SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19 #include <linux/types.h>
  20 #include <linux/string.h>
  21 #include <asm/page.h>
  22 #include <linux/mm.h>
  23 #include <linux/highmem.h>
  24 #include <linux/module.h>
  25
  26 #include "vmx.h"
  27 #include "kvm.h"
  28
  29 #undef MMU_DEBUG
  30
  31 #undef AUDIT
  32
  33 #ifdef AUDIT
  34 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  35 #else
  36 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  37 #endif
  38
  39 #ifdef MMU_DEBUG
  40
  41 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  42 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  43
  44 #else
  45
  46 #define pgprintk(x...) do { } while (0)
  47 #define rmap_printk(x...) do { } while (0)
  48
  49 #endif
  50
  51 #if defined(MMU_DEBUG) || defined(AUDIT)
  52 static int dbg = 1;
  53 #endif
  54
  55 #define ASSERT(x)                                                       \
  56         if (!(x)) {                                                     \
  57                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  58                        __FILE__, __LINE__, #x);                         \
  59         }
  60
  61 #define PT64_PT_BITS 9
  62 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  63 #define PT32_PT_BITS 10
  64 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  65
  66 #define PT_WRITABLE_SHIFT 1
  67
  68 #define PT_PRESENT_MASK (1ULL << 0)
  69 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  70 #define PT_USER_MASK (1ULL << 2)
  71 #define PT_PWT_MASK (1ULL << 3)
  72 #define PT_PCD_MASK (1ULL << 4)
  73 #define PT_ACCESSED_MASK (1ULL << 5)
  74 #define PT_DIRTY_MASK (1ULL << 6)
  75 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  76 #define PT_PAT_MASK (1ULL << 7)
  77 #define PT_GLOBAL_MASK (1ULL << 8)
  78 #define PT64_NX_MASK (1ULL << 63)
  79
  80 #define PT_PAT_SHIFT 7
  81 #define PT_DIR_PAT_SHIFT 12
  82 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  83
  84 #define PT32_DIR_PSE36_SIZE 4
  85 #define PT32_DIR_PSE36_SHIFT 13
  86 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  87
  88
  89 #define PT32_PTE_COPY_MASK \
  90         (PT_PRESENT_MASK | PT_ACCESSED_MASK | PT_DIRTY_MASK | PT_GLOBAL_MASK)
  91
  92 #define PT64_PTE_COPY_MASK (PT64_NX_MASK | PT32_PTE_COPY_MASK)
  93
  94 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  95 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  96
  97 #define PT_SHADOW_PS_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
  98 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
  99
 100 #define PT_SHADOW_WRITABLE_SHIFT (PT_FIRST_AVAIL_BITS_SHIFT + 1)
 101 #define PT_SHADOW_WRITABLE_MASK (1ULL << PT_SHADOW_WRITABLE_SHIFT)
 102
 103 #define PT_SHADOW_USER_SHIFT (PT_SHADOW_WRITABLE_SHIFT + 1)
 104 #define PT_SHADOW_USER_MASK (1ULL << (PT_SHADOW_USER_SHIFT))
 105
 106 #define PT_SHADOW_BITS_OFFSET (PT_SHADOW_WRITABLE_SHIFT - PT_WRITABLE_SHIFT)
 107
 108 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 109
 110 #define PT64_LEVEL_BITS 9
 111
 112 #define PT64_LEVEL_SHIFT(level) \
 113                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
 114
 115 #define PT64_LEVEL_MASK(level) \
 116                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 117
 118 #define PT64_INDEX(address, level)\
 119         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 120
 121
 122 #define PT32_LEVEL_BITS 10
 123
 124 #define PT32_LEVEL_SHIFT(level) \
 125                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
 126
 127 #define PT32_LEVEL_MASK(level) \
 128                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 129
 130 #define PT32_INDEX(address, level)\
 131         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 132
 133
 134 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 135 #define PT64_DIR_BASE_ADDR_MASK \
 136         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 137
 138 #define PT32_BASE_ADDR_MASK PAGE_MASK
 139 #define PT32_DIR_BASE_ADDR_MASK \
 140         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 141
 142
 143 #define PFERR_PRESENT_MASK (1U << 0)
 144 #define PFERR_WRITE_MASK (1U << 1)
 145 #define PFERR_USER_MASK (1U << 2)
 146 #define PFERR_FETCH_MASK (1U << 4)
 147
 148 #define PT64_ROOT_LEVEL 4
 149 #define PT32_ROOT_LEVEL 2
 150 #define PT32E_ROOT_LEVEL 3
 151
 152 #define PT_DIRECTORY_LEVEL 2
 153 #define PT_PAGE_TABLE_LEVEL 1
 154
 155 #define RMAP_EXT 4
 156
 157 struct kvm_rmap_desc {
 158         u64 *shadow_ptes[RMAP_EXT];
 159         struct kvm_rmap_desc *more;
 160 };
 161
 162 static struct kmem_cache *pte_chain_cache;
 163 static struct kmem_cache *rmap_desc_cache;
 164
 165 static int is_write_protection(struct kvm_vcpu *vcpu)
 166 {
 167         return vcpu->cr0 & CR0_WP_MASK;
 168 }
 169
 170 static int is_cpuid_PSE36(void)
 171 {
 172         return 1;
 173 }
 174
 175 static int is_nx(struct kvm_vcpu *vcpu)
 176 {
 177         return vcpu->shadow_efer & EFER_NX;
 178 }
 179
 180 static int is_present_pte(unsigned long pte)
 181 {
 182         return pte & PT_PRESENT_MASK;
 183 }
 184
 185 static int is_writeble_pte(unsigned long pte)
 186 {
 187         return pte & PT_WRITABLE_MASK;
 188 }
 189
 190 static int is_io_pte(unsigned long pte)
 191 {
 192         return pte & PT_SHADOW_IO_MARK;
 193 }
 194
 195 static int is_rmap_pte(u64 pte)
 196 {
 197         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 198                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 199 }
 200
 201 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 202                                   struct kmem_cache *base_cache, int min)
 203 {
 204         void *obj;
 205
 206         if (cache->nobjs >= min)
 207                 return 0;
 208         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 209                 obj = kmem_cache_zalloc(base_cache, GFP_NOWAIT);
 210                 if (!obj)
 211                         return -ENOMEM;
 212                 cache->objects[cache->nobjs++] = obj;
 213         }
 214         return 0;
 215 }
 216
 217 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 218 {
 219         while (mc->nobjs)
 220                 kfree(mc->objects[--mc->nobjs]);
 221 }
 222
 223 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 224 {
 225         int r;
 226
 227         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 228                                    pte_chain_cache, 4);
 229         if (r)
 230                 goto out;
 231         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 232                                    rmap_desc_cache, 1);
 233 out:
 234         return r;
 235 }
 236
 237 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 238 {
 239         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 240         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 241 }
 242
 243 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 244                                     size_t size)
 245 {
 246         void *p;
 247
 248         BUG_ON(!mc->nobjs);
 249         p = mc->objects[--mc->nobjs];
 250         memset(p, 0, size);
 251         return p;
 252 }
 253
 254 static void mmu_memory_cache_free(struct kvm_mmu_memory_cache *mc, void *obj)
 255 {
 256         if (mc->nobjs < KVM_NR_MEM_OBJS)
 257                 mc->objects[mc->nobjs++] = obj;
 258         else
 259                 kfree(obj);
 260 }
 261
 262 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 263 {
 264         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 265                                       sizeof(struct kvm_pte_chain));
 266 }
 267
 268 static void mmu_free_pte_chain(struct kvm_vcpu *vcpu,
 269                                struct kvm_pte_chain *pc)
 270 {
 271         mmu_memory_cache_free(&vcpu->mmu_pte_chain_cache, pc);
 272 }
 273
 274 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 275 {
 276         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 277                                       sizeof(struct kvm_rmap_desc));
 278 }
 279
 280 static void mmu_free_rmap_desc(struct kvm_vcpu *vcpu,
 281                                struct kvm_rmap_desc *rd)
 282 {
 283         mmu_memory_cache_free(&vcpu->mmu_rmap_desc_cache, rd);
 284 }
 285
 286 /*
 287  * Reverse mapping data structures:
 288  *
 289  * If page->private bit zero is zero, then page->private points to the
 290  * shadow page table entry that points to page_address(page).
 291  *
 292  * If page->private bit zero is one, (then page->private & ~1) points
 293  * to a struct kvm_rmap_desc containing more mappings.
 294  */
 295 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 296 {
 297         struct page *page;
 298         struct kvm_rmap_desc *desc;
 299         int i;
 300
 301         if (!is_rmap_pte(*spte))
 302                 return;
 303         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 304         if (!page_private(page)) {
 305                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 306                 set_page_private(page,(unsigned long)spte);
 307         } else if (!(page_private(page) & 1)) {
 308                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 309                 desc = mmu_alloc_rmap_desc(vcpu);
 310                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 311                 desc->shadow_ptes[1] = spte;
 312                 set_page_private(page,(unsigned long)desc | 1);
 313         } else {
 314                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 315                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 316                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 317                         desc = desc->more;
 318                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 319                         desc->more = mmu_alloc_rmap_desc(vcpu);
 320                         desc = desc->more;
 321                 }
 322                 for (i = 0; desc->shadow_ptes[i]; ++i)
 323                         ;
 324                 desc->shadow_ptes[i] = spte;
 325         }
 326 }
 327
 328 static void rmap_desc_remove_entry(struct kvm_vcpu *vcpu,
 329                                    struct page *page,
 330                                    struct kvm_rmap_desc *desc,
 331                                    int i,
 332                                    struct kvm_rmap_desc *prev_desc)
 333 {
 334         int j;
 335
 336         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 337                 ;
 338         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 339         desc->shadow_ptes[j] = NULL;
 340         if (j != 0)
 341                 return;
 342         if (!prev_desc && !desc->more)
 343                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 344         else
 345                 if (prev_desc)
 346                         prev_desc->more = desc->more;
 347                 else
 348                         set_page_private(page,(unsigned long)desc->more | 1);
 349         mmu_free_rmap_desc(vcpu, desc);
 350 }
 351
 352 static void rmap_remove(struct kvm_vcpu *vcpu, u64 *spte)
 353 {
 354         struct page *page;
 355         struct kvm_rmap_desc *desc;
 356         struct kvm_rmap_desc *prev_desc;
 357         int i;
 358
 359         if (!is_rmap_pte(*spte))
 360                 return;
 361         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 362         if (!page_private(page)) {
 363                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 364                 BUG();
 365         } else if (!(page_private(page) & 1)) {
 366                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 367                 if ((u64 *)page_private(page) != spte) {
 368                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 369                                spte, *spte);
 370                         BUG();
 371                 }
 372                 set_page_private(page,0);
 373         } else {
 374                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 375                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 376                 prev_desc = NULL;
 377                 while (desc) {
 378                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 379                                 if (desc->shadow_ptes[i] == spte) {
 380                                         rmap_desc_remove_entry(vcpu, page,
 381                                                                desc, i,
 382                                                                prev_desc);
 383                                         return;
 384                                 }
 385                         prev_desc = desc;
 386                         desc = desc->more;
 387                 }
 388                 BUG();
 389         }
 390 }
 391
 392 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 393 {
 394         struct kvm *kvm = vcpu->kvm;
 395         struct page *page;
 396         struct kvm_rmap_desc *desc;
 397         u64 *spte;
 398
 399         page = gfn_to_page(kvm, gfn);
 400         BUG_ON(!page);
 401
 402         while (page_private(page)) {
 403                 if (!(page_private(page) & 1))
 404                         spte = (u64 *)page_private(page);
 405                 else {
 406                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 407                         spte = desc->shadow_ptes[0];
 408                 }
 409                 BUG_ON(!spte);
 410                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 411                        != page_to_pfn(page));
 412                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 413                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 414                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 415                 rmap_remove(vcpu, spte);
 416                 kvm_arch_ops->tlb_flush(vcpu);
 417                 *spte &= ~(u64)PT_WRITABLE_MASK;
 418         }
 419 }
 420
 421 static int is_empty_shadow_page(hpa_t page_hpa)
 422 {
 423         u64 *pos;
 424         u64 *end;
 425
 426         for (pos = __va(page_hpa), end = pos + PAGE_SIZE / sizeof(u64);
 427                       pos != end; pos++)
 428                 if (*pos != 0) {
 429                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 430                                pos, *pos);
 431                         return 0;
 432                 }
 433         return 1;
 434 }
 435
 436 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu, hpa_t page_hpa)
 437 {
 438         struct kvm_mmu_page *page_head = page_header(page_hpa);
 439
 440         ASSERT(is_empty_shadow_page(page_hpa));
 441         page_head->page_hpa = page_hpa;
 442         list_move(&page_head->link, &vcpu->free_pages);
 443         ++vcpu->kvm->n_free_mmu_pages;
 444 }
 445
 446 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 447 {
 448         return gfn;
 449 }
 450
 451 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 452                                                u64 *parent_pte)
 453 {
 454         struct kvm_mmu_page *page;
 455
 456         if (list_empty(&vcpu->free_pages))
 457                 return NULL;
 458
 459         page = list_entry(vcpu->free_pages.next, struct kvm_mmu_page, link);
 460         list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 461         ASSERT(is_empty_shadow_page(page->page_hpa));
 462         page->slot_bitmap = 0;
 463         page->multimapped = 0;
 464         page->parent_pte = parent_pte;
 465         --vcpu->kvm->n_free_mmu_pages;
 466         return page;
 467 }
 468
 469 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 470                                     struct kvm_mmu_page *page, u64 *parent_pte)
 471 {
 472         struct kvm_pte_chain *pte_chain;
 473         struct hlist_node *node;
 474         int i;
 475
 476         if (!parent_pte)
 477                 return;
 478         if (!page->multimapped) {
 479                 u64 *old = page->parent_pte;
 480
 481                 if (!old) {
 482                         page->parent_pte = parent_pte;
 483                         return;
 484                 }
 485                 page->multimapped = 1;
 486                 pte_chain = mmu_alloc_pte_chain(vcpu);
 487                 INIT_HLIST_HEAD(&page->parent_ptes);
 488                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 489                 pte_chain->parent_ptes[0] = old;
 490         }
 491         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 492                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 493                         continue;
 494                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 495                         if (!pte_chain->parent_ptes[i]) {
 496                                 pte_chain->parent_ptes[i] = parent_pte;
 497                                 return;
 498                         }
 499         }
 500         pte_chain = mmu_alloc_pte_chain(vcpu);
 501         BUG_ON(!pte_chain);
 502         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 503         pte_chain->parent_ptes[0] = parent_pte;
 504 }
 505
 506 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
 507                                        struct kvm_mmu_page *page,
 508                                        u64 *parent_pte)
 509 {
 510         struct kvm_pte_chain *pte_chain;
 511         struct hlist_node *node;
 512         int i;
 513
 514         if (!page->multimapped) {
 515                 BUG_ON(page->parent_pte != parent_pte);
 516                 page->parent_pte = NULL;
 517                 return;
 518         }
 519         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 520                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 521                         if (!pte_chain->parent_ptes[i])
 522                                 break;
 523                         if (pte_chain->parent_ptes[i] != parent_pte)
 524                                 continue;
 525                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 526                                 && pte_chain->parent_ptes[i + 1]) {
 527                                 pte_chain->parent_ptes[i]
 528                                         = pte_chain->parent_ptes[i + 1];
 529                                 ++i;
 530                         }
 531                         pte_chain->parent_ptes[i] = NULL;
 532                         if (i == 0) {
 533                                 hlist_del(&pte_chain->link);
 534                                 mmu_free_pte_chain(vcpu, pte_chain);
 535                                 if (hlist_empty(&page->parent_ptes)) {
 536                                         page->multimapped = 0;
 537                                         page->parent_pte = NULL;
 538                                 }
 539                         }
 540                         return;
 541                 }
 542         BUG();
 543 }
 544
 545 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 546                                                 gfn_t gfn)
 547 {
 548         unsigned index;
 549         struct hlist_head *bucket;
 550         struct kvm_mmu_page *page;
 551         struct hlist_node *node;
 552
 553         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 554         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 555         bucket = &vcpu->kvm->mmu_page_hash[index];
 556         hlist_for_each_entry(page, node, bucket, hash_link)
 557                 if (page->gfn == gfn && !page->role.metaphysical) {
 558                         pgprintk("%s: found role %x\n",
 559                                  __FUNCTION__, page->role.word);
 560                         return page;
 561                 }
 562         return NULL;
 563 }
 564
 565 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 566                                              gfn_t gfn,
 567                                              gva_t gaddr,
 568                                              unsigned level,
 569                                              int metaphysical,
 570                                              unsigned hugepage_access,
 571                                              u64 *parent_pte)
 572 {
 573         union kvm_mmu_page_role role;
 574         unsigned index;
 575         unsigned quadrant;
 576         struct hlist_head *bucket;
 577         struct kvm_mmu_page *page;
 578         struct hlist_node *node;
 579
 580         role.word = 0;
 581         role.glevels = vcpu->mmu.root_level;
 582         role.level = level;
 583         role.metaphysical = metaphysical;
 584         role.hugepage_access = hugepage_access;
 585         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 586                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 587                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 588                 role.quadrant = quadrant;
 589         }
 590         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 591                  gfn, role.word);
 592         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 593         bucket = &vcpu->kvm->mmu_page_hash[index];
 594         hlist_for_each_entry(page, node, bucket, hash_link)
 595                 if (page->gfn == gfn && page->role.word == role.word) {
 596                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 597                         pgprintk("%s: found\n", __FUNCTION__);
 598                         return page;
 599                 }
 600         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 601         if (!page)
 602                 return page;
 603         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 604         page->gfn = gfn;
 605         page->role = role;
 606         hlist_add_head(&page->hash_link, bucket);
 607         if (!metaphysical)
 608                 rmap_write_protect(vcpu, gfn);
 609         return page;
 610 }
 611
 612 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 613                                          struct kvm_mmu_page *page)
 614 {
 615         unsigned i;
 616         u64 *pt;
 617         u64 ent;
 618
 619         pt = __va(page->page_hpa);
 620
 621         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 622                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 623                         if (pt[i] & PT_PRESENT_MASK)
 624                                 rmap_remove(vcpu, &pt[i]);
 625                         pt[i] = 0;
 626                 }
 627                 kvm_arch_ops->tlb_flush(vcpu);
 628                 return;
 629         }
 630
 631         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 632                 ent = pt[i];
 633
 634                 pt[i] = 0;
 635                 if (!(ent & PT_PRESENT_MASK))
 636                         continue;
 637                 ent &= PT64_BASE_ADDR_MASK;
 638                 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
 639         }
 640 }
 641
 642 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 643                              struct kvm_mmu_page *page,
 644                              u64 *parent_pte)
 645 {
 646         mmu_page_remove_parent_pte(vcpu, page, parent_pte);
 647 }
 648
 649 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 650                              struct kvm_mmu_page *page)
 651 {
 652         u64 *parent_pte;
 653
 654         while (page->multimapped || page->parent_pte) {
 655                 if (!page->multimapped)
 656                         parent_pte = page->parent_pte;
 657                 else {
 658                         struct kvm_pte_chain *chain;
 659
 660                         chain = container_of(page->parent_ptes.first,
 661                                              struct kvm_pte_chain, link);
 662                         parent_pte = chain->parent_ptes[0];
 663                 }
 664                 BUG_ON(!parent_pte);
 665                 kvm_mmu_put_page(vcpu, page, parent_pte);
 666                 *parent_pte = 0;
 667         }
 668         kvm_mmu_page_unlink_children(vcpu, page);
 669         if (!page->root_count) {
 670                 hlist_del(&page->hash_link);
 671                 kvm_mmu_free_page(vcpu, page->page_hpa);
 672         } else
 673                 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 674 }
 675
 676 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 677 {
 678         unsigned index;
 679         struct hlist_head *bucket;
 680         struct kvm_mmu_page *page;
 681         struct hlist_node *node, *n;
 682         int r;
 683
 684         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 685         r = 0;
 686         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 687         bucket = &vcpu->kvm->mmu_page_hash[index];
 688         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 689                 if (page->gfn == gfn && !page->role.metaphysical) {
 690                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 691                                  page->role.word);
 692                         kvm_mmu_zap_page(vcpu, page);
 693                         r = 1;
 694                 }
 695         return r;
 696 }
 697
 698 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 699 {
 700         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 701         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 702
 703         __set_bit(slot, &page_head->slot_bitmap);
 704 }
 705
 706 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 707 {
 708         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 709
 710         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 711 }
 712
 713 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 714 {
 715         struct page *page;
 716
 717         ASSERT((gpa & HPA_ERR_MASK) == 0);
 718         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 719         if (!page)
 720                 return gpa | HPA_ERR_MASK;
 721         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 722                 | (gpa & (PAGE_SIZE-1));
 723 }
 724
 725 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 726 {
 727         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 728
 729         if (gpa == UNMAPPED_GVA)
 730                 return UNMAPPED_GVA;
 731         return gpa_to_hpa(vcpu, gpa);
 732 }
 733
 734 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 735 {
 736         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 737
 738         if (gpa == UNMAPPED_GVA)
 739                 return NULL;
 740         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 741 }
 742
 743 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 744 {
 745 }
 746
 747 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 748 {
 749         int level = PT32E_ROOT_LEVEL;
 750         hpa_t table_addr = vcpu->mmu.root_hpa;
 751
 752         for (; ; level--) {
 753                 u32 index = PT64_INDEX(v, level);
 754                 u64 *table;
 755                 u64 pte;
 756
 757                 ASSERT(VALID_PAGE(table_addr));
 758                 table = __va(table_addr);
 759
 760                 if (level == 1) {
 761                         pte = table[index];
 762                         if (is_present_pte(pte) && is_writeble_pte(pte))
 763                                 return 0;
 764                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 765                         page_header_update_slot(vcpu->kvm, table, v);
 766                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 767                                                                 PT_USER_MASK;
 768                         rmap_add(vcpu, &table[index]);
 769                         return 0;
 770                 }
 771
 772                 if (table[index] == 0) {
 773                         struct kvm_mmu_page *new_table;
 774                         gfn_t pseudo_gfn;
 775
 776                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 777                                 >> PAGE_SHIFT;
 778                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 779                                                      v, level - 1,
 780                                                      1, 0, &table[index]);
 781                         if (!new_table) {
 782                                 pgprintk("nonpaging_map: ENOMEM\n");
 783                                 return -ENOMEM;
 784                         }
 785
 786                         table[index] = new_table->page_hpa | PT_PRESENT_MASK
 787                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 788                 }
 789                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 790         }
 791 }
 792
 793 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 794 {
 795         int i;
 796         struct kvm_mmu_page *page;
 797
 798 #ifdef CONFIG_X86_64
 799         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 800                 hpa_t root = vcpu->mmu.root_hpa;
 801
 802                 ASSERT(VALID_PAGE(root));
 803                 page = page_header(root);
 804                 --page->root_count;
 805                 vcpu->mmu.root_hpa = INVALID_PAGE;
 806                 return;
 807         }
 808 #endif
 809         for (i = 0; i < 4; ++i) {
 810                 hpa_t root = vcpu->mmu.pae_root[i];
 811
 812                 if (root) {
 813                         ASSERT(VALID_PAGE(root));
 814                         root &= PT64_BASE_ADDR_MASK;
 815                         page = page_header(root);
 816                         --page->root_count;
 817                 }
 818                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 819         }
 820         vcpu->mmu.root_hpa = INVALID_PAGE;
 821 }
 822
 823 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 824 {
 825         int i;
 826         gfn_t root_gfn;
 827         struct kvm_mmu_page *page;
 828
 829         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 830
 831 #ifdef CONFIG_X86_64
 832         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 833                 hpa_t root = vcpu->mmu.root_hpa;
 834
 835                 ASSERT(!VALID_PAGE(root));
 836                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 837                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 838                 root = page->page_hpa;
 839                 ++page->root_count;
 840                 vcpu->mmu.root_hpa = root;
 841                 return;
 842         }
 843 #endif
 844         for (i = 0; i < 4; ++i) {
 845                 hpa_t root = vcpu->mmu.pae_root[i];
 846
 847                 ASSERT(!VALID_PAGE(root));
 848                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 849                         if (!is_present_pte(vcpu->pdptrs[i])) {
 850                                 vcpu->mmu.pae_root[i] = 0;
 851                                 continue;
 852                         }
 853                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 854                 } else if (vcpu->mmu.root_level == 0)
 855                         root_gfn = 0;
 856                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 857                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 858                                         0, NULL);
 859                 root = page->page_hpa;
 860                 ++page->root_count;
 861                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 862         }
 863         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 864 }
 865
 866 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 867 {
 868         return vaddr;
 869 }
 870
 871 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 872                                u32 error_code)
 873 {
 874         gpa_t addr = gva;
 875         hpa_t paddr;
 876         int r;
 877
 878         r = mmu_topup_memory_caches(vcpu);
 879         if (r)
 880                 return r;
 881
 882         ASSERT(vcpu);
 883         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 884
 885
 886         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 887
 888         if (is_error_hpa(paddr))
 889                 return 1;
 890
 891         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 892 }
 893
 894 static void nonpaging_free(struct kvm_vcpu *vcpu)
 895 {
 896         mmu_free_roots(vcpu);
 897 }
 898
 899 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 900 {
 901         struct kvm_mmu *context = &vcpu->mmu;
 902
 903         context->new_cr3 = nonpaging_new_cr3;
 904         context->page_fault = nonpaging_page_fault;
 905         context->gva_to_gpa = nonpaging_gva_to_gpa;
 906         context->free = nonpaging_free;
 907         context->root_level = 0;
 908         context->shadow_root_level = PT32E_ROOT_LEVEL;
 909         mmu_alloc_roots(vcpu);
 910         ASSERT(VALID_PAGE(context->root_hpa));
 911         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 912         return 0;
 913 }
 914
 915 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 916 {
 917         ++kvm_stat.tlb_flush;
 918         kvm_arch_ops->tlb_flush(vcpu);
 919 }
 920
 921 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 922 {
 923         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 924         mmu_free_roots(vcpu);
 925         if (unlikely(vcpu->kvm->n_free_mmu_pages < KVM_MIN_FREE_MMU_PAGES))
 926                 kvm_mmu_free_some_pages(vcpu);
 927         mmu_alloc_roots(vcpu);
 928         kvm_mmu_flush_tlb(vcpu);
 929         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 930 }
 931
 932 static inline void set_pte_common(struct kvm_vcpu *vcpu,
 933                              u64 *shadow_pte,
 934                              gpa_t gaddr,
 935                              int dirty,
 936                              u64 access_bits,
 937                              gfn_t gfn)
 938 {
 939         hpa_t paddr;
 940
 941         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 942         if (!dirty)
 943                 access_bits &= ~PT_WRITABLE_MASK;
 944
 945         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 946
 947         *shadow_pte |= access_bits;
 948
 949         if (is_error_hpa(paddr)) {
 950                 *shadow_pte |= gaddr;
 951                 *shadow_pte |= PT_SHADOW_IO_MARK;
 952                 *shadow_pte &= ~PT_PRESENT_MASK;
 953                 return;
 954         }
 955
 956         *shadow_pte |= paddr;
 957
 958         if (access_bits & PT_WRITABLE_MASK) {
 959                 struct kvm_mmu_page *shadow;
 960
 961                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 962                 if (shadow) {
 963                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 964                                  __FUNCTION__, gfn);
 965                         access_bits &= ~PT_WRITABLE_MASK;
 966                         if (is_writeble_pte(*shadow_pte)) {
 967                                     *shadow_pte &= ~PT_WRITABLE_MASK;
 968                                     kvm_arch_ops->tlb_flush(vcpu);
 969                         }
 970                 }
 971         }
 972
 973         if (access_bits & PT_WRITABLE_MASK)
 974                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 975
 976         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 977         rmap_add(vcpu, shadow_pte);
 978 }
 979
 980 static void inject_page_fault(struct kvm_vcpu *vcpu,
 981                               u64 addr,
 982                               u32 err_code)
 983 {
 984         kvm_arch_ops->inject_page_fault(vcpu, addr, err_code);
 985 }
 986
 987 static inline int fix_read_pf(u64 *shadow_ent)
 988 {
 989         if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
 990             !(*shadow_ent & PT_USER_MASK)) {
 991                 /*
 992                  * If supervisor write protect is disabled, we shadow kernel
 993                  * pages as user pages so we can trap the write access.
 994                  */
 995                 *shadow_ent |= PT_USER_MASK;
 996                 *shadow_ent &= ~PT_WRITABLE_MASK;
 997
 998                 return 1;
 999
1000         }
1001         return 0;
1002 }
1003
1004 static void paging_free(struct kvm_vcpu *vcpu)
1005 {
1006         nonpaging_free(vcpu);
1007 }
1008
1009 #define PTTYPE 64
1010 #include "paging_tmpl.h"
1011 #undef PTTYPE
1012
1013 #define PTTYPE 32
1014 #include "paging_tmpl.h"
1015 #undef PTTYPE
1016
1017 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1018 {
1019         struct kvm_mmu *context = &vcpu->mmu;
1020
1021         ASSERT(is_pae(vcpu));
1022         context->new_cr3 = paging_new_cr3;
1023         context->page_fault = paging64_page_fault;
1024         context->gva_to_gpa = paging64_gva_to_gpa;
1025         context->free = paging_free;
1026         context->root_level = level;
1027         context->shadow_root_level = 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 paging64_init_context(struct kvm_vcpu *vcpu)
1036 {
1037         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1038 }
1039
1040 static int paging32_init_context(struct kvm_vcpu *vcpu)
1041 {
1042         struct kvm_mmu *context = &vcpu->mmu;
1043
1044         context->new_cr3 = paging_new_cr3;
1045         context->page_fault = paging32_page_fault;
1046         context->gva_to_gpa = paging32_gva_to_gpa;
1047         context->free = paging_free;
1048         context->root_level = PT32_ROOT_LEVEL;
1049         context->shadow_root_level = PT32E_ROOT_LEVEL;
1050         mmu_alloc_roots(vcpu);
1051         ASSERT(VALID_PAGE(context->root_hpa));
1052         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1053                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1054         return 0;
1055 }
1056
1057 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1058 {
1059         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1060 }
1061
1062 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1063 {
1064         ASSERT(vcpu);
1065         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1066
1067         if (!is_paging(vcpu))
1068                 return nonpaging_init_context(vcpu);
1069         else if (is_long_mode(vcpu))
1070                 return paging64_init_context(vcpu);
1071         else if (is_pae(vcpu))
1072                 return paging32E_init_context(vcpu);
1073         else
1074                 return paging32_init_context(vcpu);
1075 }
1076
1077 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1078 {
1079         ASSERT(vcpu);
1080         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1081                 vcpu->mmu.free(vcpu);
1082                 vcpu->mmu.root_hpa = INVALID_PAGE;
1083         }
1084 }
1085
1086 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1087 {
1088         int r;
1089
1090         destroy_kvm_mmu(vcpu);
1091         r = init_kvm_mmu(vcpu);
1092         if (r < 0)
1093                 goto out;
1094         r = mmu_topup_memory_caches(vcpu);
1095 out:
1096         return r;
1097 }
1098
1099 static void mmu_pre_write_zap_pte(struct kvm_vcpu *vcpu,
1100                                   struct kvm_mmu_page *page,
1101                                   u64 *spte)
1102 {
1103         u64 pte;
1104         struct kvm_mmu_page *child;
1105
1106         pte = *spte;
1107         if (is_present_pte(pte)) {
1108                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1109                         rmap_remove(vcpu, spte);
1110                 else {
1111                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1112                         mmu_page_remove_parent_pte(vcpu, child, spte);
1113                 }
1114         }
1115         *spte = 0;
1116 }
1117
1118 void kvm_mmu_pre_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1119 {
1120         gfn_t gfn = gpa >> PAGE_SHIFT;
1121         struct kvm_mmu_page *page;
1122         struct hlist_node *node, *n;
1123         struct hlist_head *bucket;
1124         unsigned index;
1125         u64 *spte;
1126         unsigned offset = offset_in_page(gpa);
1127         unsigned pte_size;
1128         unsigned page_offset;
1129         unsigned misaligned;
1130         int level;
1131         int flooded = 0;
1132         int npte;
1133
1134         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1135         if (gfn == vcpu->last_pt_write_gfn) {
1136                 ++vcpu->last_pt_write_count;
1137                 if (vcpu->last_pt_write_count >= 3)
1138                         flooded = 1;
1139         } else {
1140                 vcpu->last_pt_write_gfn = gfn;
1141                 vcpu->last_pt_write_count = 1;
1142         }
1143         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1144         bucket = &vcpu->kvm->mmu_page_hash[index];
1145         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1146                 if (page->gfn != gfn || page->role.metaphysical)
1147                         continue;
1148                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1149                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1150                 if (misaligned || flooded) {
1151                         /*
1152                          * Misaligned accesses are too much trouble to fix
1153                          * up; also, they usually indicate a page is not used
1154                          * as a page table.
1155                          *
1156                          * If we're seeing too many writes to a page,
1157                          * it may no longer be a page table, or we may be
1158                          * forking, in which case it is better to unmap the
1159                          * page.
1160                          */
1161                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1162                                  gpa, bytes, page->role.word);
1163                         kvm_mmu_zap_page(vcpu, page);
1164                         continue;
1165                 }
1166                 page_offset = offset;
1167                 level = page->role.level;
1168                 npte = 1;
1169                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1170                         page_offset <<= 1;      /* 32->64 */
1171                         /*
1172                          * A 32-bit pde maps 4MB while the shadow pdes map
1173                          * only 2MB.  So we need to double the offset again
1174                          * and zap two pdes instead of one.
1175                          */
1176                         if (level == PT32_ROOT_LEVEL) {
1177                                 page_offset &= ~7; /* kill rounding error */
1178                                 page_offset <<= 1;
1179                                 npte = 2;
1180                         }
1181                         page_offset &= ~PAGE_MASK;
1182                 }
1183                 spte = __va(page->page_hpa);
1184                 spte += page_offset / sizeof(*spte);
1185                 while (npte--) {
1186                         mmu_pre_write_zap_pte(vcpu, page, spte);
1187                         ++spte;
1188                 }
1189         }
1190 }
1191
1192 void kvm_mmu_post_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1193 {
1194 }
1195
1196 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1197 {
1198         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1199
1200         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1201 }
1202
1203 void kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1204 {
1205         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1206                 struct kvm_mmu_page *page;
1207
1208                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1209                                     struct kvm_mmu_page, link);
1210                 kvm_mmu_zap_page(vcpu, page);
1211         }
1212 }
1213 EXPORT_SYMBOL_GPL(kvm_mmu_free_some_pages);
1214
1215 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1216 {
1217         struct kvm_mmu_page *page;
1218
1219         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1220                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1221                                     struct kvm_mmu_page, link);
1222                 kvm_mmu_zap_page(vcpu, page);
1223         }
1224         while (!list_empty(&vcpu->free_pages)) {
1225                 page = list_entry(vcpu->free_pages.next,
1226                                   struct kvm_mmu_page, link);
1227                 list_del(&page->link);
1228                 __free_page(pfn_to_page(page->page_hpa >> PAGE_SHIFT));
1229                 page->page_hpa = INVALID_PAGE;
1230         }
1231         free_page((unsigned long)vcpu->mmu.pae_root);
1232 }
1233
1234 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1235 {
1236         struct page *page;
1237         int i;
1238
1239         ASSERT(vcpu);
1240
1241         for (i = 0; i < KVM_NUM_MMU_PAGES; i++) {
1242                 struct kvm_mmu_page *page_header = &vcpu->page_header_buf[i];
1243
1244                 INIT_LIST_HEAD(&page_header->link);
1245                 if ((page = alloc_page(GFP_KERNEL)) == NULL)
1246                         goto error_1;
1247                 set_page_private(page, (unsigned long)page_header);
1248                 page_header->page_hpa = (hpa_t)page_to_pfn(page) << PAGE_SHIFT;
1249                 memset(__va(page_header->page_hpa), 0, PAGE_SIZE);
1250                 list_add(&page_header->link, &vcpu->free_pages);
1251                 ++vcpu->kvm->n_free_mmu_pages;
1252         }
1253
1254         /*
1255          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1256          * Therefore we need to allocate shadow page tables in the first
1257          * 4GB of memory, which happens to fit the DMA32 zone.
1258          */
1259         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1260         if (!page)
1261                 goto error_1;
1262         vcpu->mmu.pae_root = page_address(page);
1263         for (i = 0; i < 4; ++i)
1264                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1265
1266         return 0;
1267
1268 error_1:
1269         free_mmu_pages(vcpu);
1270         return -ENOMEM;
1271 }
1272
1273 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1274 {
1275         ASSERT(vcpu);
1276         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1277         ASSERT(list_empty(&vcpu->free_pages));
1278
1279         return alloc_mmu_pages(vcpu);
1280 }
1281
1282 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1283 {
1284         ASSERT(vcpu);
1285         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1286         ASSERT(!list_empty(&vcpu->free_pages));
1287
1288         return init_kvm_mmu(vcpu);
1289 }
1290
1291 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1292 {
1293         ASSERT(vcpu);
1294
1295         destroy_kvm_mmu(vcpu);
1296         free_mmu_pages(vcpu);
1297         mmu_free_memory_caches(vcpu);
1298 }
1299
1300 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1301 {
1302         struct kvm *kvm = vcpu->kvm;
1303         struct kvm_mmu_page *page;
1304
1305         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1306                 int i;
1307                 u64 *pt;
1308
1309                 if (!test_bit(slot, &page->slot_bitmap))
1310                         continue;
1311
1312                 pt = __va(page->page_hpa);
1313                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1314                         /* avoid RMW */
1315                         if (pt[i] & PT_WRITABLE_MASK) {
1316                                 rmap_remove(vcpu, &pt[i]);
1317                                 pt[i] &= ~PT_WRITABLE_MASK;
1318                         }
1319         }
1320 }
1321
1322 void kvm_mmu_zap_all(struct kvm_vcpu *vcpu)
1323 {
1324         destroy_kvm_mmu(vcpu);
1325
1326         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1327                 struct kvm_mmu_page *page;
1328
1329                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1330                                     struct kvm_mmu_page, link);
1331                 kvm_mmu_zap_page(vcpu, page);
1332         }
1333
1334         mmu_free_memory_caches(vcpu);
1335         kvm_arch_ops->tlb_flush(vcpu);
1336         init_kvm_mmu(vcpu);
1337 }
1338
1339 void kvm_mmu_module_exit(void)
1340 {
1341         if (pte_chain_cache)
1342                 kmem_cache_destroy(pte_chain_cache);
1343         if (rmap_desc_cache)
1344                 kmem_cache_destroy(rmap_desc_cache);
1345 }
1346
1347 int kvm_mmu_module_init(void)
1348 {
1349         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1350                                             sizeof(struct kvm_pte_chain),
1351                                             0, 0, NULL, NULL);
1352         if (!pte_chain_cache)
1353                 goto nomem;
1354         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1355                                             sizeof(struct kvm_rmap_desc),
1356                                             0, 0, NULL, NULL);
1357         if (!rmap_desc_cache)
1358                 goto nomem;
1359
1360         return 0;
1361
1362 nomem:
1363         kvm_mmu_module_exit();
1364         return -ENOMEM;
1365 }
1366
1367 #ifdef AUDIT
1368
1369 static const char *audit_msg;
1370
1371 static gva_t canonicalize(gva_t gva)
1372 {
1373 #ifdef CONFIG_X86_64
1374         gva = (long long)(gva << 16) >> 16;
1375 #endif
1376         return gva;
1377 }
1378
1379 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1380                                 gva_t va, int level)
1381 {
1382         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1383         int i;
1384         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1385
1386         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1387                 u64 ent = pt[i];
1388
1389                 if (!ent & PT_PRESENT_MASK)
1390                         continue;
1391
1392                 va = canonicalize(va);
1393                 if (level > 1)
1394                         audit_mappings_page(vcpu, ent, va, level - 1);
1395                 else {
1396                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1397                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1398
1399                         if ((ent & PT_PRESENT_MASK)
1400                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1401                                 printk(KERN_ERR "audit error: (%s) levels %d"
1402                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1403                                        audit_msg, vcpu->mmu.root_level,
1404                                        va, gpa, hpa, ent);
1405                 }
1406         }
1407 }
1408
1409 static void audit_mappings(struct kvm_vcpu *vcpu)
1410 {
1411         unsigned i;
1412
1413         if (vcpu->mmu.root_level == 4)
1414                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1415         else
1416                 for (i = 0; i < 4; ++i)
1417                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1418                                 audit_mappings_page(vcpu,
1419                                                     vcpu->mmu.pae_root[i],
1420                                                     i << 30,
1421                                                     2);
1422 }
1423
1424 static int count_rmaps(struct kvm_vcpu *vcpu)
1425 {
1426         int nmaps = 0;
1427         int i, j, k;
1428
1429         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1430                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1431                 struct kvm_rmap_desc *d;
1432
1433                 for (j = 0; j < m->npages; ++j) {
1434                         struct page *page = m->phys_mem[j];
1435
1436                         if (!page->private)
1437                                 continue;
1438                         if (!(page->private & 1)) {
1439                                 ++nmaps;
1440                                 continue;
1441                         }
1442                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1443                         while (d) {
1444                                 for (k = 0; k < RMAP_EXT; ++k)
1445                                         if (d->shadow_ptes[k])
1446                                                 ++nmaps;
1447                                         else
1448                                                 break;
1449                                 d = d->more;
1450                         }
1451                 }
1452         }
1453         return nmaps;
1454 }
1455
1456 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1457 {
1458         int nmaps = 0;
1459         struct kvm_mmu_page *page;
1460         int i;
1461
1462         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1463                 u64 *pt = __va(page->page_hpa);
1464
1465                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1466                         continue;
1467
1468                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1469                         u64 ent = pt[i];
1470
1471                         if (!(ent & PT_PRESENT_MASK))
1472                                 continue;
1473                         if (!(ent & PT_WRITABLE_MASK))
1474                                 continue;
1475                         ++nmaps;
1476                 }
1477         }
1478         return nmaps;
1479 }
1480
1481 static void audit_rmap(struct kvm_vcpu *vcpu)
1482 {
1483         int n_rmap = count_rmaps(vcpu);
1484         int n_actual = count_writable_mappings(vcpu);
1485
1486         if (n_rmap != n_actual)
1487                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1488                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1489 }
1490
1491 static void audit_write_protection(struct kvm_vcpu *vcpu)
1492 {
1493         struct kvm_mmu_page *page;
1494
1495         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1496                 hfn_t hfn;
1497                 struct page *pg;
1498
1499                 if (page->role.metaphysical)
1500                         continue;
1501
1502                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1503                         >> PAGE_SHIFT;
1504                 pg = pfn_to_page(hfn);
1505                 if (pg->private)
1506                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1507                                " mappings: gfn %lx role %x\n",
1508                                __FUNCTION__, audit_msg, page->gfn,
1509                                page->role.word);
1510         }
1511 }
1512
1513 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1514 {
1515         int olddbg = dbg;
1516
1517         dbg = 0;
1518         audit_msg = msg;
1519         audit_rmap(vcpu);
1520         audit_write_protection(vcpu);
1521         audit_mappings(vcpu);
1522         dbg = olddbg;
1523 }
1524
1525 #endif