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_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
  98
  99 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 100
 101 #define PT64_LEVEL_BITS 9
 102
 103 #define PT64_LEVEL_SHIFT(level) \
 104                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
 105
 106 #define PT64_LEVEL_MASK(level) \
 107                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 108
 109 #define PT64_INDEX(address, level)\
 110         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 111
 112
 113 #define PT32_LEVEL_BITS 10
 114
 115 #define PT32_LEVEL_SHIFT(level) \
 116                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
 117
 118 #define PT32_LEVEL_MASK(level) \
 119                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 120
 121 #define PT32_INDEX(address, level)\
 122         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 123
 124
 125 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 126 #define PT64_DIR_BASE_ADDR_MASK \
 127         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 128
 129 #define PT32_BASE_ADDR_MASK PAGE_MASK
 130 #define PT32_DIR_BASE_ADDR_MASK \
 131         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 132
 133
 134 #define PFERR_PRESENT_MASK (1U << 0)
 135 #define PFERR_WRITE_MASK (1U << 1)
 136 #define PFERR_USER_MASK (1U << 2)
 137 #define PFERR_FETCH_MASK (1U << 4)
 138
 139 #define PT64_ROOT_LEVEL 4
 140 #define PT32_ROOT_LEVEL 2
 141 #define PT32E_ROOT_LEVEL 3
 142
 143 #define PT_DIRECTORY_LEVEL 2
 144 #define PT_PAGE_TABLE_LEVEL 1
 145
 146 #define RMAP_EXT 4
 147
 148 struct kvm_rmap_desc {
 149         u64 *shadow_ptes[RMAP_EXT];
 150         struct kvm_rmap_desc *more;
 151 };
 152
 153 static struct kmem_cache *pte_chain_cache;
 154 static struct kmem_cache *rmap_desc_cache;
 155 static struct kmem_cache *mmu_page_cache;
 156 static struct kmem_cache *mmu_page_header_cache;
 157
 158 static int is_write_protection(struct kvm_vcpu *vcpu)
 159 {
 160         return vcpu->cr0 & CR0_WP_MASK;
 161 }
 162
 163 static int is_cpuid_PSE36(void)
 164 {
 165         return 1;
 166 }
 167
 168 static int is_nx(struct kvm_vcpu *vcpu)
 169 {
 170         return vcpu->shadow_efer & EFER_NX;
 171 }
 172
 173 static int is_present_pte(unsigned long pte)
 174 {
 175         return pte & PT_PRESENT_MASK;
 176 }
 177
 178 static int is_writeble_pte(unsigned long pte)
 179 {
 180         return pte & PT_WRITABLE_MASK;
 181 }
 182
 183 static int is_io_pte(unsigned long pte)
 184 {
 185         return pte & PT_SHADOW_IO_MARK;
 186 }
 187
 188 static int is_rmap_pte(u64 pte)
 189 {
 190         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 191                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 192 }
 193
 194 static void set_shadow_pte(u64 *sptep, u64 spte)
 195 {
 196 #ifdef CONFIG_X86_64
 197         set_64bit((unsigned long *)sptep, spte);
 198 #else
 199         set_64bit((unsigned long long *)sptep, spte);
 200 #endif
 201 }
 202
 203 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 204                                   struct kmem_cache *base_cache, int min,
 205                                   gfp_t gfp_flags)
 206 {
 207         void *obj;
 208
 209         if (cache->nobjs >= min)
 210                 return 0;
 211         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 212                 obj = kmem_cache_zalloc(base_cache, gfp_flags);
 213                 if (!obj)
 214                         return -ENOMEM;
 215                 cache->objects[cache->nobjs++] = obj;
 216         }
 217         return 0;
 218 }
 219
 220 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 221 {
 222         while (mc->nobjs)
 223                 kfree(mc->objects[--mc->nobjs]);
 224 }
 225
 226 static int __mmu_topup_memory_caches(struct kvm_vcpu *vcpu, gfp_t gfp_flags)
 227 {
 228         int r;
 229
 230         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 231                                    pte_chain_cache, 4, gfp_flags);
 232         if (r)
 233                 goto out;
 234         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 235                                    rmap_desc_cache, 1, gfp_flags);
 236         if (r)
 237                 goto out;
 238         r = mmu_topup_memory_cache(&vcpu->mmu_page_cache,
 239                                    mmu_page_cache, 4, gfp_flags);
 240         if (r)
 241                 goto out;
 242         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
 243                                    mmu_page_header_cache, 4, gfp_flags);
 244 out:
 245         return r;
 246 }
 247
 248 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 249 {
 250         int r;
 251
 252         r = __mmu_topup_memory_caches(vcpu, GFP_NOWAIT);
 253         if (r < 0) {
 254                 spin_unlock(&vcpu->kvm->lock);
 255                 kvm_arch_ops->vcpu_put(vcpu);
 256                 r = __mmu_topup_memory_caches(vcpu, GFP_KERNEL);
 257                 kvm_arch_ops->vcpu_load(vcpu);
 258                 spin_lock(&vcpu->kvm->lock);
 259         }
 260         return r;
 261 }
 262
 263 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 264 {
 265         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 266         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 267         mmu_free_memory_cache(&vcpu->mmu_page_cache);
 268         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
 269 }
 270
 271 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 272                                     size_t size)
 273 {
 274         void *p;
 275
 276         BUG_ON(!mc->nobjs);
 277         p = mc->objects[--mc->nobjs];
 278         memset(p, 0, size);
 279         return p;
 280 }
 281
 282 static void mmu_memory_cache_free(struct kvm_mmu_memory_cache *mc, void *obj)
 283 {
 284         if (mc->nobjs < KVM_NR_MEM_OBJS)
 285                 mc->objects[mc->nobjs++] = obj;
 286         else
 287                 kfree(obj);
 288 }
 289
 290 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 291 {
 292         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 293                                       sizeof(struct kvm_pte_chain));
 294 }
 295
 296 static void mmu_free_pte_chain(struct kvm_vcpu *vcpu,
 297                                struct kvm_pte_chain *pc)
 298 {
 299         mmu_memory_cache_free(&vcpu->mmu_pte_chain_cache, pc);
 300 }
 301
 302 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 303 {
 304         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 305                                       sizeof(struct kvm_rmap_desc));
 306 }
 307
 308 static void mmu_free_rmap_desc(struct kvm_vcpu *vcpu,
 309                                struct kvm_rmap_desc *rd)
 310 {
 311         mmu_memory_cache_free(&vcpu->mmu_rmap_desc_cache, rd);
 312 }
 313
 314 /*
 315  * Reverse mapping data structures:
 316  *
 317  * If page->private bit zero is zero, then page->private points to the
 318  * shadow page table entry that points to page_address(page).
 319  *
 320  * If page->private bit zero is one, (then page->private & ~1) points
 321  * to a struct kvm_rmap_desc containing more mappings.
 322  */
 323 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 324 {
 325         struct page *page;
 326         struct kvm_rmap_desc *desc;
 327         int i;
 328
 329         if (!is_rmap_pte(*spte))
 330                 return;
 331         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 332         if (!page_private(page)) {
 333                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 334                 set_page_private(page,(unsigned long)spte);
 335         } else if (!(page_private(page) & 1)) {
 336                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 337                 desc = mmu_alloc_rmap_desc(vcpu);
 338                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 339                 desc->shadow_ptes[1] = spte;
 340                 set_page_private(page,(unsigned long)desc | 1);
 341         } else {
 342                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 343                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 344                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 345                         desc = desc->more;
 346                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 347                         desc->more = mmu_alloc_rmap_desc(vcpu);
 348                         desc = desc->more;
 349                 }
 350                 for (i = 0; desc->shadow_ptes[i]; ++i)
 351                         ;
 352                 desc->shadow_ptes[i] = spte;
 353         }
 354 }
 355
 356 static void rmap_desc_remove_entry(struct kvm_vcpu *vcpu,
 357                                    struct page *page,
 358                                    struct kvm_rmap_desc *desc,
 359                                    int i,
 360                                    struct kvm_rmap_desc *prev_desc)
 361 {
 362         int j;
 363
 364         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 365                 ;
 366         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 367         desc->shadow_ptes[j] = NULL;
 368         if (j != 0)
 369                 return;
 370         if (!prev_desc && !desc->more)
 371                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 372         else
 373                 if (prev_desc)
 374                         prev_desc->more = desc->more;
 375                 else
 376                         set_page_private(page,(unsigned long)desc->more | 1);
 377         mmu_free_rmap_desc(vcpu, desc);
 378 }
 379
 380 static void rmap_remove(struct kvm_vcpu *vcpu, u64 *spte)
 381 {
 382         struct page *page;
 383         struct kvm_rmap_desc *desc;
 384         struct kvm_rmap_desc *prev_desc;
 385         int i;
 386
 387         if (!is_rmap_pte(*spte))
 388                 return;
 389         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 390         if (!page_private(page)) {
 391                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 392                 BUG();
 393         } else if (!(page_private(page) & 1)) {
 394                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 395                 if ((u64 *)page_private(page) != spte) {
 396                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 397                                spte, *spte);
 398                         BUG();
 399                 }
 400                 set_page_private(page,0);
 401         } else {
 402                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 403                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 404                 prev_desc = NULL;
 405                 while (desc) {
 406                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 407                                 if (desc->shadow_ptes[i] == spte) {
 408                                         rmap_desc_remove_entry(vcpu, page,
 409                                                                desc, i,
 410                                                                prev_desc);
 411                                         return;
 412                                 }
 413                         prev_desc = desc;
 414                         desc = desc->more;
 415                 }
 416                 BUG();
 417         }
 418 }
 419
 420 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 421 {
 422         struct kvm *kvm = vcpu->kvm;
 423         struct page *page;
 424         struct kvm_rmap_desc *desc;
 425         u64 *spte;
 426
 427         page = gfn_to_page(kvm, gfn);
 428         BUG_ON(!page);
 429
 430         while (page_private(page)) {
 431                 if (!(page_private(page) & 1))
 432                         spte = (u64 *)page_private(page);
 433                 else {
 434                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 435                         spte = desc->shadow_ptes[0];
 436                 }
 437                 BUG_ON(!spte);
 438                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 439                        != page_to_pfn(page));
 440                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 441                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 442                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 443                 rmap_remove(vcpu, spte);
 444                 kvm_arch_ops->tlb_flush(vcpu);
 445                 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 446         }
 447 }
 448
 449 #ifdef MMU_DEBUG
 450 static int is_empty_shadow_page(u64 *spt)
 451 {
 452         u64 *pos;
 453         u64 *end;
 454
 455         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 456                 if (*pos != 0) {
 457                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 458                                pos, *pos);
 459                         return 0;
 460                 }
 461         return 1;
 462 }
 463 #endif
 464
 465 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu,
 466                               struct kvm_mmu_page *page_head)
 467 {
 468         ASSERT(is_empty_shadow_page(page_head->spt));
 469         list_del(&page_head->link);
 470         mmu_memory_cache_free(&vcpu->mmu_page_cache, page_head->spt);
 471         mmu_memory_cache_free(&vcpu->mmu_page_header_cache, page_head);
 472         ++vcpu->kvm->n_free_mmu_pages;
 473 }
 474
 475 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 476 {
 477         return gfn;
 478 }
 479
 480 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 481                                                u64 *parent_pte)
 482 {
 483         struct kvm_mmu_page *page;
 484
 485         if (!vcpu->kvm->n_free_mmu_pages)
 486                 return NULL;
 487
 488         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 489                                       sizeof *page);
 490         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 491         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 492         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 493         ASSERT(is_empty_shadow_page(page->spt));
 494         page->slot_bitmap = 0;
 495         page->multimapped = 0;
 496         page->parent_pte = parent_pte;
 497         --vcpu->kvm->n_free_mmu_pages;
 498         return page;
 499 }
 500
 501 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 502                                     struct kvm_mmu_page *page, u64 *parent_pte)
 503 {
 504         struct kvm_pte_chain *pte_chain;
 505         struct hlist_node *node;
 506         int i;
 507
 508         if (!parent_pte)
 509                 return;
 510         if (!page->multimapped) {
 511                 u64 *old = page->parent_pte;
 512
 513                 if (!old) {
 514                         page->parent_pte = parent_pte;
 515                         return;
 516                 }
 517                 page->multimapped = 1;
 518                 pte_chain = mmu_alloc_pte_chain(vcpu);
 519                 INIT_HLIST_HEAD(&page->parent_ptes);
 520                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 521                 pte_chain->parent_ptes[0] = old;
 522         }
 523         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 524                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 525                         continue;
 526                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 527                         if (!pte_chain->parent_ptes[i]) {
 528                                 pte_chain->parent_ptes[i] = parent_pte;
 529                                 return;
 530                         }
 531         }
 532         pte_chain = mmu_alloc_pte_chain(vcpu);
 533         BUG_ON(!pte_chain);
 534         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 535         pte_chain->parent_ptes[0] = parent_pte;
 536 }
 537
 538 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
 539                                        struct kvm_mmu_page *page,
 540                                        u64 *parent_pte)
 541 {
 542         struct kvm_pte_chain *pte_chain;
 543         struct hlist_node *node;
 544         int i;
 545
 546         if (!page->multimapped) {
 547                 BUG_ON(page->parent_pte != parent_pte);
 548                 page->parent_pte = NULL;
 549                 return;
 550         }
 551         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 552                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 553                         if (!pte_chain->parent_ptes[i])
 554                                 break;
 555                         if (pte_chain->parent_ptes[i] != parent_pte)
 556                                 continue;
 557                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 558                                 && pte_chain->parent_ptes[i + 1]) {
 559                                 pte_chain->parent_ptes[i]
 560                                         = pte_chain->parent_ptes[i + 1];
 561                                 ++i;
 562                         }
 563                         pte_chain->parent_ptes[i] = NULL;
 564                         if (i == 0) {
 565                                 hlist_del(&pte_chain->link);
 566                                 mmu_free_pte_chain(vcpu, pte_chain);
 567                                 if (hlist_empty(&page->parent_ptes)) {
 568                                         page->multimapped = 0;
 569                                         page->parent_pte = NULL;
 570                                 }
 571                         }
 572                         return;
 573                 }
 574         BUG();
 575 }
 576
 577 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 578                                                 gfn_t gfn)
 579 {
 580         unsigned index;
 581         struct hlist_head *bucket;
 582         struct kvm_mmu_page *page;
 583         struct hlist_node *node;
 584
 585         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 586         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 587         bucket = &vcpu->kvm->mmu_page_hash[index];
 588         hlist_for_each_entry(page, node, bucket, hash_link)
 589                 if (page->gfn == gfn && !page->role.metaphysical) {
 590                         pgprintk("%s: found role %x\n",
 591                                  __FUNCTION__, page->role.word);
 592                         return page;
 593                 }
 594         return NULL;
 595 }
 596
 597 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 598                                              gfn_t gfn,
 599                                              gva_t gaddr,
 600                                              unsigned level,
 601                                              int metaphysical,
 602                                              unsigned hugepage_access,
 603                                              u64 *parent_pte)
 604 {
 605         union kvm_mmu_page_role role;
 606         unsigned index;
 607         unsigned quadrant;
 608         struct hlist_head *bucket;
 609         struct kvm_mmu_page *page;
 610         struct hlist_node *node;
 611
 612         role.word = 0;
 613         role.glevels = vcpu->mmu.root_level;
 614         role.level = level;
 615         role.metaphysical = metaphysical;
 616         role.hugepage_access = hugepage_access;
 617         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 618                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 619                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 620                 role.quadrant = quadrant;
 621         }
 622         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 623                  gfn, role.word);
 624         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 625         bucket = &vcpu->kvm->mmu_page_hash[index];
 626         hlist_for_each_entry(page, node, bucket, hash_link)
 627                 if (page->gfn == gfn && page->role.word == role.word) {
 628                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 629                         pgprintk("%s: found\n", __FUNCTION__);
 630                         return page;
 631                 }
 632         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 633         if (!page)
 634                 return page;
 635         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 636         page->gfn = gfn;
 637         page->role = role;
 638         hlist_add_head(&page->hash_link, bucket);
 639         if (!metaphysical)
 640                 rmap_write_protect(vcpu, gfn);
 641         return page;
 642 }
 643
 644 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 645                                          struct kvm_mmu_page *page)
 646 {
 647         unsigned i;
 648         u64 *pt;
 649         u64 ent;
 650
 651         pt = page->spt;
 652
 653         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 654                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 655                         if (pt[i] & PT_PRESENT_MASK)
 656                                 rmap_remove(vcpu, &pt[i]);
 657                         pt[i] = 0;
 658                 }
 659                 kvm_arch_ops->tlb_flush(vcpu);
 660                 return;
 661         }
 662
 663         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 664                 ent = pt[i];
 665
 666                 pt[i] = 0;
 667                 if (!(ent & PT_PRESENT_MASK))
 668                         continue;
 669                 ent &= PT64_BASE_ADDR_MASK;
 670                 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
 671         }
 672 }
 673
 674 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 675                              struct kvm_mmu_page *page,
 676                              u64 *parent_pte)
 677 {
 678         mmu_page_remove_parent_pte(vcpu, page, parent_pte);
 679 }
 680
 681 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 682                              struct kvm_mmu_page *page)
 683 {
 684         u64 *parent_pte;
 685
 686         while (page->multimapped || page->parent_pte) {
 687                 if (!page->multimapped)
 688                         parent_pte = page->parent_pte;
 689                 else {
 690                         struct kvm_pte_chain *chain;
 691
 692                         chain = container_of(page->parent_ptes.first,
 693                                              struct kvm_pte_chain, link);
 694                         parent_pte = chain->parent_ptes[0];
 695                 }
 696                 BUG_ON(!parent_pte);
 697                 kvm_mmu_put_page(vcpu, page, parent_pte);
 698                 set_shadow_pte(parent_pte, 0);
 699         }
 700         kvm_mmu_page_unlink_children(vcpu, page);
 701         if (!page->root_count) {
 702                 hlist_del(&page->hash_link);
 703                 kvm_mmu_free_page(vcpu, page);
 704         } else
 705                 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 706 }
 707
 708 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 709 {
 710         unsigned index;
 711         struct hlist_head *bucket;
 712         struct kvm_mmu_page *page;
 713         struct hlist_node *node, *n;
 714         int r;
 715
 716         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 717         r = 0;
 718         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 719         bucket = &vcpu->kvm->mmu_page_hash[index];
 720         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 721                 if (page->gfn == gfn && !page->role.metaphysical) {
 722                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 723                                  page->role.word);
 724                         kvm_mmu_zap_page(vcpu, page);
 725                         r = 1;
 726                 }
 727         return r;
 728 }
 729
 730 static void mmu_unshadow(struct kvm_vcpu *vcpu, gfn_t gfn)
 731 {
 732         struct kvm_mmu_page *page;
 733
 734         while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
 735                 pgprintk("%s: zap %lx %x\n",
 736                          __FUNCTION__, gfn, page->role.word);
 737                 kvm_mmu_zap_page(vcpu, page);
 738         }
 739 }
 740
 741 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 742 {
 743         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 744         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 745
 746         __set_bit(slot, &page_head->slot_bitmap);
 747 }
 748
 749 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 750 {
 751         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 752
 753         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 754 }
 755
 756 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 757 {
 758         struct page *page;
 759
 760         ASSERT((gpa & HPA_ERR_MASK) == 0);
 761         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 762         if (!page)
 763                 return gpa | HPA_ERR_MASK;
 764         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 765                 | (gpa & (PAGE_SIZE-1));
 766 }
 767
 768 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 769 {
 770         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 771
 772         if (gpa == UNMAPPED_GVA)
 773                 return UNMAPPED_GVA;
 774         return gpa_to_hpa(vcpu, gpa);
 775 }
 776
 777 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 778 {
 779         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 780
 781         if (gpa == UNMAPPED_GVA)
 782                 return NULL;
 783         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 784 }
 785
 786 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 787 {
 788 }
 789
 790 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 791 {
 792         int level = PT32E_ROOT_LEVEL;
 793         hpa_t table_addr = vcpu->mmu.root_hpa;
 794
 795         for (; ; level--) {
 796                 u32 index = PT64_INDEX(v, level);
 797                 u64 *table;
 798                 u64 pte;
 799
 800                 ASSERT(VALID_PAGE(table_addr));
 801                 table = __va(table_addr);
 802
 803                 if (level == 1) {
 804                         pte = table[index];
 805                         if (is_present_pte(pte) && is_writeble_pte(pte))
 806                                 return 0;
 807                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 808                         page_header_update_slot(vcpu->kvm, table, v);
 809                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 810                                                                 PT_USER_MASK;
 811                         rmap_add(vcpu, &table[index]);
 812                         return 0;
 813                 }
 814
 815                 if (table[index] == 0) {
 816                         struct kvm_mmu_page *new_table;
 817                         gfn_t pseudo_gfn;
 818
 819                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 820                                 >> PAGE_SHIFT;
 821                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 822                                                      v, level - 1,
 823                                                      1, 0, &table[index]);
 824                         if (!new_table) {
 825                                 pgprintk("nonpaging_map: ENOMEM\n");
 826                                 return -ENOMEM;
 827                         }
 828
 829                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 830                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 831                 }
 832                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 833         }
 834 }
 835
 836 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 837 {
 838         int i;
 839         struct kvm_mmu_page *page;
 840
 841 #ifdef CONFIG_X86_64
 842         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 843                 hpa_t root = vcpu->mmu.root_hpa;
 844
 845                 ASSERT(VALID_PAGE(root));
 846                 page = page_header(root);
 847                 --page->root_count;
 848                 vcpu->mmu.root_hpa = INVALID_PAGE;
 849                 return;
 850         }
 851 #endif
 852         for (i = 0; i < 4; ++i) {
 853                 hpa_t root = vcpu->mmu.pae_root[i];
 854
 855                 if (root) {
 856                         ASSERT(VALID_PAGE(root));
 857                         root &= PT64_BASE_ADDR_MASK;
 858                         page = page_header(root);
 859                         --page->root_count;
 860                 }
 861                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 862         }
 863         vcpu->mmu.root_hpa = INVALID_PAGE;
 864 }
 865
 866 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 867 {
 868         int i;
 869         gfn_t root_gfn;
 870         struct kvm_mmu_page *page;
 871
 872         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 873
 874 #ifdef CONFIG_X86_64
 875         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 876                 hpa_t root = vcpu->mmu.root_hpa;
 877
 878                 ASSERT(!VALID_PAGE(root));
 879                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 880                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 881                 root = __pa(page->spt);
 882                 ++page->root_count;
 883                 vcpu->mmu.root_hpa = root;
 884                 return;
 885         }
 886 #endif
 887         for (i = 0; i < 4; ++i) {
 888                 hpa_t root = vcpu->mmu.pae_root[i];
 889
 890                 ASSERT(!VALID_PAGE(root));
 891                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 892                         if (!is_present_pte(vcpu->pdptrs[i])) {
 893                                 vcpu->mmu.pae_root[i] = 0;
 894                                 continue;
 895                         }
 896                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 897                 } else if (vcpu->mmu.root_level == 0)
 898                         root_gfn = 0;
 899                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 900                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 901                                         0, NULL);
 902                 root = __pa(page->spt);
 903                 ++page->root_count;
 904                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 905         }
 906         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 907 }
 908
 909 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 910 {
 911         return vaddr;
 912 }
 913
 914 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 915                                u32 error_code)
 916 {
 917         gpa_t addr = gva;
 918         hpa_t paddr;
 919         int r;
 920
 921         r = mmu_topup_memory_caches(vcpu);
 922         if (r)
 923                 return r;
 924
 925         ASSERT(vcpu);
 926         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 927
 928
 929         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 930
 931         if (is_error_hpa(paddr))
 932                 return 1;
 933
 934         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 935 }
 936
 937 static void nonpaging_free(struct kvm_vcpu *vcpu)
 938 {
 939         mmu_free_roots(vcpu);
 940 }
 941
 942 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 943 {
 944         struct kvm_mmu *context = &vcpu->mmu;
 945
 946         context->new_cr3 = nonpaging_new_cr3;
 947         context->page_fault = nonpaging_page_fault;
 948         context->gva_to_gpa = nonpaging_gva_to_gpa;
 949         context->free = nonpaging_free;
 950         context->root_level = 0;
 951         context->shadow_root_level = PT32E_ROOT_LEVEL;
 952         mmu_alloc_roots(vcpu);
 953         ASSERT(VALID_PAGE(context->root_hpa));
 954         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 955         return 0;
 956 }
 957
 958 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 959 {
 960         ++vcpu->stat.tlb_flush;
 961         kvm_arch_ops->tlb_flush(vcpu);
 962 }
 963
 964 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 965 {
 966         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 967         mmu_free_roots(vcpu);
 968         if (unlikely(vcpu->kvm->n_free_mmu_pages < KVM_MIN_FREE_MMU_PAGES))
 969                 kvm_mmu_free_some_pages(vcpu);
 970         mmu_alloc_roots(vcpu);
 971         kvm_mmu_flush_tlb(vcpu);
 972         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 973 }
 974
 975 static void inject_page_fault(struct kvm_vcpu *vcpu,
 976                               u64 addr,
 977                               u32 err_code)
 978 {
 979         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
 980 }
 981
 982 static void paging_free(struct kvm_vcpu *vcpu)
 983 {
 984         nonpaging_free(vcpu);
 985 }
 986
 987 #define PTTYPE 64
 988 #include "paging_tmpl.h"
 989 #undef PTTYPE
 990
 991 #define PTTYPE 32
 992 #include "paging_tmpl.h"
 993 #undef PTTYPE
 994
 995 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
 996 {
 997         struct kvm_mmu *context = &vcpu->mmu;
 998
 999         ASSERT(is_pae(vcpu));
1000         context->new_cr3 = paging_new_cr3;
1001         context->page_fault = paging64_page_fault;
1002         context->gva_to_gpa = paging64_gva_to_gpa;
1003         context->free = paging_free;
1004         context->root_level = level;
1005         context->shadow_root_level = level;
1006         mmu_alloc_roots(vcpu);
1007         ASSERT(VALID_PAGE(context->root_hpa));
1008         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1009                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1010         return 0;
1011 }
1012
1013 static int paging64_init_context(struct kvm_vcpu *vcpu)
1014 {
1015         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1016 }
1017
1018 static int paging32_init_context(struct kvm_vcpu *vcpu)
1019 {
1020         struct kvm_mmu *context = &vcpu->mmu;
1021
1022         context->new_cr3 = paging_new_cr3;
1023         context->page_fault = paging32_page_fault;
1024         context->gva_to_gpa = paging32_gva_to_gpa;
1025         context->free = paging_free;
1026         context->root_level = PT32_ROOT_LEVEL;
1027         context->shadow_root_level = PT32E_ROOT_LEVEL;
1028         mmu_alloc_roots(vcpu);
1029         ASSERT(VALID_PAGE(context->root_hpa));
1030         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1031                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1032         return 0;
1033 }
1034
1035 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1036 {
1037         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1038 }
1039
1040 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1041 {
1042         ASSERT(vcpu);
1043         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1044
1045         mmu_topup_memory_caches(vcpu);
1046         if (!is_paging(vcpu))
1047                 return nonpaging_init_context(vcpu);
1048         else if (is_long_mode(vcpu))
1049                 return paging64_init_context(vcpu);
1050         else if (is_pae(vcpu))
1051                 return paging32E_init_context(vcpu);
1052         else
1053                 return paging32_init_context(vcpu);
1054 }
1055
1056 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1057 {
1058         ASSERT(vcpu);
1059         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1060                 vcpu->mmu.free(vcpu);
1061                 vcpu->mmu.root_hpa = INVALID_PAGE;
1062         }
1063 }
1064
1065 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1066 {
1067         int r;
1068
1069         destroy_kvm_mmu(vcpu);
1070         r = init_kvm_mmu(vcpu);
1071         if (r < 0)
1072                 goto out;
1073         r = mmu_topup_memory_caches(vcpu);
1074 out:
1075         return r;
1076 }
1077
1078 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1079                                   struct kvm_mmu_page *page,
1080                                   u64 *spte)
1081 {
1082         u64 pte;
1083         struct kvm_mmu_page *child;
1084
1085         pte = *spte;
1086         if (is_present_pte(pte)) {
1087                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1088                         rmap_remove(vcpu, spte);
1089                 else {
1090                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1091                         mmu_page_remove_parent_pte(vcpu, child, spte);
1092                 }
1093         }
1094         *spte = 0;
1095 }
1096
1097 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1098                                   struct kvm_mmu_page *page,
1099                                   u64 *spte,
1100                                   const void *new, int bytes)
1101 {
1102         if (page->role.level != PT_PAGE_TABLE_LEVEL)
1103                 return;
1104
1105         if (page->role.glevels == PT32_ROOT_LEVEL)
1106                 paging32_update_pte(vcpu, page, spte, new, bytes);
1107         else
1108                 paging64_update_pte(vcpu, page, spte, new, bytes);
1109 }
1110
1111 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1112                        const u8 *old, const u8 *new, int bytes)
1113 {
1114         gfn_t gfn = gpa >> PAGE_SHIFT;
1115         struct kvm_mmu_page *page;
1116         struct hlist_node *node, *n;
1117         struct hlist_head *bucket;
1118         unsigned index;
1119         u64 *spte;
1120         unsigned offset = offset_in_page(gpa);
1121         unsigned pte_size;
1122         unsigned page_offset;
1123         unsigned misaligned;
1124         unsigned quadrant;
1125         int level;
1126         int flooded = 0;
1127         int npte;
1128
1129         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1130         if (gfn == vcpu->last_pt_write_gfn) {
1131                 ++vcpu->last_pt_write_count;
1132                 if (vcpu->last_pt_write_count >= 3)
1133                         flooded = 1;
1134         } else {
1135                 vcpu->last_pt_write_gfn = gfn;
1136                 vcpu->last_pt_write_count = 1;
1137         }
1138         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1139         bucket = &vcpu->kvm->mmu_page_hash[index];
1140         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1141                 if (page->gfn != gfn || page->role.metaphysical)
1142                         continue;
1143                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1144                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1145                 misaligned |= bytes < 4;
1146                 if (misaligned || flooded) {
1147                         /*
1148                          * Misaligned accesses are too much trouble to fix
1149                          * up; also, they usually indicate a page is not used
1150                          * as a page table.
1151                          *
1152                          * If we're seeing too many writes to a page,
1153                          * it may no longer be a page table, or we may be
1154                          * forking, in which case it is better to unmap the
1155                          * page.
1156                          */
1157                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1158                                  gpa, bytes, page->role.word);
1159                         kvm_mmu_zap_page(vcpu, page);
1160                         continue;
1161                 }
1162                 page_offset = offset;
1163                 level = page->role.level;
1164                 npte = 1;
1165                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1166                         page_offset <<= 1;      /* 32->64 */
1167                         /*
1168                          * A 32-bit pde maps 4MB while the shadow pdes map
1169                          * only 2MB.  So we need to double the offset again
1170                          * and zap two pdes instead of one.
1171                          */
1172                         if (level == PT32_ROOT_LEVEL) {
1173                                 page_offset &= ~7; /* kill rounding error */
1174                                 page_offset <<= 1;
1175                                 npte = 2;
1176                         }
1177                         quadrant = page_offset >> PAGE_SHIFT;
1178                         page_offset &= ~PAGE_MASK;
1179                         if (quadrant != page->role.quadrant)
1180                                 continue;
1181                 }
1182                 spte = &page->spt[page_offset / sizeof(*spte)];
1183                 while (npte--) {
1184                         mmu_pte_write_zap_pte(vcpu, page, spte);
1185                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes);
1186                         ++spte;
1187                 }
1188         }
1189 }
1190
1191 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1192 {
1193         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1194
1195         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1196 }
1197
1198 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1199 {
1200         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1201                 struct kvm_mmu_page *page;
1202
1203                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1204                                     struct kvm_mmu_page, link);
1205                 kvm_mmu_zap_page(vcpu, page);
1206         }
1207 }
1208 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1209
1210 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1211 {
1212         struct kvm_mmu_page *page;
1213
1214         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1215                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1216                                     struct kvm_mmu_page, link);
1217                 kvm_mmu_zap_page(vcpu, page);
1218         }
1219         free_page((unsigned long)vcpu->mmu.pae_root);
1220 }
1221
1222 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1223 {
1224         struct page *page;
1225         int i;
1226
1227         ASSERT(vcpu);
1228
1229         vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1230
1231         /*
1232          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1233          * Therefore we need to allocate shadow page tables in the first
1234          * 4GB of memory, which happens to fit the DMA32 zone.
1235          */
1236         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1237         if (!page)
1238                 goto error_1;
1239         vcpu->mmu.pae_root = page_address(page);
1240         for (i = 0; i < 4; ++i)
1241                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1242
1243         return 0;
1244
1245 error_1:
1246         free_mmu_pages(vcpu);
1247         return -ENOMEM;
1248 }
1249
1250 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1251 {
1252         ASSERT(vcpu);
1253         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1254
1255         return alloc_mmu_pages(vcpu);
1256 }
1257
1258 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1259 {
1260         ASSERT(vcpu);
1261         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1262
1263         return init_kvm_mmu(vcpu);
1264 }
1265
1266 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1267 {
1268         ASSERT(vcpu);
1269
1270         destroy_kvm_mmu(vcpu);
1271         free_mmu_pages(vcpu);
1272         mmu_free_memory_caches(vcpu);
1273 }
1274
1275 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1276 {
1277         struct kvm *kvm = vcpu->kvm;
1278         struct kvm_mmu_page *page;
1279
1280         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1281                 int i;
1282                 u64 *pt;
1283
1284                 if (!test_bit(slot, &page->slot_bitmap))
1285                         continue;
1286
1287                 pt = page->spt;
1288                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1289                         /* avoid RMW */
1290                         if (pt[i] & PT_WRITABLE_MASK) {
1291                                 rmap_remove(vcpu, &pt[i]);
1292                                 pt[i] &= ~PT_WRITABLE_MASK;
1293                         }
1294         }
1295 }
1296
1297 void kvm_mmu_zap_all(struct kvm_vcpu *vcpu)
1298 {
1299         destroy_kvm_mmu(vcpu);
1300
1301         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1302                 struct kvm_mmu_page *page;
1303
1304                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1305                                     struct kvm_mmu_page, link);
1306                 kvm_mmu_zap_page(vcpu, page);
1307         }
1308
1309         mmu_free_memory_caches(vcpu);
1310         kvm_arch_ops->tlb_flush(vcpu);
1311         init_kvm_mmu(vcpu);
1312 }
1313
1314 void kvm_mmu_module_exit(void)
1315 {
1316         if (pte_chain_cache)
1317                 kmem_cache_destroy(pte_chain_cache);
1318         if (rmap_desc_cache)
1319                 kmem_cache_destroy(rmap_desc_cache);
1320         if (mmu_page_cache)
1321                 kmem_cache_destroy(mmu_page_cache);
1322         if (mmu_page_header_cache)
1323                 kmem_cache_destroy(mmu_page_header_cache);
1324 }
1325
1326 int kvm_mmu_module_init(void)
1327 {
1328         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1329                                             sizeof(struct kvm_pte_chain),
1330                                             0, 0, NULL, NULL);
1331         if (!pte_chain_cache)
1332                 goto nomem;
1333         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1334                                             sizeof(struct kvm_rmap_desc),
1335                                             0, 0, NULL, NULL);
1336         if (!rmap_desc_cache)
1337                 goto nomem;
1338
1339         mmu_page_cache = kmem_cache_create("kvm_mmu_page",
1340                                            PAGE_SIZE,
1341                                            PAGE_SIZE, 0, NULL, NULL);
1342         if (!mmu_page_cache)
1343                 goto nomem;
1344
1345         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1346                                                   sizeof(struct kvm_mmu_page),
1347                                                   0, 0, NULL, NULL);
1348         if (!mmu_page_header_cache)
1349                 goto nomem;
1350
1351         return 0;
1352
1353 nomem:
1354         kvm_mmu_module_exit();
1355         return -ENOMEM;
1356 }
1357
1358 #ifdef AUDIT
1359
1360 static const char *audit_msg;
1361
1362 static gva_t canonicalize(gva_t gva)
1363 {
1364 #ifdef CONFIG_X86_64
1365         gva = (long long)(gva << 16) >> 16;
1366 #endif
1367         return gva;
1368 }
1369
1370 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1371                                 gva_t va, int level)
1372 {
1373         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1374         int i;
1375         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1376
1377         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1378                 u64 ent = pt[i];
1379
1380                 if (!(ent & PT_PRESENT_MASK))
1381                         continue;
1382
1383                 va = canonicalize(va);
1384                 if (level > 1)
1385                         audit_mappings_page(vcpu, ent, va, level - 1);
1386                 else {
1387                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1388                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1389
1390                         if ((ent & PT_PRESENT_MASK)
1391                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1392                                 printk(KERN_ERR "audit error: (%s) levels %d"
1393                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1394                                        audit_msg, vcpu->mmu.root_level,
1395                                        va, gpa, hpa, ent);
1396                 }
1397         }
1398 }
1399
1400 static void audit_mappings(struct kvm_vcpu *vcpu)
1401 {
1402         unsigned i;
1403
1404         if (vcpu->mmu.root_level == 4)
1405                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1406         else
1407                 for (i = 0; i < 4; ++i)
1408                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1409                                 audit_mappings_page(vcpu,
1410                                                     vcpu->mmu.pae_root[i],
1411                                                     i << 30,
1412                                                     2);
1413 }
1414
1415 static int count_rmaps(struct kvm_vcpu *vcpu)
1416 {
1417         int nmaps = 0;
1418         int i, j, k;
1419
1420         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1421                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1422                 struct kvm_rmap_desc *d;
1423
1424                 for (j = 0; j < m->npages; ++j) {
1425                         struct page *page = m->phys_mem[j];
1426
1427                         if (!page->private)
1428                                 continue;
1429                         if (!(page->private & 1)) {
1430                                 ++nmaps;
1431                                 continue;
1432                         }
1433                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1434                         while (d) {
1435                                 for (k = 0; k < RMAP_EXT; ++k)
1436                                         if (d->shadow_ptes[k])
1437                                                 ++nmaps;
1438                                         else
1439                                                 break;
1440                                 d = d->more;
1441                         }
1442                 }
1443         }
1444         return nmaps;
1445 }
1446
1447 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1448 {
1449         int nmaps = 0;
1450         struct kvm_mmu_page *page;
1451         int i;
1452
1453         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1454                 u64 *pt = page->spt;
1455
1456                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1457                         continue;
1458
1459                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1460                         u64 ent = pt[i];
1461
1462                         if (!(ent & PT_PRESENT_MASK))
1463                                 continue;
1464                         if (!(ent & PT_WRITABLE_MASK))
1465                                 continue;
1466                         ++nmaps;
1467                 }
1468         }
1469         return nmaps;
1470 }
1471
1472 static void audit_rmap(struct kvm_vcpu *vcpu)
1473 {
1474         int n_rmap = count_rmaps(vcpu);
1475         int n_actual = count_writable_mappings(vcpu);
1476
1477         if (n_rmap != n_actual)
1478                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1479                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1480 }
1481
1482 static void audit_write_protection(struct kvm_vcpu *vcpu)
1483 {
1484         struct kvm_mmu_page *page;
1485
1486         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1487                 hfn_t hfn;
1488                 struct page *pg;
1489
1490                 if (page->role.metaphysical)
1491                         continue;
1492
1493                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1494                         >> PAGE_SHIFT;
1495                 pg = pfn_to_page(hfn);
1496                 if (pg->private)
1497                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1498                                " mappings: gfn %lx role %x\n",
1499                                __FUNCTION__, audit_msg, page->gfn,
1500                                page->role.word);
1501         }
1502 }
1503
1504 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1505 {
1506         int olddbg = dbg;
1507
1508         dbg = 0;
1509         audit_msg = msg;
1510         audit_rmap(vcpu);
1511         audit_write_protection(vcpu);
1512         audit_mappings(vcpu);
1513         dbg = olddbg;
1514 }
1515
1516 #endif