drivers/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "vmx.h"
  21 #include "kvm.h"
  22
  23 #include <linux/types.h>
  24 #include <linux/string.h>
  25 #include <linux/mm.h>
  26 #include <linux/highmem.h>
  27 #include <linux/module.h>
  28
  29 #include <asm/page.h>
  30 #include <asm/cmpxchg.h>
  31
  32 #undef MMU_DEBUG
  33
  34 #undef AUDIT
  35
  36 #ifdef AUDIT
  37 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  38 #else
  39 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  40 #endif
  41
  42 #ifdef MMU_DEBUG
  43
  44 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  45 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  46
  47 #else
  48
  49 #define pgprintk(x...) do { } while (0)
  50 #define rmap_printk(x...) do { } while (0)
  51
  52 #endif
  53
  54 #if defined(MMU_DEBUG) || defined(AUDIT)
  55 static int dbg = 1;
  56 #endif
  57
  58 #ifndef MMU_DEBUG
  59 #define ASSERT(x) do { } while (0)
  60 #else
  61 #define ASSERT(x)                                                       \
  62         if (!(x)) {                                                     \
  63                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  64                        __FILE__, __LINE__, #x);                         \
  65         }
  66 #endif
  67
  68 #define PT64_PT_BITS 9
  69 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  70 #define PT32_PT_BITS 10
  71 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  72
  73 #define PT_WRITABLE_SHIFT 1
  74
  75 #define PT_PRESENT_MASK (1ULL << 0)
  76 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  77 #define PT_USER_MASK (1ULL << 2)
  78 #define PT_PWT_MASK (1ULL << 3)
  79 #define PT_PCD_MASK (1ULL << 4)
  80 #define PT_ACCESSED_MASK (1ULL << 5)
  81 #define PT_DIRTY_MASK (1ULL << 6)
  82 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  83 #define PT_PAT_MASK (1ULL << 7)
  84 #define PT_GLOBAL_MASK (1ULL << 8)
  85 #define PT64_NX_MASK (1ULL << 63)
  86
  87 #define PT_PAT_SHIFT 7
  88 #define PT_DIR_PAT_SHIFT 12
  89 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  90
  91 #define PT32_DIR_PSE36_SIZE 4
  92 #define PT32_DIR_PSE36_SHIFT 13
  93 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  94
  95
  96 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  97 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  98
  99 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 100
 101 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 102
 103 #define PT64_LEVEL_BITS 9
 104
 105 #define PT64_LEVEL_SHIFT(level) \
 106                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
 107
 108 #define PT64_LEVEL_MASK(level) \
 109                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 110
 111 #define PT64_INDEX(address, level)\
 112         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 113
 114
 115 #define PT32_LEVEL_BITS 10
 116
 117 #define PT32_LEVEL_SHIFT(level) \
 118                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
 119
 120 #define PT32_LEVEL_MASK(level) \
 121                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 122
 123 #define PT32_INDEX(address, level)\
 124         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 125
 126
 127 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 128 #define PT64_DIR_BASE_ADDR_MASK \
 129         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 130
 131 #define PT32_BASE_ADDR_MASK PAGE_MASK
 132 #define PT32_DIR_BASE_ADDR_MASK \
 133         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 134
 135
 136 #define PFERR_PRESENT_MASK (1U << 0)
 137 #define PFERR_WRITE_MASK (1U << 1)
 138 #define PFERR_USER_MASK (1U << 2)
 139 #define PFERR_FETCH_MASK (1U << 4)
 140
 141 #define PT64_ROOT_LEVEL 4
 142 #define PT32_ROOT_LEVEL 2
 143 #define PT32E_ROOT_LEVEL 3
 144
 145 #define PT_DIRECTORY_LEVEL 2
 146 #define PT_PAGE_TABLE_LEVEL 1
 147
 148 #define RMAP_EXT 4
 149
 150 struct kvm_rmap_desc {
 151         u64 *shadow_ptes[RMAP_EXT];
 152         struct kvm_rmap_desc *more;
 153 };
 154
 155 static struct kmem_cache *pte_chain_cache;
 156 static struct kmem_cache *rmap_desc_cache;
 157 static struct kmem_cache *mmu_page_header_cache;
 158
 159 static int is_write_protection(struct kvm_vcpu *vcpu)
 160 {
 161         return vcpu->cr0 & CR0_WP_MASK;
 162 }
 163
 164 static int is_cpuid_PSE36(void)
 165 {
 166         return 1;
 167 }
 168
 169 static int is_nx(struct kvm_vcpu *vcpu)
 170 {
 171         return vcpu->shadow_efer & EFER_NX;
 172 }
 173
 174 static int is_present_pte(unsigned long pte)
 175 {
 176         return pte & PT_PRESENT_MASK;
 177 }
 178
 179 static int is_writeble_pte(unsigned long pte)
 180 {
 181         return pte & PT_WRITABLE_MASK;
 182 }
 183
 184 static int is_io_pte(unsigned long pte)
 185 {
 186         return pte & PT_SHADOW_IO_MARK;
 187 }
 188
 189 static int is_rmap_pte(u64 pte)
 190 {
 191         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 192                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 193 }
 194
 195 static void set_shadow_pte(u64 *sptep, u64 spte)
 196 {
 197 #ifdef CONFIG_X86_64
 198         set_64bit((unsigned long *)sptep, spte);
 199 #else
 200         set_64bit((unsigned long long *)sptep, spte);
 201 #endif
 202 }
 203
 204 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 205                                   struct kmem_cache *base_cache, int min,
 206                                   gfp_t gfp_flags)
 207 {
 208         void *obj;
 209
 210         if (cache->nobjs >= min)
 211                 return 0;
 212         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 213                 obj = kmem_cache_zalloc(base_cache, gfp_flags);
 214                 if (!obj)
 215                         return -ENOMEM;
 216                 cache->objects[cache->nobjs++] = obj;
 217         }
 218         return 0;
 219 }
 220
 221 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 222 {
 223         while (mc->nobjs)
 224                 kfree(mc->objects[--mc->nobjs]);
 225 }
 226
 227 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 228                                        int min, gfp_t gfp_flags)
 229 {
 230         struct page *page;
 231
 232         if (cache->nobjs >= min)
 233                 return 0;
 234         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 235                 page = alloc_page(gfp_flags);
 236                 if (!page)
 237                         return -ENOMEM;
 238                 set_page_private(page, 0);
 239                 cache->objects[cache->nobjs++] = page_address(page);
 240         }
 241         return 0;
 242 }
 243
 244 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 245 {
 246         while (mc->nobjs)
 247                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 248 }
 249
 250 static int __mmu_topup_memory_caches(struct kvm_vcpu *vcpu, gfp_t gfp_flags)
 251 {
 252         int r;
 253
 254         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 255                                    pte_chain_cache, 4, gfp_flags);
 256         if (r)
 257                 goto out;
 258         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 259                                    rmap_desc_cache, 1, gfp_flags);
 260         if (r)
 261                 goto out;
 262         r = mmu_topup_memory_cache_page(&vcpu->mmu_page_cache, 4, gfp_flags);
 263         if (r)
 264                 goto out;
 265         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
 266                                    mmu_page_header_cache, 4, gfp_flags);
 267 out:
 268         return r;
 269 }
 270
 271 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 272 {
 273         int r;
 274
 275         r = __mmu_topup_memory_caches(vcpu, GFP_NOWAIT);
 276         kvm_mmu_free_some_pages(vcpu);
 277         if (r < 0) {
 278                 spin_unlock(&vcpu->kvm->lock);
 279                 kvm_arch_ops->vcpu_put(vcpu);
 280                 r = __mmu_topup_memory_caches(vcpu, GFP_KERNEL);
 281                 kvm_arch_ops->vcpu_load(vcpu);
 282                 spin_lock(&vcpu->kvm->lock);
 283                 kvm_mmu_free_some_pages(vcpu);
 284         }
 285         return r;
 286 }
 287
 288 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 289 {
 290         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 291         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 292         mmu_free_memory_cache_page(&vcpu->mmu_page_cache);
 293         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
 294 }
 295
 296 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 297                                     size_t size)
 298 {
 299         void *p;
 300
 301         BUG_ON(!mc->nobjs);
 302         p = mc->objects[--mc->nobjs];
 303         memset(p, 0, size);
 304         return p;
 305 }
 306
 307 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 308 {
 309         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 310                                       sizeof(struct kvm_pte_chain));
 311 }
 312
 313 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 314 {
 315         kfree(pc);
 316 }
 317
 318 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 319 {
 320         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 321                                       sizeof(struct kvm_rmap_desc));
 322 }
 323
 324 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 325 {
 326         kfree(rd);
 327 }
 328
 329 /*
 330  * Reverse mapping data structures:
 331  *
 332  * If page->private bit zero is zero, then page->private points to the
 333  * shadow page table entry that points to page_address(page).
 334  *
 335  * If page->private bit zero is one, (then page->private & ~1) points
 336  * to a struct kvm_rmap_desc containing more mappings.
 337  */
 338 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 339 {
 340         struct page *page;
 341         struct kvm_rmap_desc *desc;
 342         int i;
 343
 344         if (!is_rmap_pte(*spte))
 345                 return;
 346         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 347         if (!page_private(page)) {
 348                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 349                 set_page_private(page,(unsigned long)spte);
 350         } else if (!(page_private(page) & 1)) {
 351                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 352                 desc = mmu_alloc_rmap_desc(vcpu);
 353                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 354                 desc->shadow_ptes[1] = spte;
 355                 set_page_private(page,(unsigned long)desc | 1);
 356         } else {
 357                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 358                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 359                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 360                         desc = desc->more;
 361                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 362                         desc->more = mmu_alloc_rmap_desc(vcpu);
 363                         desc = desc->more;
 364                 }
 365                 for (i = 0; desc->shadow_ptes[i]; ++i)
 366                         ;
 367                 desc->shadow_ptes[i] = spte;
 368         }
 369 }
 370
 371 static void rmap_desc_remove_entry(struct page *page,
 372                                    struct kvm_rmap_desc *desc,
 373                                    int i,
 374                                    struct kvm_rmap_desc *prev_desc)
 375 {
 376         int j;
 377
 378         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 379                 ;
 380         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 381         desc->shadow_ptes[j] = NULL;
 382         if (j != 0)
 383                 return;
 384         if (!prev_desc && !desc->more)
 385                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 386         else
 387                 if (prev_desc)
 388                         prev_desc->more = desc->more;
 389                 else
 390                         set_page_private(page,(unsigned long)desc->more | 1);
 391         mmu_free_rmap_desc(desc);
 392 }
 393
 394 static void rmap_remove(u64 *spte)
 395 {
 396         struct page *page;
 397         struct kvm_rmap_desc *desc;
 398         struct kvm_rmap_desc *prev_desc;
 399         int i;
 400
 401         if (!is_rmap_pte(*spte))
 402                 return;
 403         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 404         if (!page_private(page)) {
 405                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 406                 BUG();
 407         } else if (!(page_private(page) & 1)) {
 408                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 409                 if ((u64 *)page_private(page) != spte) {
 410                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 411                                spte, *spte);
 412                         BUG();
 413                 }
 414                 set_page_private(page,0);
 415         } else {
 416                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 417                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 418                 prev_desc = NULL;
 419                 while (desc) {
 420                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 421                                 if (desc->shadow_ptes[i] == spte) {
 422                                         rmap_desc_remove_entry(page,
 423                                                                desc, i,
 424                                                                prev_desc);
 425                                         return;
 426                                 }
 427                         prev_desc = desc;
 428                         desc = desc->more;
 429                 }
 430                 BUG();
 431         }
 432 }
 433
 434 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 435 {
 436         struct kvm *kvm = vcpu->kvm;
 437         struct page *page;
 438         struct kvm_rmap_desc *desc;
 439         u64 *spte;
 440
 441         page = gfn_to_page(kvm, gfn);
 442         BUG_ON(!page);
 443
 444         while (page_private(page)) {
 445                 if (!(page_private(page) & 1))
 446                         spte = (u64 *)page_private(page);
 447                 else {
 448                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 449                         spte = desc->shadow_ptes[0];
 450                 }
 451                 BUG_ON(!spte);
 452                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 453                        != page_to_pfn(page));
 454                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 455                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 456                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 457                 rmap_remove(spte);
 458                 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 459                 kvm_flush_remote_tlbs(vcpu->kvm);
 460         }
 461 }
 462
 463 #ifdef MMU_DEBUG
 464 static int is_empty_shadow_page(u64 *spt)
 465 {
 466         u64 *pos;
 467         u64 *end;
 468
 469         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 470                 if (*pos != 0) {
 471                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 472                                pos, *pos);
 473                         return 0;
 474                 }
 475         return 1;
 476 }
 477 #endif
 478
 479 static void kvm_mmu_free_page(struct kvm *kvm,
 480                               struct kvm_mmu_page *page_head)
 481 {
 482         ASSERT(is_empty_shadow_page(page_head->spt));
 483         list_del(&page_head->link);
 484         __free_page(virt_to_page(page_head->spt));
 485         kfree(page_head);
 486         ++kvm->n_free_mmu_pages;
 487 }
 488
 489 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 490 {
 491         return gfn;
 492 }
 493
 494 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 495                                                u64 *parent_pte)
 496 {
 497         struct kvm_mmu_page *page;
 498
 499         if (!vcpu->kvm->n_free_mmu_pages)
 500                 return NULL;
 501
 502         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 503                                       sizeof *page);
 504         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 505         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 506         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 507         ASSERT(is_empty_shadow_page(page->spt));
 508         page->slot_bitmap = 0;
 509         page->multimapped = 0;
 510         page->parent_pte = parent_pte;
 511         --vcpu->kvm->n_free_mmu_pages;
 512         return page;
 513 }
 514
 515 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 516                                     struct kvm_mmu_page *page, u64 *parent_pte)
 517 {
 518         struct kvm_pte_chain *pte_chain;
 519         struct hlist_node *node;
 520         int i;
 521
 522         if (!parent_pte)
 523                 return;
 524         if (!page->multimapped) {
 525                 u64 *old = page->parent_pte;
 526
 527                 if (!old) {
 528                         page->parent_pte = parent_pte;
 529                         return;
 530                 }
 531                 page->multimapped = 1;
 532                 pte_chain = mmu_alloc_pte_chain(vcpu);
 533                 INIT_HLIST_HEAD(&page->parent_ptes);
 534                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 535                 pte_chain->parent_ptes[0] = old;
 536         }
 537         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 538                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 539                         continue;
 540                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 541                         if (!pte_chain->parent_ptes[i]) {
 542                                 pte_chain->parent_ptes[i] = parent_pte;
 543                                 return;
 544                         }
 545         }
 546         pte_chain = mmu_alloc_pte_chain(vcpu);
 547         BUG_ON(!pte_chain);
 548         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 549         pte_chain->parent_ptes[0] = parent_pte;
 550 }
 551
 552 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
 553                                        u64 *parent_pte)
 554 {
 555         struct kvm_pte_chain *pte_chain;
 556         struct hlist_node *node;
 557         int i;
 558
 559         if (!page->multimapped) {
 560                 BUG_ON(page->parent_pte != parent_pte);
 561                 page->parent_pte = NULL;
 562                 return;
 563         }
 564         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 565                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 566                         if (!pte_chain->parent_ptes[i])
 567                                 break;
 568                         if (pte_chain->parent_ptes[i] != parent_pte)
 569                                 continue;
 570                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 571                                 && pte_chain->parent_ptes[i + 1]) {
 572                                 pte_chain->parent_ptes[i]
 573                                         = pte_chain->parent_ptes[i + 1];
 574                                 ++i;
 575                         }
 576                         pte_chain->parent_ptes[i] = NULL;
 577                         if (i == 0) {
 578                                 hlist_del(&pte_chain->link);
 579                                 mmu_free_pte_chain(pte_chain);
 580                                 if (hlist_empty(&page->parent_ptes)) {
 581                                         page->multimapped = 0;
 582                                         page->parent_pte = NULL;
 583                                 }
 584                         }
 585                         return;
 586                 }
 587         BUG();
 588 }
 589
 590 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 591                                                 gfn_t gfn)
 592 {
 593         unsigned index;
 594         struct hlist_head *bucket;
 595         struct kvm_mmu_page *page;
 596         struct hlist_node *node;
 597
 598         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 599         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 600         bucket = &vcpu->kvm->mmu_page_hash[index];
 601         hlist_for_each_entry(page, node, bucket, hash_link)
 602                 if (page->gfn == gfn && !page->role.metaphysical) {
 603                         pgprintk("%s: found role %x\n",
 604                                  __FUNCTION__, page->role.word);
 605                         return page;
 606                 }
 607         return NULL;
 608 }
 609
 610 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 611                                              gfn_t gfn,
 612                                              gva_t gaddr,
 613                                              unsigned level,
 614                                              int metaphysical,
 615                                              unsigned hugepage_access,
 616                                              u64 *parent_pte)
 617 {
 618         union kvm_mmu_page_role role;
 619         unsigned index;
 620         unsigned quadrant;
 621         struct hlist_head *bucket;
 622         struct kvm_mmu_page *page;
 623         struct hlist_node *node;
 624
 625         role.word = 0;
 626         role.glevels = vcpu->mmu.root_level;
 627         role.level = level;
 628         role.metaphysical = metaphysical;
 629         role.hugepage_access = hugepage_access;
 630         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 631                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 632                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 633                 role.quadrant = quadrant;
 634         }
 635         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 636                  gfn, role.word);
 637         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 638         bucket = &vcpu->kvm->mmu_page_hash[index];
 639         hlist_for_each_entry(page, node, bucket, hash_link)
 640                 if (page->gfn == gfn && page->role.word == role.word) {
 641                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 642                         pgprintk("%s: found\n", __FUNCTION__);
 643                         return page;
 644                 }
 645         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 646         if (!page)
 647                 return page;
 648         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 649         page->gfn = gfn;
 650         page->role = role;
 651         hlist_add_head(&page->hash_link, bucket);
 652         if (!metaphysical)
 653                 rmap_write_protect(vcpu, gfn);
 654         return page;
 655 }
 656
 657 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 658                                          struct kvm_mmu_page *page)
 659 {
 660         unsigned i;
 661         u64 *pt;
 662         u64 ent;
 663
 664         pt = page->spt;
 665
 666         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 667                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 668                         if (pt[i] & PT_PRESENT_MASK)
 669                                 rmap_remove(&pt[i]);
 670                         pt[i] = 0;
 671                 }
 672                 kvm_flush_remote_tlbs(kvm);
 673                 return;
 674         }
 675
 676         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 677                 ent = pt[i];
 678
 679                 pt[i] = 0;
 680                 if (!(ent & PT_PRESENT_MASK))
 681                         continue;
 682                 ent &= PT64_BASE_ADDR_MASK;
 683                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 684         }
 685         kvm_flush_remote_tlbs(kvm);
 686 }
 687
 688 static void kvm_mmu_put_page(struct kvm_mmu_page *page,
 689                              u64 *parent_pte)
 690 {
 691         mmu_page_remove_parent_pte(page, parent_pte);
 692 }
 693
 694 static void kvm_mmu_zap_page(struct kvm *kvm,
 695                              struct kvm_mmu_page *page)
 696 {
 697         u64 *parent_pte;
 698
 699         while (page->multimapped || page->parent_pte) {
 700                 if (!page->multimapped)
 701                         parent_pte = page->parent_pte;
 702                 else {
 703                         struct kvm_pte_chain *chain;
 704
 705                         chain = container_of(page->parent_ptes.first,
 706                                              struct kvm_pte_chain, link);
 707                         parent_pte = chain->parent_ptes[0];
 708                 }
 709                 BUG_ON(!parent_pte);
 710                 kvm_mmu_put_page(page, parent_pte);
 711                 set_shadow_pte(parent_pte, 0);
 712         }
 713         kvm_mmu_page_unlink_children(kvm, page);
 714         if (!page->root_count) {
 715                 hlist_del(&page->hash_link);
 716                 kvm_mmu_free_page(kvm, page);
 717         } else
 718                 list_move(&page->link, &kvm->active_mmu_pages);
 719 }
 720
 721 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 722 {
 723         unsigned index;
 724         struct hlist_head *bucket;
 725         struct kvm_mmu_page *page;
 726         struct hlist_node *node, *n;
 727         int r;
 728
 729         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 730         r = 0;
 731         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 732         bucket = &vcpu->kvm->mmu_page_hash[index];
 733         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 734                 if (page->gfn == gfn && !page->role.metaphysical) {
 735                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 736                                  page->role.word);
 737                         kvm_mmu_zap_page(vcpu->kvm, page);
 738                         r = 1;
 739                 }
 740         return r;
 741 }
 742
 743 static void mmu_unshadow(struct kvm_vcpu *vcpu, gfn_t gfn)
 744 {
 745         struct kvm_mmu_page *page;
 746
 747         while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
 748                 pgprintk("%s: zap %lx %x\n",
 749                          __FUNCTION__, gfn, page->role.word);
 750                 kvm_mmu_zap_page(vcpu->kvm, page);
 751         }
 752 }
 753
 754 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 755 {
 756         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 757         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 758
 759         __set_bit(slot, &page_head->slot_bitmap);
 760 }
 761
 762 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 763 {
 764         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 765
 766         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 767 }
 768
 769 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 770 {
 771         struct page *page;
 772
 773         ASSERT((gpa & HPA_ERR_MASK) == 0);
 774         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 775         if (!page)
 776                 return gpa | HPA_ERR_MASK;
 777         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 778                 | (gpa & (PAGE_SIZE-1));
 779 }
 780
 781 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 782 {
 783         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 784
 785         if (gpa == UNMAPPED_GVA)
 786                 return UNMAPPED_GVA;
 787         return gpa_to_hpa(vcpu, gpa);
 788 }
 789
 790 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 791 {
 792         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 793
 794         if (gpa == UNMAPPED_GVA)
 795                 return NULL;
 796         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 797 }
 798
 799 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 800 {
 801 }
 802
 803 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 804 {
 805         int level = PT32E_ROOT_LEVEL;
 806         hpa_t table_addr = vcpu->mmu.root_hpa;
 807
 808         for (; ; level--) {
 809                 u32 index = PT64_INDEX(v, level);
 810                 u64 *table;
 811                 u64 pte;
 812
 813                 ASSERT(VALID_PAGE(table_addr));
 814                 table = __va(table_addr);
 815
 816                 if (level == 1) {
 817                         pte = table[index];
 818                         if (is_present_pte(pte) && is_writeble_pte(pte))
 819                                 return 0;
 820                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 821                         page_header_update_slot(vcpu->kvm, table, v);
 822                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 823                                                                 PT_USER_MASK;
 824                         rmap_add(vcpu, &table[index]);
 825                         return 0;
 826                 }
 827
 828                 if (table[index] == 0) {
 829                         struct kvm_mmu_page *new_table;
 830                         gfn_t pseudo_gfn;
 831
 832                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 833                                 >> PAGE_SHIFT;
 834                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 835                                                      v, level - 1,
 836                                                      1, 0, &table[index]);
 837                         if (!new_table) {
 838                                 pgprintk("nonpaging_map: ENOMEM\n");
 839                                 return -ENOMEM;
 840                         }
 841
 842                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 843                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 844                 }
 845                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 846         }
 847 }
 848
 849 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 850 {
 851         int i;
 852         struct kvm_mmu_page *page;
 853
 854         if (!VALID_PAGE(vcpu->mmu.root_hpa))
 855                 return;
 856 #ifdef CONFIG_X86_64
 857         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 858                 hpa_t root = vcpu->mmu.root_hpa;
 859
 860                 page = page_header(root);
 861                 --page->root_count;
 862                 vcpu->mmu.root_hpa = INVALID_PAGE;
 863                 return;
 864         }
 865 #endif
 866         for (i = 0; i < 4; ++i) {
 867                 hpa_t root = vcpu->mmu.pae_root[i];
 868
 869                 if (root) {
 870                         root &= PT64_BASE_ADDR_MASK;
 871                         page = page_header(root);
 872                         --page->root_count;
 873                 }
 874                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 875         }
 876         vcpu->mmu.root_hpa = INVALID_PAGE;
 877 }
 878
 879 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 880 {
 881         int i;
 882         gfn_t root_gfn;
 883         struct kvm_mmu_page *page;
 884
 885         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 886
 887 #ifdef CONFIG_X86_64
 888         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 889                 hpa_t root = vcpu->mmu.root_hpa;
 890
 891                 ASSERT(!VALID_PAGE(root));
 892                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 893                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 894                 root = __pa(page->spt);
 895                 ++page->root_count;
 896                 vcpu->mmu.root_hpa = root;
 897                 return;
 898         }
 899 #endif
 900         for (i = 0; i < 4; ++i) {
 901                 hpa_t root = vcpu->mmu.pae_root[i];
 902
 903                 ASSERT(!VALID_PAGE(root));
 904                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 905                         if (!is_present_pte(vcpu->pdptrs[i])) {
 906                                 vcpu->mmu.pae_root[i] = 0;
 907                                 continue;
 908                         }
 909                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 910                 } else if (vcpu->mmu.root_level == 0)
 911                         root_gfn = 0;
 912                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 913                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 914                                         0, NULL);
 915                 root = __pa(page->spt);
 916                 ++page->root_count;
 917                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 918         }
 919         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 920 }
 921
 922 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 923 {
 924         return vaddr;
 925 }
 926
 927 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 928                                u32 error_code)
 929 {
 930         gpa_t addr = gva;
 931         hpa_t paddr;
 932         int r;
 933
 934         r = mmu_topup_memory_caches(vcpu);
 935         if (r)
 936                 return r;
 937
 938         ASSERT(vcpu);
 939         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 940
 941
 942         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 943
 944         if (is_error_hpa(paddr))
 945                 return 1;
 946
 947         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 948 }
 949
 950 static void nonpaging_free(struct kvm_vcpu *vcpu)
 951 {
 952         mmu_free_roots(vcpu);
 953 }
 954
 955 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 956 {
 957         struct kvm_mmu *context = &vcpu->mmu;
 958
 959         context->new_cr3 = nonpaging_new_cr3;
 960         context->page_fault = nonpaging_page_fault;
 961         context->gva_to_gpa = nonpaging_gva_to_gpa;
 962         context->free = nonpaging_free;
 963         context->root_level = 0;
 964         context->shadow_root_level = PT32E_ROOT_LEVEL;
 965         context->root_hpa = INVALID_PAGE;
 966         return 0;
 967 }
 968
 969 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 970 {
 971         ++vcpu->stat.tlb_flush;
 972         kvm_arch_ops->tlb_flush(vcpu);
 973 }
 974
 975 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 976 {
 977         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 978         mmu_free_roots(vcpu);
 979 }
 980
 981 static void inject_page_fault(struct kvm_vcpu *vcpu,
 982                               u64 addr,
 983                               u32 err_code)
 984 {
 985         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
 986 }
 987
 988 static void paging_free(struct kvm_vcpu *vcpu)
 989 {
 990         nonpaging_free(vcpu);
 991 }
 992
 993 #define PTTYPE 64
 994 #include "paging_tmpl.h"
 995 #undef PTTYPE
 996
 997 #define PTTYPE 32
 998 #include "paging_tmpl.h"
 999 #undef PTTYPE
1000
1001 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1002 {
1003         struct kvm_mmu *context = &vcpu->mmu;
1004
1005         ASSERT(is_pae(vcpu));
1006         context->new_cr3 = paging_new_cr3;
1007         context->page_fault = paging64_page_fault;
1008         context->gva_to_gpa = paging64_gva_to_gpa;
1009         context->free = paging_free;
1010         context->root_level = level;
1011         context->shadow_root_level = level;
1012         context->root_hpa = INVALID_PAGE;
1013         return 0;
1014 }
1015
1016 static int paging64_init_context(struct kvm_vcpu *vcpu)
1017 {
1018         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1019 }
1020
1021 static int paging32_init_context(struct kvm_vcpu *vcpu)
1022 {
1023         struct kvm_mmu *context = &vcpu->mmu;
1024
1025         context->new_cr3 = paging_new_cr3;
1026         context->page_fault = paging32_page_fault;
1027         context->gva_to_gpa = paging32_gva_to_gpa;
1028         context->free = paging_free;
1029         context->root_level = PT32_ROOT_LEVEL;
1030         context->shadow_root_level = PT32E_ROOT_LEVEL;
1031         context->root_hpa = INVALID_PAGE;
1032         return 0;
1033 }
1034
1035 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1036 {
1037         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1038 }
1039
1040 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1041 {
1042         ASSERT(vcpu);
1043         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1044
1045         if (!is_paging(vcpu))
1046                 return nonpaging_init_context(vcpu);
1047         else if (is_long_mode(vcpu))
1048                 return paging64_init_context(vcpu);
1049         else if (is_pae(vcpu))
1050                 return paging32E_init_context(vcpu);
1051         else
1052                 return paging32_init_context(vcpu);
1053 }
1054
1055 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1056 {
1057         ASSERT(vcpu);
1058         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1059                 vcpu->mmu.free(vcpu);
1060                 vcpu->mmu.root_hpa = INVALID_PAGE;
1061         }
1062 }
1063
1064 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1065 {
1066         destroy_kvm_mmu(vcpu);
1067         return init_kvm_mmu(vcpu);
1068 }
1069
1070 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1071 {
1072         int r;
1073
1074         spin_lock(&vcpu->kvm->lock);
1075         r = mmu_topup_memory_caches(vcpu);
1076         if (r)
1077                 goto out;
1078         mmu_alloc_roots(vcpu);
1079         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1080         kvm_mmu_flush_tlb(vcpu);
1081 out:
1082         spin_unlock(&vcpu->kvm->lock);
1083         return r;
1084 }
1085 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1086
1087 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1088 {
1089         mmu_free_roots(vcpu);
1090 }
1091
1092 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1093                                   struct kvm_mmu_page *page,
1094                                   u64 *spte)
1095 {
1096         u64 pte;
1097         struct kvm_mmu_page *child;
1098
1099         pte = *spte;
1100         if (is_present_pte(pte)) {
1101                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1102                         rmap_remove(spte);
1103                 else {
1104                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1105                         mmu_page_remove_parent_pte(child, spte);
1106                 }
1107         }
1108         *spte = 0;
1109         kvm_flush_remote_tlbs(vcpu->kvm);
1110 }
1111
1112 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1113                                   struct kvm_mmu_page *page,
1114                                   u64 *spte,
1115                                   const void *new, int bytes)
1116 {
1117         if (page->role.level != PT_PAGE_TABLE_LEVEL)
1118                 return;
1119
1120         if (page->role.glevels == PT32_ROOT_LEVEL)
1121                 paging32_update_pte(vcpu, page, spte, new, bytes);
1122         else
1123                 paging64_update_pte(vcpu, page, spte, new, bytes);
1124 }
1125
1126 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1127                        const u8 *old, const u8 *new, int bytes)
1128 {
1129         gfn_t gfn = gpa >> PAGE_SHIFT;
1130         struct kvm_mmu_page *page;
1131         struct hlist_node *node, *n;
1132         struct hlist_head *bucket;
1133         unsigned index;
1134         u64 *spte;
1135         unsigned offset = offset_in_page(gpa);
1136         unsigned pte_size;
1137         unsigned page_offset;
1138         unsigned misaligned;
1139         unsigned quadrant;
1140         int level;
1141         int flooded = 0;
1142         int npte;
1143
1144         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1145         if (gfn == vcpu->last_pt_write_gfn) {
1146                 ++vcpu->last_pt_write_count;
1147                 if (vcpu->last_pt_write_count >= 3)
1148                         flooded = 1;
1149         } else {
1150                 vcpu->last_pt_write_gfn = gfn;
1151                 vcpu->last_pt_write_count = 1;
1152         }
1153         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1154         bucket = &vcpu->kvm->mmu_page_hash[index];
1155         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1156                 if (page->gfn != gfn || page->role.metaphysical)
1157                         continue;
1158                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1159                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1160                 misaligned |= bytes < 4;
1161                 if (misaligned || flooded) {
1162                         /*
1163                          * Misaligned accesses are too much trouble to fix
1164                          * up; also, they usually indicate a page is not used
1165                          * as a page table.
1166                          *
1167                          * If we're seeing too many writes to a page,
1168                          * it may no longer be a page table, or we may be
1169                          * forking, in which case it is better to unmap the
1170                          * page.
1171                          */
1172                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1173                                  gpa, bytes, page->role.word);
1174                         kvm_mmu_zap_page(vcpu->kvm, page);
1175                         continue;
1176                 }
1177                 page_offset = offset;
1178                 level = page->role.level;
1179                 npte = 1;
1180                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1181                         page_offset <<= 1;      /* 32->64 */
1182                         /*
1183                          * A 32-bit pde maps 4MB while the shadow pdes map
1184                          * only 2MB.  So we need to double the offset again
1185                          * and zap two pdes instead of one.
1186                          */
1187                         if (level == PT32_ROOT_LEVEL) {
1188                                 page_offset &= ~7; /* kill rounding error */
1189                                 page_offset <<= 1;
1190                                 npte = 2;
1191                         }
1192                         quadrant = page_offset >> PAGE_SHIFT;
1193                         page_offset &= ~PAGE_MASK;
1194                         if (quadrant != page->role.quadrant)
1195                                 continue;
1196                 }
1197                 spte = &page->spt[page_offset / sizeof(*spte)];
1198                 while (npte--) {
1199                         mmu_pte_write_zap_pte(vcpu, page, spte);
1200                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes);
1201                         ++spte;
1202                 }
1203         }
1204 }
1205
1206 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1207 {
1208         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1209
1210         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1211 }
1212
1213 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1214 {
1215         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1216                 struct kvm_mmu_page *page;
1217
1218                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1219                                     struct kvm_mmu_page, link);
1220                 kvm_mmu_zap_page(vcpu->kvm, page);
1221         }
1222 }
1223
1224 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1225 {
1226         struct kvm_mmu_page *page;
1227
1228         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1229                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1230                                     struct kvm_mmu_page, link);
1231                 kvm_mmu_zap_page(vcpu->kvm, page);
1232         }
1233         free_page((unsigned long)vcpu->mmu.pae_root);
1234 }
1235
1236 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1237 {
1238         struct page *page;
1239         int i;
1240
1241         ASSERT(vcpu);
1242
1243         vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1244
1245         /*
1246          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1247          * Therefore we need to allocate shadow page tables in the first
1248          * 4GB of memory, which happens to fit the DMA32 zone.
1249          */
1250         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1251         if (!page)
1252                 goto error_1;
1253         vcpu->mmu.pae_root = page_address(page);
1254         for (i = 0; i < 4; ++i)
1255                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1256
1257         return 0;
1258
1259 error_1:
1260         free_mmu_pages(vcpu);
1261         return -ENOMEM;
1262 }
1263
1264 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1265 {
1266         ASSERT(vcpu);
1267         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1268
1269         return alloc_mmu_pages(vcpu);
1270 }
1271
1272 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1273 {
1274         ASSERT(vcpu);
1275         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1276
1277         return init_kvm_mmu(vcpu);
1278 }
1279
1280 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1281 {
1282         ASSERT(vcpu);
1283
1284         destroy_kvm_mmu(vcpu);
1285         free_mmu_pages(vcpu);
1286         mmu_free_memory_caches(vcpu);
1287 }
1288
1289 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1290 {
1291         struct kvm_mmu_page *page;
1292
1293         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1294                 int i;
1295                 u64 *pt;
1296
1297                 if (!test_bit(slot, &page->slot_bitmap))
1298                         continue;
1299
1300                 pt = page->spt;
1301                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1302                         /* avoid RMW */
1303                         if (pt[i] & PT_WRITABLE_MASK) {
1304                                 rmap_remove(&pt[i]);
1305                                 pt[i] &= ~PT_WRITABLE_MASK;
1306                         }
1307         }
1308 }
1309
1310 void kvm_mmu_zap_all(struct kvm *kvm)
1311 {
1312         struct kvm_mmu_page *page, *node;
1313
1314         list_for_each_entry_safe(page, node, &kvm->active_mmu_pages, link)
1315                 kvm_mmu_zap_page(kvm, page);
1316
1317         kvm_flush_remote_tlbs(kvm);
1318 }
1319
1320 void kvm_mmu_module_exit(void)
1321 {
1322         if (pte_chain_cache)
1323                 kmem_cache_destroy(pte_chain_cache);
1324         if (rmap_desc_cache)
1325                 kmem_cache_destroy(rmap_desc_cache);
1326         if (mmu_page_header_cache)
1327                 kmem_cache_destroy(mmu_page_header_cache);
1328 }
1329
1330 int kvm_mmu_module_init(void)
1331 {
1332         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1333                                             sizeof(struct kvm_pte_chain),
1334                                             0, 0, NULL);
1335         if (!pte_chain_cache)
1336                 goto nomem;
1337         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1338                                             sizeof(struct kvm_rmap_desc),
1339                                             0, 0, NULL);
1340         if (!rmap_desc_cache)
1341                 goto nomem;
1342
1343         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1344                                                   sizeof(struct kvm_mmu_page),
1345                                                   0, 0, NULL);
1346         if (!mmu_page_header_cache)
1347                 goto nomem;
1348
1349         return 0;
1350
1351 nomem:
1352         kvm_mmu_module_exit();
1353         return -ENOMEM;
1354 }
1355
1356 #ifdef AUDIT
1357
1358 static const char *audit_msg;
1359
1360 static gva_t canonicalize(gva_t gva)
1361 {
1362 #ifdef CONFIG_X86_64
1363         gva = (long long)(gva << 16) >> 16;
1364 #endif
1365         return gva;
1366 }
1367
1368 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1369                                 gva_t va, int level)
1370 {
1371         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1372         int i;
1373         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1374
1375         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1376                 u64 ent = pt[i];
1377
1378                 if (!(ent & PT_PRESENT_MASK))
1379                         continue;
1380
1381                 va = canonicalize(va);
1382                 if (level > 1)
1383                         audit_mappings_page(vcpu, ent, va, level - 1);
1384                 else {
1385                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1386                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1387
1388                         if ((ent & PT_PRESENT_MASK)
1389                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1390                                 printk(KERN_ERR "audit error: (%s) levels %d"
1391                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1392                                        audit_msg, vcpu->mmu.root_level,
1393                                        va, gpa, hpa, ent);
1394                 }
1395         }
1396 }
1397
1398 static void audit_mappings(struct kvm_vcpu *vcpu)
1399 {
1400         unsigned i;
1401
1402         if (vcpu->mmu.root_level == 4)
1403                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1404         else
1405                 for (i = 0; i < 4; ++i)
1406                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1407                                 audit_mappings_page(vcpu,
1408                                                     vcpu->mmu.pae_root[i],
1409                                                     i << 30,
1410                                                     2);
1411 }
1412
1413 static int count_rmaps(struct kvm_vcpu *vcpu)
1414 {
1415         int nmaps = 0;
1416         int i, j, k;
1417
1418         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1419                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1420                 struct kvm_rmap_desc *d;
1421
1422                 for (j = 0; j < m->npages; ++j) {
1423                         struct page *page = m->phys_mem[j];
1424
1425                         if (!page->private)
1426                                 continue;
1427                         if (!(page->private & 1)) {
1428                                 ++nmaps;
1429                                 continue;
1430                         }
1431                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1432                         while (d) {
1433                                 for (k = 0; k < RMAP_EXT; ++k)
1434                                         if (d->shadow_ptes[k])
1435                                                 ++nmaps;
1436                                         else
1437                                                 break;
1438                                 d = d->more;
1439                         }
1440                 }
1441         }
1442         return nmaps;
1443 }
1444
1445 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1446 {
1447         int nmaps = 0;
1448         struct kvm_mmu_page *page;
1449         int i;
1450
1451         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1452                 u64 *pt = page->spt;
1453
1454                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1455                         continue;
1456
1457                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1458                         u64 ent = pt[i];
1459
1460                         if (!(ent & PT_PRESENT_MASK))
1461                                 continue;
1462                         if (!(ent & PT_WRITABLE_MASK))
1463                                 continue;
1464                         ++nmaps;
1465                 }
1466         }
1467         return nmaps;
1468 }
1469
1470 static void audit_rmap(struct kvm_vcpu *vcpu)
1471 {
1472         int n_rmap = count_rmaps(vcpu);
1473         int n_actual = count_writable_mappings(vcpu);
1474
1475         if (n_rmap != n_actual)
1476                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1477                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1478 }
1479
1480 static void audit_write_protection(struct kvm_vcpu *vcpu)
1481 {
1482         struct kvm_mmu_page *page;
1483
1484         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1485                 hfn_t hfn;
1486                 struct page *pg;
1487
1488                 if (page->role.metaphysical)
1489                         continue;
1490
1491                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1492                         >> PAGE_SHIFT;
1493                 pg = pfn_to_page(hfn);
1494                 if (pg->private)
1495                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1496                                " mappings: gfn %lx role %x\n",
1497                                __FUNCTION__, audit_msg, page->gfn,
1498                                page->role.word);
1499         }
1500 }
1501
1502 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1503 {
1504         int olddbg = dbg;
1505
1506         dbg = 0;
1507         audit_msg = msg;
1508         audit_rmap(vcpu);
1509         audit_write_protection(vcpu);
1510         audit_mappings(vcpu);
1511         dbg = olddbg;
1512 }
1513
1514 #endif