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