git.oblomov.eu Git - linux-2.6/blob - 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) & PAGE_MASK)
 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_memory_slot *slot;
 394         struct kvm_rmap_desc *desc;
 395         u64 *spte;
 396
 397         slot = gfn_to_memslot(kvm, gfn);
 398         BUG_ON(!slot);
 399         page = gfn_to_page(slot, gfn);
 400
 401         while (page_private(page)) {
 402                 if (!(page_private(page) & 1))
 403                         spte = (u64 *)page_private(page);
 404                 else {
 405                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 406                         spte = desc->shadow_ptes[0];
 407                 }
 408                 BUG_ON(!spte);
 409                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) !=
 410                        page_to_pfn(page) << PAGE_SHIFT);
 411                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 412                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 413                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 414                 rmap_remove(vcpu, spte);
 415                 kvm_arch_ops->tlb_flush(vcpu);
 416                 *spte &= ~(u64)PT_WRITABLE_MASK;
 417         }
 418 }
 419
 420 static int is_empty_shadow_page(hpa_t page_hpa)
 421 {
 422         u64 *pos;
 423         u64 *end;
 424
 425         for (pos = __va(page_hpa), end = pos + PAGE_SIZE / sizeof(u64);
 426                       pos != end; pos++)
 427                 if (*pos != 0) {
 428                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 429                                pos, *pos);
 430                         return 0;
 431                 }
 432         return 1;
 433 }
 434
 435 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu, hpa_t page_hpa)
 436 {
 437         struct kvm_mmu_page *page_head = page_header(page_hpa);
 438
 439         ASSERT(is_empty_shadow_page(page_hpa));
 440         list_del(&page_head->link);
 441         page_head->page_hpa = page_hpa;
 442         list_add(&page_head->link, &vcpu->free_pages);
 443         ++vcpu->kvm->n_free_mmu_pages;
 444 }
 445
 446 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 447 {
 448         return gfn;
 449 }
 450
 451 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 452                                                u64 *parent_pte)
 453 {
 454         struct kvm_mmu_page *page;
 455
 456         if (list_empty(&vcpu->free_pages))
 457                 return NULL;
 458
 459         page = list_entry(vcpu->free_pages.next, struct kvm_mmu_page, link);
 460         list_del(&page->link);
 461         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 462         ASSERT(is_empty_shadow_page(page->page_hpa));
 463         page->slot_bitmap = 0;
 464         page->global = 1;
 465         page->multimapped = 0;
 466         page->parent_pte = parent_pte;
 467         --vcpu->kvm->n_free_mmu_pages;
 468         return page;
 469 }
 470
 471 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 472                                     struct kvm_mmu_page *page, u64 *parent_pte)
 473 {
 474         struct kvm_pte_chain *pte_chain;
 475         struct hlist_node *node;
 476         int i;
 477
 478         if (!parent_pte)
 479                 return;
 480         if (!page->multimapped) {
 481                 u64 *old = page->parent_pte;
 482
 483                 if (!old) {
 484                         page->parent_pte = parent_pte;
 485                         return;
 486                 }
 487                 page->multimapped = 1;
 488                 pte_chain = mmu_alloc_pte_chain(vcpu);
 489                 INIT_HLIST_HEAD(&page->parent_ptes);
 490                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 491                 pte_chain->parent_ptes[0] = old;
 492         }
 493         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 494                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 495                         continue;
 496                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 497                         if (!pte_chain->parent_ptes[i]) {
 498                                 pte_chain->parent_ptes[i] = parent_pte;
 499                                 return;
 500                         }
 501         }
 502         pte_chain = mmu_alloc_pte_chain(vcpu);
 503         BUG_ON(!pte_chain);
 504         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 505         pte_chain->parent_ptes[0] = parent_pte;
 506 }
 507
 508 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
 509                                        struct kvm_mmu_page *page,
 510                                        u64 *parent_pte)
 511 {
 512         struct kvm_pte_chain *pte_chain;
 513         struct hlist_node *node;
 514         int i;
 515
 516         if (!page->multimapped) {
 517                 BUG_ON(page->parent_pte != parent_pte);
 518                 page->parent_pte = NULL;
 519                 return;
 520         }
 521         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 522                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 523                         if (!pte_chain->parent_ptes[i])
 524                                 break;
 525                         if (pte_chain->parent_ptes[i] != parent_pte)
 526                                 continue;
 527                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 528                                 && pte_chain->parent_ptes[i + 1]) {
 529                                 pte_chain->parent_ptes[i]
 530                                         = pte_chain->parent_ptes[i + 1];
 531                                 ++i;
 532                         }
 533                         pte_chain->parent_ptes[i] = NULL;
 534                         if (i == 0) {
 535                                 hlist_del(&pte_chain->link);
 536                                 mmu_free_pte_chain(vcpu, pte_chain);
 537                                 if (hlist_empty(&page->parent_ptes)) {
 538                                         page->multimapped = 0;
 539                                         page->parent_pte = NULL;
 540                                 }
 541                         }
 542                         return;
 543                 }
 544         BUG();
 545 }
 546
 547 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 548                                                 gfn_t gfn)
 549 {
 550         unsigned index;
 551         struct hlist_head *bucket;
 552         struct kvm_mmu_page *page;
 553         struct hlist_node *node;
 554
 555         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 556         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 557         bucket = &vcpu->kvm->mmu_page_hash[index];
 558         hlist_for_each_entry(page, node, bucket, hash_link)
 559                 if (page->gfn == gfn && !page->role.metaphysical) {
 560                         pgprintk("%s: found role %x\n",
 561                                  __FUNCTION__, page->role.word);
 562                         return page;
 563                 }
 564         return NULL;
 565 }
 566
 567 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 568                                              gfn_t gfn,
 569                                              gva_t gaddr,
 570                                              unsigned level,
 571                                              int metaphysical,
 572                                              u64 *parent_pte)
 573 {
 574         union kvm_mmu_page_role role;
 575         unsigned index;
 576         unsigned quadrant;
 577         struct hlist_head *bucket;
 578         struct kvm_mmu_page *page;
 579         struct hlist_node *node;
 580
 581         role.word = 0;
 582         role.glevels = vcpu->mmu.root_level;
 583         role.level = level;
 584         role.metaphysical = metaphysical;
 585         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 586                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 587                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 588                 role.quadrant = quadrant;
 589         }
 590         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 591                  gfn, role.word);
 592         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 593         bucket = &vcpu->kvm->mmu_page_hash[index];
 594         hlist_for_each_entry(page, node, bucket, hash_link)
 595                 if (page->gfn == gfn && page->role.word == role.word) {
 596                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 597                         pgprintk("%s: found\n", __FUNCTION__);
 598                         return page;
 599                 }
 600         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 601         if (!page)
 602                 return page;
 603         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 604         page->gfn = gfn;
 605         page->role = role;
 606         hlist_add_head(&page->hash_link, bucket);
 607         if (!metaphysical)
 608                 rmap_write_protect(vcpu, gfn);
 609         return page;
 610 }
 611
 612 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 613                                          struct kvm_mmu_page *page)
 614 {
 615         unsigned i;
 616         u64 *pt;
 617         u64 ent;
 618
 619         pt = __va(page->page_hpa);
 620
 621         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 622                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 623                         if (pt[i] & PT_PRESENT_MASK)
 624                                 rmap_remove(vcpu, &pt[i]);
 625                         pt[i] = 0;
 626                 }
 627                 kvm_arch_ops->tlb_flush(vcpu);
 628                 return;
 629         }
 630
 631         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 632                 ent = pt[i];
 633
 634                 pt[i] = 0;
 635                 if (!(ent & PT_PRESENT_MASK))
 636                         continue;
 637                 ent &= PT64_BASE_ADDR_MASK;
 638                 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
 639         }
 640 }
 641
 642 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 643                              struct kvm_mmu_page *page,
 644                              u64 *parent_pte)
 645 {
 646         mmu_page_remove_parent_pte(vcpu, page, parent_pte);
 647 }
 648
 649 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 650                              struct kvm_mmu_page *page)
 651 {
 652         u64 *parent_pte;
 653
 654         while (page->multimapped || page->parent_pte) {
 655                 if (!page->multimapped)
 656                         parent_pte = page->parent_pte;
 657                 else {
 658                         struct kvm_pte_chain *chain;
 659
 660                         chain = container_of(page->parent_ptes.first,
 661                                              struct kvm_pte_chain, link);
 662                         parent_pte = chain->parent_ptes[0];
 663                 }
 664                 BUG_ON(!parent_pte);
 665                 kvm_mmu_put_page(vcpu, page, parent_pte);
 666                 *parent_pte = 0;
 667         }
 668         kvm_mmu_page_unlink_children(vcpu, page);
 669         if (!page->root_count) {
 670                 hlist_del(&page->hash_link);
 671                 kvm_mmu_free_page(vcpu, page->page_hpa);
 672         } else {
 673                 list_del(&page->link);
 674                 list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 675         }
 676 }
 677
 678 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 679 {
 680         unsigned index;
 681         struct hlist_head *bucket;
 682         struct kvm_mmu_page *page;
 683         struct hlist_node *node, *n;
 684         int r;
 685
 686         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 687         r = 0;
 688         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 689         bucket = &vcpu->kvm->mmu_page_hash[index];
 690         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 691                 if (page->gfn == gfn && !page->role.metaphysical) {
 692                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 693                                  page->role.word);
 694                         kvm_mmu_zap_page(vcpu, page);
 695                         r = 1;
 696                 }
 697         return r;
 698 }
 699
 700 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 701 {
 702         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 703         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 704
 705         __set_bit(slot, &page_head->slot_bitmap);
 706 }
 707
 708 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 709 {
 710         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 711
 712         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 713 }
 714
 715 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 716 {
 717         struct kvm_memory_slot *slot;
 718         struct page *page;
 719
 720         ASSERT((gpa & HPA_ERR_MASK) == 0);
 721         slot = gfn_to_memslot(vcpu->kvm, gpa >> PAGE_SHIFT);
 722         if (!slot)
 723                 return gpa | HPA_ERR_MASK;
 724         page = gfn_to_page(slot, gpa >> PAGE_SHIFT);
 725         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 726                 | (gpa & (PAGE_SIZE-1));
 727 }
 728
 729 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 730 {
 731         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 732
 733         if (gpa == UNMAPPED_GVA)
 734                 return UNMAPPED_GVA;
 735         return gpa_to_hpa(vcpu, gpa);
 736 }
 737
 738 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 739 {
 740 }
 741
 742 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 743 {
 744         int level = PT32E_ROOT_LEVEL;
 745         hpa_t table_addr = vcpu->mmu.root_hpa;
 746
 747         for (; ; level--) {
 748                 u32 index = PT64_INDEX(v, level);
 749                 u64 *table;
 750                 u64 pte;
 751
 752                 ASSERT(VALID_PAGE(table_addr));
 753                 table = __va(table_addr);
 754
 755                 if (level == 1) {
 756                         pte = table[index];
 757                         if (is_present_pte(pte) && is_writeble_pte(pte))
 758                                 return 0;
 759                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 760                         page_header_update_slot(vcpu->kvm, table, v);
 761                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 762                                                                 PT_USER_MASK;
 763                         rmap_add(vcpu, &table[index]);
 764                         return 0;
 765                 }
 766
 767                 if (table[index] == 0) {
 768                         struct kvm_mmu_page *new_table;
 769                         gfn_t pseudo_gfn;
 770
 771                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 772                                 >> PAGE_SHIFT;
 773                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 774                                                      v, level - 1,
 775                                                      1, &table[index]);
 776                         if (!new_table) {
 777                                 pgprintk("nonpaging_map: ENOMEM\n");
 778                                 return -ENOMEM;
 779                         }
 780
 781                         table[index] = new_table->page_hpa | PT_PRESENT_MASK
 782                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 783                 }
 784                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 785         }
 786 }
 787
 788 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 789 {
 790         int i;
 791         struct kvm_mmu_page *page;
 792
 793 #ifdef CONFIG_X86_64
 794         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 795                 hpa_t root = vcpu->mmu.root_hpa;
 796
 797                 ASSERT(VALID_PAGE(root));
 798                 page = page_header(root);
 799                 --page->root_count;
 800                 vcpu->mmu.root_hpa = INVALID_PAGE;
 801                 return;
 802         }
 803 #endif
 804         for (i = 0; i < 4; ++i) {
 805                 hpa_t root = vcpu->mmu.pae_root[i];
 806
 807                 ASSERT(VALID_PAGE(root));
 808                 root &= PT64_BASE_ADDR_MASK;
 809                 page = page_header(root);
 810                 --page->root_count;
 811                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 812         }
 813         vcpu->mmu.root_hpa = INVALID_PAGE;
 814 }
 815
 816 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 817 {
 818         int i;
 819         gfn_t root_gfn;
 820         struct kvm_mmu_page *page;
 821
 822         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 823
 824 #ifdef CONFIG_X86_64
 825         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 826                 hpa_t root = vcpu->mmu.root_hpa;
 827
 828                 ASSERT(!VALID_PAGE(root));
 829                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 830                                         PT64_ROOT_LEVEL, 0, NULL);
 831                 root = page->page_hpa;
 832                 ++page->root_count;
 833                 vcpu->mmu.root_hpa = root;
 834                 return;
 835         }
 836 #endif
 837         for (i = 0; i < 4; ++i) {
 838                 hpa_t root = vcpu->mmu.pae_root[i];
 839
 840                 ASSERT(!VALID_PAGE(root));
 841                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL)
 842                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 843                 else if (vcpu->mmu.root_level == 0)
 844                         root_gfn = 0;
 845                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 846                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 847                                         NULL);
 848                 root = page->page_hpa;
 849                 ++page->root_count;
 850                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 851         }
 852         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 853 }
 854
 855 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 856 {
 857         return vaddr;
 858 }
 859
 860 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 861                                u32 error_code)
 862 {
 863         gpa_t addr = gva;
 864         hpa_t paddr;
 865         int r;
 866
 867         r = mmu_topup_memory_caches(vcpu);
 868         if (r)
 869                 return r;
 870
 871         ASSERT(vcpu);
 872         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 873
 874
 875         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 876
 877         if (is_error_hpa(paddr))
 878                 return 1;
 879
 880         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 881 }
 882
 883 static void nonpaging_free(struct kvm_vcpu *vcpu)
 884 {
 885         mmu_free_roots(vcpu);
 886 }
 887
 888 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 889 {
 890         struct kvm_mmu *context = &vcpu->mmu;
 891
 892         context->new_cr3 = nonpaging_new_cr3;
 893         context->page_fault = nonpaging_page_fault;
 894         context->gva_to_gpa = nonpaging_gva_to_gpa;
 895         context->free = nonpaging_free;
 896         context->root_level = 0;
 897         context->shadow_root_level = PT32E_ROOT_LEVEL;
 898         mmu_alloc_roots(vcpu);
 899         ASSERT(VALID_PAGE(context->root_hpa));
 900         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 901         return 0;
 902 }
 903
 904 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 905 {
 906         ++kvm_stat.tlb_flush;
 907         kvm_arch_ops->tlb_flush(vcpu);
 908 }
 909
 910 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 911 {
 912         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 913         mmu_free_roots(vcpu);
 914         if (unlikely(vcpu->kvm->n_free_mmu_pages < KVM_MIN_FREE_MMU_PAGES))
 915                 kvm_mmu_free_some_pages(vcpu);
 916         mmu_alloc_roots(vcpu);
 917         kvm_mmu_flush_tlb(vcpu);
 918         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 919 }
 920
 921 static void mark_pagetable_nonglobal(void *shadow_pte)
 922 {
 923         page_header(__pa(shadow_pte))->global = 0;
 924 }
 925
 926 static inline void set_pte_common(struct kvm_vcpu *vcpu,
 927                              u64 *shadow_pte,
 928                              gpa_t gaddr,
 929                              int dirty,
 930                              u64 access_bits,
 931                              gfn_t gfn)
 932 {
 933         hpa_t paddr;
 934
 935         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 936         if (!dirty)
 937                 access_bits &= ~PT_WRITABLE_MASK;
 938
 939         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 940
 941         *shadow_pte |= access_bits;
 942
 943         if (!(*shadow_pte & PT_GLOBAL_MASK))
 944                 mark_pagetable_nonglobal(shadow_pte);
 945
 946         if (is_error_hpa(paddr)) {
 947                 *shadow_pte |= gaddr;
 948                 *shadow_pte |= PT_SHADOW_IO_MARK;
 949                 *shadow_pte &= ~PT_PRESENT_MASK;
 950                 return;
 951         }
 952
 953         *shadow_pte |= paddr;
 954
 955         if (access_bits & PT_WRITABLE_MASK) {
 956                 struct kvm_mmu_page *shadow;
 957
 958                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 959                 if (shadow) {
 960                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 961                                  __FUNCTION__, gfn);
 962                         access_bits &= ~PT_WRITABLE_MASK;
 963                         if (is_writeble_pte(*shadow_pte)) {
 964                                     *shadow_pte &= ~PT_WRITABLE_MASK;
 965                                     kvm_arch_ops->tlb_flush(vcpu);
 966                         }
 967                 }
 968         }
 969
 970         if (access_bits & PT_WRITABLE_MASK)
 971                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 972
 973         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 974         rmap_add(vcpu, shadow_pte);
 975 }
 976
 977 static void inject_page_fault(struct kvm_vcpu *vcpu,
 978                               u64 addr,
 979                               u32 err_code)
 980 {
 981         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
 982 }
 983
 984 static inline int fix_read_pf(u64 *shadow_ent)
 985 {
 986         if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
 987             !(*shadow_ent & PT_USER_MASK)) {
 988                 /*
 989                  * If supervisor write protect is disabled, we shadow kernel
 990                  * pages as user pages so we can trap the write access.
 991                  */
 992                 *shadow_ent |= PT_USER_MASK;
 993                 *shadow_ent &= ~PT_WRITABLE_MASK;
 994
 995                 return 1;
 996
 997         }
 998         return 0;
 999 }
1000
1001 static void paging_free(struct kvm_vcpu *vcpu)
1002 {
1003         nonpaging_free(vcpu);
1004 }
1005
1006 #define PTTYPE 64
1007 #include "paging_tmpl.h"
1008 #undef PTTYPE
1009
1010 #define PTTYPE 32
1011 #include "paging_tmpl.h"
1012 #undef PTTYPE
1013
1014 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1015 {
1016         struct kvm_mmu *context = &vcpu->mmu;
1017
1018         ASSERT(is_pae(vcpu));
1019         context->new_cr3 = paging_new_cr3;
1020         context->page_fault = paging64_page_fault;
1021         context->gva_to_gpa = paging64_gva_to_gpa;
1022         context->free = paging_free;
1023         context->root_level = level;
1024         context->shadow_root_level = level;
1025         mmu_alloc_roots(vcpu);
1026         ASSERT(VALID_PAGE(context->root_hpa));
1027         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1028                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1029         return 0;
1030 }
1031
1032 static int paging64_init_context(struct kvm_vcpu *vcpu)
1033 {
1034         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1035 }
1036
1037 static int paging32_init_context(struct kvm_vcpu *vcpu)
1038 {
1039         struct kvm_mmu *context = &vcpu->mmu;
1040
1041         context->new_cr3 = paging_new_cr3;
1042         context->page_fault = paging32_page_fault;
1043         context->gva_to_gpa = paging32_gva_to_gpa;
1044         context->free = paging_free;
1045         context->root_level = PT32_ROOT_LEVEL;
1046         context->shadow_root_level = PT32E_ROOT_LEVEL;
1047         mmu_alloc_roots(vcpu);
1048         ASSERT(VALID_PAGE(context->root_hpa));
1049         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1050                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1051         return 0;
1052 }
1053
1054 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1055 {
1056         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1057 }
1058
1059 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1060 {
1061         ASSERT(vcpu);
1062         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1063
1064         if (!is_paging(vcpu))
1065                 return nonpaging_init_context(vcpu);
1066         else if (is_long_mode(vcpu))
1067                 return paging64_init_context(vcpu);
1068         else if (is_pae(vcpu))
1069                 return paging32E_init_context(vcpu);
1070         else
1071                 return paging32_init_context(vcpu);
1072 }
1073
1074 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1075 {
1076         ASSERT(vcpu);
1077         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1078                 vcpu->mmu.free(vcpu);
1079                 vcpu->mmu.root_hpa = INVALID_PAGE;
1080         }
1081 }
1082
1083 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1084 {
1085         int r;
1086
1087         destroy_kvm_mmu(vcpu);
1088         r = init_kvm_mmu(vcpu);
1089         if (r < 0)
1090                 goto out;
1091         r = mmu_topup_memory_caches(vcpu);
1092 out:
1093         return r;
1094 }
1095
1096 void kvm_mmu_pre_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1097 {
1098         gfn_t gfn = gpa >> PAGE_SHIFT;
1099         struct kvm_mmu_page *page;
1100         struct kvm_mmu_page *child;
1101         struct hlist_node *node, *n;
1102         struct hlist_head *bucket;
1103         unsigned index;
1104         u64 *spte;
1105         u64 pte;
1106         unsigned offset = offset_in_page(gpa);
1107         unsigned pte_size;
1108         unsigned page_offset;
1109         unsigned misaligned;
1110         int level;
1111         int flooded = 0;
1112
1113         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1114         if (gfn == vcpu->last_pt_write_gfn) {
1115                 ++vcpu->last_pt_write_count;
1116                 if (vcpu->last_pt_write_count >= 3)
1117                         flooded = 1;
1118         } else {
1119                 vcpu->last_pt_write_gfn = gfn;
1120                 vcpu->last_pt_write_count = 1;
1121         }
1122         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1123         bucket = &vcpu->kvm->mmu_page_hash[index];
1124         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1125                 if (page->gfn != gfn || page->role.metaphysical)
1126                         continue;
1127                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1128                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1129                 if (misaligned || flooded) {
1130                         /*
1131                          * Misaligned accesses are too much trouble to fix
1132                          * up; also, they usually indicate a page is not used
1133                          * as a page table.
1134                          *
1135                          * If we're seeing too many writes to a page,
1136                          * it may no longer be a page table, or we may be
1137                          * forking, in which case it is better to unmap the
1138                          * page.
1139                          */
1140                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1141                                  gpa, bytes, page->role.word);
1142                         kvm_mmu_zap_page(vcpu, page);
1143                         continue;
1144                 }
1145                 page_offset = offset;
1146                 level = page->role.level;
1147                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1148                         page_offset <<= 1;          /* 32->64 */
1149                         page_offset &= ~PAGE_MASK;
1150                 }
1151                 spte = __va(page->page_hpa);
1152                 spte += page_offset / sizeof(*spte);
1153                 pte = *spte;
1154                 if (is_present_pte(pte)) {
1155                         if (level == PT_PAGE_TABLE_LEVEL)
1156                                 rmap_remove(vcpu, spte);
1157                         else {
1158                                 child = page_header(pte & PT64_BASE_ADDR_MASK);
1159                                 mmu_page_remove_parent_pte(vcpu, child, spte);
1160                         }
1161                 }
1162                 *spte = 0;
1163         }
1164 }
1165
1166 void kvm_mmu_post_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1167 {
1168 }
1169
1170 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1171 {
1172         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1173
1174         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1175 }
1176
1177 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1178 {
1179         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1180                 struct kvm_mmu_page *page;
1181
1182                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1183                                     struct kvm_mmu_page, link);
1184                 kvm_mmu_zap_page(vcpu, page);
1185         }
1186 }
1187 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1188
1189 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1190 {
1191         struct kvm_mmu_page *page;
1192
1193         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1194                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1195                                     struct kvm_mmu_page, link);
1196                 kvm_mmu_zap_page(vcpu, page);
1197         }
1198         while (!list_empty(&vcpu->free_pages)) {
1199                 page = list_entry(vcpu->free_pages.next,
1200                                   struct kvm_mmu_page, link);
1201                 list_del(&page->link);
1202                 __free_page(pfn_to_page(page->page_hpa >> PAGE_SHIFT));
1203                 page->page_hpa = INVALID_PAGE;
1204         }
1205         free_page((unsigned long)vcpu->mmu.pae_root);
1206 }
1207
1208 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1209 {
1210         struct page *page;
1211         int i;
1212
1213         ASSERT(vcpu);
1214
1215         for (i = 0; i < KVM_NUM_MMU_PAGES; i++) {
1216                 struct kvm_mmu_page *page_header = &vcpu->page_header_buf[i];
1217
1218                 INIT_LIST_HEAD(&page_header->link);
1219                 if ((page = alloc_page(GFP_KERNEL)) == NULL)
1220                         goto error_1;
1221                 set_page_private(page, (unsigned long)page_header);
1222                 page_header->page_hpa = (hpa_t)page_to_pfn(page) << PAGE_SHIFT;
1223                 memset(__va(page_header->page_hpa), 0, PAGE_SIZE);
1224                 list_add(&page_header->link, &vcpu->free_pages);
1225                 ++vcpu->kvm->n_free_mmu_pages;
1226         }
1227
1228         /*
1229          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1230          * Therefore we need to allocate shadow page tables in the first
1231          * 4GB of memory, which happens to fit the DMA32 zone.
1232          */
1233         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1234         if (!page)
1235                 goto error_1;
1236         vcpu->mmu.pae_root = page_address(page);
1237         for (i = 0; i < 4; ++i)
1238                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1239
1240         return 0;
1241
1242 error_1:
1243         free_mmu_pages(vcpu);
1244         return -ENOMEM;
1245 }
1246
1247 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1248 {
1249         ASSERT(vcpu);
1250         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1251         ASSERT(list_empty(&vcpu->free_pages));
1252
1253         return alloc_mmu_pages(vcpu);
1254 }
1255
1256 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1257 {
1258         ASSERT(vcpu);
1259         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1260         ASSERT(!list_empty(&vcpu->free_pages));
1261
1262         return init_kvm_mmu(vcpu);
1263 }
1264
1265 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1266 {
1267         ASSERT(vcpu);
1268
1269         destroy_kvm_mmu(vcpu);
1270         free_mmu_pages(vcpu);
1271         mmu_free_memory_caches(vcpu);
1272 }
1273
1274 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1275 {
1276         struct kvm *kvm = vcpu->kvm;
1277         struct kvm_mmu_page *page;
1278
1279         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1280                 int i;
1281                 u64 *pt;
1282
1283                 if (!test_bit(slot, &page->slot_bitmap))
1284                         continue;
1285
1286                 pt = __va(page->page_hpa);
1287                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1288                         /* avoid RMW */
1289                         if (pt[i] & PT_WRITABLE_MASK) {
1290                                 rmap_remove(vcpu, &pt[i]);
1291                                 pt[i] &= ~PT_WRITABLE_MASK;
1292                         }
1293         }
1294 }
1295
1296 #ifdef AUDIT
1297
1298 static const char *audit_msg;
1299
1300 static gva_t canonicalize(gva_t gva)
1301 {
1302 #ifdef CONFIG_X86_64
1303         gva = (long long)(gva << 16) >> 16;
1304 #endif
1305         return gva;
1306 }
1307
1308 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1309                                 gva_t va, int level)
1310 {
1311         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1312         int i;
1313         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1314
1315         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1316                 u64 ent = pt[i];
1317
1318                 if (!ent & PT_PRESENT_MASK)
1319                         continue;
1320
1321                 va = canonicalize(va);
1322                 if (level > 1)
1323                         audit_mappings_page(vcpu, ent, va, level - 1);
1324                 else {
1325                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1326                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1327
1328                         if ((ent & PT_PRESENT_MASK)
1329                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1330                                 printk(KERN_ERR "audit error: (%s) levels %d"
1331                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1332                                        audit_msg, vcpu->mmu.root_level,
1333                                        va, gpa, hpa, ent);
1334                 }
1335         }
1336 }
1337
1338 static void audit_mappings(struct kvm_vcpu *vcpu)
1339 {
1340         int i;
1341
1342         if (vcpu->mmu.root_level == 4)
1343                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1344         else
1345                 for (i = 0; i < 4; ++i)
1346                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1347                                 audit_mappings_page(vcpu,
1348                                                     vcpu->mmu.pae_root[i],
1349                                                     i << 30,
1350                                                     2);
1351 }
1352
1353 static int count_rmaps(struct kvm_vcpu *vcpu)
1354 {
1355         int nmaps = 0;
1356         int i, j, k;
1357
1358         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1359                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1360                 struct kvm_rmap_desc *d;
1361
1362                 for (j = 0; j < m->npages; ++j) {
1363                         struct page *page = m->phys_mem[j];
1364
1365                         if (!page->private)
1366                                 continue;
1367                         if (!(page->private & 1)) {
1368                                 ++nmaps;
1369                                 continue;
1370                         }
1371                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1372                         while (d) {
1373                                 for (k = 0; k < RMAP_EXT; ++k)
1374                                         if (d->shadow_ptes[k])
1375                                                 ++nmaps;
1376                                         else
1377                                                 break;
1378                                 d = d->more;
1379                         }
1380                 }
1381         }
1382         return nmaps;
1383 }
1384
1385 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1386 {
1387         int nmaps = 0;
1388         struct kvm_mmu_page *page;
1389         int i;
1390
1391         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1392                 u64 *pt = __va(page->page_hpa);
1393
1394                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1395                         continue;
1396
1397                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1398                         u64 ent = pt[i];
1399
1400                         if (!(ent & PT_PRESENT_MASK))
1401                                 continue;
1402                         if (!(ent & PT_WRITABLE_MASK))
1403                                 continue;
1404                         ++nmaps;
1405                 }
1406         }
1407         return nmaps;
1408 }
1409
1410 static void audit_rmap(struct kvm_vcpu *vcpu)
1411 {
1412         int n_rmap = count_rmaps(vcpu);
1413         int n_actual = count_writable_mappings(vcpu);
1414
1415         if (n_rmap != n_actual)
1416                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1417                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1418 }
1419
1420 static void audit_write_protection(struct kvm_vcpu *vcpu)
1421 {
1422         struct kvm_mmu_page *page;
1423
1424         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1425                 hfn_t hfn;
1426                 struct page *pg;
1427
1428                 if (page->role.metaphysical)
1429                         continue;
1430
1431                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1432                         >> PAGE_SHIFT;
1433                 pg = pfn_to_page(hfn);
1434                 if (pg->private)
1435                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1436                                " mappings: gfn %lx role %x\n",
1437                                __FUNCTION__, audit_msg, page->gfn,
1438                                page->role.word);
1439         }
1440 }
1441
1442 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1443 {
1444         int olddbg = dbg;
1445
1446         dbg = 0;
1447         audit_msg = msg;
1448         audit_rmap(vcpu);
1449         audit_write_protection(vcpu);
1450         audit_mappings(vcpu);
1451         dbg = olddbg;
1452 }
1453
1454 #endif