64 *spt)
 465 {
 466         u64 *pos;
 467         u64 *end;
 468
 469         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 470                 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
 471                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 472                                pos, *pos);
 473                         return 0;
 474                 }
 475         return 1;
 476 }
 477 #endif
 478
 479 static void kvm_mmu_free_page(struct kvm *kvm,
 480                               struct kvm_mmu_page *page_head)
 481 {
 482         ASSERT(is_empty_shadow_page(page_head->spt));
 483         list_del(&page_head->link);
 484         __free_page(virt_to_page(page_head->spt));
 485         kfree(page_head);
 486         ++kvm->n_free_mmu_pages;
 487 }
 488
 489 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 490 {
 491         return gfn;
 492 }
 493
 494 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 495                                                u64 *parent_pte)
 496 {
 497         struct kvm_mmu_page *page;
 498
 499         if (!vcpu->kvm->n_free_mmu_pages)
 500                 return NULL;
 501
 502         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 503                                       sizeof *page);
 504         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 505         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 506         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 507         ASSERT(is_empty_shadow_page(page->spt));
 508         page->slot_bitmap = 0;
 509         page->multimapped = 0;
 510         page->parent_pte = parent_pte;
 511         --vcpu->kvm->n_free_mmu_pages;
 512         return page;
 513 }
 514
 515 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 516                                     struct kvm_mmu_page *page, u64 *parent_pte)
 517 {
 518         struct kvm_pte_chain *pte_chain;
 519         struct hlist_node *node;
 520         int i;
 521
 522         if (!parent_pte)
 523                 return;
 524         if (!page->multimapped) {
 525                 u64 *old = page->parent_pte;
 526
 527                 if (!old) {
 528                         page->parent_pte = parent_pte;
 529                         return;
 530                 }
 531                 page->multimapped = 1;
 532                 pte_chain = mmu_alloc_pte_chain(vcpu);
 533                 INIT_HLIST_HEAD(&page->parent_ptes);
 534                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 535                 pte_chain->parent_ptes[0] = old;
 536         }
 537         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 538                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 539                         continue;
 540                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 541                         if (!pte_chain->parent_ptes[i]) {
 542                                 pte_chain->parent_ptes[i] = parent_pte;
 543                                 return;
 544                         }
 545         }
 546         pte_chain = mmu_alloc_pte_chain(vcpu);
 547         BUG_ON(!pte_chain);
 548         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 549         pte_chain->parent_ptes[0] = parent_pte;
 550 }
 551
 552 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
 553                                        u64 *parent_pte)
 554 {
 555         struct kvm_pte_chain *pte_chain;
 556         struct hlist_node *node;
 557         int i;
 558
 559         if (!page->multimapped) {
 560                 BUG_ON(page->parent_pte != parent_pte);
 561                 page->parent_pte = NULL;
 562                 return;
 563         }
 564         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 565                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 566                         if (!pte_chain->parent_ptes[i])
 567                                 break;
 568                         if (pte_chain->parent_ptes[i] != parent_pte)
 569                                 continue;
 570                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 571                                 && pte_chain->parent_ptes[i + 1]) {
 572                                 pte_chain->parent_ptes[i]
 573                                         = pte_chain->parent_ptes[i + 1];
 574                                 ++i;
 575                         }
 576                         pte_chain->parent_ptes[i] = NULL;
 577                         if (i == 0) {
 578                                 hlist_del(&pte_chain->link);
 579                                 mmu_free_pte_chain(pte_chain);
 580                                 if (hlist_empty(&page->parent_ptes)) {
 581                                         page->multimapped = 0;
 582                                         page->parent_pte = NULL;
 583                                 }
 584                         }
 585                         return;
 586                 }
 587         BUG();
 588 }
 589
 590 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 591                                                 gfn_t gfn)
 592 {
 593         unsigned index;
 594         struct hlist_head *bucket;
 595         struct kvm_mmu_page *page;
 596         struct hlist_node *node;
 597
 598         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 599         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 600         bucket = &vcpu->kvm->mmu_page_hash[index];
 601         hlist_for_each_entry(page, node, bucket, hash_link)
 602                 if (page->gfn == gfn && !page->role.metaphysical) {
 603                         pgprintk("%s: found role %x\n",
 604                                  __FUNCTION__, page->role.word);
 605                         return page;
 606                 }
 607         return NULL;
 608 }
 609
 610 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 611                                              gfn_t gfn,
 612                                              gva_t gaddr,
 613                                              unsigned level,
 614                                              int metaphysical,
 615                                              unsigned hugepage_access,
 616                                              u64 *parent_pte)
 617 {
 618         union kvm_mmu_page_role role;
 619         unsigned index;
 620         unsigned quadrant;
 621         struct hlist_head *bucket;
 622         struct kvm_mmu_page *page;
 623         struct hlist_node *node;
 624
 625         role.word = 0;
 626         role.glevels = vcpu->mmu.root_level;
 627         role.level = level;
 628         role.metaphysical = metaphysical;
 629         role.hugepage_access = hugepage_access;
 630         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 631                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 632                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 633                 role.quadrant = quadrant;
 634         }
 635         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 636                  gfn, role.word);
 637         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 638         bucket = &vcpu->kvm->mmu_page_hash[index];
 639         hlist_for_each_entry(page, node, bucket, hash_link)
 640                 if (page->gfn == gfn && page->role.word == role.word) {
 641                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 642                         pgprintk("%s: found\n", __FUNCTION__);
 643                         return page;
 644                 }
 645         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 646         if (!page)
 647                 return page;
 648         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 649         page->gfn = gfn;
 650         page->role = role;
 651         hlist_add_head(&page->hash_link, bucket);
 652         vcpu->mmu.prefetch_page(vcpu, page);
 653         if (!metaphysical)
 654                 rmap_write_protect(vcpu, gfn);
 655         return page;
 656 }
 657
 658 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 659                                          struct kvm_mmu_page *page)
 660 {
 661         unsigned i;
 662         u64 *pt;
 663         u64 ent;
 664
 665         pt = page->spt;
 666
 667         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 668                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 669                         if (is_shadow_present_pte(pt[i]))
 670                                 rmap_remove(&pt[i]);
 671                         pt[i] = shadow_trap_nonpresent_pte;
 672                 }
 673                 kvm_flush_remote_tlbs(kvm);
 674                 return;
 675         }
 676
 677         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 678                 ent = pt[i];
 679
 680                 pt[i] = shadow_trap_nonpresent_pte;
 681                 if (!is_shadow_present_pte(ent))
 682                         continue;
 683                 ent &= PT64_BASE_ADDR_MASK;
 684                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 685         }
 686         kvm_flush_remote_tlbs(kvm);
 687 }
 688
 689 static void kvm_mmu_put_page(struct kvm_mmu_page *page,
 690                              u64 *parent_pte)
 691 {
 692         mmu_page_remove_parent_pte(page, parent_pte);
 693 }
 694
 695 static void kvm_mmu_zap_page(struct kvm *kvm,
 696                              struct kvm_mmu_page *page)
 697 {
 698         u64 *parent_pte;
 699
 700         while (page->multimapped || page->parent_pte) {
 701                 if (!page->multimapped)
 702                         parent_pte = page->parent_pte;
 703                 else {
 704                         struct kvm_pte_chain *chain;
 705
 706                         chain = container_of(page->parent_ptes.first,
 707                                              struct kvm_pte_chain, link);
 708                         parent_pte = chain->parent_ptes[0];
 709                 }
 710                 BUG_ON(!parent_pte);
 711                 kvm_mmu_put_page(page, parent_pte);
 712                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 713         }
 714         kvm_mmu_page_unlink_children(kvm, page);
 715         if (!page->root_count) {
 716                 hlist_del(&page->hash_link);
 717                 kvm_mmu_free_page(kvm, page);
 718         } else
 719                 list_move(&page->link, &kvm->active_mmu_pages);
 720 }
 721
 722 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 723 {
 724         unsigned index;
 725         struct hlist_head *bucket;
 726         struct kvm_mmu_page *page;
 727         struct hlist_node *node, *n;
 728         int r;
 729
 730         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 731         r = 0;
 732         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 733         bucket = &vcpu->kvm->mmu_page_hash[index];
 734         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 735                 if (page->gfn == gfn && !page->role.metaphysical) {
 736                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 737                                  page->role.word);
 738                         kvm_mmu_zap_page(vcpu->kvm, page);
 739                         r = 1;
 740                 }
 741         return r;
 742 }
 743
 744 static void mmu_unshadow(struct kvm_vcpu *vcpu, gfn_t gfn)
 745 {
 746         struct kvm_mmu_page *page;
 747
 748         while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
 749                 pgprintk("%s: zap %lx %x\n",
 750                          __FUNCTION__, gfn, page->role.word);
 751                 kvm_mmu_zap_page(vcpu->kvm, page);
 752         }
 753 }
 754
 755 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 756 {
 757         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 758         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 759
 760         __set_bit(slot, &page_head->slot_bitmap);
 761 }
 762
 763 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 764 {
 765         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 766
 767         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 768 }
 769
 770 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 771 {
 772         struct page *page;
 773
 774         ASSERT((gpa & HPA_ERR_MASK) == 0);
 775         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 776         if (!page)
 777                 return gpa | HPA_ERR_MASK;
 778         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 779                 | (gpa & (PAGE_SIZE-1));
 780 }
 781
 782 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 783 {
 784         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 785
 786         if (gpa == UNMAPPED_GVA)
 787                 return UNMAPPED_GVA;
 788         return gpa_to_hpa(vcpu, gpa);
 789 }
 790
 791 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 792 {
 793         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 794
 795         if (gpa == UNMAPPED_GVA)
 796                 return NULL;
 797         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 798 }
 799
 800 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 801 {
 802 }
 803
 804 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 805 {
 806         int level = PT32E_ROOT_LEVEL;
 807         hpa_t table_addr = vcpu->mmu.root_hpa;
 808
 809         for (; ; level--) {
 810                 u32 index = PT64_INDEX(v, level);
 811                 u64 *table;
 812                 u64 pte;
 813
 814                 ASSERT(VALID_PAGE(table_addr));
 815                 table = __va(table_addr);
 816
 817                 if (level == 1) {
 818                         pte = table[index];
 819                         if (is_shadow_present_pte(pte) && is_writeble_pte(pte))
 820                                 return 0;
 821                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 822                         page_header_update_slot(vcpu->kvm, table, v);
 823                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 824                                                                 PT_USER_MASK;
 825                         rmap_add(vcpu, &table[index]);
 826                         return 0;
 827                 }
 828
 829                 if (table[index] == shadow_trap_nonpresent_pte) {
 830                         struct kvm_mmu_page *new_table;
 831                         gfn_t pseudo_gfn;
 832
 833                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 834                                 >> PAGE_SHIFT;
 835                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 836                                                      v, level - 1,
 837                                                      1, 0, &table[index]);
 838                         if (!new_table) {
 839                                 pgprintk("nonpaging_map: ENOMEM\n");
 840                                 return -ENOMEM;
 841                         }
 842
 843                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 844                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 845                 }
 846                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 847         }
 848 }
 849
 850 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 851                                     struct kvm_mmu_page *sp)
 852 {
 853         int i;
 854
 855         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 856                 sp->spt[i] = shadow_trap_nonpresent_pte;
 857 }
 858
 859 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 860 {
 861         int i;
 862         struct kvm_mmu_page *page;
 863
 864         if (!VALID_PAGE(vcpu->mmu.root_hpa))
 865                 return;
 866 #ifdef CONFIG_X86_64
 867         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 868                 hpa_t root = vcpu->mmu.root_hpa;
 869
 870                 page = page_header(root);
 871                 --page->root_count;
 872                 vcpu->mmu.root_hpa = INVALID_PAGE;
 873                 return;
 874         }
 875 #endif
 876         for (i = 0; i < 4; ++i) {
 877                 hpa_t root = vcpu->mmu.pae_root[i];
 878
 879                 if (root) {
 880                         root &= PT64_BASE_ADDR_MASK;
 881                         page = page_header(root);
 882                         --page->root_count;
 883                 }
 884                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 885         }
 886         vcpu->mmu.root_hpa = INVALID_PAGE;
 887 }
 888
 889 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 890 {
 891         int i;
 892         gfn_t root_gfn;
 893         struct kvm_mmu_page *page;
 894
 895         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 896
 897 #ifdef CONFIG_X86_64
 898         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 899                 hpa_t root = vcpu->mmu.root_hpa;
 900
 901                 ASSERT(!VALID_PAGE(root));
 902                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 903                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 904                 root = __pa(page->spt);
 905                 ++page->root_count;
 906                 vcpu->mmu.root_hpa = root;
 907                 return;
 908         }
 909 #endif
 910         for (i = 0; i < 4; ++i) {
 911                 hpa_t root = vcpu->mmu.pae_root[i];
 912
 913                 ASSERT(!VALID_PAGE(root));
 914                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 915                         if (!is_present_pte(vcpu->pdptrs[i])) {
 916                                 vcpu->mmu.pae_root[i] = 0;
 917                                 continue;
 918                         }
 919                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 920                 } else if (vcpu->mmu.root_level == 0)
 921                         root_gfn = 0;
 922                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 923                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 924                                         0, NULL);
 925                 root = __pa(page->spt);
 926                 ++page->root_count;
 927                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 928         }
 929         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 930 }
 931
 932 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 933 {
 934         return vaddr;
 935 }
 936
 937 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 938                                u32 error_code)
 939 {
 940         gpa_t addr = gva;
 941         hpa_t paddr;
 942         int r;
 943
 944         r = mmu_topup_memory_caches(vcpu);
 945         if (r)
 946                 return r;
 947
 948         ASSERT(vcpu);
 949         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 950
 951
 952         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 953
 954         if (is_error_hpa(paddr))
 955                 return 1;
 956
 957         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 958 }
 959
 960 static void nonpaging_free(struct kvm_vcpu *vcpu)
 961 {
 962         mmu_free_roots(vcpu);
 963 }
 964
 965 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 966 {
 967         struct kvm_mmu *context = &vcpu->mmu;
 968
 969         context->new_cr3 = nonpaging_new_cr3;
 970         context->page_fault = nonpaging_page_fault;
 971         context->gva_to_gpa = nonpaging_gva_to_gpa;
 972         context->free = nonpaging_free;
 973         context->prefetch_page = nonpaging_prefetch_page;
 974         context->root_level = 0;
 975         context->shadow_root_level = PT32E_ROOT_LEVEL;
 976         context->root_hpa = INVALID_PAGE;
 977         return 0;
 978 }
 979
 980 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 981 {
 982         ++vcpu->stat.tlb_flush;
 983         kvm_x86_ops->tlb_flush(vcpu);
 984 }
 985
 986 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 987 {
 988         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 989         mmu_free_roots(vcpu);
 990 }
 991
 992 static void inject_page_fault(struct kvm_vcpu *vcpu,
 993                               u64 addr,
 994                               u32 err_code)
 995 {
 996         kvm_x86_ops->inject_page_fault(vcpu, addr, err_code);
 997 }
 998
 999 static void paging_free(struct kvm_vcpu *vcpu)
1000 {
1001         nonpaging_free(vcpu);
1002 }
1003
1004 #define PTTYPE 64
1005 #include "paging_tmpl.h"
1006 #undef PTTYPE
1007
1008 #define PTTYPE 32
1009 #include "paging_tmpl.h"
1010 #undef PTTYPE
1011
1012 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1013 {
1014         struct kvm_mmu *context = &vcpu->mmu;
1015
1016         ASSERT(is_pae(vcpu));
1017         context->new_cr3 = paging_new_cr3;
1018         context->page_fault = paging64_page_fault;
1019         context->gva_to_gpa = paging64_gva_to_gpa;
1020         context->prefetch_page = paging64_prefetch_page;
1021         context->free = paging_free;
1022         context->root_level = level;
1023         context->shadow_root_level = level;
1024         context->root_hpa = INVALID_PAGE;
1025         return 0;
1026 }
1027
1028 static int paging64_init_context(struct kvm_vcpu *vcpu)
1029 {
1030         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1031 }
1032
1033 static int paging32_init_context(struct kvm_vcpu *vcpu)
1034 {
1035         struct kvm_mmu *context = &vcpu->mmu;
1036
1037         context->new_cr3 = paging_new_cr3;
1038         context->page_fault = paging32_page_fault;
1039         context->gva_to_gpa = paging32_gva_to_gpa;
1040         context->free = paging_free;
1041         context->prefetch_page = paging32_prefetch_page;
1042         context->root_level = PT32_ROOT_LEVEL;
1043         context->shadow_root_level = PT32E_ROOT_LEVEL;
1044         context->root_hpa = INVALID_PAGE;
1045         return 0;
1046 }
1047
1048 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1049 {
1050         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1051 }
1052
1053 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1054 {
1055         ASSERT(vcpu);
1056         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1057
1058         if (!is_paging(vcpu))
1059                 return nonpaging_init_context(vcpu);
1060         else if (is_long_mode(vcpu))
1061                 return paging64_init_context(vcpu);
1062         else if (is_pae(vcpu))
1063                 return paging32E_init_context(vcpu);
1064         else
1065                 return paging32_init_context(vcpu);
1066 }
1067
1068 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1069 {
1070         ASSERT(vcpu);
1071         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1072                 vcpu->mmu.free(vcpu);
1073                 vcpu->mmu.root_hpa = INVALID_PAGE;
1074         }
1075 }
1076
1077 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1078 {
1079         destroy_kvm_mmu(vcpu);
1080         return init_kvm_mmu(vcpu);
1081 }
1082 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1083
1084 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1085 {
1086         int r;
1087
1088         mutex_lock(&vcpu->kvm->lock);
1089         r = mmu_topup_memory_caches(vcpu);
1090         if (r)
1091                 goto out;
1092         mmu_alloc_roots(vcpu);
1093         kvm_x86_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1094         kvm_mmu_flush_tlb(vcpu);
1095 out:
1096         mutex_unlock(&vcpu->kvm->lock);
1097         return r;
1098 }
1099 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1100
1101 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1102 {
1103         mmu_free_roots(vcpu);
1104 }
1105
1106 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1107                                   struct kvm_mmu_page *page,
1108                                   u64 *spte)
1109 {
1110         u64 pte;
1111         struct kvm_mmu_page *child;
1112
1113         pte = *spte;
1114         if (is_shadow_present_pte(pte)) {
1115                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1116                         rmap_remove(spte);
1117                 else {
1118                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1119                         mmu_page_remove_parent_pte(child, spte);
1120                 }
1121         }
1122         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1123         kvm_flush_remote_tlbs(vcpu->kvm);
1124 }
1125
1126 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1127                                   struct kvm_mmu_page *page,
1128                                   u64 *spte,
1129                                   const void *new, int bytes,
1130                                   int offset_in_pte)
1131 {
1132         if (page->role.level != PT_PAGE_TABLE_LEVEL)
1133                 return;
1134
1135         if (page->role.glevels == PT32_ROOT_LEVEL)
1136                 paging32_update_pte(vcpu, page, spte, new, bytes,
1137                                     offset_in_pte);
1138         else
1139                 paging64_update_pte(vcpu, page, spte, new, bytes,
1140                                     offset_in_pte);
1141 }
1142
1143 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1144                        const u8 *new, int bytes)
1145 {
1146         gfn_t gfn = gpa >> PAGE_SHIFT;
1147         struct kvm_mmu_page *page;
1148         struct hlist_node *node, *n;
1149         struct hlist_head *bucket;
1150         unsigned index;
1151         u64 *spte;
1152         unsigned offset = offset_in_page(gpa);
1153         unsigned pte_size;
1154         unsigned page_offset;
1155         unsigned misaligned;
1156         unsigned quadrant;
1157         int level;
1158         int flooded = 0;
1159         int npte;
1160
1161         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1162         kvm_mmu_audit(vcpu, "pre pte write");
1163         if (gfn == vcpu->last_pt_write_gfn) {
1164                 ++vcpu->last_pt_write_count;
1165                 if (vcpu->last_pt_write_count >= 3)
1166                         flooded = 1;
1167         } else {
1168                 vcpu->last_pt_write_gfn = gfn;
1169                 vcpu->last_pt_write_count = 1;
1170         }
1171         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1172         bucket = &vcpu->kvm->mmu_page_hash[index];
1173         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1174                 if (page->gfn != gfn || page->role.metaphysical)
1175                         continue;
1176                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1177                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1178                 misaligned |= bytes < 4;
1179                 if (misaligned || flooded) {
1180                         /*
1181                          * Misaligned accesses are too much trouble to fix
1182                          * up; also, they usually indicate a page is not used
1183                          * as a page table.
1184                          *
1185                          * If we're seeing too many writes to a page,
1186                          * it may no longer be a page table, or we may be
1187                          * forking, in which case it is better to unmap the
1188                          * page.
1189                          */
1190                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1191                                  gpa, bytes, page->role.word);
1192                         kvm_mmu_zap_page(vcpu->kvm, page);
1193                         continue;
1194                 }
1195                 page_offset = offset;
1196                 level = page->role.level;
1197                 npte = 1;
1198                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1199                         page_offset <<= 1;      /* 32->64 */
1200                         /*
1201                          * A 32-bit pde maps 4MB while the shadow pdes map
1202                          * only 2MB.  So we need to double the offset again
1203                          * and zap two pdes instead of one.
1204                          */
1205                         if (level == PT32_ROOT_LEVEL) {
1206                                 page_offset &= ~7; /* kill rounding error */
1207                                 page_offset <<= 1;
1208                                 npte = 2;
1209                         }
1210                         quadrant = page_offset >> PAGE_SHIFT;
1211                         page_offset &= ~PAGE_MASK;
1212                         if (quadrant != page->role.quadrant)
1213                                 continue;
1214                 }
1215                 spte = &page->spt[page_offset / sizeof(*spte)];
1216                 while (npte--) {
1217                         mmu_pte_write_zap_pte(vcpu, page, spte);
1218                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes,
1219                                               page_offset & (pte_size - 1));
1220                         ++spte;
1221                 }
1222         }
1223         kvm_mmu_audit(vcpu, "post pte write");
1224 }
1225
1226 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1227 {
1228         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1229
1230         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1231 }
1232
1233 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1234 {
1235         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1236                 struct kvm_mmu_page *page;
1237
1238                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1239                                     struct kvm_mmu_page, link);
1240                 kvm_mmu_zap_page(vcpu->kvm, page);
1241         }
1242 }
1243
1244 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1245 {
1246         struct kvm_mmu_page *page;
1247
1248         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1249                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1250                                     struct kvm_mmu_page, link);
1251                 kvm_mmu_zap_page(vcpu->kvm, page);
1252         }
1253         free_page((unsigned long)vcpu->mmu.pae_root);
1254 }
1255
1256 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1257 {
1258         struct page *page;
1259         int i;
1260
1261         ASSERT(vcpu);
1262
1263         vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1264
1265         /*
1266          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1267          * Therefore we need to allocate shadow page tables in the first
1268          * 4GB of memory, which happens to fit the DMA32 zone.
1269          */
1270         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1271         if (!page)
1272                 goto error_1;
1273         vcpu->mmu.pae_root = page_address(page);
1274         for (i = 0; i < 4; ++i)
1275                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1276
1277         return 0;
1278
1279 error_1:
1280         free_mmu_pages(vcpu);
1281         return -ENOMEM;
1282 }
1283
1284 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1285 {
1286         ASSERT(vcpu);
1287         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1288
1289         return alloc_mmu_pages(vcpu);
1290 }
1291
1292 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1293 {
1294         ASSERT(vcpu);
1295         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1296
1297         return init_kvm_mmu(vcpu);
1298 }
1299
1300 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1301 {
1302         ASSERT(vcpu);
1303
1304         destroy_kvm_mmu(vcpu);
1305         free_mmu_pages(vcpu);
1306         mmu_free_memory_caches(vcpu);
1307 }
1308
1309 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1310 {
1311         struct kvm_mmu_page *page;
1312
1313         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1314                 int i;
1315                 u64 *pt;
1316
1317                 if (!test_bit(slot, &page->slot_bitmap))
1318                         continue;
1319
1320                 pt = page->spt;
1321                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1322                         /* avoid RMW */
1323                         if (pt[i] & PT_WRITABLE_MASK) {
1324                                 rmap_remove(&pt[i]);
1325                                 pt[i] &= ~PT_WRITABLE_MASK;
1326                         }
1327         }
1328 }
1329
1330 void kvm_mmu_zap_all(struct kvm *kvm)
1331 {
1332         struct kvm_mmu_page *page, *node;
1333
1334         list_for_each_entry_safe(page, node, &kvm->active_mmu_pages, link)
1335                 kvm_mmu_zap_page(kvm, page);
1336
1337         kvm_flush_remote_tlbs(kvm);
1338 }
1339
1340 void kvm_mmu_module_exit(void)
1341 {
1342         if (pte_chain_cache)
1343                 kmem_cache_destroy(pte_chain_cache);
1344         if (rmap_desc_cache)
1345                 kmem_cache_destroy(rmap_desc_cache);
1346         if (mmu_page_header_cache)
1347                 kmem_cache_destroy(mmu_page_header_cache);
1348 }
1349
1350 int kvm_mmu_module_init(void)
1351 {
1352         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1353                                             sizeof(struct kvm_pte_chain),
1354                                             0, 0, NULL);
1355         if (!pte_chain_cache)
1356                 goto nomem;
1357         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1358                                             sizeof(struct kvm_rmap_desc),
1359                                             0, 0, NULL);
1360         if (!rmap_desc_cache)
1361                 goto nomem;
1362
1363         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1364                                                   sizeof(struct kvm_mmu_page),
1365                                                   0, 0, NULL);
1366         if (!mmu_page_header_cache)
1367                 goto nomem;
1368
1369         return 0;
1370
1371 nomem:
1372         kvm_mmu_module_exit();
1373         return -ENOMEM;
1374 }
1375
1376 #ifdef AUDIT
1377
1378 static const char *audit_msg;
1379
1380 static gva_t canonicalize(gva_t gva)
1381 {
1382 #ifdef CONFIG_X86_64
1383         gva = (long long)(gva << 16) >> 16;
1384 #endif
1385         return gva;
1386 }
1387
1388 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1389                                 gva_t va, int level)
1390 {
1391         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1392         int i;
1393         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1394
1395         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1396                 u64 ent = pt[i];
1397
1398                 if (ent == shadow_trap_nonpresent_pte)
1399                         continue;
1400
1401                 va = canonicalize(va);
1402                 if (level > 1) {
1403                         if (ent == shadow_notrap_nonpresent_pte)
1404                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1405                                        " in nonleaf level: levels %d gva %lx"
1406                                        " level %d pte %llx\n", audit_msg,
1407                                        vcpu->mmu.root_level, va, level, ent);
1408
1409                         audit_mappings_page(vcpu, ent, va, level - 1);
1410                 } else {
1411                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1412                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1413
1414                         if (is_shadow_present_pte(ent)
1415                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1416                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1417                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1418                                        audit_msg, vcpu->mmu.root_level,
1419                                        va, gpa, hpa, ent, is_shadow_present_pte(ent));
1420                         else if (ent == shadow_notrap_nonpresent_pte
1421                                  && !is_error_hpa(hpa))
1422                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1423                                        " valid guest gva %lx\n", audit_msg, va);
1424
1425                 }
1426         }
1427 }
1428
1429 static void audit_mappings(struct kvm_vcpu *vcpu)
1430 {
1431         unsigned i;
1432
1433         if (vcpu->mmu.root_level == 4)
1434                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1435         else
1436                 for (i = 0; i < 4; ++i)
1437                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1438                                 audit_mappings_page(vcpu,
1439                                                     vcpu->mmu.pae_root[i],
1440                                                     i << 30,
1441                                                     2);
1442 }
1443
1444 static int count_rmaps(struct kvm_vcpu *vcpu)
1445 {
1446         int nmaps = 0;
1447         int i, j, k;
1448
1449         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1450                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1451                 struct kvm_rmap_desc *d;
1452
1453                 for (j = 0; j < m->npages; ++j) {
1454                         struct page *page = m->phys_mem[j];
1455
1456                         if (!page->private)
1457                                 continue;
1458                         if (!(page->private & 1)) {
1459                                 ++nmaps;
1460                                 continue;
1461                         }
1462                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1463                         while (d) {
1464                                 for (k = 0; k < RMAP_EXT; ++k)
1465                                         if (d->shadow_ptes[k])
1466                                                 ++nmaps;
1467                                         else
1468                                                 break;
1469                                 d = d->more;
1470                         }
1471                 }
1472         }
1473         return nmaps;
1474 }
1475
1476 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1477 {
1478         int nmaps = 0;
1479         struct kvm_mmu_page *page;
1480         int i;
1481
1482         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1483                 u64 *pt = page->spt;
1484
1485                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1486                         continue;
1487
1488                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1489                         u64 ent = pt[i];
1490
1491                         if (!(ent & PT_PRESENT_MASK))
1492                                 continue;
1493                         if (!(ent & PT_WRITABLE_MASK))
1494                                 continue;
1495                         ++nmaps;
1496                 }
1497         }
1498         return nmaps;
1499 }
1500
1501 static void audit_rmap(struct kvm_vcpu *vcpu)
1502 {
1503         int n_rmap = count_rmaps(vcpu);
1504         int n_actual = count_writable_mappings(vcpu);
1505
1506         if (n_rmap != n_actual)
1507                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1508                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1509 }
1510
1511 static void audit_write_protection(struct kvm_vcpu *vcpu)
1512 {
1513         struct kvm_mmu_page *page;
1514
1515         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1516                 hfn_t hfn;
1517                 struct page *pg;
1518
1519                 if (page->role.metaphysical)
1520                         continue;
1521
1522                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1523                         >> PAGE_SHIFT;
1524                 pg = pfn_to_page(hfn);
1525                 if (pg->private)
1526                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1527                                " mappings: gfn %lx role %x\n",
1528                                __FUNCTION__, audit_msg, page->gfn,
1529                                page->role.word);
1530         }
1531 }
1532
1533 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1534 {
1535         int olddbg = dbg;
1536
1537         dbg = 0;
1538         audit_msg = msg;
1539         audit_rmap(vcpu);
1540         audit_write_protection(vcpu);
1541         audit_mappings(vcpu);
1542         dbg = olddbg;
1543 }
1544
1545 #endif