drivers/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19 #include <linux/types.h>
  20 #include <linux/string.h>
  21 #include <asm/page.h>
  22 #include <linux/mm.h>
  23 #include <linux/highmem.h>
  24 #include <linux/module.h>
  25
  26 #include "vmx.h"
  27 #include "kvm.h"
  28
  29 #undef MMU_DEBUG
  30
  31 #undef AUDIT
  32
  33 #ifdef AUDIT
  34 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  35 #else
  36 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  37 #endif
  38
  39 #ifdef MMU_DEBUG
  40
  41 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  42 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  43
  44 #else
  45
  46 #define pgprintk(x...) do { } while (0)
  47 #define rmap_printk(x...) do { } while (0)
  48
  49 #endif
  50
  51 #if defined(MMU_DEBUG) || defined(AUDIT)
  52 static int dbg = 1;
  53 #endif
  54
  55 #define ASSERT(x)                                                       \
  56         if (!(x)) {                                                     \
  57                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  58                        __FILE__, __LINE__, #x);                         \
  59         }
  60
  61 #define PT64_PT_BITS 9
  62 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  63 #define PT32_PT_BITS 10
  64 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  65
  66 #define PT_WRITABLE_SHIFT 1
  67
  68 #define PT_PRESENT_MASK (1ULL << 0)
  69 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  70 #define PT_USER_MASK (1ULL << 2)
  71 #define PT_PWT_MASK (1ULL << 3)
  72 #define PT_PCD_MASK (1ULL << 4)
  73 #define PT_ACCESSED_MASK (1ULL << 5)
  74 #define PT_DIRTY_MASK (1ULL << 6)
  75 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  76 #define PT_PAT_MASK (1ULL << 7)
  77 #define PT_GLOBAL_MASK (1ULL << 8)
  78 #define PT64_NX_MASK (1ULL << 63)
  79
  80 #define PT_PAT_SHIFT 7
  81 #define PT_DIR_PAT_SHIFT 12
  82 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  83
  84 #define PT32_DIR_PSE36_SIZE 4
  85 #define PT32_DIR_PSE36_SHIFT 13
  86 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  87
  88
  89 #define PT32_PTE_COPY_MASK \
  90         (PT_PRESENT_MASK | PT_ACCESSED_MASK | PT_DIRTY_MASK | PT_GLOBAL_MASK)
  91
  92 #define PT64_PTE_COPY_MASK (PT64_NX_MASK | PT32_PTE_COPY_MASK)
  93
  94 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  95 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  96
  97 #define PT_SHADOW_PS_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
  98 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
  99
 100 #define PT_SHADOW_WRITABLE_SHIFT (PT_FIRST_AVAIL_BITS_SHIFT + 1)
 101 #define PT_SHADOW_WRITABLE_MASK (1ULL << PT_SHADOW_WRITABLE_SHIFT)
 102
 103 #define PT_SHADOW_USER_SHIFT (PT_SHADOW_WRITABLE_SHIFT + 1)
 104 #define PT_SHADOW_USER_MASK (1ULL << (PT_SHADOW_USER_SHIFT))
 105
 106 #define PT_SHADOW_BITS_OFFSET (PT_SHADOW_WRITABLE_SHIFT - PT_WRITABLE_SHIFT)
 107
 108 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 109
 110 #define PT64_LEVEL_BITS 9
 111
 112 #define PT64_LEVEL_SHIFT(level) \
 113                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
 114
 115 #define PT64_LEVEL_MASK(level) \
 116                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 117
 118 #define PT64_INDEX(address, level)\
 119         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 120
 121
 122 #define PT32_LEVEL_BITS 10
 123
 124 #define PT32_LEVEL_SHIFT(level) \
 125                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
 126
 127 #define PT32_LEVEL_MASK(level) \
 128                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 129
 130 #define PT32_INDEX(address, level)\
 131         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 132
 133
 134 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 135 #define PT64_DIR_BASE_ADDR_MASK \
 136         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 137
 138 #define PT32_BASE_ADDR_MASK PAGE_MASK
 139 #define PT32_DIR_BASE_ADDR_MASK \
 140         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 141
 142
 143 #define PFERR_PRESENT_MASK (1U << 0)
 144 #define PFERR_WRITE_MASK (1U << 1)
 145 #define PFERR_USER_MASK (1U << 2)
 146 #define PFERR_FETCH_MASK (1U << 4)
 147
 148 #define PT64_ROOT_LEVEL 4
 149 #define PT32_ROOT_LEVEL 2
 150 #define PT32E_ROOT_LEVEL 3
 151
 152 #define PT_DIRECTORY_LEVEL 2
 153 #define PT_PAGE_TABLE_LEVEL 1
 154
 155 #define RMAP_EXT 4
 156
 157 struct kvm_rmap_desc {
 158         u64 *shadow_ptes[RMAP_EXT];
 159         struct kvm_rmap_desc *more;
 160 };
 161
 162 static int is_write_protection(struct kvm_vcpu *vcpu)
 163 {
 164         return vcpu->cr0 & CR0_WP_MASK;
 165 }
 166
 167 static int is_cpuid_PSE36(void)
 168 {
 169         return 1;
 170 }
 171
 172 static int is_nx(struct kvm_vcpu *vcpu)
 173 {
 174         return vcpu->shadow_efer & EFER_NX;
 175 }
 176
 177 static int is_present_pte(unsigned long pte)
 178 {
 179         return pte & PT_PRESENT_MASK;
 180 }
 181
 182 static int is_writeble_pte(unsigned long pte)
 183 {
 184         return pte & PT_WRITABLE_MASK;
 185 }
 186
 187 static int is_io_pte(unsigned long pte)
 188 {
 189         return pte & PT_SHADOW_IO_MARK;
 190 }
 191
 192 static int is_rmap_pte(u64 pte)
 193 {
 194         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 195                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 196 }
 197
 198 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 199                                   size_t objsize, int min)
 200 {
 201         void *obj;
 202
 203         if (cache->nobjs >= min)
 204                 return 0;
 205         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 206                 obj = kzalloc(objsize, GFP_NOWAIT);
 207                 if (!obj)
 208                         return -ENOMEM;
 209                 cache->objects[cache->nobjs++] = obj;
 210         }
 211         return 0;
 212 }
 213
 214 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 215 {
 216         while (mc->nobjs)
 217                 kfree(mc->objects[--mc->nobjs]);
 218 }
 219
 220 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 221 {
 222         int r;
 223
 224         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 225                                    sizeof(struct kvm_pte_chain), 4);
 226         if (r)
 227                 goto out;
 228         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 229                                    sizeof(struct kvm_rmap_desc), 1);
 230 out:
 231         return r;
 232 }
 233
 234 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 235 {
 236         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 237         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 238 }
 239
 240 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 241                                     size_t size)
 242 {
 243         void *p;
 244
 245         BUG_ON(!mc->nobjs);
 246         p = mc->objects[--mc->nobjs];
 247         memset(p, 0, size);
 248         return p;
 249 }
 250
 251 static void mmu_memory_cache_free(struct kvm_mmu_memory_cache *mc, void *obj)
 252 {
 253         if (mc->nobjs < KVM_NR_MEM_OBJS)
 254                 mc->objects[mc->nobjs++] = obj;
 255         else
 256                 kfree(obj);
 257 }
 258
 259 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 260 {
 261         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 262                                       sizeof(struct kvm_pte_chain));
 263 }
 264
 265 static void mmu_free_pte_chain(struct kvm_vcpu *vcpu,
 266                                struct kvm_pte_chain *pc)
 267 {
 268         mmu_memory_cache_free(&vcpu->mmu_pte_chain_cache, pc);
 269 }
 270
 271 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 272 {
 273         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 274                                       sizeof(struct kvm_rmap_desc));
 275 }
 276
 277 static void mmu_free_rmap_desc(struct kvm_vcpu *vcpu,
 278                                struct kvm_rmap_desc *rd)
 279 {
 280         mmu_memory_cache_free(&vcpu->mmu_rmap_desc_cache, rd);
 281 }
 282
 283 /*
 284  * Reverse mapping data structures:
 285  *
 286  * If page->private bit zero is zero, then page->private points to the
 287  * shadow page table entry that points to page_address(page).
 288  *
 289  * If page->private bit zero is one, (then page->private & ~1) points
 290  * to a struct kvm_rmap_desc containing more mappings.
 291  */
 292 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 293 {
 294         struct page *page;
 295         struct kvm_rmap_desc *desc;
 296         int i;
 297
 298         if (!is_rmap_pte(*spte))
 299                 return;
 300         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 301         if (!page_private(page)) {
 302                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 303                 set_page_private(page,(unsigned long)spte);
 304         } else if (!(page_private(page) & 1)) {
 305                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 306                 desc = mmu_alloc_rmap_desc(vcpu);
 307                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 308                 desc->shadow_ptes[1] = spte;
 309                 set_page_private(page,(unsigned long)desc | 1);
 310         } else {
 311                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 312                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 313                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 314                         desc = desc->more;
 315                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 316                         desc->more = mmu_alloc_rmap_desc(vcpu);
 317                         desc = desc->more;
 318                 }
 319                 for (i = 0; desc->shadow_ptes[i]; ++i)
 320                         ;
 321                 desc->shadow_ptes[i] = spte;
 322         }
 323 }
 324
 325 static void rmap_desc_remove_entry(struct kvm_vcpu *vcpu,
 326                                    struct page *page,
 327                                    struct kvm_rmap_desc *desc,
 328                                    int i,
 329                                    struct kvm_rmap_desc *prev_desc)
 330 {
 331         int j;
 332
 333         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 334                 ;
 335         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 336         desc->shadow_ptes[j] = NULL;
 337         if (j != 0)
 338                 return;
 339         if (!prev_desc && !desc->more)
 340                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 341         else
 342                 if (prev_desc)
 343                         prev_desc->more = desc->more;
 344                 else
 345                         set_page_private(page,(unsigned long)desc->more | 1);
 346         mmu_free_rmap_desc(vcpu, desc);
 347 }
 348
 349 static void rmap_remove(struct kvm_vcpu *vcpu, u64 *spte)
 350 {
 351         struct page *page;
 352         struct kvm_rmap_desc *desc;
 353         struct kvm_rmap_desc *prev_desc;
 354         int i;
 355
 356         if (!is_rmap_pte(*spte))
 357                 return;
 358         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 359         if (!page_private(page)) {
 360                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 361                 BUG();
 362         } else if (!(page_private(page) & 1)) {
 363                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 364                 if ((u64 *)page_private(page) != spte) {
 365                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 366                                spte, *spte);
 367                         BUG();
 368                 }
 369                 set_page_private(page,0);
 370         } else {
 371                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 372                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 373                 prev_desc = NULL;
 374                 while (desc) {
 375                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 376                                 if (desc->shadow_ptes[i] == spte) {
 377                                         rmap_desc_remove_entry(vcpu, page,
 378                                                                desc, i,
 379                                                                prev_desc);
 380                                         return;
 381                                 }
 382                         prev_desc = desc;
 383                         desc = desc->more;
 384                 }
 385                 BUG();
 386         }
 387 }
 388
 389 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 390 {
 391         struct kvm *kvm = vcpu->kvm;
 392         struct page *page;
 393         struct kvm_rmap_desc *desc;
 394         u64 *spte;
 395
 396         page = gfn_to_page(kvm, gfn);
 397         BUG_ON(!page);
 398
 399         while (page_private(page)) {
 400                 if (!(page_private(page) & 1))
 401                         spte = (u64 *)page_private(page);
 402                 else {
 403                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 404                         spte = desc->shadow_ptes[0];
 405                 }
 406                 BUG_ON(!spte);
 407                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 408                        != page_to_pfn(page));
 409                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 410                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 411                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 412                 rmap_remove(vcpu, spte);
 413                 kvm_arch_ops->tlb_flush(vcpu);
 414                 *spte &= ~(u64)PT_WRITABLE_MASK;
 415         }
 416 }
 417
 418 static int is_empty_shadow_page(hpa_t page_hpa)
 419 {
 420         u64 *pos;
 421         u64 *end;
 422
 423         for (pos = __va(page_hpa), end = pos + PAGE_SIZE / sizeof(u64);
 424                       pos != end; pos++)
 425                 if (*pos != 0) {
 426                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 427                                pos, *pos);
 428                         return 0;
 429                 }
 430         return 1;
 431 }
 432
 433 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu, hpa_t page_hpa)
 434 {
 435         struct kvm_mmu_page *page_head = page_header(page_hpa);
 436
 437         ASSERT(is_empty_shadow_page(page_hpa));
 438         page_head->page_hpa = page_hpa;
 439         list_move(&page_head->link, &vcpu->free_pages);
 440         ++vcpu->kvm->n_free_mmu_pages;
 441 }
 442
 443 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 444 {
 445         return gfn;
 446 }
 447
 448 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 449                                                u64 *parent_pte)
 450 {
 451         struct kvm_mmu_page *page;
 452
 453         if (list_empty(&vcpu->free_pages))
 454                 return NULL;
 455
 456         page = list_entry(vcpu->free_pages.next, struct kvm_mmu_page, link);
 457         list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 458         ASSERT(is_empty_shadow_page(page->page_hpa));
 459         page->slot_bitmap = 0;
 460         page->multimapped = 0;
 461         page->parent_pte = parent_pte;
 462         --vcpu->kvm->n_free_mmu_pages;
 463         return page;
 464 }
 465
 466 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 467                                     struct kvm_mmu_page *page, u64 *parent_pte)
 468 {
 469         struct kvm_pte_chain *pte_chain;
 470         struct hlist_node *node;
 471         int i;
 472
 473         if (!parent_pte)
 474                 return;
 475         if (!page->multimapped) {
 476                 u64 *old = page->parent_pte;
 477
 478                 if (!old) {
 479                         page->parent_pte = parent_pte;
 480                         return;
 481                 }
 482                 page->multimapped = 1;
 483                 pte_chain = mmu_alloc_pte_chain(vcpu);
 484                 INIT_HLIST_HEAD(&page->parent_ptes);
 485                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 486                 pte_chain->parent_ptes[0] = old;
 487         }
 488         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 489                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 490                         continue;
 491                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 492                         if (!pte_chain->parent_ptes[i]) {
 493                                 pte_chain->parent_ptes[i] = parent_pte;
 494                                 return;
 495                         }
 496         }
 497         pte_chain = mmu_alloc_pte_chain(vcpu);
 498         BUG_ON(!pte_chain);
 499         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 500         pte_chain->parent_ptes[0] = parent_pte;
 501 }
 502
 503 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
 504                                        struct kvm_mmu_page *page,
 505                                        u64 *parent_pte)
 506 {
 507         struct kvm_pte_chain *pte_chain;
 508         struct hlist_node *node;
 509         int i;
 510
 511         if (!page->multimapped) {
 512                 BUG_ON(page->parent_pte != parent_pte);
 513                 page->parent_pte = NULL;
 514                 return;
 515         }
 516         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 517                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 518                         if (!pte_chain->parent_ptes[i])
 519                                 break;
 520                         if (pte_chain->parent_ptes[i] != parent_pte)
 521                                 continue;
 522                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 523                                 && pte_chain->parent_ptes[i + 1]) {
 524                                 pte_chain->parent_ptes[i]
 525                                         = pte_chain->parent_ptes[i + 1];
 526                                 ++i;
 527                         }
 528                         pte_chain->parent_ptes[i] = NULL;
 529                         if (i == 0) {
 530                                 hlist_del(&pte_chain->link);
 531                                 mmu_free_pte_chain(vcpu, pte_chain);
 532                                 if (hlist_empty(&page->parent_ptes)) {
 533                                         page->multimapped = 0;
 534                                         page->parent_pte = NULL;
 535                                 }
 536                         }
 537                         return;
 538                 }
 539         BUG();
 540 }
 541
 542 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 543                                                 gfn_t gfn)
 544 {
 545         unsigned index;
 546         struct hlist_head *bucket;
 547         struct kvm_mmu_page *page;
 548         struct hlist_node *node;
 549
 550         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 551         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 552         bucket = &vcpu->kvm->mmu_page_hash[index];
 553         hlist_for_each_entry(page, node, bucket, hash_link)
 554                 if (page->gfn == gfn && !page->role.metaphysical) {
 555                         pgprintk("%s: found role %x\n",
 556                                  __FUNCTION__, page->role.word);
 557                         return page;
 558                 }
 559         return NULL;
 560 }
 561
 562 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 563                                              gfn_t gfn,
 564                                              gva_t gaddr,
 565                                              unsigned level,
 566                                              int metaphysical,
 567                                              unsigned hugepage_access,
 568                                              u64 *parent_pte)
 569 {
 570         union kvm_mmu_page_role role;
 571         unsigned index;
 572         unsigned quadrant;
 573         struct hlist_head *bucket;
 574         struct kvm_mmu_page *page;
 575         struct hlist_node *node;
 576
 577         role.word = 0;
 578         role.glevels = vcpu->mmu.root_level;
 579         role.level = level;
 580         role.metaphysical = metaphysical;
 581         role.hugepage_access = hugepage_access;
 582         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 583                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 584                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 585                 role.quadrant = quadrant;
 586         }
 587         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 588                  gfn, role.word);
 589         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 590         bucket = &vcpu->kvm->mmu_page_hash[index];
 591         hlist_for_each_entry(page, node, bucket, hash_link)
 592                 if (page->gfn == gfn && page->role.word == role.word) {
 593                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 594                         pgprintk("%s: found\n", __FUNCTION__);
 595                         return page;
 596                 }
 597         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 598         if (!page)
 599                 return page;
 600         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 601         page->gfn = gfn;
 602         page->role = role;
 603         hlist_add_head(&page->hash_link, bucket);
 604         if (!metaphysical)
 605                 rmap_write_protect(vcpu, gfn);
 606         return page;
 607 }
 608
 609 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 610                                          struct kvm_mmu_page *page)
 611 {
 612         unsigned i;
 613         u64 *pt;
 614         u64 ent;
 615
 616         pt = __va(page->page_hpa);
 617
 618         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 619                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 620                         if (pt[i] & PT_PRESENT_MASK)
 621                                 rmap_remove(vcpu, &pt[i]);
 622                         pt[i] = 0;
 623                 }
 624                 kvm_arch_ops->tlb_flush(vcpu);
 625                 return;
 626         }
 627
 628         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 629                 ent = pt[i];
 630
 631                 pt[i] = 0;
 632                 if (!(ent & PT_PRESENT_MASK))
 633                         continue;
 634                 ent &= PT64_BASE_ADDR_MASK;
 635                 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
 636         }
 637 }
 638
 639 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 640                              struct kvm_mmu_page *page,
 641                              u64 *parent_pte)
 642 {
 643         mmu_page_remove_parent_pte(vcpu, page, parent_pte);
 644 }
 645
 646 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 647                              struct kvm_mmu_page *page)
 648 {
 649         u64 *parent_pte;
 650
 651         while (page->multimapped || page->parent_pte) {
 652                 if (!page->multimapped)
 653                         parent_pte = page->parent_pte;
 654                 else {
 655                         struct kvm_pte_chain *chain;
 656
 657                         chain = container_of(page->parent_ptes.first,
 658                                              struct kvm_pte_chain, link);
 659                         parent_pte = chain->parent_ptes[0];
 660                 }
 661                 BUG_ON(!parent_pte);
 662                 kvm_mmu_put_page(vcpu, page, parent_pte);
 663                 *parent_pte = 0;
 664         }
 665         kvm_mmu_page_unlink_children(vcpu, page);
 666         if (!page->root_count) {
 667                 hlist_del(&page->hash_link);
 668                 kvm_mmu_free_page(vcpu, page->page_hpa);
 669         } else
 670                 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 671 }
 672
 673 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 674 {
 675         unsigned index;
 676         struct hlist_head *bucket;
 677         struct kvm_mmu_page *page;
 678         struct hlist_node *node, *n;
 679         int r;
 680
 681         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 682         r = 0;
 683         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 684         bucket = &vcpu->kvm->mmu_page_hash[index];
 685         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 686                 if (page->gfn == gfn && !page->role.metaphysical) {
 687                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 688                                  page->role.word);
 689                         kvm_mmu_zap_page(vcpu, page);
 690                         r = 1;
 691                 }
 692         return r;
 693 }
 694
 695 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 696 {
 697         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 698         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 699
 700         __set_bit(slot, &page_head->slot_bitmap);
 701 }
 702
 703 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 704 {
 705         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 706
 707         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 708 }
 709
 710 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 711 {
 712         struct page *page;
 713
 714         ASSERT((gpa & HPA_ERR_MASK) == 0);
 715         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 716         if (!page)
 717                 return gpa | HPA_ERR_MASK;
 718         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 719                 | (gpa & (PAGE_SIZE-1));
 720 }
 721
 722 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 723 {
 724         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 725
 726         if (gpa == UNMAPPED_GVA)
 727                 return UNMAPPED_GVA;
 728         return gpa_to_hpa(vcpu, gpa);
 729 }
 730
 731 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 732 {
 733         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 734
 735         if (gpa == UNMAPPED_GVA)
 736                 return NULL;
 737         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 738 }
 739
 740 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 741 {
 742 }
 743
 744 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 745 {
 746         int level = PT32E_ROOT_LEVEL;
 747         hpa_t table_addr = vcpu->mmu.root_hpa;
 748
 749         for (; ; level--) {
 750                 u32 index = PT64_INDEX(v, level);
 751                 u64 *table;
 752                 u64 pte;
 753
 754                 ASSERT(VALID_PAGE(table_addr));
 755                 table = __va(table_addr);
 756
 757                 if (level == 1) {
 758                         pte = table[index];
 759                         if (is_present_pte(pte) && is_writeble_pte(pte))
 760                                 return 0;
 761                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 762                         page_header_update_slot(vcpu->kvm, table, v);
 763                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 764                                                                 PT_USER_MASK;
 765                         rmap_add(vcpu, &table[index]);
 766                         return 0;
 767                 }
 768
 769                 if (table[index] == 0) {
 770                         struct kvm_mmu_page *new_table;
 771                         gfn_t pseudo_gfn;
 772
 773                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 774                                 >> PAGE_SHIFT;
 775                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 776                                                      v, level - 1,
 777                                                      1, 0, &table[index]);
 778                         if (!new_table) {
 779                                 pgprintk("nonpaging_map: ENOMEM\n");
 780                                 return -ENOMEM;
 781                         }
 782
 783                         table[index] = new_table->page_hpa | PT_PRESENT_MASK
 784                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 785                 }
 786                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 787         }
 788 }
 789
 790 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 791 {
 792         int i;
 793         struct kvm_mmu_page *page;
 794
 795 #ifdef CONFIG_X86_64
 796         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 797                 hpa_t root = vcpu->mmu.root_hpa;
 798
 799                 ASSERT(VALID_PAGE(root));
 800                 page = page_header(root);
 801                 --page->root_count;
 802                 vcpu->mmu.root_hpa = INVALID_PAGE;
 803                 return;
 804         }
 805 #endif
 806         for (i = 0; i < 4; ++i) {
 807                 hpa_t root = vcpu->mmu.pae_root[i];
 808
 809                 ASSERT(VALID_PAGE(root));
 810                 root &= PT64_BASE_ADDR_MASK;
 811                 page = page_header(root);
 812                 --page->root_count;
 813                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 814         }
 815         vcpu->mmu.root_hpa = INVALID_PAGE;
 816 }
 817
 818 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 819 {
 820         int i;
 821         gfn_t root_gfn;
 822         struct kvm_mmu_page *page;
 823
 824         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 825
 826 #ifdef CONFIG_X86_64
 827         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 828                 hpa_t root = vcpu->mmu.root_hpa;
 829
 830                 ASSERT(!VALID_PAGE(root));
 831                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 832                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 833                 root = page->page_hpa;
 834                 ++page->root_count;
 835                 vcpu->mmu.root_hpa = root;
 836                 return;
 837         }
 838 #endif
 839         for (i = 0; i < 4; ++i) {
 840                 hpa_t root = vcpu->mmu.pae_root[i];
 841
 842                 ASSERT(!VALID_PAGE(root));
 843                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL)
 844                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 845                 else if (vcpu->mmu.root_level == 0)
 846                         root_gfn = 0;
 847                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 848                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 849                                         0, NULL);
 850                 root = page->page_hpa;
 851                 ++page->root_count;
 852                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 853         }
 854         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 855 }
 856
 857 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 858 {
 859         return vaddr;
 860 }
 861
 862 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 863                                u32 error_code)
 864 {
 865         gpa_t addr = gva;
 866         hpa_t paddr;
 867         int r;
 868
 869         r = mmu_topup_memory_caches(vcpu);
 870         if (r)
 871                 return r;
 872
 873         ASSERT(vcpu);
 874         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 875
 876
 877         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 878
 879         if (is_error_hpa(paddr))
 880                 return 1;
 881
 882         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 883 }
 884
 885 static void nonpaging_free(struct kvm_vcpu *vcpu)
 886 {
 887         mmu_free_roots(vcpu);
 888 }
 889
 890 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 891 {
 892         struct kvm_mmu *context = &vcpu->mmu;
 893
 894         context->new_cr3 = nonpaging_new_cr3;
 895         context->page_fault = nonpaging_page_fault;
 896         context->gva_to_gpa = nonpaging_gva_to_gpa;
 897         context->free = nonpaging_free;
 898         context->root_level = 0;
 899         context->shadow_root_level = PT32E_ROOT_LEVEL;
 900         mmu_alloc_roots(vcpu);
 901         ASSERT(VALID_PAGE(context->root_hpa));
 902         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 903         return 0;
 904 }
 905
 906 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 907 {
 908         ++kvm_stat.tlb_flush;
 909         kvm_arch_ops->tlb_flush(vcpu);
 910 }
 911
 912 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 913 {
 914         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 915         mmu_free_roots(vcpu);
 916         if (unlikely(vcpu->kvm->n_free_mmu_pages < KVM_MIN_FREE_MMU_PAGES))
 917                 kvm_mmu_free_some_pages(vcpu);
 918         mmu_alloc_roots(vcpu);
 919         kvm_mmu_flush_tlb(vcpu);
 920         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 921 }
 922
 923 static inline void set_pte_common(struct kvm_vcpu *vcpu,
 924                              u64 *shadow_pte,
 925                              gpa_t gaddr,
 926                              int dirty,
 927                              u64 access_bits,
 928                              gfn_t gfn)
 929 {
 930         hpa_t paddr;
 931
 932         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 933         if (!dirty)
 934                 access_bits &= ~PT_WRITABLE_MASK;
 935
 936         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 937
 938         *shadow_pte |= access_bits;
 939
 940         if (is_error_hpa(paddr)) {
 941                 *shadow_pte |= gaddr;
 942                 *shadow_pte |= PT_SHADOW_IO_MARK;
 943                 *shadow_pte &= ~PT_PRESENT_MASK;
 944                 return;
 945         }
 946
 947         *shadow_pte |= paddr;
 948
 949         if (access_bits & PT_WRITABLE_MASK) {
 950                 struct kvm_mmu_page *shadow;
 951
 952                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 953                 if (shadow) {
 954                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 955                                  __FUNCTION__, gfn);
 956                         access_bits &= ~PT_WRITABLE_MASK;
 957                         if (is_writeble_pte(*shadow_pte)) {
 958                                     *shadow_pte &= ~PT_WRITABLE_MASK;
 959                                     kvm_arch_ops->tlb_flush(vcpu);
 960                         }
 961                 }
 962         }
 963
 964         if (access_bits & PT_WRITABLE_MASK)
 965                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 966
 967         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 968         rmap_add(vcpu, shadow_pte);
 969 }
 970
 971 static void inject_page_fault(struct kvm_vcpu *vcpu,
 972                               u64 addr,
 973                               u32 err_code)
 974 {
 975         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
 976 }
 977
 978 static inline int fix_read_pf(u64 *shadow_ent)
 979 {
 980         if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
 981             !(*shadow_ent & PT_USER_MASK)) {
 982                 /*
 983                  * If supervisor write protect is disabled, we shadow kernel
 984                  * pages as user pages so we can trap the write access.
 985                  */
 986                 *shadow_ent |= PT_USER_MASK;
 987                 *shadow_ent &= ~PT_WRITABLE_MASK;
 988
 989                 return 1;
 990
 991         }
 992         return 0;
 993 }
 994
 995 static void paging_free(struct kvm_vcpu *vcpu)
 996 {
 997         nonpaging_free(vcpu);
 998 }
 999
1000 #define PTTYPE 64
1001 #include "paging_tmpl.h"
1002 #undef PTTYPE
1003
1004 #define PTTYPE 32
1005 #include "paging_tmpl.h"
1006 #undef PTTYPE
1007
1008 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1009 {
1010         struct kvm_mmu *context = &vcpu->mmu;
1011
1012         ASSERT(is_pae(vcpu));
1013         context->new_cr3 = paging_new_cr3;
1014         context->page_fault = paging64_page_fault;
1015         context->gva_to_gpa = paging64_gva_to_gpa;
1016         context->free = paging_free;
1017         context->root_level = level;
1018         context->shadow_root_level = level;
1019         mmu_alloc_roots(vcpu);
1020         ASSERT(VALID_PAGE(context->root_hpa));
1021         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1022                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1023         return 0;
1024 }
1025
1026 static int paging64_init_context(struct kvm_vcpu *vcpu)
1027 {
1028         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1029 }
1030
1031 static int paging32_init_context(struct kvm_vcpu *vcpu)
1032 {
1033         struct kvm_mmu *context = &vcpu->mmu;
1034
1035         context->new_cr3 = paging_new_cr3;
1036         context->page_fault = paging32_page_fault;
1037         context->gva_to_gpa = paging32_gva_to_gpa;
1038         context->free = paging_free;
1039         context->root_level = PT32_ROOT_LEVEL;
1040         context->shadow_root_level = PT32E_ROOT_LEVEL;
1041         mmu_alloc_roots(vcpu);
1042         ASSERT(VALID_PAGE(context->root_hpa));
1043         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1044                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1045         return 0;
1046 }
1047
1048 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1049 {
1050         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1051 }
1052
1053 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1054 {
1055         ASSERT(vcpu);
1056         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1057
1058         if (!is_paging(vcpu))
1059                 return nonpaging_init_context(vcpu);
1060         else if (is_long_mode(vcpu))
1061                 return paging64_init_context(vcpu);
1062         else if (is_pae(vcpu))
1063                 return paging32E_init_context(vcpu);
1064         else
1065                 return paging32_init_context(vcpu);
1066 }
1067
1068 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1069 {
1070         ASSERT(vcpu);
1071         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1072                 vcpu->mmu.free(vcpu);
1073                 vcpu->mmu.root_hpa = INVALID_PAGE;
1074         }
1075 }
1076
1077 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1078 {
1079         int r;
1080
1081         destroy_kvm_mmu(vcpu);
1082         r = init_kvm_mmu(vcpu);
1083         if (r < 0)
1084                 goto out;
1085         r = mmu_topup_memory_caches(vcpu);
1086 out:
1087         return r;
1088 }
1089
1090 static void mmu_pre_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(vcpu, spte);
1101                 else {
1102                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1103                         mmu_page_remove_parent_pte(vcpu, child, spte);
1104                 }
1105         }
1106         *spte = 0;
1107 }
1108
1109 void kvm_mmu_pre_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1110 {
1111         gfn_t gfn = gpa >> PAGE_SHIFT;
1112         struct kvm_mmu_page *page;
1113         struct hlist_node *node, *n;
1114         struct hlist_head *bucket;
1115         unsigned index;
1116         u64 *spte;
1117         unsigned offset = offset_in_page(gpa);
1118         unsigned pte_size;
1119         unsigned page_offset;
1120         unsigned misaligned;
1121         int level;
1122         int flooded = 0;
1123         int npte;
1124
1125         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1126         if (gfn == vcpu->last_pt_write_gfn) {
1127                 ++vcpu->last_pt_write_count;
1128                 if (vcpu->last_pt_write_count >= 3)
1129                         flooded = 1;
1130         } else {
1131                 vcpu->last_pt_write_gfn = gfn;
1132                 vcpu->last_pt_write_count = 1;
1133         }
1134         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1135         bucket = &vcpu->kvm->mmu_page_hash[index];
1136         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1137                 if (page->gfn != gfn || page->role.metaphysical)
1138                         continue;
1139                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1140                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1141                 if (misaligned || flooded) {
1142                         /*
1143                          * Misaligned accesses are too much trouble to fix
1144                          * up; also, they usually indicate a page is not used
1145                          * as a page table.
1146                          *
1147                          * If we're seeing too many writes to a page,
1148                          * it may no longer be a page table, or we may be
1149                          * forking, in which case it is better to unmap the
1150                          * page.
1151                          */
1152                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1153                                  gpa, bytes, page->role.word);
1154                         kvm_mmu_zap_page(vcpu, page);
1155                         continue;
1156                 }
1157                 page_offset = offset;
1158                 level = page->role.level;
1159                 npte = 1;
1160                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1161                         page_offset <<= 1;      /* 32->64 */
1162                         /*
1163                          * A 32-bit pde maps 4MB while the shadow pdes map
1164                          * only 2MB.  So we need to double the offset again
1165                          * and zap two pdes instead of one.
1166                          */
1167                         if (level == PT32_ROOT_LEVEL) {
1168                                 page_offset &= ~7; /* kill rounding error */
1169                                 page_offset <<= 1;
1170                                 npte = 2;
1171                         }
1172                         page_offset &= ~PAGE_MASK;
1173                 }
1174                 spte = __va(page->page_hpa);
1175                 spte += page_offset / sizeof(*spte);
1176                 while (npte--) {
1177                         mmu_pre_write_zap_pte(vcpu, page, spte);
1178                         ++spte;
1179                 }
1180         }
1181 }
1182
1183 void kvm_mmu_post_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1184 {
1185 }
1186
1187 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1188 {
1189         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1190
1191         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1192 }
1193
1194 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1195 {
1196         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1197                 struct kvm_mmu_page *page;
1198
1199                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1200                                     struct kvm_mmu_page, link);
1201                 kvm_mmu_zap_page(vcpu, page);
1202         }
1203 }
1204 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1205
1206 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1207 {
1208         struct kvm_mmu_page *page;
1209
1210         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1211                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1212                                     struct kvm_mmu_page, link);
1213                 kvm_mmu_zap_page(vcpu, page);
1214         }
1215         while (!list_empty(&vcpu->free_pages)) {
1216                 page = list_entry(vcpu->free_pages.next,
1217                                   struct kvm_mmu_page, link);
1218                 list_del(&page->link);
1219                 __free_page(pfn_to_page(page->page_hpa >> PAGE_SHIFT));
1220                 page->page_hpa = INVALID_PAGE;
1221         }
1222         free_page((unsigned long)vcpu->mmu.pae_root);
1223 }
1224
1225 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1226 {
1227         struct page *page;
1228         int i;
1229
1230         ASSERT(vcpu);
1231
1232         for (i = 0; i < KVM_NUM_MMU_PAGES; i++) {
1233                 struct kvm_mmu_page *page_header = &vcpu->page_header_buf[i];
1234
1235                 INIT_LIST_HEAD(&page_header->link);
1236                 if ((page = alloc_page(GFP_KERNEL)) == NULL)
1237                         goto error_1;
1238                 set_page_private(page, (unsigned long)page_header);
1239                 page_header->page_hpa = (hpa_t)page_to_pfn(page) << PAGE_SHIFT;
1240                 memset(__va(page_header->page_hpa), 0, PAGE_SIZE);
1241                 list_add(&page_header->link, &vcpu->free_pages);
1242                 ++vcpu->kvm->n_free_mmu_pages;
1243         }
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         ASSERT(list_empty(&vcpu->free_pages));
1269
1270         return alloc_mmu_pages(vcpu);
1271 }
1272
1273 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1274 {
1275         ASSERT(vcpu);
1276         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1277         ASSERT(!list_empty(&vcpu->free_pages));
1278
1279         return init_kvm_mmu(vcpu);
1280 }
1281
1282 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1283 {
1284         ASSERT(vcpu);
1285
1286         destroy_kvm_mmu(vcpu);
1287         free_mmu_pages(vcpu);
1288         mmu_free_memory_caches(vcpu);
1289 }
1290
1291 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1292 {
1293         struct kvm *kvm = vcpu->kvm;
1294         struct kvm_mmu_page *page;
1295
1296         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1297                 int i;
1298                 u64 *pt;
1299
1300                 if (!test_bit(slot, &page->slot_bitmap))
1301                         continue;
1302
1303                 pt = __va(page->page_hpa);
1304                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1305                         /* avoid RMW */
1306                         if (pt[i] & PT_WRITABLE_MASK) {
1307                                 rmap_remove(vcpu, &pt[i]);
1308                                 pt[i] &= ~PT_WRITABLE_MASK;
1309                         }
1310         }
1311 }
1312
1313 void kvm_mmu_zap_all(struct kvm_vcpu *vcpu)
1314 {
1315         destroy_kvm_mmu(vcpu);
1316
1317         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1318                 struct kvm_mmu_page *page;
1319
1320                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1321                                     struct kvm_mmu_page, link);
1322                 kvm_mmu_zap_page(vcpu, page);
1323         }
1324
1325         mmu_free_memory_caches(vcpu);
1326         kvm_arch_ops->tlb_flush(vcpu);
1327         init_kvm_mmu(vcpu);
1328 }
1329
1330 #ifdef AUDIT
1331
1332 static const char *audit_msg;
1333
1334 static gva_t canonicalize(gva_t gva)
1335 {
1336 #ifdef CONFIG_X86_64
1337         gva = (long long)(gva << 16) >> 16;
1338 #endif
1339         return gva;
1340 }
1341
1342 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1343                                 gva_t va, int level)
1344 {
1345         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1346         int i;
1347         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1348
1349         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1350                 u64 ent = pt[i];
1351
1352                 if (!ent & PT_PRESENT_MASK)
1353                         continue;
1354
1355                 va = canonicalize(va);
1356                 if (level > 1)
1357                         audit_mappings_page(vcpu, ent, va, level - 1);
1358                 else {
1359                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1360                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1361
1362                         if ((ent & PT_PRESENT_MASK)
1363                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1364                                 printk(KERN_ERR "audit error: (%s) levels %d"
1365                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1366                                        audit_msg, vcpu->mmu.root_level,
1367                                        va, gpa, hpa, ent);
1368                 }
1369         }
1370 }
1371
1372 static void audit_mappings(struct kvm_vcpu *vcpu)
1373 {
1374         unsigned i;
1375
1376         if (vcpu->mmu.root_level == 4)
1377                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1378         else
1379                 for (i = 0; i < 4; ++i)
1380                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1381                                 audit_mappings_page(vcpu,
1382                                                     vcpu->mmu.pae_root[i],
1383                                                     i << 30,
1384                                                     2);
1385 }
1386
1387 static int count_rmaps(struct kvm_vcpu *vcpu)
1388 {
1389         int nmaps = 0;
1390         int i, j, k;
1391
1392         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1393                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1394                 struct kvm_rmap_desc *d;
1395
1396                 for (j = 0; j < m->npages; ++j) {
1397                         struct page *page = m->phys_mem[j];
1398
1399                         if (!page->private)
1400                                 continue;
1401                         if (!(page->private & 1)) {
1402                                 ++nmaps;
1403                                 continue;
1404                         }
1405                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1406                         while (d) {
1407                                 for (k = 0; k < RMAP_EXT; ++k)
1408                                         if (d->shadow_ptes[k])
1409                                                 ++nmaps;
1410                                         else
1411                                                 break;
1412                                 d = d->more;
1413                         }
1414                 }
1415         }
1416         return nmaps;
1417 }
1418
1419 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1420 {
1421         int nmaps = 0;
1422         struct kvm_mmu_page *page;
1423         int i;
1424
1425         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1426                 u64 *pt = __va(page->page_hpa);
1427
1428                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1429                         continue;
1430
1431                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1432                         u64 ent = pt[i];
1433
1434                         if (!(ent & PT_PRESENT_MASK))
1435                                 continue;
1436                         if (!(ent & PT_WRITABLE_MASK))
1437                                 continue;
1438                         ++nmaps;
1439                 }
1440         }
1441         return nmaps;
1442 }
1443
1444 static void audit_rmap(struct kvm_vcpu *vcpu)
1445 {
1446         int n_rmap = count_rmaps(vcpu);
1447         int n_actual = count_writable_mappings(vcpu);
1448
1449         if (n_rmap != n_actual)
1450                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1451                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1452 }
1453
1454 static void audit_write_protection(struct kvm_vcpu *vcpu)
1455 {
1456         struct kvm_mmu_page *page;
1457
1458         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1459                 hfn_t hfn;
1460                 struct page *pg;
1461
1462                 if (page->role.metaphysical)
1463                         continue;
1464
1465                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1466                         >> PAGE_SHIFT;
1467                 pg = pfn_to_page(hfn);
1468                 if (pg->private)
1469                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1470                                " mappings: gfn %lx role %x\n",
1471                                __FUNCTION__, audit_msg, page->gfn,
1472                                page->role.word);
1473         }
1474 }
1475
1476 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1477 {
1478         int olddbg = dbg;
1479
1480         dbg = 0;
1481         audit_msg = msg;
1482         audit_rmap(vcpu);
1483         audit_write_protection(vcpu);
1484         audit_mappings(vcpu);
1485         dbg = olddbg;
1486 }
1487
1488 #endif