drivers/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19 #include <linux/types.h>
  20 #include <linux/string.h>
  21 #include <asm/page.h>
  22 #include <linux/mm.h>
  23 #include <linux/highmem.h>
  24 #include <linux/module.h>
  25
  26 #include "vmx.h"
  27 #include "kvm.h"
  28
  29 #undef MMU_DEBUG
  30
  31 #undef AUDIT
  32
  33 #ifdef AUDIT
  34 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  35 #else
  36 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  37 #endif
  38
  39 #ifdef MMU_DEBUG
  40
  41 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  42 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  43
  44 #else
  45
  46 #define pgprintk(x...) do { } while (0)
  47 #define rmap_printk(x...) do { } while (0)
  48
  49 #endif
  50
  51 #if defined(MMU_DEBUG) || defined(AUDIT)
  52 static int dbg = 1;
  53 #endif
  54
  55 #ifndef MMU_DEBUG
  56 #define ASSERT(x) do { } while (0)
  57 #else
  58 #define ASSERT(x)                                                       \
  59         if (!(x)) {                                                     \
  60                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  61                        __FILE__, __LINE__, #x);                         \
  62         }
  63 #endif
  64
  65 #define PT64_PT_BITS 9
  66 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  67 #define PT32_PT_BITS 10
  68 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  69
  70 #define PT_WRITABLE_SHIFT 1
  71
  72 #define PT_PRESENT_MASK (1ULL << 0)
  73 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  74 #define PT_USER_MASK (1ULL << 2)
  75 #define PT_PWT_MASK (1ULL << 3)
  76 #define PT_PCD_MASK (1ULL << 4)
  77 #define PT_ACCESSED_MASK (1ULL << 5)
  78 #define PT_DIRTY_MASK (1ULL << 6)
  79 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  80 #define PT_PAT_MASK (1ULL << 7)
  81 #define PT_GLOBAL_MASK (1ULL << 8)
  82 #define PT64_NX_MASK (1ULL << 63)
  83
  84 #define PT_PAT_SHIFT 7
  85 #define PT_DIR_PAT_SHIFT 12
  86 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  87
  88 #define PT32_DIR_PSE36_SIZE 4
  89 #define PT32_DIR_PSE36_SHIFT 13
  90 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  91
  92
  93 #define PT32_PTE_COPY_MASK \
  94         (PT_PRESENT_MASK | PT_ACCESSED_MASK | PT_DIRTY_MASK | PT_GLOBAL_MASK)
  95
  96 #define PT64_PTE_COPY_MASK (PT64_NX_MASK | PT32_PTE_COPY_MASK)
  97
  98 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  99 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
 100
 101 #define PT_SHADOW_PS_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 102 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 103
 104 #define PT_SHADOW_WRITABLE_SHIFT (PT_FIRST_AVAIL_BITS_SHIFT + 1)
 105 #define PT_SHADOW_WRITABLE_MASK (1ULL << PT_SHADOW_WRITABLE_SHIFT)
 106
 107 #define PT_SHADOW_USER_SHIFT (PT_SHADOW_WRITABLE_SHIFT + 1)
 108 #define PT_SHADOW_USER_MASK (1ULL << (PT_SHADOW_USER_SHIFT))
 109
 110 #define PT_SHADOW_BITS_OFFSET (PT_SHADOW_WRITABLE_SHIFT - PT_WRITABLE_SHIFT)
 111
 112 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 113
 114 #define PT64_LEVEL_BITS 9
 115
 116 #define PT64_LEVEL_SHIFT(level) \
 117                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
 118
 119 #define PT64_LEVEL_MASK(level) \
 120                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 121
 122 #define PT64_INDEX(address, level)\
 123         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 124
 125
 126 #define PT32_LEVEL_BITS 10
 127
 128 #define PT32_LEVEL_SHIFT(level) \
 129                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
 130
 131 #define PT32_LEVEL_MASK(level) \
 132                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 133
 134 #define PT32_INDEX(address, level)\
 135         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 136
 137
 138 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 139 #define PT64_DIR_BASE_ADDR_MASK \
 140         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 141
 142 #define PT32_BASE_ADDR_MASK PAGE_MASK
 143 #define PT32_DIR_BASE_ADDR_MASK \
 144         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 145
 146
 147 #define PFERR_PRESENT_MASK (1U << 0)
 148 #define PFERR_WRITE_MASK (1U << 1)
 149 #define PFERR_USER_MASK (1U << 2)
 150 #define PFERR_FETCH_MASK (1U << 4)
 151
 152 #define PT64_ROOT_LEVEL 4
 153 #define PT32_ROOT_LEVEL 2
 154 #define PT32E_ROOT_LEVEL 3
 155
 156 #define PT_DIRECTORY_LEVEL 2
 157 #define PT_PAGE_TABLE_LEVEL 1
 158
 159 #define RMAP_EXT 4
 160
 161 struct kvm_rmap_desc {
 162         u64 *shadow_ptes[RMAP_EXT];
 163         struct kvm_rmap_desc *more;
 164 };
 165
 166 static struct kmem_cache *pte_chain_cache;
 167 static struct kmem_cache *rmap_desc_cache;
 168 static struct kmem_cache *mmu_page_cache;
 169 static struct kmem_cache *mmu_page_header_cache;
 170
 171 static int is_write_protection(struct kvm_vcpu *vcpu)
 172 {
 173         return vcpu->cr0 & CR0_WP_MASK;
 174 }
 175
 176 static int is_cpuid_PSE36(void)
 177 {
 178         return 1;
 179 }
 180
 181 static int is_nx(struct kvm_vcpu *vcpu)
 182 {
 183         return vcpu->shadow_efer & EFER_NX;
 184 }
 185
 186 static int is_present_pte(unsigned long pte)
 187 {
 188         return pte & PT_PRESENT_MASK;
 189 }
 190
 191 static int is_writeble_pte(unsigned long pte)
 192 {
 193         return pte & PT_WRITABLE_MASK;
 194 }
 195
 196 static int is_io_pte(unsigned long pte)
 197 {
 198         return pte & PT_SHADOW_IO_MARK;
 199 }
 200
 201 static int is_rmap_pte(u64 pte)
 202 {
 203         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 204                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 205 }
 206
 207 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 208                                   struct kmem_cache *base_cache, int min,
 209                                   gfp_t gfp_flags)
 210 {
 211         void *obj;
 212
 213         if (cache->nobjs >= min)
 214                 return 0;
 215         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 216                 obj = kmem_cache_zalloc(base_cache, gfp_flags);
 217                 if (!obj)
 218                         return -ENOMEM;
 219                 cache->objects[cache->nobjs++] = obj;
 220         }
 221         return 0;
 222 }
 223
 224 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 225 {
 226         while (mc->nobjs)
 227                 kfree(mc->objects[--mc->nobjs]);
 228 }
 229
 230 static int __mmu_topup_memory_caches(struct kvm_vcpu *vcpu, gfp_t gfp_flags)
 231 {
 232         int r;
 233
 234         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 235                                    pte_chain_cache, 4, gfp_flags);
 236         if (r)
 237                 goto out;
 238         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 239                                    rmap_desc_cache, 1, gfp_flags);
 240         if (r)
 241                 goto out;
 242         r = mmu_topup_memory_cache(&vcpu->mmu_page_cache,
 243                                    mmu_page_cache, 4, gfp_flags);
 244         if (r)
 245                 goto out;
 246         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
 247                                    mmu_page_header_cache, 4, gfp_flags);
 248 out:
 249         return r;
 250 }
 251
 252 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 253 {
 254         int r;
 255
 256         r = __mmu_topup_memory_caches(vcpu, GFP_NOWAIT);
 257         if (r < 0) {
 258                 spin_unlock(&vcpu->kvm->lock);
 259                 kvm_arch_ops->vcpu_put(vcpu);
 260                 r = __mmu_topup_memory_caches(vcpu, GFP_KERNEL);
 261                 kvm_arch_ops->vcpu_load(vcpu);
 262                 spin_lock(&vcpu->kvm->lock);
 263         }
 264         return r;
 265 }
 266
 267 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 268 {
 269         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 270         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 271         mmu_free_memory_cache(&vcpu->mmu_page_cache);
 272         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
 273 }
 274
 275 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 276                                     size_t size)
 277 {
 278         void *p;
 279
 280         BUG_ON(!mc->nobjs);
 281         p = mc->objects[--mc->nobjs];
 282         memset(p, 0, size);
 283         return p;
 284 }
 285
 286 static void mmu_memory_cache_free(struct kvm_mmu_memory_cache *mc, void *obj)
 287 {
 288         if (mc->nobjs < KVM_NR_MEM_OBJS)
 289                 mc->objects[mc->nobjs++] = obj;
 290         else
 291                 kfree(obj);
 292 }
 293
 294 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 295 {
 296         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 297                                       sizeof(struct kvm_pte_chain));
 298 }
 299
 300 static void mmu_free_pte_chain(struct kvm_vcpu *vcpu,
 301                                struct kvm_pte_chain *pc)
 302 {
 303         mmu_memory_cache_free(&vcpu->mmu_pte_chain_cache, pc);
 304 }
 305
 306 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 307 {
 308         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 309                                       sizeof(struct kvm_rmap_desc));
 310 }
 311
 312 static void mmu_free_rmap_desc(struct kvm_vcpu *vcpu,
 313                                struct kvm_rmap_desc *rd)
 314 {
 315         mmu_memory_cache_free(&vcpu->mmu_rmap_desc_cache, rd);
 316 }
 317
 318 /*
 319  * Reverse mapping data structures:
 320  *
 321  * If page->private bit zero is zero, then page->private points to the
 322  * shadow page table entry that points to page_address(page).
 323  *
 324  * If page->private bit zero is one, (then page->private & ~1) points
 325  * to a struct kvm_rmap_desc containing more mappings.
 326  */
 327 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 328 {
 329         struct page *page;
 330         struct kvm_rmap_desc *desc;
 331         int i;
 332
 333         if (!is_rmap_pte(*spte))
 334                 return;
 335         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 336         if (!page_private(page)) {
 337                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 338                 set_page_private(page,(unsigned long)spte);
 339         } else if (!(page_private(page) & 1)) {
 340                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 341                 desc = mmu_alloc_rmap_desc(vcpu);
 342                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 343                 desc->shadow_ptes[1] = spte;
 344                 set_page_private(page,(unsigned long)desc | 1);
 345         } else {
 346                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 347                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 348                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 349                         desc = desc->more;
 350                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 351                         desc->more = mmu_alloc_rmap_desc(vcpu);
 352                         desc = desc->more;
 353                 }
 354                 for (i = 0; desc->shadow_ptes[i]; ++i)
 355                         ;
 356                 desc->shadow_ptes[i] = spte;
 357         }
 358 }
 359
 360 static void rmap_desc_remove_entry(struct kvm_vcpu *vcpu,
 361                                    struct page *page,
 362                                    struct kvm_rmap_desc *desc,
 363                                    int i,
 364                                    struct kvm_rmap_desc *prev_desc)
 365 {
 366         int j;
 367
 368         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 369                 ;
 370         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 371         desc->shadow_ptes[j] = NULL;
 372         if (j != 0)
 373                 return;
 374         if (!prev_desc && !desc->more)
 375                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 376         else
 377                 if (prev_desc)
 378                         prev_desc->more = desc->more;
 379                 else
 380                         set_page_private(page,(unsigned long)desc->more | 1);
 381         mmu_free_rmap_desc(vcpu, desc);
 382 }
 383
 384 static void rmap_remove(struct kvm_vcpu *vcpu, u64 *spte)
 385 {
 386         struct page *page;
 387         struct kvm_rmap_desc *desc;
 388         struct kvm_rmap_desc *prev_desc;
 389         int i;
 390
 391         if (!is_rmap_pte(*spte))
 392                 return;
 393         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 394         if (!page_private(page)) {
 395                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 396                 BUG();
 397         } else if (!(page_private(page) & 1)) {
 398                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 399                 if ((u64 *)page_private(page) != spte) {
 400                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 401                                spte, *spte);
 402                         BUG();
 403                 }
 404                 set_page_private(page,0);
 405         } else {
 406                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 407                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 408                 prev_desc = NULL;
 409                 while (desc) {
 410                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 411                                 if (desc->shadow_ptes[i] == spte) {
 412                                         rmap_desc_remove_entry(vcpu, page,
 413                                                                desc, i,
 414                                                                prev_desc);
 415                                         return;
 416                                 }
 417                         prev_desc = desc;
 418                         desc = desc->more;
 419                 }
 420                 BUG();
 421         }
 422 }
 423
 424 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 425 {
 426         struct kvm *kvm = vcpu->kvm;
 427         struct page *page;
 428         struct kvm_rmap_desc *desc;
 429         u64 *spte;
 430
 431         page = gfn_to_page(kvm, gfn);
 432         BUG_ON(!page);
 433
 434         while (page_private(page)) {
 435                 if (!(page_private(page) & 1))
 436                         spte = (u64 *)page_private(page);
 437                 else {
 438                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 439                         spte = desc->shadow_ptes[0];
 440                 }
 441                 BUG_ON(!spte);
 442                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 443                        != page_to_pfn(page));
 444                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 445                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 446                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 447                 rmap_remove(vcpu, spte);
 448                 kvm_arch_ops->tlb_flush(vcpu);
 449                 *spte &= ~(u64)PT_WRITABLE_MASK;
 450         }
 451 }
 452
 453 #ifdef MMU_DEBUG
 454 static int is_empty_shadow_page(u64 *spt)
 455 {
 456         u64 *pos;
 457         u64 *end;
 458
 459         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 460                 if (*pos != 0) {
 461                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 462                                pos, *pos);
 463                         return 0;
 464                 }
 465         return 1;
 466 }
 467 #endif
 468
 469 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu,
 470                               struct kvm_mmu_page *page_head)
 471 {
 472         ASSERT(is_empty_shadow_page(page_head->spt));
 473         list_del(&page_head->link);
 474         mmu_memory_cache_free(&vcpu->mmu_page_cache, page_head->spt);
 475         mmu_memory_cache_free(&vcpu->mmu_page_header_cache, page_head);
 476         ++vcpu->kvm->n_free_mmu_pages;
 477 }
 478
 479 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 480 {
 481         return gfn;
 482 }
 483
 484 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 485                                                u64 *parent_pte)
 486 {
 487         struct kvm_mmu_page *page;
 488
 489         if (!vcpu->kvm->n_free_mmu_pages)
 490                 return NULL;
 491
 492         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 493                                       sizeof *page);
 494         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 495         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 496         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 497         ASSERT(is_empty_shadow_page(page->spt));
 498         page->slot_bitmap = 0;
 499         page->multimapped = 0;
 500         page->parent_pte = parent_pte;
 501         --vcpu->kvm->n_free_mmu_pages;
 502         return page;
 503 }
 504
 505 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 506                                     struct kvm_mmu_page *page, u64 *parent_pte)
 507 {
 508         struct kvm_pte_chain *pte_chain;
 509         struct hlist_node *node;
 510         int i;
 511
 512         if (!parent_pte)
 513                 return;
 514         if (!page->multimapped) {
 515                 u64 *old = page->parent_pte;
 516
 517                 if (!old) {
 518                         page->parent_pte = parent_pte;
 519                         return;
 520                 }
 521                 page->multimapped = 1;
 522                 pte_chain = mmu_alloc_pte_chain(vcpu);
 523                 INIT_HLIST_HEAD(&page->parent_ptes);
 524                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 525                 pte_chain->parent_ptes[0] = old;
 526         }
 527         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 528                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 529                         continue;
 530                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 531                         if (!pte_chain->parent_ptes[i]) {
 532                                 pte_chain->parent_ptes[i] = parent_pte;
 533                                 return;
 534                         }
 535         }
 536         pte_chain = mmu_alloc_pte_chain(vcpu);
 537         BUG_ON(!pte_chain);
 538         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 539         pte_chain->parent_ptes[0] = parent_pte;
 540 }
 541
 542 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
 543                                        struct kvm_mmu_page *page,
 544                                        u64 *parent_pte)
 545 {
 546         struct kvm_pte_chain *pte_chain;
 547         struct hlist_node *node;
 548         int i;
 549
 550         if (!page->multimapped) {
 551                 BUG_ON(page->parent_pte != parent_pte);
 552                 page->parent_pte = NULL;
 553                 return;
 554         }
 555         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 556                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 557                         if (!pte_chain->parent_ptes[i])
 558                                 break;
 559                         if (pte_chain->parent_ptes[i] != parent_pte)
 560                                 continue;
 561                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 562                                 && pte_chain->parent_ptes[i + 1]) {
 563                                 pte_chain->parent_ptes[i]
 564                                         = pte_chain->parent_ptes[i + 1];
 565                                 ++i;
 566                         }
 567                         pte_chain->parent_ptes[i] = NULL;
 568                         if (i == 0) {
 569                                 hlist_del(&pte_chain->link);
 570                                 mmu_free_pte_chain(vcpu, pte_chain);
 571                                 if (hlist_empty(&page->parent_ptes)) {
 572                                         page->multimapped = 0;
 573                                         page->parent_pte = NULL;
 574                                 }
 575                         }
 576                         return;
 577                 }
 578         BUG();
 579 }
 580
 581 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 582                                                 gfn_t gfn)
 583 {
 584         unsigned index;
 585         struct hlist_head *bucket;
 586         struct kvm_mmu_page *page;
 587         struct hlist_node *node;
 588
 589         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 590         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 591         bucket = &vcpu->kvm->mmu_page_hash[index];
 592         hlist_for_each_entry(page, node, bucket, hash_link)
 593                 if (page->gfn == gfn && !page->role.metaphysical) {
 594                         pgprintk("%s: found role %x\n",
 595                                  __FUNCTION__, page->role.word);
 596                         return page;
 597                 }
 598         return NULL;
 599 }
 600
 601 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 602                                              gfn_t gfn,
 603                                              gva_t gaddr,
 604                                              unsigned level,
 605                                              int metaphysical,
 606                                              unsigned hugepage_access,
 607                                              u64 *parent_pte)
 608 {
 609         union kvm_mmu_page_role role;
 610         unsigned index;
 611         unsigned quadrant;
 612         struct hlist_head *bucket;
 613         struct kvm_mmu_page *page;
 614         struct hlist_node *node;
 615
 616         role.word = 0;
 617         role.glevels = vcpu->mmu.root_level;
 618         role.level = level;
 619         role.metaphysical = metaphysical;
 620         role.hugepage_access = hugepage_access;
 621         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 622                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 623                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 624                 role.quadrant = quadrant;
 625         }
 626         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 627                  gfn, role.word);
 628         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 629         bucket = &vcpu->kvm->mmu_page_hash[index];
 630         hlist_for_each_entry(page, node, bucket, hash_link)
 631                 if (page->gfn == gfn && page->role.word == role.word) {
 632                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 633                         pgprintk("%s: found\n", __FUNCTION__);
 634                         return page;
 635                 }
 636         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 637         if (!page)
 638                 return page;
 639         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 640         page->gfn = gfn;
 641         page->role = role;
 642         hlist_add_head(&page->hash_link, bucket);
 643         if (!metaphysical)
 644                 rmap_write_protect(vcpu, gfn);
 645         return page;
 646 }
 647
 648 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 649                                          struct kvm_mmu_page *page)
 650 {
 651         unsigned i;
 652         u64 *pt;
 653         u64 ent;
 654
 655         pt = page->spt;
 656
 657         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 658                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 659                         if (pt[i] & PT_PRESENT_MASK)
 660                                 rmap_remove(vcpu, &pt[i]);
 661                         pt[i] = 0;
 662                 }
 663                 kvm_arch_ops->tlb_flush(vcpu);
 664                 return;
 665         }
 666
 667         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 668                 ent = pt[i];
 669
 670                 pt[i] = 0;
 671                 if (!(ent & PT_PRESENT_MASK))
 672                         continue;
 673                 ent &= PT64_BASE_ADDR_MASK;
 674                 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
 675         }
 676 }
 677
 678 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 679                              struct kvm_mmu_page *page,
 680                              u64 *parent_pte)
 681 {
 682         mmu_page_remove_parent_pte(vcpu, page, parent_pte);
 683 }
 684
 685 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 686                              struct kvm_mmu_page *page)
 687 {
 688         u64 *parent_pte;
 689
 690         while (page->multimapped || page->parent_pte) {
 691                 if (!page->multimapped)
 692                         parent_pte = page->parent_pte;
 693                 else {
 694                         struct kvm_pte_chain *chain;
 695
 696                         chain = container_of(page->parent_ptes.first,
 697                                              struct kvm_pte_chain, link);
 698                         parent_pte = chain->parent_ptes[0];
 699                 }
 700                 BUG_ON(!parent_pte);
 701                 kvm_mmu_put_page(vcpu, page, parent_pte);
 702                 *parent_pte = 0;
 703         }
 704         kvm_mmu_page_unlink_children(vcpu, page);
 705         if (!page->root_count) {
 706                 hlist_del(&page->hash_link);
 707                 kvm_mmu_free_page(vcpu, page);
 708         } else
 709                 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 710 }
 711
 712 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 713 {
 714         unsigned index;
 715         struct hlist_head *bucket;
 716         struct kvm_mmu_page *page;
 717         struct hlist_node *node, *n;
 718         int r;
 719
 720         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 721         r = 0;
 722         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 723         bucket = &vcpu->kvm->mmu_page_hash[index];
 724         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 725                 if (page->gfn == gfn && !page->role.metaphysical) {
 726                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 727                                  page->role.word);
 728                         kvm_mmu_zap_page(vcpu, page);
 729                         r = 1;
 730                 }
 731         return r;
 732 }
 733
 734 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 735 {
 736         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 737         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 738
 739         __set_bit(slot, &page_head->slot_bitmap);
 740 }
 741
 742 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 743 {
 744         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 745
 746         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 747 }
 748
 749 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 750 {
 751         struct page *page;
 752
 753         ASSERT((gpa & HPA_ERR_MASK) == 0);
 754         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 755         if (!page)
 756                 return gpa | HPA_ERR_MASK;
 757         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 758                 | (gpa & (PAGE_SIZE-1));
 759 }
 760
 761 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 762 {
 763         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 764
 765         if (gpa == UNMAPPED_GVA)
 766                 return UNMAPPED_GVA;
 767         return gpa_to_hpa(vcpu, gpa);
 768 }
 769
 770 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 771 {
 772         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 773
 774         if (gpa == UNMAPPED_GVA)
 775                 return NULL;
 776         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 777 }
 778
 779 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 780 {
 781 }
 782
 783 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 784 {
 785         int level = PT32E_ROOT_LEVEL;
 786         hpa_t table_addr = vcpu->mmu.root_hpa;
 787
 788         for (; ; level--) {
 789                 u32 index = PT64_INDEX(v, level);
 790                 u64 *table;
 791                 u64 pte;
 792
 793                 ASSERT(VALID_PAGE(table_addr));
 794                 table = __va(table_addr);
 795
 796                 if (level == 1) {
 797                         pte = table[index];
 798                         if (is_present_pte(pte) && is_writeble_pte(pte))
 799                                 return 0;
 800                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 801                         page_header_update_slot(vcpu->kvm, table, v);
 802                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 803                                                                 PT_USER_MASK;
 804                         rmap_add(vcpu, &table[index]);
 805                         return 0;
 806                 }
 807
 808                 if (table[index] == 0) {
 809                         struct kvm_mmu_page *new_table;
 810                         gfn_t pseudo_gfn;
 811
 812                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 813                                 >> PAGE_SHIFT;
 814                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 815                                                      v, level - 1,
 816                                                      1, 0, &table[index]);
 817                         if (!new_table) {
 818                                 pgprintk("nonpaging_map: ENOMEM\n");
 819                                 return -ENOMEM;
 820                         }
 821
 822                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 823                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 824                 }
 825                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 826         }
 827 }
 828
 829 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 830 {
 831         int i;
 832         struct kvm_mmu_page *page;
 833
 834 #ifdef CONFIG_X86_64
 835         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 836                 hpa_t root = vcpu->mmu.root_hpa;
 837
 838                 ASSERT(VALID_PAGE(root));
 839                 page = page_header(root);
 840                 --page->root_count;
 841                 vcpu->mmu.root_hpa = INVALID_PAGE;
 842                 return;
 843         }
 844 #endif
 845         for (i = 0; i < 4; ++i) {
 846                 hpa_t root = vcpu->mmu.pae_root[i];
 847
 848                 if (root) {
 849                         ASSERT(VALID_PAGE(root));
 850                         root &= PT64_BASE_ADDR_MASK;
 851                         page = page_header(root);
 852                         --page->root_count;
 853                 }
 854                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 855         }
 856         vcpu->mmu.root_hpa = INVALID_PAGE;
 857 }
 858
 859 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 860 {
 861         int i;
 862         gfn_t root_gfn;
 863         struct kvm_mmu_page *page;
 864
 865         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 866
 867 #ifdef CONFIG_X86_64
 868         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 869                 hpa_t root = vcpu->mmu.root_hpa;
 870
 871                 ASSERT(!VALID_PAGE(root));
 872                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 873                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 874                 root = __pa(page->spt);
 875                 ++page->root_count;
 876                 vcpu->mmu.root_hpa = root;
 877                 return;
 878         }
 879 #endif
 880         for (i = 0; i < 4; ++i) {
 881                 hpa_t root = vcpu->mmu.pae_root[i];
 882
 883                 ASSERT(!VALID_PAGE(root));
 884                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 885                         if (!is_present_pte(vcpu->pdptrs[i])) {
 886                                 vcpu->mmu.pae_root[i] = 0;
 887                                 continue;
 888                         }
 889                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 890                 } else if (vcpu->mmu.root_level == 0)
 891                         root_gfn = 0;
 892                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 893                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 894                                         0, NULL);
 895                 root = __pa(page->spt);
 896                 ++page->root_count;
 897                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 898         }
 899         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 900 }
 901
 902 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 903 {
 904         return vaddr;
 905 }
 906
 907 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 908                                u32 error_code)
 909 {
 910         gpa_t addr = gva;
 911         hpa_t paddr;
 912         int r;
 913
 914         r = mmu_topup_memory_caches(vcpu);
 915         if (r)
 916                 return r;
 917
 918         ASSERT(vcpu);
 919         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 920
 921
 922         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 923
 924         if (is_error_hpa(paddr))
 925                 return 1;
 926
 927         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 928 }
 929
 930 static void nonpaging_free(struct kvm_vcpu *vcpu)
 931 {
 932         mmu_free_roots(vcpu);
 933 }
 934
 935 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 936 {
 937         struct kvm_mmu *context = &vcpu->mmu;
 938
 939         context->new_cr3 = nonpaging_new_cr3;
 940         context->page_fault = nonpaging_page_fault;
 941         context->gva_to_gpa = nonpaging_gva_to_gpa;
 942         context->free = nonpaging_free;
 943         context->root_level = 0;
 944         context->shadow_root_level = PT32E_ROOT_LEVEL;
 945         mmu_alloc_roots(vcpu);
 946         ASSERT(VALID_PAGE(context->root_hpa));
 947         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 948         return 0;
 949 }
 950
 951 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 952 {
 953         ++vcpu->stat.tlb_flush;
 954         kvm_arch_ops->tlb_flush(vcpu);
 955 }
 956
 957 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 958 {
 959         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 960         mmu_free_roots(vcpu);
 961         if (unlikely(vcpu->kvm->n_free_mmu_pages < KVM_MIN_FREE_MMU_PAGES))
 962                 kvm_mmu_free_some_pages(vcpu);
 963         mmu_alloc_roots(vcpu);
 964         kvm_mmu_flush_tlb(vcpu);
 965         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 966 }
 967
 968 static inline void set_pte_common(struct kvm_vcpu *vcpu,
 969                              u64 *shadow_pte,
 970                              gpa_t gaddr,
 971                              int dirty,
 972                              u64 access_bits,
 973                              gfn_t gfn)
 974 {
 975         hpa_t paddr;
 976
 977         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 978         if (!dirty)
 979                 access_bits &= ~PT_WRITABLE_MASK;
 980
 981         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 982
 983         *shadow_pte |= access_bits;
 984
 985         if (is_error_hpa(paddr)) {
 986                 *shadow_pte |= gaddr;
 987                 *shadow_pte |= PT_SHADOW_IO_MARK;
 988                 *shadow_pte &= ~PT_PRESENT_MASK;
 989                 return;
 990         }
 991
 992         *shadow_pte |= paddr;
 993
 994         if (access_bits & PT_WRITABLE_MASK) {
 995                 struct kvm_mmu_page *shadow;
 996
 997                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 998                 if (shadow) {
 999                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1000                                  __FUNCTION__, gfn);
1001                         access_bits &= ~PT_WRITABLE_MASK;
1002                         if (is_writeble_pte(*shadow_pte)) {
1003                                     *shadow_pte &= ~PT_WRITABLE_MASK;
1004                                     kvm_arch_ops->tlb_flush(vcpu);
1005                         }
1006                 }
1007         }
1008
1009         if (access_bits & PT_WRITABLE_MASK)
1010                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
1011
1012         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
1013         rmap_add(vcpu, shadow_pte);
1014 }
1015
1016 static void inject_page_fault(struct kvm_vcpu *vcpu,
1017                               u64 addr,
1018                               u32 err_code)
1019 {
1020         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
1021 }
1022
1023 static inline int fix_read_pf(u64 *shadow_ent)
1024 {
1025         if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
1026             !(*shadow_ent & PT_USER_MASK)) {
1027                 /*
1028                  * If supervisor write protect is disabled, we shadow kernel
1029                  * pages as user pages so we can trap the write access.
1030                  */
1031                 *shadow_ent |= PT_USER_MASK;
1032                 *shadow_ent &= ~PT_WRITABLE_MASK;
1033
1034                 return 1;
1035
1036         }
1037         return 0;
1038 }
1039
1040 static void paging_free(struct kvm_vcpu *vcpu)
1041 {
1042         nonpaging_free(vcpu);
1043 }
1044
1045 #define PTTYPE 64
1046 #include "paging_tmpl.h"
1047 #undef PTTYPE
1048
1049 #define PTTYPE 32
1050 #include "paging_tmpl.h"
1051 #undef PTTYPE
1052
1053 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1054 {
1055         struct kvm_mmu *context = &vcpu->mmu;
1056
1057         ASSERT(is_pae(vcpu));
1058         context->new_cr3 = paging_new_cr3;
1059         context->page_fault = paging64_page_fault;
1060         context->gva_to_gpa = paging64_gva_to_gpa;
1061         context->free = paging_free;
1062         context->root_level = level;
1063         context->shadow_root_level = level;
1064         mmu_alloc_roots(vcpu);
1065         ASSERT(VALID_PAGE(context->root_hpa));
1066         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1067                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1068         return 0;
1069 }
1070
1071 static int paging64_init_context(struct kvm_vcpu *vcpu)
1072 {
1073         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1074 }
1075
1076 static int paging32_init_context(struct kvm_vcpu *vcpu)
1077 {
1078         struct kvm_mmu *context = &vcpu->mmu;
1079
1080         context->new_cr3 = paging_new_cr3;
1081         context->page_fault = paging32_page_fault;
1082         context->gva_to_gpa = paging32_gva_to_gpa;
1083         context->free = paging_free;
1084         context->root_level = PT32_ROOT_LEVEL;
1085         context->shadow_root_level = PT32E_ROOT_LEVEL;
1086         mmu_alloc_roots(vcpu);
1087         ASSERT(VALID_PAGE(context->root_hpa));
1088         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1089                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1090         return 0;
1091 }
1092
1093 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1094 {
1095         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1096 }
1097
1098 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1099 {
1100         ASSERT(vcpu);
1101         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1102
1103         mmu_topup_memory_caches(vcpu);
1104         if (!is_paging(vcpu))
1105                 return nonpaging_init_context(vcpu);
1106         else if (is_long_mode(vcpu))
1107                 return paging64_init_context(vcpu);
1108         else if (is_pae(vcpu))
1109                 return paging32E_init_context(vcpu);
1110         else
1111                 return paging32_init_context(vcpu);
1112 }
1113
1114 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1115 {
1116         ASSERT(vcpu);
1117         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1118                 vcpu->mmu.free(vcpu);
1119                 vcpu->mmu.root_hpa = INVALID_PAGE;
1120         }
1121 }
1122
1123 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1124 {
1125         int r;
1126
1127         destroy_kvm_mmu(vcpu);
1128         r = init_kvm_mmu(vcpu);
1129         if (r < 0)
1130                 goto out;
1131         r = mmu_topup_memory_caches(vcpu);
1132 out:
1133         return r;
1134 }
1135
1136 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1137                                   struct kvm_mmu_page *page,
1138                                   u64 *spte)
1139 {
1140         u64 pte;
1141         struct kvm_mmu_page *child;
1142
1143         pte = *spte;
1144         if (is_present_pte(pte)) {
1145                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1146                         rmap_remove(vcpu, spte);
1147                 else {
1148                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1149                         mmu_page_remove_parent_pte(vcpu, child, spte);
1150                 }
1151         }
1152         *spte = 0;
1153 }
1154
1155 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1156                                   struct kvm_mmu_page *page,
1157                                   u64 *spte,
1158                                   const void *new, int bytes)
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         else
1166                 paging64_update_pte(vcpu, page, spte, new, bytes);
1167 }
1168
1169 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1170                        const u8 *old, const u8 *new, int bytes)
1171 {
1172         gfn_t gfn = gpa >> PAGE_SHIFT;
1173         struct kvm_mmu_page *page;
1174         struct hlist_node *node, *n;
1175         struct hlist_head *bucket;
1176         unsigned index;
1177         u64 *spte;
1178         unsigned offset = offset_in_page(gpa);
1179         unsigned pte_size;
1180         unsigned page_offset;
1181         unsigned misaligned;
1182         unsigned quadrant;
1183         int level;
1184         int flooded = 0;
1185         int npte;
1186
1187         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1188         if (gfn == vcpu->last_pt_write_gfn) {
1189                 ++vcpu->last_pt_write_count;
1190                 if (vcpu->last_pt_write_count >= 3)
1191                         flooded = 1;
1192         } else {
1193                 vcpu->last_pt_write_gfn = gfn;
1194                 vcpu->last_pt_write_count = 1;
1195         }
1196         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1197         bucket = &vcpu->kvm->mmu_page_hash[index];
1198         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1199                 if (page->gfn != gfn || page->role.metaphysical)
1200                         continue;
1201                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1202                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1203                 misaligned |= bytes < 4;
1204                 if (misaligned || flooded) {
1205                         /*
1206                          * Misaligned accesses are too much trouble to fix
1207                          * up; also, they usually indicate a page is not used
1208                          * as a page table.
1209                          *
1210                          * If we're seeing too many writes to a page,
1211                          * it may no longer be a page table, or we may be
1212                          * forking, in which case it is better to unmap the
1213                          * page.
1214                          */
1215                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1216                                  gpa, bytes, page->role.word);
1217                         kvm_mmu_zap_page(vcpu, page);
1218                         continue;
1219                 }
1220                 page_offset = offset;
1221                 level = page->role.level;
1222                 npte = 1;
1223                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1224                         page_offset <<= 1;      /* 32->64 */
1225                         /*
1226                          * A 32-bit pde maps 4MB while the shadow pdes map
1227                          * only 2MB.  So we need to double the offset again
1228                          * and zap two pdes instead of one.
1229                          */
1230                         if (level == PT32_ROOT_LEVEL) {
1231                                 page_offset &= ~7; /* kill rounding error */
1232                                 page_offset <<= 1;
1233                                 npte = 2;
1234                         }
1235                         quadrant = page_offset >> PAGE_SHIFT;
1236                         page_offset &= ~PAGE_MASK;
1237                         if (quadrant != page->role.quadrant)
1238                                 continue;
1239                 }
1240                 spte = &page->spt[page_offset / sizeof(*spte)];
1241                 while (npte--) {
1242                         mmu_pte_write_zap_pte(vcpu, page, spte);
1243                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes);
1244                         ++spte;
1245                 }
1246         }
1247 }
1248
1249 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1250 {
1251         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1252
1253         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1254 }
1255
1256 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1257 {
1258         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1259                 struct kvm_mmu_page *page;
1260
1261                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1262                                     struct kvm_mmu_page, link);
1263                 kvm_mmu_zap_page(vcpu, page);
1264         }
1265 }
1266 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1267
1268 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1269 {
1270         struct kvm_mmu_page *page;
1271
1272         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1273                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1274                                     struct kvm_mmu_page, link);
1275                 kvm_mmu_zap_page(vcpu, page);
1276         }
1277         free_page((unsigned long)vcpu->mmu.pae_root);
1278 }
1279
1280 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1281 {
1282         struct page *page;
1283         int i;
1284
1285         ASSERT(vcpu);
1286
1287         vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1288
1289         /*
1290          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1291          * Therefore we need to allocate shadow page tables in the first
1292          * 4GB of memory, which happens to fit the DMA32 zone.
1293          */
1294         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1295         if (!page)
1296                 goto error_1;
1297         vcpu->mmu.pae_root = page_address(page);
1298         for (i = 0; i < 4; ++i)
1299                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1300
1301         return 0;
1302
1303 error_1:
1304         free_mmu_pages(vcpu);
1305         return -ENOMEM;
1306 }
1307
1308 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1309 {
1310         ASSERT(vcpu);
1311         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1312
1313         return alloc_mmu_pages(vcpu);
1314 }
1315
1316 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1317 {
1318         ASSERT(vcpu);
1319         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1320
1321         return init_kvm_mmu(vcpu);
1322 }
1323
1324 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1325 {
1326         ASSERT(vcpu);
1327
1328         destroy_kvm_mmu(vcpu);
1329         free_mmu_pages(vcpu);
1330         mmu_free_memory_caches(vcpu);
1331 }
1332
1333 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1334 {
1335         struct kvm *kvm = vcpu->kvm;
1336         struct kvm_mmu_page *page;
1337
1338         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1339                 int i;
1340                 u64 *pt;
1341
1342                 if (!test_bit(slot, &page->slot_bitmap))
1343                         continue;
1344
1345                 pt = page->spt;
1346                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1347                         /* avoid RMW */
1348                         if (pt[i] & PT_WRITABLE_MASK) {
1349                                 rmap_remove(vcpu, &pt[i]);
1350                                 pt[i] &= ~PT_WRITABLE_MASK;
1351                         }
1352         }
1353 }
1354
1355 void kvm_mmu_zap_all(struct kvm_vcpu *vcpu)
1356 {
1357         destroy_kvm_mmu(vcpu);
1358
1359         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1360                 struct kvm_mmu_page *page;
1361
1362                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1363                                     struct kvm_mmu_page, link);
1364                 kvm_mmu_zap_page(vcpu, page);
1365         }
1366
1367         mmu_free_memory_caches(vcpu);
1368         kvm_arch_ops->tlb_flush(vcpu);
1369         init_kvm_mmu(vcpu);
1370 }
1371
1372 void kvm_mmu_module_exit(void)
1373 {
1374         if (pte_chain_cache)
1375                 kmem_cache_destroy(pte_chain_cache);
1376         if (rmap_desc_cache)
1377                 kmem_cache_destroy(rmap_desc_cache);
1378         if (mmu_page_cache)
1379                 kmem_cache_destroy(mmu_page_cache);
1380         if (mmu_page_header_cache)
1381                 kmem_cache_destroy(mmu_page_header_cache);
1382 }
1383
1384 int kvm_mmu_module_init(void)
1385 {
1386         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1387                                             sizeof(struct kvm_pte_chain),
1388                                             0, 0, NULL, NULL);
1389         if (!pte_chain_cache)
1390                 goto nomem;
1391         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1392                                             sizeof(struct kvm_rmap_desc),
1393                                             0, 0, NULL, NULL);
1394         if (!rmap_desc_cache)
1395                 goto nomem;
1396
1397         mmu_page_cache = kmem_cache_create("kvm_mmu_page",
1398                                            PAGE_SIZE,
1399                                            PAGE_SIZE, 0, NULL, NULL);
1400         if (!mmu_page_cache)
1401                 goto nomem;
1402
1403         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1404                                                   sizeof(struct kvm_mmu_page),
1405                                                   0, 0, NULL, NULL);
1406         if (!mmu_page_header_cache)
1407                 goto nomem;
1408
1409         return 0;
1410
1411 nomem:
1412         kvm_mmu_module_exit();
1413         return -ENOMEM;
1414 }
1415
1416 #ifdef AUDIT
1417
1418 static const char *audit_msg;
1419
1420 static gva_t canonicalize(gva_t gva)
1421 {
1422 #ifdef CONFIG_X86_64
1423         gva = (long long)(gva << 16) >> 16;
1424 #endif
1425         return gva;
1426 }
1427
1428 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1429                                 gva_t va, int level)
1430 {
1431         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1432         int i;
1433         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1434
1435         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1436                 u64 ent = pt[i];
1437
1438                 if (!(ent & PT_PRESENT_MASK))
1439                         continue;
1440
1441                 va = canonicalize(va);
1442                 if (level > 1)
1443                         audit_mappings_page(vcpu, ent, va, level - 1);
1444                 else {
1445                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1446                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1447
1448                         if ((ent & PT_PRESENT_MASK)
1449                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1450                                 printk(KERN_ERR "audit error: (%s) levels %d"
1451                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1452                                        audit_msg, vcpu->mmu.root_level,
1453                                        va, gpa, hpa, ent);
1454                 }
1455         }
1456 }
1457
1458 static void audit_mappings(struct kvm_vcpu *vcpu)
1459 {
1460         unsigned i;
1461
1462         if (vcpu->mmu.root_level == 4)
1463                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1464         else
1465                 for (i = 0; i < 4; ++i)
1466                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1467                                 audit_mappings_page(vcpu,
1468                                                     vcpu->mmu.pae_root[i],
1469                                                     i << 30,
1470                                                     2);
1471 }
1472
1473 static int count_rmaps(struct kvm_vcpu *vcpu)
1474 {
1475         int nmaps = 0;
1476         int i, j, k;
1477
1478         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1479                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1480                 struct kvm_rmap_desc *d;
1481
1482                 for (j = 0; j < m->npages; ++j) {
1483                         struct page *page = m->phys_mem[j];
1484
1485                         if (!page->private)
1486                                 continue;
1487                         if (!(page->private & 1)) {
1488                                 ++nmaps;
1489                                 continue;
1490                         }
1491                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1492                         while (d) {
1493                                 for (k = 0; k < RMAP_EXT; ++k)
1494                                         if (d->shadow_ptes[k])
1495                                                 ++nmaps;
1496                                         else
1497                                                 break;
1498                                 d = d->more;
1499                         }
1500                 }
1501         }
1502         return nmaps;
1503 }
1504
1505 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1506 {
1507         int nmaps = 0;
1508         struct kvm_mmu_page *page;
1509         int i;
1510
1511         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1512                 u64 *pt = page->spt;
1513
1514                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1515                         continue;
1516
1517                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1518                         u64 ent = pt[i];
1519
1520                         if (!(ent & PT_PRESENT_MASK))
1521                                 continue;
1522                         if (!(ent & PT_WRITABLE_MASK))
1523                                 continue;
1524                         ++nmaps;
1525                 }
1526         }
1527         return nmaps;
1528 }
1529
1530 static void audit_rmap(struct kvm_vcpu *vcpu)
1531 {
1532         int n_rmap = count_rmaps(vcpu);
1533         int n_actual = count_writable_mappings(vcpu);
1534
1535         if (n_rmap != n_actual)
1536                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1537                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1538 }
1539
1540 static void audit_write_protection(struct kvm_vcpu *vcpu)
1541 {
1542         struct kvm_mmu_page *page;
1543
1544         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1545                 hfn_t hfn;
1546                 struct page *pg;
1547
1548                 if (page->role.metaphysical)
1549                         continue;
1550
1551                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1552                         >> PAGE_SHIFT;
1553                 pg = pfn_to_page(hfn);
1554                 if (pg->private)
1555                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1556                                " mappings: gfn %lx role %x\n",
1557                                __FUNCTION__, audit_msg, page->gfn,
1558                                page->role.word);
1559         }
1560 }
1561
1562 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1563 {
1564         int olddbg = dbg;
1565
1566         dbg = 0;
1567         audit_msg = msg;
1568         audit_rmap(vcpu);
1569         audit_write_protection(vcpu);
1570         audit_mappings(vcpu);
1571         dbg = olddbg;
1572 }
1573
1574 #endif