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, 8);
 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  * Take gfn and return the reverse mapping to it.
 331  * Note: gfn must be unaliased before this function get called
 332  */
 333
 334 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
 335 {
 336         struct kvm_memory_slot *slot;
 337
 338         slot = gfn_to_memslot(kvm, gfn);
 339         return &slot->rmap[gfn - slot->base_gfn];
 340 }
 341
 342 /*
 343  * Reverse mapping data structures:
 344  *
 345  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 346  * that points to page_address(page).
 347  *
 348  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 349  * containing more mappings.
 350  */
 351 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 352 {
 353         struct kvm_mmu_page *page;
 354         struct kvm_rmap_desc *desc;
 355         unsigned long *rmapp;
 356         int i;
 357
 358         if (!is_rmap_pte(*spte))
 359                 return;
 360         gfn = unalias_gfn(vcpu->kvm, gfn);
 361         page = page_header(__pa(spte));
 362         page->gfns[spte - page->spt] = gfn;
 363         rmapp = gfn_to_rmap(vcpu->kvm, gfn);
 364         if (!*rmapp) {
 365                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 366                 *rmapp = (unsigned long)spte;
 367         } else if (!(*rmapp & 1)) {
 368                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 369                 desc = mmu_alloc_rmap_desc(vcpu);
 370                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 371                 desc->shadow_ptes[1] = spte;
 372                 *rmapp = (unsigned long)desc | 1;
 373         } else {
 374                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 375                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 376                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 377                         desc = desc->more;
 378                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 379                         desc->more = mmu_alloc_rmap_desc(vcpu);
 380                         desc = desc->more;
 381                 }
 382                 for (i = 0; desc->shadow_ptes[i]; ++i)
 383                         ;
 384                 desc->shadow_ptes[i] = spte;
 385         }
 386 }
 387
 388 static void rmap_desc_remove_entry(unsigned long *rmapp,
 389                                    struct kvm_rmap_desc *desc,
 390                                    int i,
 391                                    struct kvm_rmap_desc *prev_desc)
 392 {
 393         int j;
 394
 395         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 396                 ;
 397         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 398         desc->shadow_ptes[j] = NULL;
 399         if (j != 0)
 400                 return;
 401         if (!prev_desc && !desc->more)
 402                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 403         else
 404                 if (prev_desc)
 405                         prev_desc->more = desc->more;
 406                 else
 407                         *rmapp = (unsigned long)desc->more | 1;
 408         mmu_free_rmap_desc(desc);
 409 }
 410
 411 static void rmap_remove(struct kvm *kvm, u64 *spte)
 412 {
 413         struct kvm_rmap_desc *desc;
 414         struct kvm_rmap_desc *prev_desc;
 415         struct kvm_mmu_page *page;
 416         unsigned long *rmapp;
 417         int i;
 418
 419         if (!is_rmap_pte(*spte))
 420                 return;
 421         page = page_header(__pa(spte));
 422         rmapp = gfn_to_rmap(kvm, page->gfns[spte - page->spt]);
 423         if (!*rmapp) {
 424                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 425                 BUG();
 426         } else if (!(*rmapp & 1)) {
 427                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 428                 if ((u64 *)*rmapp != spte) {
 429                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 430                                spte, *spte);
 431                         BUG();
 432                 }
 433                 *rmapp = 0;
 434         } else {
 435                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 436                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 437                 prev_desc = NULL;
 438                 while (desc) {
 439                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 440                                 if (desc->shadow_ptes[i] == spte) {
 441                                         rmap_desc_remove_entry(rmapp,
 442                                                                desc, i,
 443                                                                prev_desc);
 444                                         return;
 445                                 }
 446                         prev_desc = desc;
 447                         desc = desc->more;
 448                 }
 449                 BUG();
 450         }
 451 }
 452
 453 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 454 {
 455         struct kvm_rmap_desc *desc;
 456         unsigned long *rmapp;
 457         u64 *spte;
 458
 459         gfn = unalias_gfn(vcpu->kvm, gfn);
 460         rmapp = gfn_to_rmap(vcpu->kvm, gfn);
 461
 462         while (*rmapp) {
 463                 if (!(*rmapp & 1))
 464                         spte = (u64 *)*rmapp;
 465                 else {
 466                         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 467                         spte = desc->shadow_ptes[0];
 468                 }
 469                 BUG_ON(!spte);
 470                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 471                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 472                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 473                 rmap_remove(vcpu->kvm, spte);
 474                 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 475                 kvm_flush_remote_tlbs(vcpu->kvm);
 476         }
 477 }
 478
 479 #ifdef MMU_DEBUG
 480 static int is_empty_shadow_page(u64 *spt)
 481 {
 482         u64 *pos;
 483         u64 *end;
 484
 485         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 486                 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
 487                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 488                                pos, *pos);
 489                         return 0;
 490                 }
 491         return 1;
 492 }
 493 #endif
 494
 495 static void kvm_mmu_free_page(struct kvm *kvm,
 496                               struct kvm_mmu_page *page_head)
 497 {
 498         ASSERT(is_empty_shadow_page(page_head->spt));
 499         list_del(&page_head->link);
 500         __free_page(virt_to_page(page_head->spt));
 501         __free_page(virt_to_page(page_head->gfns));
 502         kfree(page_head);
 503         ++kvm->n_free_mmu_pages;
 504 }
 505
 506 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 507 {
 508         return gfn;
 509 }
 510
 511 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 512                                                u64 *parent_pte)
 513 {
 514         struct kvm_mmu_page *page;
 515
 516         if (!vcpu->kvm->n_free_mmu_pages)
 517                 return NULL;
 518
 519         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 520                                       sizeof *page);
 521         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 522         page->gfns = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 523         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 524         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 525         ASSERT(is_empty_shadow_page(page->spt));
 526         page->slot_bitmap = 0;
 527         page->multimapped = 0;
 528         page->parent_pte = parent_pte;
 529         --vcpu->kvm->n_free_mmu_pages;
 530         return page;
 531 }
 532
 533 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 534                                     struct kvm_mmu_page *page, u64 *parent_pte)
 535 {
 536         struct kvm_pte_chain *pte_chain;
 537         struct hlist_node *node;
 538         int i;
 539
 540         if (!parent_pte)
 541                 return;
 542         if (!page->multimapped) {
 543                 u64 *old = page->parent_pte;
 544
 545                 if (!old) {
 546                         page->parent_pte = parent_pte;
 547                         return;
 548                 }
 549                 page->multimapped = 1;
 550                 pte_chain = mmu_alloc_pte_chain(vcpu);
 551                 INIT_HLIST_HEAD(&page->parent_ptes);
 552                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 553                 pte_chain->parent_ptes[0] = old;
 554         }
 555         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 556                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 557                         continue;
 558                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 559                         if (!pte_chain->parent_ptes[i]) {
 560                                 pte_chain->parent_ptes[i] = parent_pte;
 561                                 return;
 562                         }
 563         }
 564         pte_chain = mmu_alloc_pte_chain(vcpu);
 565         BUG_ON(!pte_chain);
 566         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 567         pte_chain->parent_ptes[0] = parent_pte;
 568 }
 569
 570 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
 571                                        u64 *parent_pte)
 572 {
 573         struct kvm_pte_chain *pte_chain;
 574         struct hlist_node *node;
 575         int i;
 576
 577         if (!page->multimapped) {
 578                 BUG_ON(page->parent_pte != parent_pte);
 579                 page->parent_pte = NULL;
 580                 return;
 581         }
 582         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 583                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 584                         if (!pte_chain->parent_ptes[i])
 585                                 break;
 586                         if (pte_chain->parent_ptes[i] != parent_pte)
 587                                 continue;
 588                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 589                                 && pte_chain->parent_ptes[i + 1]) {
 590                                 pte_chain->parent_ptes[i]
 591                                         = pte_chain->parent_ptes[i + 1];
 592                                 ++i;
 593                         }
 594                         pte_chain->parent_ptes[i] = NULL;
 595                         if (i == 0) {
 596                                 hlist_del(&pte_chain->link);
 597                                 mmu_free_pte_chain(pte_chain);
 598                                 if (hlist_empty(&page->parent_ptes)) {
 599                                         page->multimapped = 0;
 600                                         page->parent_pte = NULL;
 601                                 }
 602                         }
 603                         return;
 604                 }
 605         BUG();
 606 }
 607
 608 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 609                                                 gfn_t gfn)
 610 {
 611         unsigned index;
 612         struct hlist_head *bucket;
 613         struct kvm_mmu_page *page;
 614         struct hlist_node *node;
 615
 616         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 617         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 618         bucket = &vcpu->kvm->mmu_page_hash[index];
 619         hlist_for_each_entry(page, node, bucket, hash_link)
 620                 if (page->gfn == gfn && !page->role.metaphysical) {
 621                         pgprintk("%s: found role %x\n",
 622                                  __FUNCTION__, page->role.word);
 623                         return page;
 624                 }
 625         return NULL;
 626 }
 627
 628 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 629                                              gfn_t gfn,
 630                                              gva_t gaddr,
 631                                              unsigned level,
 632                                              int metaphysical,
 633                                              unsigned hugepage_access,
 634                                              u64 *parent_pte)
 635 {
 636         union kvm_mmu_page_role role;
 637         unsigned index;
 638         unsigned quadrant;
 639         struct hlist_head *bucket;
 640         struct kvm_mmu_page *page;
 641         struct hlist_node *node;
 642
 643         role.word = 0;
 644         role.glevels = vcpu->mmu.root_level;
 645         role.level = level;
 646         role.metaphysical = metaphysical;
 647         role.hugepage_access = hugepage_access;
 648         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 649                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 650                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 651                 role.quadrant = quadrant;
 652         }
 653         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 654                  gfn, role.word);
 655         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 656         bucket = &vcpu->kvm->mmu_page_hash[index];
 657         hlist_for_each_entry(page, node, bucket, hash_link)
 658                 if (page->gfn == gfn && page->role.word == role.word) {
 659                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 660                         pgprintk("%s: found\n", __FUNCTION__);
 661                         return page;
 662                 }
 663         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 664         if (!page)
 665                 return page;
 666         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 667         page->gfn = gfn;
 668         page->role = role;
 669         hlist_add_head(&page->hash_link, bucket);
 670         vcpu->mmu.prefetch_page(vcpu, page);
 671         if (!metaphysical)
 672                 rmap_write_protect(vcpu, gfn);
 673         return page;
 674 }
 675
 676 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 677                                          struct kvm_mmu_page *page)
 678 {
 679         unsigned i;
 680         u64 *pt;
 681         u64 ent;
 682
 683         pt = page->spt;
 684
 685         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 686                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 687                         if (is_shadow_present_pte(pt[i]))
 688                                 rmap_remove(kvm, &pt[i]);
 689                         pt[i] = shadow_trap_nonpresent_pte;
 690                 }
 691                 kvm_flush_remote_tlbs(kvm);
 692                 return;
 693         }
 694
 695         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 696                 ent = pt[i];
 697
 698                 pt[i] = shadow_trap_nonpresent_pte;
 699                 if (!is_shadow_present_pte(ent))
 700                         continue;
 701                 ent &= PT64_BASE_ADDR_MASK;
 702                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 703         }
 704         kvm_flush_remote_tlbs(kvm);
 705 }
 706
 707 static void kvm_mmu_put_page(struct kvm_mmu_page *page,
 708                              u64 *parent_pte)
 709 {
 710         mmu_page_remove_parent_pte(page, parent_pte);
 711 }
 712
 713 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 714 {
 715         int i;
 716
 717         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 718                 if (kvm->vcpus[i])
 719                         kvm->vcpus[i]->last_pte_updated = NULL;
 720 }
 721
 722 static void kvm_mmu_zap_page(struct kvm *kvm,
 723                              struct kvm_mmu_page *page)
 724 {
 725         u64 *parent_pte;
 726
 727         while (page->multimapped || page->parent_pte) {
 728                 if (!page->multimapped)
 729                         parent_pte = page->parent_pte;
 730                 else {
 731                         struct kvm_pte_chain *chain;
 732
 733                         chain = container_of(page->parent_ptes.first,
 734                                              struct kvm_pte_chain, link);
 735                         parent_pte = chain->parent_ptes[0];
 736                 }
 737                 BUG_ON(!parent_pte);
 738                 kvm_mmu_put_page(page, parent_pte);
 739                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 740         }
 741         kvm_mmu_page_unlink_children(kvm, page);
 742         if (!page->root_count) {
 743                 hlist_del(&page->hash_link);
 744                 kvm_mmu_free_page(kvm, page);
 745         } else
 746                 list_move(&page->link, &kvm->active_mmu_pages);
 747         kvm_mmu_reset_last_pte_updated(kvm);
 748 }
 749
 750 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 751 {
 752         unsigned index;
 753         struct hlist_head *bucket;
 754         struct kvm_mmu_page *page;
 755         struct hlist_node *node, *n;
 756         int r;
 757
 758         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 759         r = 0;
 760         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 761         bucket = &vcpu->kvm->mmu_page_hash[index];
 762         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 763                 if (page->gfn == gfn && !page->role.metaphysical) {
 764                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 765                                  page->role.word);
 766                         kvm_mmu_zap_page(vcpu->kvm, page);
 767                         r = 1;
 768                 }
 769         return r;
 770 }
 771
 772 static void mmu_unshadow(struct kvm_vcpu *vcpu, gfn_t gfn)
 773 {
 774         struct kvm_mmu_page *page;
 775
 776         while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
 777                 pgprintk("%s: zap %lx %x\n",
 778                          __FUNCTION__, gfn, page->role.word);
 779                 kvm_mmu_zap_page(vcpu->kvm, page);
 780         }
 781 }
 782
 783 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 784 {
 785         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 786         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 787
 788         __set_bit(slot, &page_head->slot_bitmap);
 789 }
 790
 791 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 792 {
 793         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 794
 795         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 796 }
 797
 798 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 799 {
 800         struct page *page;
 801
 802         ASSERT((gpa & HPA_ERR_MASK) == 0);
 803         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 804         if (!page)
 805                 return gpa | HPA_ERR_MASK;
 806         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 807                 | (gpa & (PAGE_SIZE-1));
 808 }
 809
 810 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 811 {
 812         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 813
 814         if (gpa == UNMAPPED_GVA)
 815                 return UNMAPPED_GVA;
 816         return gpa_to_hpa(vcpu, gpa);
 817 }
 818
 819 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 820 {
 821         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 822
 823         if (gpa == UNMAPPED_GVA)
 824                 return NULL;
 825         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 826 }
 827
 828 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 829 {
 830 }
 831
 832 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 833 {
 834         int level = PT32E_ROOT_LEVEL;
 835         hpa_t table_addr = vcpu->mmu.root_hpa;
 836
 837         for (; ; level--) {
 838                 u32 index = PT64_INDEX(v, level);
 839                 u64 *table;
 840                 u64 pte;
 841
 842                 ASSERT(VALID_PAGE(table_addr));
 843                 table = __va(table_addr);
 844
 845                 if (level == 1) {
 846                         pte = table[index];
 847                         if (is_shadow_present_pte(pte) && is_writeble_pte(pte))
 848                                 return 0;
 849                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 850                         page_header_update_slot(vcpu->kvm, table, v);
 851                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 852                                                                 PT_USER_MASK;
 853                         rmap_add(vcpu, &table[index], v >> PAGE_SHIFT);
 854                         return 0;
 855                 }
 856
 857                 if (table[index] == shadow_trap_nonpresent_pte) {
 858                         struct kvm_mmu_page *new_table;
 859                         gfn_t pseudo_gfn;
 860
 861                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 862                                 >> PAGE_SHIFT;
 863                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 864                                                      v, level - 1,
 865                                                      1, 0, &table[index]);
 866                         if (!new_table) {
 867                                 pgprintk("nonpaging_map: ENOMEM\n");
 868                                 return -ENOMEM;
 869                         }
 870
 871                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 872                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 873                 }
 874                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 875         }
 876 }
 877
 878 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 879                                     struct kvm_mmu_page *sp)
 880 {
 881         int i;
 882
 883         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 884                 sp->spt[i] = shadow_trap_nonpresent_pte;
 885 }
 886
 887 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 888 {
 889         int i;
 890         struct kvm_mmu_page *page;
 891
 892         if (!VALID_PAGE(vcpu->mmu.root_hpa))
 893                 return;
 894 #ifdef CONFIG_X86_64
 895         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 896                 hpa_t root = vcpu->mmu.root_hpa;
 897
 898                 page = page_header(root);
 899                 --page->root_count;
 900                 vcpu->mmu.root_hpa = INVALID_PAGE;
 901                 return;
 902         }
 903 #endif
 904         for (i = 0; i < 4; ++i) {
 905                 hpa_t root = vcpu->mmu.pae_root[i];
 906
 907                 if (root) {
 908                         root &= PT64_BASE_ADDR_MASK;
 909                         page = page_header(root);
 910                         --page->root_count;
 911                 }
 912                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 913         }
 914         vcpu->mmu.root_hpa = INVALID_PAGE;
 915 }
 916
 917 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 918 {
 919         int i;
 920         gfn_t root_gfn;
 921         struct kvm_mmu_page *page;
 922
 923         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 924
 925 #ifdef CONFIG_X86_64
 926         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 927                 hpa_t root = vcpu->mmu.root_hpa;
 928
 929                 ASSERT(!VALID_PAGE(root));
 930                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 931                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 932                 root = __pa(page->spt);
 933                 ++page->root_count;
 934                 vcpu->mmu.root_hpa = root;
 935                 return;
 936         }
 937 #endif
 938         for (i = 0; i < 4; ++i) {
 939                 hpa_t root = vcpu->mmu.pae_root[i];
 940
 941                 ASSERT(!VALID_PAGE(root));
 942                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 943                         if (!is_present_pte(vcpu->pdptrs[i])) {
 944                                 vcpu->mmu.pae_root[i] = 0;
 945                                 continue;
 946                         }
 947                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 948                 } else if (vcpu->mmu.root_level == 0)
 949                         root_gfn = 0;
 950                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 951                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 952                                         0, NULL);
 953                 root = __pa(page->spt);
 954                 ++page->root_count;
 955                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 956         }
 957         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 958 }
 959
 960 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 961 {
 962         return vaddr;
 963 }
 964
 965 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 966                                u32 error_code)
 967 {
 968         gpa_t addr = gva;
 969         hpa_t paddr;
 970         int r;
 971
 972         r = mmu_topup_memory_caches(vcpu);
 973         if (r)
 974                 return r;
 975
 976         ASSERT(vcpu);
 977         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 978
 979
 980         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 981
 982         if (is_error_hpa(paddr))
 983                 return 1;
 984
 985         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 986 }
 987
 988 static void nonpaging_free(struct kvm_vcpu *vcpu)
 989 {
 990         mmu_free_roots(vcpu);
 991 }
 992
 993 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 994 {
 995         struct kvm_mmu *context = &vcpu->mmu;
 996
 997         context->new_cr3 = nonpaging_new_cr3;
 998         context->page_fault = nonpaging_page_fault;
 999         context->gva_to_gpa = nonpaging_gva_to_gpa;
1000         context->free = nonpaging_free;
1001         context->prefetch_page = nonpaging_prefetch_page;
1002         context->root_level = 0;
1003         context->shadow_root_level = PT32E_ROOT_LEVEL;
1004         context->root_hpa = INVALID_PAGE;
1005         return 0;
1006 }
1007
1008 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1009 {
1010         ++vcpu->stat.tlb_flush;
1011         kvm_x86_ops->tlb_flush(vcpu);
1012 }
1013
1014 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1015 {
1016         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
1017         mmu_free_roots(vcpu);
1018 }
1019
1020 static void inject_page_fault(struct kvm_vcpu *vcpu,
1021                               u64 addr,
1022                               u32 err_code)
1023 {
1024         kvm_x86_ops->inject_page_fault(vcpu, addr, err_code);
1025 }
1026
1027 static void paging_free(struct kvm_vcpu *vcpu)
1028 {
1029         nonpaging_free(vcpu);
1030 }
1031
1032 #define PTTYPE 64
1033 #include "paging_tmpl.h"
1034 #undef PTTYPE
1035
1036 #define PTTYPE 32
1037 #include "paging_tmpl.h"
1038 #undef PTTYPE
1039
1040 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1041 {
1042         struct kvm_mmu *context = &vcpu->mmu;
1043
1044         ASSERT(is_pae(vcpu));
1045         context->new_cr3 = paging_new_cr3;
1046         context->page_fault = paging64_page_fault;
1047         context->gva_to_gpa = paging64_gva_to_gpa;
1048         context->prefetch_page = paging64_prefetch_page;
1049         context->free = paging_free;
1050         context->root_level = level;
1051         context->shadow_root_level = level;
1052         context->root_hpa = INVALID_PAGE;
1053         return 0;
1054 }
1055
1056 static int paging64_init_context(struct kvm_vcpu *vcpu)
1057 {
1058         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1059 }
1060
1061 static int paging32_init_context(struct kvm_vcpu *vcpu)
1062 {
1063         struct kvm_mmu *context = &vcpu->mmu;
1064
1065         context->new_cr3 = paging_new_cr3;
1066         context->page_fault = paging32_page_fault;
1067         context->gva_to_gpa = paging32_gva_to_gpa;
1068         context->free = paging_free;
1069         context->prefetch_page = paging32_prefetch_page;
1070         context->root_level = PT32_ROOT_LEVEL;
1071         context->shadow_root_level = PT32E_ROOT_LEVEL;
1072         context->root_hpa = INVALID_PAGE;
1073         return 0;
1074 }
1075
1076 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1077 {
1078         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1079 }
1080
1081 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1082 {
1083         ASSERT(vcpu);
1084         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1085
1086         if (!is_paging(vcpu))
1087                 return nonpaging_init_context(vcpu);
1088         else if (is_long_mode(vcpu))
1089                 return paging64_init_context(vcpu);
1090         else if (is_pae(vcpu))
1091                 return paging32E_init_context(vcpu);
1092         else
1093                 return paging32_init_context(vcpu);
1094 }
1095
1096 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1097 {
1098         ASSERT(vcpu);
1099         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1100                 vcpu->mmu.free(vcpu);
1101                 vcpu->mmu.root_hpa = INVALID_PAGE;
1102         }
1103 }
1104
1105 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1106 {
1107         destroy_kvm_mmu(vcpu);
1108         return init_kvm_mmu(vcpu);
1109 }
1110 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1111
1112 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1113 {
1114         int r;
1115
1116         mutex_lock(&vcpu->kvm->lock);
1117         r = mmu_topup_memory_caches(vcpu);
1118         if (r)
1119                 goto out;
1120         mmu_alloc_roots(vcpu);
1121         kvm_x86_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1122         kvm_mmu_flush_tlb(vcpu);
1123 out:
1124         mutex_unlock(&vcpu->kvm->lock);
1125         return r;
1126 }
1127 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1128
1129 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1130 {
1131         mmu_free_roots(vcpu);
1132 }
1133
1134 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1135                                   struct kvm_mmu_page *page,
1136                                   u64 *spte)
1137 {
1138         u64 pte;
1139         struct kvm_mmu_page *child;
1140
1141         pte = *spte;
1142         if (is_shadow_present_pte(pte)) {
1143                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1144                         rmap_remove(vcpu->kvm, spte);
1145                 else {
1146                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1147                         mmu_page_remove_parent_pte(child, spte);
1148                 }
1149         }
1150         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1151         kvm_flush_remote_tlbs(vcpu->kvm);
1152 }
1153
1154 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1155                                   struct kvm_mmu_page *page,
1156                                   u64 *spte,
1157                                   const void *new, int bytes,
1158                                   int offset_in_pte)
1159 {
1160         if (page->role.level != PT_PAGE_TABLE_LEVEL)
1161                 return;
1162
1163         if (page->role.glevels == PT32_ROOT_LEVEL)
1164                 paging32_update_pte(vcpu, page, spte, new, bytes,
1165                                     offset_in_pte);
1166         else
1167                 paging64_update_pte(vcpu, page, spte, new, bytes,
1168                                     offset_in_pte);
1169 }
1170
1171 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1172 {
1173         u64 *spte = vcpu->last_pte_updated;
1174
1175         return !!(spte && (*spte & PT_ACCESSED_MASK));
1176 }
1177
1178 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1179                        const u8 *new, int bytes)
1180 {
1181         gfn_t gfn = gpa >> PAGE_SHIFT;
1182         struct kvm_mmu_page *page;
1183         struct hlist_node *node, *n;
1184         struct hlist_head *bucket;
1185         unsigned index;
1186         u64 *spte;
1187         unsigned offset = offset_in_page(gpa);
1188         unsigned pte_size;
1189         unsigned page_offset;
1190         unsigned misaligned;
1191         unsigned quadrant;
1192         int level;
1193         int flooded = 0;
1194         int npte;
1195
1196         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1197         kvm_mmu_audit(vcpu, "pre pte write");
1198         if (gfn == vcpu->last_pt_write_gfn
1199             && !last_updated_pte_accessed(vcpu)) {
1200                 ++vcpu->last_pt_write_count;
1201                 if (vcpu->last_pt_write_count >= 3)
1202                         flooded = 1;
1203         } else {
1204                 vcpu->last_pt_write_gfn = gfn;
1205                 vcpu->last_pt_write_count = 1;
1206                 vcpu->last_pte_updated = NULL;
1207         }
1208         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1209         bucket = &vcpu->kvm->mmu_page_hash[index];
1210         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1211                 if (page->gfn != gfn || page->role.metaphysical)
1212                         continue;
1213                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1214                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1215                 misaligned |= bytes < 4;
1216                 if (misaligned || flooded) {
1217                         /*
1218                          * Misaligned accesses are too much trouble to fix
1219                          * up; also, they usually indicate a page is not used
1220                          * as a page table.
1221                          *
1222                          * If we're seeing too many writes to a page,
1223                          * it may no longer be a page table, or we may be
1224                          * forking, in which case it is better to unmap the
1225                          * page.
1226                          */
1227                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1228                                  gpa, bytes, page->role.word);
1229                         kvm_mmu_zap_page(vcpu->kvm, page);
1230                         continue;
1231                 }
1232                 page_offset = offset;
1233                 level = page->role.level;
1234                 npte = 1;
1235                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1236                         page_offset <<= 1;      /* 32->64 */
1237                         /*
1238                          * A 32-bit pde maps 4MB while the shadow pdes map
1239                          * only 2MB.  So we need to double the offset again
1240                          * and zap two pdes instead of one.
1241                          */
1242                         if (level == PT32_ROOT_LEVEL) {
1243                                 page_offset &= ~7; /* kill rounding error */
1244                                 page_offset <<= 1;
1245                                 npte = 2;
1246                         }
1247                         quadrant = page_offset >> PAGE_SHIFT;
1248                         page_offset &= ~PAGE_MASK;
1249                         if (quadrant != page->role.quadrant)
1250                                 continue;
1251                 }
1252                 spte = &page->spt[page_offset / sizeof(*spte)];
1253                 while (npte--) {
1254                         mmu_pte_write_zap_pte(vcpu, page, spte);
1255                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes,
1256                                               page_offset & (pte_size - 1));
1257                         ++spte;
1258                 }
1259         }
1260         kvm_mmu_audit(vcpu, "post pte write");
1261 }
1262
1263 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1264 {
1265         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1266
1267         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1268 }
1269
1270 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1271 {
1272         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1273                 struct kvm_mmu_page *page;
1274
1275                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1276                                     struct kvm_mmu_page, link);
1277                 kvm_mmu_zap_page(vcpu->kvm, page);
1278         }
1279 }
1280
1281 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1282 {
1283         struct kvm_mmu_page *page;
1284
1285         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1286                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1287                                     struct kvm_mmu_page, link);
1288                 kvm_mmu_zap_page(vcpu->kvm, page);
1289         }
1290         free_page((unsigned long)vcpu->mmu.pae_root);
1291 }
1292
1293 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1294 {
1295         struct page *page;
1296         int i;
1297
1298         ASSERT(vcpu);
1299
1300         vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1301
1302         /*
1303          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1304          * Therefore we need to allocate shadow page tables in the first
1305          * 4GB of memory, which happens to fit the DMA32 zone.
1306          */
1307         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1308         if (!page)
1309                 goto error_1;
1310         vcpu->mmu.pae_root = page_address(page);
1311         for (i = 0; i < 4; ++i)
1312                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1313
1314         return 0;
1315
1316 error_1:
1317         free_mmu_pages(vcpu);
1318         return -ENOMEM;
1319 }
1320
1321 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1322 {
1323         ASSERT(vcpu);
1324         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1325
1326         return alloc_mmu_pages(vcpu);
1327 }
1328
1329 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1330 {
1331         ASSERT(vcpu);
1332         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1333
1334         return init_kvm_mmu(vcpu);
1335 }
1336
1337 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1338 {
1339         ASSERT(vcpu);
1340
1341         destroy_kvm_mmu(vcpu);
1342         free_mmu_pages(vcpu);
1343         mmu_free_memory_caches(vcpu);
1344 }
1345
1346 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1347 {
1348         struct kvm_mmu_page *page;
1349
1350         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1351                 int i;
1352                 u64 *pt;
1353
1354                 if (!test_bit(slot, &page->slot_bitmap))
1355                         continue;
1356
1357                 pt = page->spt;
1358                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1359                         /* avoid RMW */
1360                         if (pt[i] & PT_WRITABLE_MASK) {
1361                                 rmap_remove(kvm, &pt[i]);
1362                                 pt[i] &= ~PT_WRITABLE_MASK;
1363                         }
1364         }
1365 }
1366
1367 void kvm_mmu_zap_all(struct kvm *kvm)
1368 {
1369         struct kvm_mmu_page *page, *node;
1370
1371         list_for_each_entry_safe(page, node, &kvm->active_mmu_pages, link)
1372                 kvm_mmu_zap_page(kvm, page);
1373
1374         kvm_flush_remote_tlbs(kvm);
1375 }
1376
1377 void kvm_mmu_module_exit(void)
1378 {
1379         if (pte_chain_cache)
1380                 kmem_cache_destroy(pte_chain_cache);
1381         if (rmap_desc_cache)
1382                 kmem_cache_destroy(rmap_desc_cache);
1383         if (mmu_page_header_cache)
1384                 kmem_cache_destroy(mmu_page_header_cache);
1385 }
1386
1387 int kvm_mmu_module_init(void)
1388 {
1389         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1390                                             sizeof(struct kvm_pte_chain),
1391                                             0, 0, NULL);
1392         if (!pte_chain_cache)
1393                 goto nomem;
1394         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1395                                             sizeof(struct kvm_rmap_desc),
1396                                             0, 0, NULL);
1397         if (!rmap_desc_cache)
1398                 goto nomem;
1399
1400         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1401                                                   sizeof(struct kvm_mmu_page),
1402                                                   0, 0, NULL);
1403         if (!mmu_page_header_cache)
1404                 goto nomem;
1405
1406         return 0;
1407
1408 nomem:
1409         kvm_mmu_module_exit();
1410         return -ENOMEM;
1411 }
1412
1413 #ifdef AUDIT
1414
1415 static const char *audit_msg;
1416
1417 static gva_t canonicalize(gva_t gva)
1418 {
1419 #ifdef CONFIG_X86_64
1420         gva = (long long)(gva << 16) >> 16;
1421 #endif
1422         return gva;
1423 }
1424
1425 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1426                                 gva_t va, int level)
1427 {
1428         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1429         int i;
1430         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1431
1432         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1433                 u64 ent = pt[i];
1434
1435                 if (ent == shadow_trap_nonpresent_pte)
1436                         continue;
1437
1438                 va = canonicalize(va);
1439                 if (level > 1) {
1440                         if (ent == shadow_notrap_nonpresent_pte)
1441                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1442                                        " in nonleaf level: levels %d gva %lx"
1443                                        " level %d pte %llx\n", audit_msg,
1444                                        vcpu->mmu.root_level, va, level, ent);
1445
1446                         audit_mappings_page(vcpu, ent, va, level - 1);
1447                 } else {
1448                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1449                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1450
1451                         if (is_shadow_present_pte(ent)
1452                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1453                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1454                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1455                                        audit_msg, vcpu->mmu.root_level,
1456                                        va, gpa, hpa, ent, is_shadow_present_pte(ent));
1457                         else if (ent == shadow_notrap_nonpresent_pte
1458                                  && !is_error_hpa(hpa))
1459                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1460                                        " valid guest gva %lx\n", audit_msg, va);
1461
1462                 }
1463         }
1464 }
1465
1466 static void audit_mappings(struct kvm_vcpu *vcpu)
1467 {
1468         unsigned i;
1469
1470         if (vcpu->mmu.root_level == 4)
1471                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1472         else
1473                 for (i = 0; i < 4; ++i)
1474                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1475                                 audit_mappings_page(vcpu,
1476                                                     vcpu->mmu.pae_root[i],
1477                                                     i << 30,
1478                                                     2);
1479 }
1480
1481 static int count_rmaps(struct kvm_vcpu *vcpu)
1482 {
1483         int nmaps = 0;
1484         int i, j, k;
1485
1486         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1487                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1488                 struct kvm_rmap_desc *d;
1489
1490                 for (j = 0; j < m->npages; ++j) {
1491                         unsigned long *rmapp = &m->rmap[j];
1492
1493                         if (!*rmapp)
1494                                 continue;
1495                         if (!(*rmapp & 1)) {
1496                                 ++nmaps;
1497                                 continue;
1498                         }
1499                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1500                         while (d) {
1501                                 for (k = 0; k < RMAP_EXT; ++k)
1502                                         if (d->shadow_ptes[k])
1503                                                 ++nmaps;
1504                                         else
1505                                                 break;
1506                                 d = d->more;
1507                         }
1508                 }
1509         }
1510         return nmaps;
1511 }
1512
1513 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1514 {
1515         int nmaps = 0;
1516         struct kvm_mmu_page *page;
1517         int i;
1518
1519         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1520                 u64 *pt = page->spt;
1521
1522                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1523                         continue;
1524
1525                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1526                         u64 ent = pt[i];
1527
1528                         if (!(ent & PT_PRESENT_MASK))
1529                                 continue;
1530                         if (!(ent & PT_WRITABLE_MASK))
1531                                 continue;
1532                         ++nmaps;
1533                 }
1534         }
1535         return nmaps;
1536 }
1537
1538 static void audit_rmap(struct kvm_vcpu *vcpu)
1539 {
1540         int n_rmap = count_rmaps(vcpu);
1541         int n_actual = count_writable_mappings(vcpu);
1542
1543         if (n_rmap != n_actual)
1544                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1545                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1546 }
1547
1548 static void audit_write_protection(struct kvm_vcpu *vcpu)
1549 {
1550         struct kvm_mmu_page *page;
1551         struct kvm_memory_slot *slot;
1552         unsigned long *rmapp;
1553         gfn_t gfn;
1554
1555         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1556                 if (page->role.metaphysical)
1557                         continue;
1558
1559                 slot = gfn_to_memslot(vcpu->kvm, page->gfn);
1560                 gfn = unalias_gfn(vcpu->kvm, page->gfn);
1561                 rmapp = &slot->rmap[gfn - slot->base_gfn];
1562                 if (*rmapp)
1563                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1564                                " mappings: gfn %lx role %x\n",
1565                                __FUNCTION__, audit_msg, page->gfn,
1566                                page->role.word);
1567         }
1568 }
1569
1570 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1571 {
1572         int olddbg = dbg;
1573
1574         dbg = 0;
1575         audit_msg = msg;
1576         audit_rmap(vcpu);
1577         audit_write_protection(vcpu);
1578         audit_mappings(vcpu);
1579         dbg = olddbg;
1580 }
1581
1582 #endif