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 #include <asm/cmpxchg.h>
  26
  27 #include "vmx.h"
  28 #include "kvm.h"
  29
  30 #undef MMU_DEBUG
  31
  32 #undef AUDIT
  33
  34 #ifdef AUDIT
  35 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  36 #else
  37 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  38 #endif
  39
  40 #ifdef MMU_DEBUG
  41
  42 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  43 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  44
  45 #else
  46
  47 #define pgprintk(x...) do { } while (0)
  48 #define rmap_printk(x...) do { } while (0)
  49
  50 #endif
  51
  52 #if defined(MMU_DEBUG) || defined(AUDIT)
  53 static int dbg = 1;
  54 #endif
  55
  56 #ifndef MMU_DEBUG
  57 #define ASSERT(x) do { } while (0)
  58 #else
  59 #define ASSERT(x)                                                       \
  60         if (!(x)) {                                                     \
  61                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  62                        __FILE__, __LINE__, #x);                         \
  63         }
  64 #endif
  65
  66 #define PT64_PT_BITS 9
  67 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  68 #define PT32_PT_BITS 10
  69 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  70
  71 #define PT_WRITABLE_SHIFT 1
  72
  73 #define PT_PRESENT_MASK (1ULL << 0)
  74 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  75 #define PT_USER_MASK (1ULL << 2)
  76 #define PT_PWT_MASK (1ULL << 3)
  77 #define PT_PCD_MASK (1ULL << 4)
  78 #define PT_ACCESSED_MASK (1ULL << 5)
  79 #define PT_DIRTY_MASK (1ULL << 6)
  80 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  81 #define PT_PAT_MASK (1ULL << 7)
  82 #define PT_GLOBAL_MASK (1ULL << 8)
  83 #define PT64_NX_MASK (1ULL << 63)
  84
  85 #define PT_PAT_SHIFT 7
  86 #define PT_DIR_PAT_SHIFT 12
  87 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  88
  89 #define PT32_DIR_PSE36_SIZE 4
  90 #define PT32_DIR_PSE36_SHIFT 13
  91 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  92
  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 VALID_PAGE(x) ((x) != INVALID_PAGE)
 101
 102 #define PT64_LEVEL_BITS 9
 103
 104 #define PT64_LEVEL_SHIFT(level) \
 105                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
 106
 107 #define PT64_LEVEL_MASK(level) \
 108                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 109
 110 #define PT64_INDEX(address, level)\
 111         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 112
 113
 114 #define PT32_LEVEL_BITS 10
 115
 116 #define PT32_LEVEL_SHIFT(level) \
 117                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
 118
 119 #define PT32_LEVEL_MASK(level) \
 120                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 121
 122 #define PT32_INDEX(address, level)\
 123         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 124
 125
 126 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 127 #define PT64_DIR_BASE_ADDR_MASK \
 128         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 129
 130 #define PT32_BASE_ADDR_MASK PAGE_MASK
 131 #define PT32_DIR_BASE_ADDR_MASK \
 132         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 133
 134
 135 #define PFERR_PRESENT_MASK (1U << 0)
 136 #define PFERR_WRITE_MASK (1U << 1)
 137 #define PFERR_USER_MASK (1U << 2)
 138 #define PFERR_FETCH_MASK (1U << 4)
 139
 140 #define PT64_ROOT_LEVEL 4
 141 #define PT32_ROOT_LEVEL 2
 142 #define PT32E_ROOT_LEVEL 3
 143
 144 #define PT_DIRECTORY_LEVEL 2
 145 #define PT_PAGE_TABLE_LEVEL 1
 146
 147 #define RMAP_EXT 4
 148
 149 struct kvm_rmap_desc {
 150         u64 *shadow_ptes[RMAP_EXT];
 151         struct kvm_rmap_desc *more;
 152 };
 153
 154 static struct kmem_cache *pte_chain_cache;
 155 static struct kmem_cache *rmap_desc_cache;
 156 static struct kmem_cache *mmu_page_cache;
 157 static struct kmem_cache *mmu_page_header_cache;
 158
 159 static int is_write_protection(struct kvm_vcpu *vcpu)
 160 {
 161         return vcpu->cr0 & CR0_WP_MASK;
 162 }
 163
 164 static int is_cpuid_PSE36(void)
 165 {
 166         return 1;
 167 }
 168
 169 static int is_nx(struct kvm_vcpu *vcpu)
 170 {
 171         return vcpu->shadow_efer & EFER_NX;
 172 }
 173
 174 static int is_present_pte(unsigned long pte)
 175 {
 176         return pte & PT_PRESENT_MASK;
 177 }
 178
 179 static int is_writeble_pte(unsigned long pte)
 180 {
 181         return pte & PT_WRITABLE_MASK;
 182 }
 183
 184 static int is_io_pte(unsigned long pte)
 185 {
 186         return pte & PT_SHADOW_IO_MARK;
 187 }
 188
 189 static int is_rmap_pte(u64 pte)
 190 {
 191         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 192                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 193 }
 194
 195 static void set_shadow_pte(u64 *sptep, u64 spte)
 196 {
 197 #ifdef CONFIG_X86_64
 198         set_64bit((unsigned long *)sptep, spte);
 199 #else
 200         set_64bit((unsigned long long *)sptep, spte);
 201 #endif
 202 }
 203
 204 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 205                                   struct kmem_cache *base_cache, int min,
 206                                   gfp_t gfp_flags)
 207 {
 208         void *obj;
 209
 210         if (cache->nobjs >= min)
 211                 return 0;
 212         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 213                 obj = kmem_cache_zalloc(base_cache, gfp_flags);
 214                 if (!obj)
 215                         return -ENOMEM;
 216                 cache->objects[cache->nobjs++] = obj;
 217         }
 218         return 0;
 219 }
 220
 221 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 222 {
 223         while (mc->nobjs)
 224                 kfree(mc->objects[--mc->nobjs]);
 225 }
 226
 227 static int __mmu_topup_memory_caches(struct kvm_vcpu *vcpu, gfp_t gfp_flags)
 228 {
 229         int r;
 230
 231         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 232                                    pte_chain_cache, 4, gfp_flags);
 233         if (r)
 234                 goto out;
 235         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 236                                    rmap_desc_cache, 1, gfp_flags);
 237         if (r)
 238                 goto out;
 239         r = mmu_topup_memory_cache(&vcpu->mmu_page_cache,
 240                                    mmu_page_cache, 4, gfp_flags);
 241         if (r)
 242                 goto out;
 243         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
 244                                    mmu_page_header_cache, 4, gfp_flags);
 245 out:
 246         return r;
 247 }
 248
 249 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 250 {
 251         int r;
 252
 253         r = __mmu_topup_memory_caches(vcpu, GFP_NOWAIT);
 254         if (r < 0) {
 255                 spin_unlock(&vcpu->kvm->lock);
 256                 kvm_arch_ops->vcpu_put(vcpu);
 257                 r = __mmu_topup_memory_caches(vcpu, GFP_KERNEL);
 258                 kvm_arch_ops->vcpu_load(vcpu);
 259                 spin_lock(&vcpu->kvm->lock);
 260         }
 261         return r;
 262 }
 263
 264 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 265 {
 266         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 267         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 268         mmu_free_memory_cache(&vcpu->mmu_page_cache);
 269         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
 270 }
 271
 272 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 273                                     size_t size)
 274 {
 275         void *p;
 276
 277         BUG_ON(!mc->nobjs);
 278         p = mc->objects[--mc->nobjs];
 279         memset(p, 0, size);
 280         return p;
 281 }
 282
 283 static void mmu_memory_cache_free(struct kvm_mmu_memory_cache *mc, void *obj)
 284 {
 285         if (mc->nobjs < KVM_NR_MEM_OBJS)
 286                 mc->objects[mc->nobjs++] = obj;
 287         else
 288                 kfree(obj);
 289 }
 290
 291 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 292 {
 293         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 294                                       sizeof(struct kvm_pte_chain));
 295 }
 296
 297 static void mmu_free_pte_chain(struct kvm_vcpu *vcpu,
 298                                struct kvm_pte_chain *pc)
 299 {
 300         mmu_memory_cache_free(&vcpu->mmu_pte_chain_cache, pc);
 301 }
 302
 303 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 304 {
 305         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 306                                       sizeof(struct kvm_rmap_desc));
 307 }
 308
 309 static void mmu_free_rmap_desc(struct kvm_vcpu *vcpu,
 310                                struct kvm_rmap_desc *rd)
 311 {
 312         mmu_memory_cache_free(&vcpu->mmu_rmap_desc_cache, rd);
 313 }
 314
 315 /*
 316  * Reverse mapping data structures:
 317  *
 318  * If page->private bit zero is zero, then page->private points to the
 319  * shadow page table entry that points to page_address(page).
 320  *
 321  * If page->private bit zero is one, (then page->private & ~1) points
 322  * to a struct kvm_rmap_desc containing more mappings.
 323  */
 324 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 325 {
 326         struct page *page;
 327         struct kvm_rmap_desc *desc;
 328         int i;
 329
 330         if (!is_rmap_pte(*spte))
 331                 return;
 332         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 333         if (!page_private(page)) {
 334                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 335                 set_page_private(page,(unsigned long)spte);
 336         } else if (!(page_private(page) & 1)) {
 337                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 338                 desc = mmu_alloc_rmap_desc(vcpu);
 339                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 340                 desc->shadow_ptes[1] = spte;
 341                 set_page_private(page,(unsigned long)desc | 1);
 342         } else {
 343                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 344                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 345                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 346                         desc = desc->more;
 347                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 348                         desc->more = mmu_alloc_rmap_desc(vcpu);
 349                         desc = desc->more;
 350                 }
 351                 for (i = 0; desc->shadow_ptes[i]; ++i)
 352                         ;
 353                 desc->shadow_ptes[i] = spte;
 354         }
 355 }
 356
 357 static void rmap_desc_remove_entry(struct kvm_vcpu *vcpu,
 358                                    struct page *page,
 359                                    struct kvm_rmap_desc *desc,
 360                                    int i,
 361                                    struct kvm_rmap_desc *prev_desc)
 362 {
 363         int j;
 364
 365         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 366                 ;
 367         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 368         desc->shadow_ptes[j] = NULL;
 369         if (j != 0)
 370                 return;
 371         if (!prev_desc && !desc->more)
 372                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 373         else
 374                 if (prev_desc)
 375                         prev_desc->more = desc->more;
 376                 else
 377                         set_page_private(page,(unsigned long)desc->more | 1);
 378         mmu_free_rmap_desc(vcpu, desc);
 379 }
 380
 381 static void rmap_remove(struct kvm_vcpu *vcpu, u64 *spte)
 382 {
 383         struct page *page;
 384         struct kvm_rmap_desc *desc;
 385         struct kvm_rmap_desc *prev_desc;
 386         int i;
 387
 388         if (!is_rmap_pte(*spte))
 389                 return;
 390         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 391         if (!page_private(page)) {
 392                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 393                 BUG();
 394         } else if (!(page_private(page) & 1)) {
 395                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 396                 if ((u64 *)page_private(page) != spte) {
 397                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 398                                spte, *spte);
 399                         BUG();
 400                 }
 401                 set_page_private(page,0);
 402         } else {
 403                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 404                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 405                 prev_desc = NULL;
 406                 while (desc) {
 407                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 408                                 if (desc->shadow_ptes[i] == spte) {
 409                                         rmap_desc_remove_entry(vcpu, page,
 410                                                                desc, i,
 411                                                                prev_desc);
 412                                         return;
 413                                 }
 414                         prev_desc = desc;
 415                         desc = desc->more;
 416                 }
 417                 BUG();
 418         }
 419 }
 420
 421 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 422 {
 423         struct kvm *kvm = vcpu->kvm;
 424         struct page *page;
 425         struct kvm_rmap_desc *desc;
 426         u64 *spte;
 427
 428         page = gfn_to_page(kvm, gfn);
 429         BUG_ON(!page);
 430
 431         while (page_private(page)) {
 432                 if (!(page_private(page) & 1))
 433                         spte = (u64 *)page_private(page);
 434                 else {
 435                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 436                         spte = desc->shadow_ptes[0];
 437                 }
 438                 BUG_ON(!spte);
 439                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 440                        != page_to_pfn(page));
 441                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 442                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 443                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 444                 rmap_remove(vcpu, spte);
 445                 kvm_arch_ops->tlb_flush(vcpu);
 446                 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 447         }
 448 }
 449
 450 #ifdef MMU_DEBUG
 451 static int is_empty_shadow_page(u64 *spt)
 452 {
 453         u64 *pos;
 454         u64 *end;
 455
 456         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 457                 if (*pos != 0) {
 458                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 459                                pos, *pos);
 460                         return 0;
 461                 }
 462         return 1;
 463 }
 464 #endif
 465
 466 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu,
 467                               struct kvm_mmu_page *page_head)
 468 {
 469         ASSERT(is_empty_shadow_page(page_head->spt));
 470         list_del(&page_head->link);
 471         mmu_memory_cache_free(&vcpu->mmu_page_cache, page_head->spt);
 472         mmu_memory_cache_free(&vcpu->mmu_page_header_cache, page_head);
 473         ++vcpu->kvm->n_free_mmu_pages;
 474 }
 475
 476 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 477 {
 478         return gfn;
 479 }
 480
 481 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 482                                                u64 *parent_pte)
 483 {
 484         struct kvm_mmu_page *page;
 485
 486         if (!vcpu->kvm->n_free_mmu_pages)
 487                 return NULL;
 488
 489         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 490                                       sizeof *page);
 491         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 492         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 493         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 494         ASSERT(is_empty_shadow_page(page->spt));
 495         page->slot_bitmap = 0;
 496         page->multimapped = 0;
 497         page->parent_pte = parent_pte;
 498         --vcpu->kvm->n_free_mmu_pages;
 499         return page;
 500 }
 501
 502 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 503                                     struct kvm_mmu_page *page, u64 *parent_pte)
 504 {
 505         struct kvm_pte_chain *pte_chain;
 506         struct hlist_node *node;
 507         int i;
 508
 509         if (!parent_pte)
 510                 return;
 511         if (!page->multimapped) {
 512                 u64 *old = page->parent_pte;
 513
 514                 if (!old) {
 515                         page->parent_pte = parent_pte;
 516                         return;
 517                 }
 518                 page->multimapped = 1;
 519                 pte_chain = mmu_alloc_pte_chain(vcpu);
 520                 INIT_HLIST_HEAD(&page->parent_ptes);
 521                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 522                 pte_chain->parent_ptes[0] = old;
 523         }
 524         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 525                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 526                         continue;
 527                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 528                         if (!pte_chain->parent_ptes[i]) {
 529                                 pte_chain->parent_ptes[i] = parent_pte;
 530                                 return;
 531                         }
 532         }
 533         pte_chain = mmu_alloc_pte_chain(vcpu);
 534         BUG_ON(!pte_chain);
 535         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 536         pte_chain->parent_ptes[0] = parent_pte;
 537 }
 538
 539 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
 540                                        struct kvm_mmu_page *page,
 541                                        u64 *parent_pte)
 542 {
 543         struct kvm_pte_chain *pte_chain;
 544         struct hlist_node *node;
 545         int i;
 546
 547         if (!page->multimapped) {
 548                 BUG_ON(page->parent_pte != parent_pte);
 549                 page->parent_pte = NULL;
 550                 return;
 551         }
 552         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 553                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 554                         if (!pte_chain->parent_ptes[i])
 555                                 break;
 556                         if (pte_chain->parent_ptes[i] != parent_pte)
 557                                 continue;
 558                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 559                                 && pte_chain->parent_ptes[i + 1]) {
 560                                 pte_chain->parent_ptes[i]
 561                                         = pte_chain->parent_ptes[i + 1];
 562                                 ++i;
 563                         }
 564                         pte_chain->parent_ptes[i] = NULL;
 565                         if (i == 0) {
 566                                 hlist_del(&pte_chain->link);
 567                                 mmu_free_pte_chain(vcpu, pte_chain);
 568                                 if (hlist_empty(&page->parent_ptes)) {
 569                                         page->multimapped = 0;
 570                                         page->parent_pte = NULL;
 571                                 }
 572                         }
 573                         return;
 574                 }
 575         BUG();
 576 }
 577
 578 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 579                                                 gfn_t gfn)
 580 {
 581         unsigned index;
 582         struct hlist_head *bucket;
 583         struct kvm_mmu_page *page;
 584         struct hlist_node *node;
 585
 586         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 587         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 588         bucket = &vcpu->kvm->mmu_page_hash[index];
 589         hlist_for_each_entry(page, node, bucket, hash_link)
 590                 if (page->gfn == gfn && !page->role.metaphysical) {
 591                         pgprintk("%s: found role %x\n",
 592                                  __FUNCTION__, page->role.word);
 593                         return page;
 594                 }
 595         return NULL;
 596 }
 597
 598 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 599                                              gfn_t gfn,
 600                                              gva_t gaddr,
 601                                              unsigned level,
 602                                              int metaphysical,
 603                                              unsigned hugepage_access,
 604                                              u64 *parent_pte)
 605 {
 606         union kvm_mmu_page_role role;
 607         unsigned index;
 608         unsigned quadrant;
 609         struct hlist_head *bucket;
 610         struct kvm_mmu_page *page;
 611         struct hlist_node *node;
 612
 613         role.word = 0;
 614         role.glevels = vcpu->mmu.root_level;
 615         role.level = level;
 616         role.metaphysical = metaphysical;
 617         role.hugepage_access = hugepage_access;
 618         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 619                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 620                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 621                 role.quadrant = quadrant;
 622         }
 623         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 624                  gfn, role.word);
 625         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 626         bucket = &vcpu->kvm->mmu_page_hash[index];
 627         hlist_for_each_entry(page, node, bucket, hash_link)
 628                 if (page->gfn == gfn && page->role.word == role.word) {
 629                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 630                         pgprintk("%s: found\n", __FUNCTION__);
 631                         return page;
 632                 }
 633         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 634         if (!page)
 635                 return page;
 636         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 637         page->gfn = gfn;
 638         page->role = role;
 639         hlist_add_head(&page->hash_link, bucket);
 640         if (!metaphysical)
 641                 rmap_write_protect(vcpu, gfn);
 642         return page;
 643 }
 644
 645 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 646                                          struct kvm_mmu_page *page)
 647 {
 648         unsigned i;
 649         u64 *pt;
 650         u64 ent;
 651
 652         pt = page->spt;
 653
 654         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 655                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 656                         if (pt[i] & PT_PRESENT_MASK)
 657                                 rmap_remove(vcpu, &pt[i]);
 658                         pt[i] = 0;
 659                 }
 660                 kvm_arch_ops->tlb_flush(vcpu);
 661                 return;
 662         }
 663
 664         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 665                 ent = pt[i];
 666
 667                 pt[i] = 0;
 668                 if (!(ent & PT_PRESENT_MASK))
 669                         continue;
 670                 ent &= PT64_BASE_ADDR_MASK;
 671                 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
 672         }
 673 }
 674
 675 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 676                              struct kvm_mmu_page *page,
 677                              u64 *parent_pte)
 678 {
 679         mmu_page_remove_parent_pte(vcpu, page, parent_pte);
 680 }
 681
 682 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 683                              struct kvm_mmu_page *page)
 684 {
 685         u64 *parent_pte;
 686
 687         while (page->multimapped || page->parent_pte) {
 688                 if (!page->multimapped)
 689                         parent_pte = page->parent_pte;
 690                 else {
 691                         struct kvm_pte_chain *chain;
 692
 693                         chain = container_of(page->parent_ptes.first,
 694                                              struct kvm_pte_chain, link);
 695                         parent_pte = chain->parent_ptes[0];
 696                 }
 697                 BUG_ON(!parent_pte);
 698                 kvm_mmu_put_page(vcpu, page, parent_pte);
 699                 set_shadow_pte(parent_pte, 0);
 700         }
 701         kvm_mmu_page_unlink_children(vcpu, page);
 702         if (!page->root_count) {
 703                 hlist_del(&page->hash_link);
 704                 kvm_mmu_free_page(vcpu, page);
 705         } else
 706                 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 707 }
 708
 709 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 710 {
 711         unsigned index;
 712         struct hlist_head *bucket;
 713         struct kvm_mmu_page *page;
 714         struct hlist_node *node, *n;
 715         int r;
 716
 717         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 718         r = 0;
 719         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 720         bucket = &vcpu->kvm->mmu_page_hash[index];
 721         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 722                 if (page->gfn == gfn && !page->role.metaphysical) {
 723                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 724                                  page->role.word);
 725                         kvm_mmu_zap_page(vcpu, page);
 726                         r = 1;
 727                 }
 728         return r;
 729 }
 730
 731 static void mmu_unshadow(struct kvm_vcpu *vcpu, gfn_t gfn)
 732 {
 733         struct kvm_mmu_page *page;
 734
 735         while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
 736                 pgprintk("%s: zap %lx %x\n",
 737                          __FUNCTION__, gfn, page->role.word);
 738                 kvm_mmu_zap_page(vcpu, page);
 739         }
 740 }
 741
 742 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 743 {
 744         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 745         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 746
 747         __set_bit(slot, &page_head->slot_bitmap);
 748 }
 749
 750 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 751 {
 752         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 753
 754         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 755 }
 756
 757 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 758 {
 759         struct page *page;
 760
 761         ASSERT((gpa & HPA_ERR_MASK) == 0);
 762         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 763         if (!page)
 764                 return gpa | HPA_ERR_MASK;
 765         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 766                 | (gpa & (PAGE_SIZE-1));
 767 }
 768
 769 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 770 {
 771         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 772
 773         if (gpa == UNMAPPED_GVA)
 774                 return UNMAPPED_GVA;
 775         return gpa_to_hpa(vcpu, gpa);
 776 }
 777
 778 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 779 {
 780         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 781
 782         if (gpa == UNMAPPED_GVA)
 783                 return NULL;
 784         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 785 }
 786
 787 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 788 {
 789 }
 790
 791 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 792 {
 793         int level = PT32E_ROOT_LEVEL;
 794         hpa_t table_addr = vcpu->mmu.root_hpa;
 795
 796         for (; ; level--) {
 797                 u32 index = PT64_INDEX(v, level);
 798                 u64 *table;
 799                 u64 pte;
 800
 801                 ASSERT(VALID_PAGE(table_addr));
 802                 table = __va(table_addr);
 803
 804                 if (level == 1) {
 805                         pte = table[index];
 806                         if (is_present_pte(pte) && is_writeble_pte(pte))
 807                                 return 0;
 808                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 809                         page_header_update_slot(vcpu->kvm, table, v);
 810                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 811                                                                 PT_USER_MASK;
 812                         rmap_add(vcpu, &table[index]);
 813                         return 0;
 814                 }
 815
 816                 if (table[index] == 0) {
 817                         struct kvm_mmu_page *new_table;
 818                         gfn_t pseudo_gfn;
 819
 820                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 821                                 >> PAGE_SHIFT;
 822                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 823                                                      v, level - 1,
 824                                                      1, 0, &table[index]);
 825                         if (!new_table) {
 826                                 pgprintk("nonpaging_map: ENOMEM\n");
 827                                 return -ENOMEM;
 828                         }
 829
 830                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 831                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 832                 }
 833                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 834         }
 835 }
 836
 837 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 838 {
 839         int i;
 840         struct kvm_mmu_page *page;
 841
 842 #ifdef CONFIG_X86_64
 843         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 844                 hpa_t root = vcpu->mmu.root_hpa;
 845
 846                 ASSERT(VALID_PAGE(root));
 847                 page = page_header(root);
 848                 --page->root_count;
 849                 vcpu->mmu.root_hpa = INVALID_PAGE;
 850                 return;
 851         }
 852 #endif
 853         for (i = 0; i < 4; ++i) {
 854                 hpa_t root = vcpu->mmu.pae_root[i];
 855
 856                 if (root) {
 857                         ASSERT(VALID_PAGE(root));
 858                         root &= PT64_BASE_ADDR_MASK;
 859                         page = page_header(root);
 860                         --page->root_count;
 861                 }
 862                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 863         }
 864         vcpu->mmu.root_hpa = INVALID_PAGE;
 865 }
 866
 867 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 868 {
 869         int i;
 870         gfn_t root_gfn;
 871         struct kvm_mmu_page *page;
 872
 873         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 874
 875 #ifdef CONFIG_X86_64
 876         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 877                 hpa_t root = vcpu->mmu.root_hpa;
 878
 879                 ASSERT(!VALID_PAGE(root));
 880                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 881                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 882                 root = __pa(page->spt);
 883                 ++page->root_count;
 884                 vcpu->mmu.root_hpa = root;
 885                 return;
 886         }
 887 #endif
 888         for (i = 0; i < 4; ++i) {
 889                 hpa_t root = vcpu->mmu.pae_root[i];
 890
 891                 ASSERT(!VALID_PAGE(root));
 892                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 893                         if (!is_present_pte(vcpu->pdptrs[i])) {
 894                                 vcpu->mmu.pae_root[i] = 0;
 895                                 continue;
 896                         }
 897                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 898                 } else if (vcpu->mmu.root_level == 0)
 899                         root_gfn = 0;
 900                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 901                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 902                                         0, NULL);
 903                 root = __pa(page->spt);
 904                 ++page->root_count;
 905                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 906         }
 907         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 908 }
 909
 910 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 911 {
 912         return vaddr;
 913 }
 914
 915 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 916                                u32 error_code)
 917 {
 918         gpa_t addr = gva;
 919         hpa_t paddr;
 920         int r;
 921
 922         r = mmu_topup_memory_caches(vcpu);
 923         if (r)
 924                 return r;
 925
 926         ASSERT(vcpu);
 927         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 928
 929
 930         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 931
 932         if (is_error_hpa(paddr))
 933                 return 1;
 934
 935         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 936 }
 937
 938 static void nonpaging_free(struct kvm_vcpu *vcpu)
 939 {
 940         mmu_free_roots(vcpu);
 941 }
 942
 943 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 944 {
 945         struct kvm_mmu *context = &vcpu->mmu;
 946
 947         context->new_cr3 = nonpaging_new_cr3;
 948         context->page_fault = nonpaging_page_fault;
 949         context->gva_to_gpa = nonpaging_gva_to_gpa;
 950         context->free = nonpaging_free;
 951         context->root_level = 0;
 952         context->shadow_root_level = PT32E_ROOT_LEVEL;
 953         mmu_alloc_roots(vcpu);
 954         ASSERT(VALID_PAGE(context->root_hpa));
 955         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 956         return 0;
 957 }
 958
 959 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 960 {
 961         ++vcpu->stat.tlb_flush;
 962         kvm_arch_ops->tlb_flush(vcpu);
 963 }
 964
 965 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 966 {
 967         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 968         mmu_free_roots(vcpu);
 969         if (unlikely(vcpu->kvm->n_free_mmu_pages < KVM_MIN_FREE_MMU_PAGES))
 970                 kvm_mmu_free_some_pages(vcpu);
 971         mmu_alloc_roots(vcpu);
 972         kvm_mmu_flush_tlb(vcpu);
 973         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 974 }
 975
 976 static void inject_page_fault(struct kvm_vcpu *vcpu,
 977                               u64 addr,
 978                               u32 err_code)
 979 {
 980         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
 981 }
 982
 983 static void paging_free(struct kvm_vcpu *vcpu)
 984 {
 985         nonpaging_free(vcpu);
 986 }
 987
 988 #define PTTYPE 64
 989 #include "paging_tmpl.h"
 990 #undef PTTYPE
 991
 992 #define PTTYPE 32
 993 #include "paging_tmpl.h"
 994 #undef PTTYPE
 995
 996 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
 997 {
 998         struct kvm_mmu *context = &vcpu->mmu;
 999
1000         ASSERT(is_pae(vcpu));
1001         context->new_cr3 = paging_new_cr3;
1002         context->page_fault = paging64_page_fault;
1003         context->gva_to_gpa = paging64_gva_to_gpa;
1004         context->free = paging_free;
1005         context->root_level = level;
1006         context->shadow_root_level = level;
1007         mmu_alloc_roots(vcpu);
1008         ASSERT(VALID_PAGE(context->root_hpa));
1009         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1010                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1011         return 0;
1012 }
1013
1014 static int paging64_init_context(struct kvm_vcpu *vcpu)
1015 {
1016         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1017 }
1018
1019 static int paging32_init_context(struct kvm_vcpu *vcpu)
1020 {
1021         struct kvm_mmu *context = &vcpu->mmu;
1022
1023         context->new_cr3 = paging_new_cr3;
1024         context->page_fault = paging32_page_fault;
1025         context->gva_to_gpa = paging32_gva_to_gpa;
1026         context->free = paging_free;
1027         context->root_level = PT32_ROOT_LEVEL;
1028         context->shadow_root_level = PT32E_ROOT_LEVEL;
1029         mmu_alloc_roots(vcpu);
1030         ASSERT(VALID_PAGE(context->root_hpa));
1031         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1032                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1033         return 0;
1034 }
1035
1036 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1037 {
1038         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1039 }
1040
1041 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1042 {
1043         ASSERT(vcpu);
1044         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1045
1046         mmu_topup_memory_caches(vcpu);
1047         if (!is_paging(vcpu))
1048                 return nonpaging_init_context(vcpu);
1049         else if (is_long_mode(vcpu))
1050                 return paging64_init_context(vcpu);
1051         else if (is_pae(vcpu))
1052                 return paging32E_init_context(vcpu);
1053         else
1054                 return paging32_init_context(vcpu);
1055 }
1056
1057 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1058 {
1059         ASSERT(vcpu);
1060         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1061                 vcpu->mmu.free(vcpu);
1062                 vcpu->mmu.root_hpa = INVALID_PAGE;
1063         }
1064 }
1065
1066 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1067 {
1068         int r;
1069
1070         destroy_kvm_mmu(vcpu);
1071         r = init_kvm_mmu(vcpu);
1072         if (r < 0)
1073                 goto out;
1074         r = mmu_topup_memory_caches(vcpu);
1075 out:
1076         return r;
1077 }
1078
1079 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1080                                   struct kvm_mmu_page *page,
1081                                   u64 *spte)
1082 {
1083         u64 pte;
1084         struct kvm_mmu_page *child;
1085
1086         pte = *spte;
1087         if (is_present_pte(pte)) {
1088                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1089                         rmap_remove(vcpu, spte);
1090                 else {
1091                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1092                         mmu_page_remove_parent_pte(vcpu, child, spte);
1093                 }
1094         }
1095         *spte = 0;
1096 }
1097
1098 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1099                                   struct kvm_mmu_page *page,
1100                                   u64 *spte,
1101                                   const void *new, int bytes)
1102 {
1103         if (page->role.level != PT_PAGE_TABLE_LEVEL)
1104                 return;
1105
1106         if (page->role.glevels == PT32_ROOT_LEVEL)
1107                 paging32_update_pte(vcpu, page, spte, new, bytes);
1108         else
1109                 paging64_update_pte(vcpu, page, spte, new, bytes);
1110 }
1111
1112 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1113                        const u8 *old, const u8 *new, int bytes)
1114 {
1115         gfn_t gfn = gpa >> PAGE_SHIFT;
1116         struct kvm_mmu_page *page;
1117         struct hlist_node *node, *n;
1118         struct hlist_head *bucket;
1119         unsigned index;
1120         u64 *spte;
1121         unsigned offset = offset_in_page(gpa);
1122         unsigned pte_size;
1123         unsigned page_offset;
1124         unsigned misaligned;
1125         unsigned quadrant;
1126         int level;
1127         int flooded = 0;
1128         int npte;
1129
1130         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1131         if (gfn == vcpu->last_pt_write_gfn) {
1132                 ++vcpu->last_pt_write_count;
1133                 if (vcpu->last_pt_write_count >= 3)
1134                         flooded = 1;
1135         } else {
1136                 vcpu->last_pt_write_gfn = gfn;
1137                 vcpu->last_pt_write_count = 1;
1138         }
1139         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1140         bucket = &vcpu->kvm->mmu_page_hash[index];
1141         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1142                 if (page->gfn != gfn || page->role.metaphysical)
1143                         continue;
1144                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1145                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1146                 misaligned |= bytes < 4;
1147                 if (misaligned || flooded) {
1148                         /*
1149                          * Misaligned accesses are too much trouble to fix
1150                          * up; also, they usually indicate a page is not used
1151                          * as a page table.
1152                          *
1153                          * If we're seeing too many writes to a page,
1154                          * it may no longer be a page table, or we may be
1155                          * forking, in which case it is better to unmap the
1156                          * page.
1157                          */
1158                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1159                                  gpa, bytes, page->role.word);
1160                         kvm_mmu_zap_page(vcpu, page);
1161                         continue;
1162                 }
1163                 page_offset = offset;
1164                 level = page->role.level;
1165                 npte = 1;
1166                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1167                         page_offset <<= 1;      /* 32->64 */
1168                         /*
1169                          * A 32-bit pde maps 4MB while the shadow pdes map
1170                          * only 2MB.  So we need to double the offset again
1171                          * and zap two pdes instead of one.
1172                          */
1173                         if (level == PT32_ROOT_LEVEL) {
1174                                 page_offset &= ~7; /* kill rounding error */
1175                                 page_offset <<= 1;
1176                                 npte = 2;
1177                         }
1178                         quadrant = page_offset >> PAGE_SHIFT;
1179                         page_offset &= ~PAGE_MASK;
1180                         if (quadrant != page->role.quadrant)
1181                                 continue;
1182                 }
1183                 spte = &page->spt[page_offset / sizeof(*spte)];
1184                 while (npte--) {
1185                         mmu_pte_write_zap_pte(vcpu, page, spte);
1186                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes);
1187                         ++spte;
1188                 }
1189         }
1190 }
1191
1192 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1193 {
1194         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1195
1196         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1197 }
1198
1199 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1200 {
1201         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1202                 struct kvm_mmu_page *page;
1203
1204                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1205                                     struct kvm_mmu_page, link);
1206                 kvm_mmu_zap_page(vcpu, page);
1207         }
1208 }
1209 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1210
1211 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1212 {
1213         struct kvm_mmu_page *page;
1214
1215         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1216                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1217                                     struct kvm_mmu_page, link);
1218                 kvm_mmu_zap_page(vcpu, page);
1219         }
1220         free_page((unsigned long)vcpu->mmu.pae_root);
1221 }
1222
1223 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1224 {
1225         struct page *page;
1226         int i;
1227
1228         ASSERT(vcpu);
1229
1230         vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1231
1232         /*
1233          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1234          * Therefore we need to allocate shadow page tables in the first
1235          * 4GB of memory, which happens to fit the DMA32 zone.
1236          */
1237         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1238         if (!page)
1239                 goto error_1;
1240         vcpu->mmu.pae_root = page_address(page);
1241         for (i = 0; i < 4; ++i)
1242                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1243
1244         return 0;
1245
1246 error_1:
1247         free_mmu_pages(vcpu);
1248         return -ENOMEM;
1249 }
1250
1251 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1252 {
1253         ASSERT(vcpu);
1254         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1255
1256         return alloc_mmu_pages(vcpu);
1257 }
1258
1259 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1260 {
1261         ASSERT(vcpu);
1262         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1263
1264         return init_kvm_mmu(vcpu);
1265 }
1266
1267 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1268 {
1269         ASSERT(vcpu);
1270
1271         destroy_kvm_mmu(vcpu);
1272         free_mmu_pages(vcpu);
1273         mmu_free_memory_caches(vcpu);
1274 }
1275
1276 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1277 {
1278         struct kvm *kvm = vcpu->kvm;
1279         struct kvm_mmu_page *page;
1280
1281         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1282                 int i;
1283                 u64 *pt;
1284
1285                 if (!test_bit(slot, &page->slot_bitmap))
1286                         continue;
1287
1288                 pt = page->spt;
1289                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1290                         /* avoid RMW */
1291                         if (pt[i] & PT_WRITABLE_MASK) {
1292                                 rmap_remove(vcpu, &pt[i]);
1293                                 pt[i] &= ~PT_WRITABLE_MASK;
1294                         }
1295         }
1296 }
1297
1298 void kvm_mmu_zap_all(struct kvm_vcpu *vcpu)
1299 {
1300         destroy_kvm_mmu(vcpu);
1301
1302         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1303                 struct kvm_mmu_page *page;
1304
1305                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1306                                     struct kvm_mmu_page, link);
1307                 kvm_mmu_zap_page(vcpu, page);
1308         }
1309
1310         mmu_free_memory_caches(vcpu);
1311         kvm_arch_ops->tlb_flush(vcpu);
1312         init_kvm_mmu(vcpu);
1313 }
1314
1315 void kvm_mmu_module_exit(void)
1316 {
1317         if (pte_chain_cache)
1318                 kmem_cache_destroy(pte_chain_cache);
1319         if (rmap_desc_cache)
1320                 kmem_cache_destroy(rmap_desc_cache);
1321         if (mmu_page_cache)
1322                 kmem_cache_destroy(mmu_page_cache);
1323         if (mmu_page_header_cache)
1324                 kmem_cache_destroy(mmu_page_header_cache);
1325 }
1326
1327 int kvm_mmu_module_init(void)
1328 {
1329         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1330                                             sizeof(struct kvm_pte_chain),
1331                                             0, 0, NULL, NULL);
1332         if (!pte_chain_cache)
1333                 goto nomem;
1334         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1335                                             sizeof(struct kvm_rmap_desc),
1336                                             0, 0, NULL, NULL);
1337         if (!rmap_desc_cache)
1338                 goto nomem;
1339
1340         mmu_page_cache = kmem_cache_create("kvm_mmu_page",
1341                                            PAGE_SIZE,
1342                                            PAGE_SIZE, 0, NULL, NULL);
1343         if (!mmu_page_cache)
1344                 goto nomem;
1345
1346         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1347                                                   sizeof(struct kvm_mmu_page),
1348                                                   0, 0, NULL, NULL);
1349         if (!mmu_page_header_cache)
1350                 goto nomem;
1351
1352         return 0;
1353
1354 nomem:
1355         kvm_mmu_module_exit();
1356         return -ENOMEM;
1357 }
1358
1359 #ifdef AUDIT
1360
1361 static const char *audit_msg;
1362
1363 static gva_t canonicalize(gva_t gva)
1364 {
1365 #ifdef CONFIG_X86_64
1366         gva = (long long)(gva << 16) >> 16;
1367 #endif
1368         return gva;
1369 }
1370
1371 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1372                                 gva_t va, int level)
1373 {
1374         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1375         int i;
1376         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1377
1378         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1379                 u64 ent = pt[i];
1380
1381                 if (!(ent & PT_PRESENT_MASK))
1382                         continue;
1383
1384                 va = canonicalize(va);
1385                 if (level > 1)
1386                         audit_mappings_page(vcpu, ent, va, level - 1);
1387                 else {
1388                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1389                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1390
1391                         if ((ent & PT_PRESENT_MASK)
1392                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1393                                 printk(KERN_ERR "audit error: (%s) levels %d"
1394                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1395                                        audit_msg, vcpu->mmu.root_level,
1396                                        va, gpa, hpa, ent);
1397                 }
1398         }
1399 }
1400
1401 static void audit_mappings(struct kvm_vcpu *vcpu)
1402 {
1403         unsigned i;
1404
1405         if (vcpu->mmu.root_level == 4)
1406                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1407         else
1408                 for (i = 0; i < 4; ++i)
1409                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1410                                 audit_mappings_page(vcpu,
1411                                                     vcpu->mmu.pae_root[i],
1412                                                     i << 30,
1413                                                     2);
1414 }
1415
1416 static int count_rmaps(struct kvm_vcpu *vcpu)
1417 {
1418         int nmaps = 0;
1419         int i, j, k;
1420
1421         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1422                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1423                 struct kvm_rmap_desc *d;
1424
1425                 for (j = 0; j < m->npages; ++j) {
1426                         struct page *page = m->phys_mem[j];
1427
1428                         if (!page->private)
1429                                 continue;
1430                         if (!(page->private & 1)) {
1431                                 ++nmaps;
1432                                 continue;
1433                         }
1434                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1435                         while (d) {
1436                                 for (k = 0; k < RMAP_EXT; ++k)
1437                                         if (d->shadow_ptes[k])
1438                                                 ++nmaps;
1439                                         else
1440                                                 break;
1441                                 d = d->more;
1442                         }
1443                 }
1444         }
1445         return nmaps;
1446 }
1447
1448 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1449 {
1450         int nmaps = 0;
1451         struct kvm_mmu_page *page;
1452         int i;
1453
1454         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1455                 u64 *pt = page->spt;
1456
1457                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1458                         continue;
1459
1460                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1461                         u64 ent = pt[i];
1462
1463                         if (!(ent & PT_PRESENT_MASK))
1464                                 continue;
1465                         if (!(ent & PT_WRITABLE_MASK))
1466                                 continue;
1467                         ++nmaps;
1468                 }
1469         }
1470         return nmaps;
1471 }
1472
1473 static void audit_rmap(struct kvm_vcpu *vcpu)
1474 {
1475         int n_rmap = count_rmaps(vcpu);
1476         int n_actual = count_writable_mappings(vcpu);
1477
1478         if (n_rmap != n_actual)
1479                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1480                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1481 }
1482
1483 static void audit_write_protection(struct kvm_vcpu *vcpu)
1484 {
1485         struct kvm_mmu_page *page;
1486
1487         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1488                 hfn_t hfn;
1489                 struct page *pg;
1490
1491                 if (page->role.metaphysical)
1492                         continue;
1493
1494                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1495                         >> PAGE_SHIFT;
1496                 pg = pfn_to_page(hfn);
1497                 if (pg->private)
1498                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1499                                " mappings: gfn %lx role %x\n",
1500                                __FUNCTION__, audit_msg, page->gfn,
1501                                page->role.word);
1502         }
1503 }
1504
1505 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1506 {
1507         int olddbg = dbg;
1508
1509         dbg = 0;
1510         audit_msg = msg;
1511         audit_rmap(vcpu);
1512         audit_write_protection(vcpu);
1513         audit_mappings(vcpu);
1514         dbg = olddbg;
1515 }
1516
1517 #endif