drivers/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19 #include <linux/types.h>
  20 #include <linux/string.h>
  21 #include <asm/page.h>
  22 #include <linux/mm.h>
  23 #include <linux/highmem.h>
  24 #include <linux/module.h>
  25
  26 #include "vmx.h"
  27 #include "kvm.h"
  28
  29 #undef MMU_DEBUG
  30
  31 #undef AUDIT
  32
  33 #ifdef AUDIT
  34 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  35 #else
  36 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  37 #endif
  38
  39 #ifdef MMU_DEBUG
  40
  41 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  42 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  43
  44 #else
  45
  46 #define pgprintk(x...) do { } while (0)
  47 #define rmap_printk(x...) do { } while (0)
  48
  49 #endif
  50
  51 #if defined(MMU_DEBUG) || defined(AUDIT)
  52 static int dbg = 1;
  53 #endif
  54
  55 #define ASSERT(x)                                                       \
  56         if (!(x)) {                                                     \
  57                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  58                        __FILE__, __LINE__, #x);                         \
  59         }
  60
  61 #define PT64_PT_BITS 9
  62 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  63 #define PT32_PT_BITS 10
  64 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  65
  66 #define PT_WRITABLE_SHIFT 1
  67
  68 #define PT_PRESENT_MASK (1ULL << 0)
  69 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  70 #define PT_USER_MASK (1ULL << 2)
  71 #define PT_PWT_MASK (1ULL << 3)
  72 #define PT_PCD_MASK (1ULL << 4)
  73 #define PT_ACCESSED_MASK (1ULL << 5)
  74 #define PT_DIRTY_MASK (1ULL << 6)
  75 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  76 #define PT_PAT_MASK (1ULL << 7)
  77 #define PT_GLOBAL_MASK (1ULL << 8)
  78 #define PT64_NX_MASK (1ULL << 63)
  79
  80 #define PT_PAT_SHIFT 7
  81 #define PT_DIR_PAT_SHIFT 12
  82 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  83
  84 #define PT32_DIR_PSE36_SIZE 4
  85 #define PT32_DIR_PSE36_SHIFT 13
  86 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  87
  88
  89 #define PT32_PTE_COPY_MASK \
  90         (PT_PRESENT_MASK | PT_ACCESSED_MASK | PT_DIRTY_MASK | PT_GLOBAL_MASK)
  91
  92 #define PT64_PTE_COPY_MASK (PT64_NX_MASK | PT32_PTE_COPY_MASK)
  93
  94 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  95 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  96
  97 #define PT_SHADOW_PS_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
  98 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
  99
 100 #define PT_SHADOW_WRITABLE_SHIFT (PT_FIRST_AVAIL_BITS_SHIFT + 1)
 101 #define PT_SHADOW_WRITABLE_MASK (1ULL << PT_SHADOW_WRITABLE_SHIFT)
 102
 103 #define PT_SHADOW_USER_SHIFT (PT_SHADOW_WRITABLE_SHIFT + 1)
 104 #define PT_SHADOW_USER_MASK (1ULL << (PT_SHADOW_USER_SHIFT))
 105
 106 #define PT_SHADOW_BITS_OFFSET (PT_SHADOW_WRITABLE_SHIFT - PT_WRITABLE_SHIFT)
 107
 108 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 109
 110 #define PT64_LEVEL_BITS 9
 111
 112 #define PT64_LEVEL_SHIFT(level) \
 113                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
 114
 115 #define PT64_LEVEL_MASK(level) \
 116                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 117
 118 #define PT64_INDEX(address, level)\
 119         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 120
 121
 122 #define PT32_LEVEL_BITS 10
 123
 124 #define PT32_LEVEL_SHIFT(level) \
 125                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
 126
 127 #define PT32_LEVEL_MASK(level) \
 128                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 129
 130 #define PT32_INDEX(address, level)\
 131         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 132
 133
 134 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 135 #define PT64_DIR_BASE_ADDR_MASK \
 136         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 137
 138 #define PT32_BASE_ADDR_MASK PAGE_MASK
 139 #define PT32_DIR_BASE_ADDR_MASK \
 140         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 141
 142
 143 #define PFERR_PRESENT_MASK (1U << 0)
 144 #define PFERR_WRITE_MASK (1U << 1)
 145 #define PFERR_USER_MASK (1U << 2)
 146 #define PFERR_FETCH_MASK (1U << 4)
 147
 148 #define PT64_ROOT_LEVEL 4
 149 #define PT32_ROOT_LEVEL 2
 150 #define PT32E_ROOT_LEVEL 3
 151
 152 #define PT_DIRECTORY_LEVEL 2
 153 #define PT_PAGE_TABLE_LEVEL 1
 154
 155 #define RMAP_EXT 4
 156
 157 struct kvm_rmap_desc {
 158         u64 *shadow_ptes[RMAP_EXT];
 159         struct kvm_rmap_desc *more;
 160 };
 161
 162 static int is_write_protection(struct kvm_vcpu *vcpu)
 163 {
 164         return vcpu->cr0 & CR0_WP_MASK;
 165 }
 166
 167 static int is_cpuid_PSE36(void)
 168 {
 169         return 1;
 170 }
 171
 172 static int is_nx(struct kvm_vcpu *vcpu)
 173 {
 174         return vcpu->shadow_efer & EFER_NX;
 175 }
 176
 177 static int is_present_pte(unsigned long pte)
 178 {
 179         return pte & PT_PRESENT_MASK;
 180 }
 181
 182 static int is_writeble_pte(unsigned long pte)
 183 {
 184         return pte & PT_WRITABLE_MASK;
 185 }
 186
 187 static int is_io_pte(unsigned long pte)
 188 {
 189         return pte & PT_SHADOW_IO_MARK;
 190 }
 191
 192 static int is_rmap_pte(u64 pte)
 193 {
 194         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 195                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 196 }
 197
 198 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 199                                   size_t objsize, int min)
 200 {
 201         void *obj;
 202
 203         if (cache->nobjs >= min)
 204                 return 0;
 205         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 206                 obj = kzalloc(objsize, GFP_NOWAIT);
 207                 if (!obj)
 208                         return -ENOMEM;
 209                 cache->objects[cache->nobjs++] = obj;
 210         }
 211         return 0;
 212 }
 213
 214 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 215 {
 216         while (mc->nobjs)
 217                 kfree(mc->objects[--mc->nobjs]);
 218 }
 219
 220 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 221 {
 222         int r;
 223
 224         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 225                                    sizeof(struct kvm_pte_chain), 4);
 226         if (r)
 227                 goto out;
 228         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 229                                    sizeof(struct kvm_rmap_desc), 1);
 230 out:
 231         return r;
 232 }
 233
 234 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 235 {
 236         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 237         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 238 }
 239
 240 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 241                                     size_t size)
 242 {
 243         void *p;
 244
 245         BUG_ON(!mc->nobjs);
 246         p = mc->objects[--mc->nobjs];
 247         memset(p, 0, size);
 248         return p;
 249 }
 250
 251 static void mmu_memory_cache_free(struct kvm_mmu_memory_cache *mc, void *obj)
 252 {
 253         if (mc->nobjs < KVM_NR_MEM_OBJS)
 254                 mc->objects[mc->nobjs++] = obj;
 255         else
 256                 kfree(obj);
 257 }
 258
 259 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 260 {
 261         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 262                                       sizeof(struct kvm_pte_chain));
 263 }
 264
 265 static void mmu_free_pte_chain(struct kvm_vcpu *vcpu,
 266                                struct kvm_pte_chain *pc)
 267 {
 268         mmu_memory_cache_free(&vcpu->mmu_pte_chain_cache, pc);
 269 }
 270
 271 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 272 {
 273         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 274                                       sizeof(struct kvm_rmap_desc));
 275 }
 276
 277 static void mmu_free_rmap_desc(struct kvm_vcpu *vcpu,
 278                                struct kvm_rmap_desc *rd)
 279 {
 280         mmu_memory_cache_free(&vcpu->mmu_rmap_desc_cache, rd);
 281 }
 282
 283 /*
 284  * Reverse mapping data structures:
 285  *
 286  * If page->private bit zero is zero, then page->private points to the
 287  * shadow page table entry that points to page_address(page).
 288  *
 289  * If page->private bit zero is one, (then page->private & ~1) points
 290  * to a struct kvm_rmap_desc containing more mappings.
 291  */
 292 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 293 {
 294         struct page *page;
 295         struct kvm_rmap_desc *desc;
 296         int i;
 297
 298         if (!is_rmap_pte(*spte))
 299                 return;
 300         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 301         if (!page_private(page)) {
 302                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 303                 set_page_private(page,(unsigned long)spte);
 304         } else if (!(page_private(page) & 1)) {
 305                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 306                 desc = mmu_alloc_rmap_desc(vcpu);
 307                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 308                 desc->shadow_ptes[1] = spte;
 309                 set_page_private(page,(unsigned long)desc | 1);
 310         } else {
 311                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 312                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 313                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 314                         desc = desc->more;
 315                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 316                         desc->more = mmu_alloc_rmap_desc(vcpu);
 317                         desc = desc->more;
 318                 }
 319                 for (i = 0; desc->shadow_ptes[i]; ++i)
 320                         ;
 321                 desc->shadow_ptes[i] = spte;
 322         }
 323 }
 324
 325 static void rmap_desc_remove_entry(struct kvm_vcpu *vcpu,
 326                                    struct page *page,
 327                                    struct kvm_rmap_desc *desc,
 328                                    int i,
 329                                    struct kvm_rmap_desc *prev_desc)
 330 {
 331         int j;
 332
 333         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 334                 ;
 335         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 336         desc->shadow_ptes[j] = NULL;
 337         if (j != 0)
 338                 return;
 339         if (!prev_desc && !desc->more)
 340                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 341         else
 342                 if (prev_desc)
 343                         prev_desc->more = desc->more;
 344                 else
 345                         set_page_private(page,(unsigned long)desc->more | 1);
 346         mmu_free_rmap_desc(vcpu, desc);
 347 }
 348
 349 static void rmap_remove(struct kvm_vcpu *vcpu, u64 *spte)
 350 {
 351         struct page *page;
 352         struct kvm_rmap_desc *desc;
 353         struct kvm_rmap_desc *prev_desc;
 354         int i;
 355
 356         if (!is_rmap_pte(*spte))
 357                 return;
 358         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 359         if (!page_private(page)) {
 360                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 361                 BUG();
 362         } else if (!(page_private(page) & 1)) {
 363                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 364                 if ((u64 *)page_private(page) != spte) {
 365                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 366                                spte, *spte);
 367                         BUG();
 368                 }
 369                 set_page_private(page,0);
 370         } else {
 371                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 372                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 373                 prev_desc = NULL;
 374                 while (desc) {
 375                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 376                                 if (desc->shadow_ptes[i] == spte) {
 377                                         rmap_desc_remove_entry(vcpu, page,
 378                                                                desc, i,
 379                                                                prev_desc);
 380                                         return;
 381                                 }
 382                         prev_desc = desc;
 383                         desc = desc->more;
 384                 }
 385                 BUG();
 386         }
 387 }
 388
 389 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 390 {
 391         struct kvm *kvm = vcpu->kvm;
 392         struct page *page;
 393         struct kvm_memory_slot *slot;
 394         struct kvm_rmap_desc *desc;
 395         u64 *spte;
 396
 397         slot = gfn_to_memslot(kvm, gfn);
 398         BUG_ON(!slot);
 399         page = gfn_to_page(slot, gfn);
 400
 401         while (page_private(page)) {
 402                 if (!(page_private(page) & 1))
 403                         spte = (u64 *)page_private(page);
 404                 else {
 405                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 406                         spte = desc->shadow_ptes[0];
 407                 }
 408                 BUG_ON(!spte);
 409                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 410                        != page_to_pfn(page));
 411                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 412                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 413                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 414                 rmap_remove(vcpu, spte);
 415                 kvm_arch_ops->tlb_flush(vcpu);
 416                 *spte &= ~(u64)PT_WRITABLE_MASK;
 417         }
 418 }
 419
 420 static int is_empty_shadow_page(hpa_t page_hpa)
 421 {
 422         u64 *pos;
 423         u64 *end;
 424
 425         for (pos = __va(page_hpa), end = pos + PAGE_SIZE / sizeof(u64);
 426                       pos != end; pos++)
 427                 if (*pos != 0) {
 428                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 429                                pos, *pos);
 430                         return 0;
 431                 }
 432         return 1;
 433 }
 434
 435 static void kvm_mmu_free_page(struct kvm_vcpu *vcpu, hpa_t page_hpa)
 436 {
 437         struct kvm_mmu_page *page_head = page_header(page_hpa);
 438
 439         ASSERT(is_empty_shadow_page(page_hpa));
 440         page_head->page_hpa = page_hpa;
 441         list_move(&page_head->link, &vcpu->free_pages);
 442         ++vcpu->kvm->n_free_mmu_pages;
 443 }
 444
 445 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 446 {
 447         return gfn;
 448 }
 449
 450 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 451                                                u64 *parent_pte)
 452 {
 453         struct kvm_mmu_page *page;
 454
 455         if (list_empty(&vcpu->free_pages))
 456                 return NULL;
 457
 458         page = list_entry(vcpu->free_pages.next, struct kvm_mmu_page, link);
 459         list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 460         ASSERT(is_empty_shadow_page(page->page_hpa));
 461         page->slot_bitmap = 0;
 462         page->multimapped = 0;
 463         page->parent_pte = parent_pte;
 464         --vcpu->kvm->n_free_mmu_pages;
 465         return page;
 466 }
 467
 468 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 469                                     struct kvm_mmu_page *page, u64 *parent_pte)
 470 {
 471         struct kvm_pte_chain *pte_chain;
 472         struct hlist_node *node;
 473         int i;
 474
 475         if (!parent_pte)
 476                 return;
 477         if (!page->multimapped) {
 478                 u64 *old = page->parent_pte;
 479
 480                 if (!old) {
 481                         page->parent_pte = parent_pte;
 482                         return;
 483                 }
 484                 page->multimapped = 1;
 485                 pte_chain = mmu_alloc_pte_chain(vcpu);
 486                 INIT_HLIST_HEAD(&page->parent_ptes);
 487                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 488                 pte_chain->parent_ptes[0] = old;
 489         }
 490         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 491                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 492                         continue;
 493                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 494                         if (!pte_chain->parent_ptes[i]) {
 495                                 pte_chain->parent_ptes[i] = parent_pte;
 496                                 return;
 497                         }
 498         }
 499         pte_chain = mmu_alloc_pte_chain(vcpu);
 500         BUG_ON(!pte_chain);
 501         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 502         pte_chain->parent_ptes[0] = parent_pte;
 503 }
 504
 505 static void mmu_page_remove_parent_pte(struct kvm_vcpu *vcpu,
 506                                        struct kvm_mmu_page *page,
 507                                        u64 *parent_pte)
 508 {
 509         struct kvm_pte_chain *pte_chain;
 510         struct hlist_node *node;
 511         int i;
 512
 513         if (!page->multimapped) {
 514                 BUG_ON(page->parent_pte != parent_pte);
 515                 page->parent_pte = NULL;
 516                 return;
 517         }
 518         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 519                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 520                         if (!pte_chain->parent_ptes[i])
 521                                 break;
 522                         if (pte_chain->parent_ptes[i] != parent_pte)
 523                                 continue;
 524                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 525                                 && pte_chain->parent_ptes[i + 1]) {
 526                                 pte_chain->parent_ptes[i]
 527                                         = pte_chain->parent_ptes[i + 1];
 528                                 ++i;
 529                         }
 530                         pte_chain->parent_ptes[i] = NULL;
 531                         if (i == 0) {
 532                                 hlist_del(&pte_chain->link);
 533                                 mmu_free_pte_chain(vcpu, pte_chain);
 534                                 if (hlist_empty(&page->parent_ptes)) {
 535                                         page->multimapped = 0;
 536                                         page->parent_pte = NULL;
 537                                 }
 538                         }
 539                         return;
 540                 }
 541         BUG();
 542 }
 543
 544 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 545                                                 gfn_t gfn)
 546 {
 547         unsigned index;
 548         struct hlist_head *bucket;
 549         struct kvm_mmu_page *page;
 550         struct hlist_node *node;
 551
 552         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 553         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 554         bucket = &vcpu->kvm->mmu_page_hash[index];
 555         hlist_for_each_entry(page, node, bucket, hash_link)
 556                 if (page->gfn == gfn && !page->role.metaphysical) {
 557                         pgprintk("%s: found role %x\n",
 558                                  __FUNCTION__, page->role.word);
 559                         return page;
 560                 }
 561         return NULL;
 562 }
 563
 564 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 565                                              gfn_t gfn,
 566                                              gva_t gaddr,
 567                                              unsigned level,
 568                                              int metaphysical,
 569                                              unsigned hugepage_access,
 570                                              u64 *parent_pte)
 571 {
 572         union kvm_mmu_page_role role;
 573         unsigned index;
 574         unsigned quadrant;
 575         struct hlist_head *bucket;
 576         struct kvm_mmu_page *page;
 577         struct hlist_node *node;
 578
 579         role.word = 0;
 580         role.glevels = vcpu->mmu.root_level;
 581         role.level = level;
 582         role.metaphysical = metaphysical;
 583         role.hugepage_access = hugepage_access;
 584         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 585                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 586                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 587                 role.quadrant = quadrant;
 588         }
 589         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 590                  gfn, role.word);
 591         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 592         bucket = &vcpu->kvm->mmu_page_hash[index];
 593         hlist_for_each_entry(page, node, bucket, hash_link)
 594                 if (page->gfn == gfn && page->role.word == role.word) {
 595                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 596                         pgprintk("%s: found\n", __FUNCTION__);
 597                         return page;
 598                 }
 599         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 600         if (!page)
 601                 return page;
 602         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 603         page->gfn = gfn;
 604         page->role = role;
 605         hlist_add_head(&page->hash_link, bucket);
 606         if (!metaphysical)
 607                 rmap_write_protect(vcpu, gfn);
 608         return page;
 609 }
 610
 611 static void kvm_mmu_page_unlink_children(struct kvm_vcpu *vcpu,
 612                                          struct kvm_mmu_page *page)
 613 {
 614         unsigned i;
 615         u64 *pt;
 616         u64 ent;
 617
 618         pt = __va(page->page_hpa);
 619
 620         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 621                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 622                         if (pt[i] & PT_PRESENT_MASK)
 623                                 rmap_remove(vcpu, &pt[i]);
 624                         pt[i] = 0;
 625                 }
 626                 kvm_arch_ops->tlb_flush(vcpu);
 627                 return;
 628         }
 629
 630         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 631                 ent = pt[i];
 632
 633                 pt[i] = 0;
 634                 if (!(ent & PT_PRESENT_MASK))
 635                         continue;
 636                 ent &= PT64_BASE_ADDR_MASK;
 637                 mmu_page_remove_parent_pte(vcpu, page_header(ent), &pt[i]);
 638         }
 639 }
 640
 641 static void kvm_mmu_put_page(struct kvm_vcpu *vcpu,
 642                              struct kvm_mmu_page *page,
 643                              u64 *parent_pte)
 644 {
 645         mmu_page_remove_parent_pte(vcpu, page, parent_pte);
 646 }
 647
 648 static void kvm_mmu_zap_page(struct kvm_vcpu *vcpu,
 649                              struct kvm_mmu_page *page)
 650 {
 651         u64 *parent_pte;
 652
 653         while (page->multimapped || page->parent_pte) {
 654                 if (!page->multimapped)
 655                         parent_pte = page->parent_pte;
 656                 else {
 657                         struct kvm_pte_chain *chain;
 658
 659                         chain = container_of(page->parent_ptes.first,
 660                                              struct kvm_pte_chain, link);
 661                         parent_pte = chain->parent_ptes[0];
 662                 }
 663                 BUG_ON(!parent_pte);
 664                 kvm_mmu_put_page(vcpu, page, parent_pte);
 665                 *parent_pte = 0;
 666         }
 667         kvm_mmu_page_unlink_children(vcpu, page);
 668         if (!page->root_count) {
 669                 hlist_del(&page->hash_link);
 670                 kvm_mmu_free_page(vcpu, page->page_hpa);
 671         } else
 672                 list_move(&page->link, &vcpu->kvm->active_mmu_pages);
 673 }
 674
 675 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 676 {
 677         unsigned index;
 678         struct hlist_head *bucket;
 679         struct kvm_mmu_page *page;
 680         struct hlist_node *node, *n;
 681         int r;
 682
 683         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 684         r = 0;
 685         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 686         bucket = &vcpu->kvm->mmu_page_hash[index];
 687         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 688                 if (page->gfn == gfn && !page->role.metaphysical) {
 689                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 690                                  page->role.word);
 691                         kvm_mmu_zap_page(vcpu, page);
 692                         r = 1;
 693                 }
 694         return r;
 695 }
 696
 697 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 698 {
 699         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 700         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 701
 702         __set_bit(slot, &page_head->slot_bitmap);
 703 }
 704
 705 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 706 {
 707         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 708
 709         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 710 }
 711
 712 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 713 {
 714         struct kvm_memory_slot *slot;
 715         struct page *page;
 716
 717         ASSERT((gpa & HPA_ERR_MASK) == 0);
 718         slot = gfn_to_memslot(vcpu->kvm, gpa >> PAGE_SHIFT);
 719         if (!slot)
 720                 return gpa | HPA_ERR_MASK;
 721         page = gfn_to_page(slot, gpa >> PAGE_SHIFT);
 722         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 723                 | (gpa & (PAGE_SIZE-1));
 724 }
 725
 726 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 727 {
 728         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 729
 730         if (gpa == UNMAPPED_GVA)
 731                 return UNMAPPED_GVA;
 732         return gpa_to_hpa(vcpu, gpa);
 733 }
 734
 735 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 736 {
 737         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 738
 739         if (gpa == UNMAPPED_GVA)
 740                 return NULL;
 741         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 742 }
 743
 744 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 745 {
 746 }
 747
 748 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 749 {
 750         int level = PT32E_ROOT_LEVEL;
 751         hpa_t table_addr = vcpu->mmu.root_hpa;
 752
 753         for (; ; level--) {
 754                 u32 index = PT64_INDEX(v, level);
 755                 u64 *table;
 756                 u64 pte;
 757
 758                 ASSERT(VALID_PAGE(table_addr));
 759                 table = __va(table_addr);
 760
 761                 if (level == 1) {
 762                         pte = table[index];
 763                         if (is_present_pte(pte) && is_writeble_pte(pte))
 764                                 return 0;
 765                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 766                         page_header_update_slot(vcpu->kvm, table, v);
 767                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 768                                                                 PT_USER_MASK;
 769                         rmap_add(vcpu, &table[index]);
 770                         return 0;
 771                 }
 772
 773                 if (table[index] == 0) {
 774                         struct kvm_mmu_page *new_table;
 775                         gfn_t pseudo_gfn;
 776
 777                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 778                                 >> PAGE_SHIFT;
 779                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 780                                                      v, level - 1,
 781                                                      1, 0, &table[index]);
 782                         if (!new_table) {
 783                                 pgprintk("nonpaging_map: ENOMEM\n");
 784                                 return -ENOMEM;
 785                         }
 786
 787                         table[index] = new_table->page_hpa | PT_PRESENT_MASK
 788                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 789                 }
 790                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 791         }
 792 }
 793
 794 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 795 {
 796         int i;
 797         struct kvm_mmu_page *page;
 798
 799 #ifdef CONFIG_X86_64
 800         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 801                 hpa_t root = vcpu->mmu.root_hpa;
 802
 803                 ASSERT(VALID_PAGE(root));
 804                 page = page_header(root);
 805                 --page->root_count;
 806                 vcpu->mmu.root_hpa = INVALID_PAGE;
 807                 return;
 808         }
 809 #endif
 810         for (i = 0; i < 4; ++i) {
 811                 hpa_t root = vcpu->mmu.pae_root[i];
 812
 813                 ASSERT(VALID_PAGE(root));
 814                 root &= PT64_BASE_ADDR_MASK;
 815                 page = page_header(root);
 816                 --page->root_count;
 817                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 818         }
 819         vcpu->mmu.root_hpa = INVALID_PAGE;
 820 }
 821
 822 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 823 {
 824         int i;
 825         gfn_t root_gfn;
 826         struct kvm_mmu_page *page;
 827
 828         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 829
 830 #ifdef CONFIG_X86_64
 831         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 832                 hpa_t root = vcpu->mmu.root_hpa;
 833
 834                 ASSERT(!VALID_PAGE(root));
 835                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 836                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 837                 root = page->page_hpa;
 838                 ++page->root_count;
 839                 vcpu->mmu.root_hpa = root;
 840                 return;
 841         }
 842 #endif
 843         for (i = 0; i < 4; ++i) {
 844                 hpa_t root = vcpu->mmu.pae_root[i];
 845
 846                 ASSERT(!VALID_PAGE(root));
 847                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL)
 848                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 849                 else if (vcpu->mmu.root_level == 0)
 850                         root_gfn = 0;
 851                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 852                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 853                                         0, NULL);
 854                 root = page->page_hpa;
 855                 ++page->root_count;
 856                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 857         }
 858         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 859 }
 860
 861 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 862 {
 863         return vaddr;
 864 }
 865
 866 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 867                                u32 error_code)
 868 {
 869         gpa_t addr = gva;
 870         hpa_t paddr;
 871         int r;
 872
 873         r = mmu_topup_memory_caches(vcpu);
 874         if (r)
 875                 return r;
 876
 877         ASSERT(vcpu);
 878         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 879
 880
 881         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 882
 883         if (is_error_hpa(paddr))
 884                 return 1;
 885
 886         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 887 }
 888
 889 static void nonpaging_free(struct kvm_vcpu *vcpu)
 890 {
 891         mmu_free_roots(vcpu);
 892 }
 893
 894 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 895 {
 896         struct kvm_mmu *context = &vcpu->mmu;
 897
 898         context->new_cr3 = nonpaging_new_cr3;
 899         context->page_fault = nonpaging_page_fault;
 900         context->gva_to_gpa = nonpaging_gva_to_gpa;
 901         context->free = nonpaging_free;
 902         context->root_level = 0;
 903         context->shadow_root_level = PT32E_ROOT_LEVEL;
 904         mmu_alloc_roots(vcpu);
 905         ASSERT(VALID_PAGE(context->root_hpa));
 906         kvm_arch_ops->set_cr3(vcpu, context->root_hpa);
 907         return 0;
 908 }
 909
 910 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 911 {
 912         ++kvm_stat.tlb_flush;
 913         kvm_arch_ops->tlb_flush(vcpu);
 914 }
 915
 916 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 917 {
 918         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 919         mmu_free_roots(vcpu);
 920         if (unlikely(vcpu->kvm->n_free_mmu_pages < KVM_MIN_FREE_MMU_PAGES))
 921                 kvm_mmu_free_some_pages(vcpu);
 922         mmu_alloc_roots(vcpu);
 923         kvm_mmu_flush_tlb(vcpu);
 924         kvm_arch_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
 925 }
 926
 927 static inline void set_pte_common(struct kvm_vcpu *vcpu,
 928                              u64 *shadow_pte,
 929                              gpa_t gaddr,
 930                              int dirty,
 931                              u64 access_bits,
 932                              gfn_t gfn)
 933 {
 934         hpa_t paddr;
 935
 936         *shadow_pte |= access_bits << PT_SHADOW_BITS_OFFSET;
 937         if (!dirty)
 938                 access_bits &= ~PT_WRITABLE_MASK;
 939
 940         paddr = gpa_to_hpa(vcpu, gaddr & PT64_BASE_ADDR_MASK);
 941
 942         *shadow_pte |= access_bits;
 943
 944         if (is_error_hpa(paddr)) {
 945                 *shadow_pte |= gaddr;
 946                 *shadow_pte |= PT_SHADOW_IO_MARK;
 947                 *shadow_pte &= ~PT_PRESENT_MASK;
 948                 return;
 949         }
 950
 951         *shadow_pte |= paddr;
 952
 953         if (access_bits & PT_WRITABLE_MASK) {
 954                 struct kvm_mmu_page *shadow;
 955
 956                 shadow = kvm_mmu_lookup_page(vcpu, gfn);
 957                 if (shadow) {
 958                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 959                                  __FUNCTION__, gfn);
 960                         access_bits &= ~PT_WRITABLE_MASK;
 961                         if (is_writeble_pte(*shadow_pte)) {
 962                                     *shadow_pte &= ~PT_WRITABLE_MASK;
 963                                     kvm_arch_ops->tlb_flush(vcpu);
 964                         }
 965                 }
 966         }
 967
 968         if (access_bits & PT_WRITABLE_MASK)
 969                 mark_page_dirty(vcpu->kvm, gaddr >> PAGE_SHIFT);
 970
 971         page_header_update_slot(vcpu->kvm, shadow_pte, gaddr);
 972         rmap_add(vcpu, shadow_pte);
 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 inline int fix_read_pf(u64 *shadow_ent)
 983 {
 984         if ((*shadow_ent & PT_SHADOW_USER_MASK) &&
 985             !(*shadow_ent & PT_USER_MASK)) {
 986                 /*
 987                  * If supervisor write protect is disabled, we shadow kernel
 988                  * pages as user pages so we can trap the write access.
 989                  */
 990                 *shadow_ent |= PT_USER_MASK;
 991                 *shadow_ent &= ~PT_WRITABLE_MASK;
 992
 993                 return 1;
 994
 995         }
 996         return 0;
 997 }
 998
 999 static void paging_free(struct kvm_vcpu *vcpu)
1000 {
1001         nonpaging_free(vcpu);
1002 }
1003
1004 #define PTTYPE 64
1005 #include "paging_tmpl.h"
1006 #undef PTTYPE
1007
1008 #define PTTYPE 32
1009 #include "paging_tmpl.h"
1010 #undef PTTYPE
1011
1012 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1013 {
1014         struct kvm_mmu *context = &vcpu->mmu;
1015
1016         ASSERT(is_pae(vcpu));
1017         context->new_cr3 = paging_new_cr3;
1018         context->page_fault = paging64_page_fault;
1019         context->gva_to_gpa = paging64_gva_to_gpa;
1020         context->free = paging_free;
1021         context->root_level = level;
1022         context->shadow_root_level = level;
1023         mmu_alloc_roots(vcpu);
1024         ASSERT(VALID_PAGE(context->root_hpa));
1025         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1026                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1027         return 0;
1028 }
1029
1030 static int paging64_init_context(struct kvm_vcpu *vcpu)
1031 {
1032         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1033 }
1034
1035 static int paging32_init_context(struct kvm_vcpu *vcpu)
1036 {
1037         struct kvm_mmu *context = &vcpu->mmu;
1038
1039         context->new_cr3 = paging_new_cr3;
1040         context->page_fault = paging32_page_fault;
1041         context->gva_to_gpa = paging32_gva_to_gpa;
1042         context->free = paging_free;
1043         context->root_level = PT32_ROOT_LEVEL;
1044         context->shadow_root_level = PT32E_ROOT_LEVEL;
1045         mmu_alloc_roots(vcpu);
1046         ASSERT(VALID_PAGE(context->root_hpa));
1047         kvm_arch_ops->set_cr3(vcpu, context->root_hpa |
1048                     (vcpu->cr3 & (CR3_PCD_MASK | CR3_WPT_MASK)));
1049         return 0;
1050 }
1051
1052 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1053 {
1054         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1055 }
1056
1057 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1058 {
1059         ASSERT(vcpu);
1060         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1061
1062         if (!is_paging(vcpu))
1063                 return nonpaging_init_context(vcpu);
1064         else if (is_long_mode(vcpu))
1065                 return paging64_init_context(vcpu);
1066         else if (is_pae(vcpu))
1067                 return paging32E_init_context(vcpu);
1068         else
1069                 return paging32_init_context(vcpu);
1070 }
1071
1072 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1073 {
1074         ASSERT(vcpu);
1075         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1076                 vcpu->mmu.free(vcpu);
1077                 vcpu->mmu.root_hpa = INVALID_PAGE;
1078         }
1079 }
1080
1081 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1082 {
1083         int r;
1084
1085         destroy_kvm_mmu(vcpu);
1086         r = init_kvm_mmu(vcpu);
1087         if (r < 0)
1088                 goto out;
1089         r = mmu_topup_memory_caches(vcpu);
1090 out:
1091         return r;
1092 }
1093
1094 static void mmu_pre_write_zap_pte(struct kvm_vcpu *vcpu,
1095                                   struct kvm_mmu_page *page,
1096                                   u64 *spte)
1097 {
1098         u64 pte;
1099         struct kvm_mmu_page *child;
1100
1101         pte = *spte;
1102         if (is_present_pte(pte)) {
1103                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1104                         rmap_remove(vcpu, spte);
1105                 else {
1106                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1107                         mmu_page_remove_parent_pte(vcpu, child, spte);
1108                 }
1109         }
1110         *spte = 0;
1111 }
1112
1113 void kvm_mmu_pre_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
1114 {
1115         gfn_t gfn = gpa >> PAGE_SHIFT;
1116         struct kvm_mmu_page *page;
1117         struct hlist_node *node, *n;
1118         struct hlist_head *bucket;
1119         unsigned index;
1120         u64 *spte;
1121         unsigned offset = offset_in_page(gpa);
1122         unsigned pte_size;
1123         unsigned page_offset;
1124         unsigned misaligned;
1125         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                 if (misaligned || flooded) {
1146                         /*
1147                          * Misaligned accesses are too much trouble to fix
1148                          * up; also, they usually indicate a page is not used
1149                          * as a page table.
1150                          *
1151                          * If we're seeing too many writes to a page,
1152                          * it may no longer be a page table, or we may be
1153                          * forking, in which case it is better to unmap the
1154                          * page.
1155                          */
1156                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1157                                  gpa, bytes, page->role.word);
1158                         kvm_mmu_zap_page(vcpu, page);
1159                         continue;
1160                 }
1161                 page_offset = offset;
1162                 level = page->role.level;
1163                 npte = 1;
1164                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1165                         page_offset <<= 1;      /* 32->64 */
1166                         /*
1167                          * A 32-bit pde maps 4MB while the shadow pdes map
1168                          * only 2MB.  So we need to double the offset again
1169                          * and zap two pdes instead of one.
1170                          */
1171                         if (level == PT32_ROOT_LEVEL) {
1172                                 page_offset &= ~7; /* kill rounding error */
1173                                 page_offset <<= 1;
1174                                 npte = 2;
1175                         }
1176                         page_offset &= ~PAGE_MASK;
1177                 }
1178                 spte = __va(page->page_hpa);
1179                 spte += page_offset / sizeof(*spte);
1180                 while (npte--) {
1181                         mmu_pre_write_zap_pte(vcpu, page, spte);
1182                         ++spte;
1183                 }
1184         }
1185 }
1186
1187 void kvm_mmu_post_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
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         while (!list_empty(&vcpu->free_pages)) {
1220                 page = list_entry(vcpu->free_pages.next,
1221                                   struct kvm_mmu_page, link);
1222                 list_del(&page->link);
1223                 __free_page(pfn_to_page(page->page_hpa >> PAGE_SHIFT));
1224                 page->page_hpa = INVALID_PAGE;
1225         }
1226         free_page((unsigned long)vcpu->mmu.pae_root);
1227 }
1228
1229 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1230 {
1231         struct page *page;
1232         int i;
1233
1234         ASSERT(vcpu);
1235
1236         for (i = 0; i < KVM_NUM_MMU_PAGES; i++) {
1237                 struct kvm_mmu_page *page_header = &vcpu->page_header_buf[i];
1238
1239                 INIT_LIST_HEAD(&page_header->link);
1240                 if ((page = alloc_page(GFP_KERNEL)) == NULL)
1241                         goto error_1;
1242                 set_page_private(page, (unsigned long)page_header);
1243                 page_header->page_hpa = (hpa_t)page_to_pfn(page) << PAGE_SHIFT;
1244                 memset(__va(page_header->page_hpa), 0, PAGE_SIZE);
1245                 list_add(&page_header->link, &vcpu->free_pages);
1246                 ++vcpu->kvm->n_free_mmu_pages;
1247         }
1248
1249         /*
1250          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1251          * Therefore we need to allocate shadow page tables in the first
1252          * 4GB of memory, which happens to fit the DMA32 zone.
1253          */
1254         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1255         if (!page)
1256                 goto error_1;
1257         vcpu->mmu.pae_root = page_address(page);
1258         for (i = 0; i < 4; ++i)
1259                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1260
1261         return 0;
1262
1263 error_1:
1264         free_mmu_pages(vcpu);
1265         return -ENOMEM;
1266 }
1267
1268 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1269 {
1270         ASSERT(vcpu);
1271         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1272         ASSERT(list_empty(&vcpu->free_pages));
1273
1274         return alloc_mmu_pages(vcpu);
1275 }
1276
1277 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1278 {
1279         ASSERT(vcpu);
1280         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1281         ASSERT(!list_empty(&vcpu->free_pages));
1282
1283         return init_kvm_mmu(vcpu);
1284 }
1285
1286 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1287 {
1288         ASSERT(vcpu);
1289
1290         destroy_kvm_mmu(vcpu);
1291         free_mmu_pages(vcpu);
1292         mmu_free_memory_caches(vcpu);
1293 }
1294
1295 void kvm_mmu_slot_remove_write_access(struct kvm_vcpu *vcpu, int slot)
1296 {
1297         struct kvm *kvm = vcpu->kvm;
1298         struct kvm_mmu_page *page;
1299
1300         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1301                 int i;
1302                 u64 *pt;
1303
1304                 if (!test_bit(slot, &page->slot_bitmap))
1305                         continue;
1306
1307                 pt = __va(page->page_hpa);
1308                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1309                         /* avoid RMW */
1310                         if (pt[i] & PT_WRITABLE_MASK) {
1311                                 rmap_remove(vcpu, &pt[i]);
1312                                 pt[i] &= ~PT_WRITABLE_MASK;
1313                         }
1314         }
1315 }
1316
1317 #ifdef AUDIT
1318
1319 static const char *audit_msg;
1320
1321 static gva_t canonicalize(gva_t gva)
1322 {
1323 #ifdef CONFIG_X86_64
1324         gva = (long long)(gva << 16) >> 16;
1325 #endif
1326         return gva;
1327 }
1328
1329 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1330                                 gva_t va, int level)
1331 {
1332         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1333         int i;
1334         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1335
1336         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1337                 u64 ent = pt[i];
1338
1339                 if (!ent & PT_PRESENT_MASK)
1340                         continue;
1341
1342                 va = canonicalize(va);
1343                 if (level > 1)
1344                         audit_mappings_page(vcpu, ent, va, level - 1);
1345                 else {
1346                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1347                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1348
1349                         if ((ent & PT_PRESENT_MASK)
1350                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1351                                 printk(KERN_ERR "audit error: (%s) levels %d"
1352                                        " gva %lx gpa %llx hpa %llx ent %llx\n",
1353                                        audit_msg, vcpu->mmu.root_level,
1354                                        va, gpa, hpa, ent);
1355                 }
1356         }
1357 }
1358
1359 static void audit_mappings(struct kvm_vcpu *vcpu)
1360 {
1361         unsigned i;
1362
1363         if (vcpu->mmu.root_level == 4)
1364                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1365         else
1366                 for (i = 0; i < 4; ++i)
1367                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1368                                 audit_mappings_page(vcpu,
1369                                                     vcpu->mmu.pae_root[i],
1370                                                     i << 30,
1371                                                     2);
1372 }
1373
1374 static int count_rmaps(struct kvm_vcpu *vcpu)
1375 {
1376         int nmaps = 0;
1377         int i, j, k;
1378
1379         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1380                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1381                 struct kvm_rmap_desc *d;
1382
1383                 for (j = 0; j < m->npages; ++j) {
1384                         struct page *page = m->phys_mem[j];
1385
1386                         if (!page->private)
1387                                 continue;
1388                         if (!(page->private & 1)) {
1389                                 ++nmaps;
1390                                 continue;
1391                         }
1392                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1393                         while (d) {
1394                                 for (k = 0; k < RMAP_EXT; ++k)
1395                                         if (d->shadow_ptes[k])
1396                                                 ++nmaps;
1397                                         else
1398                                                 break;
1399                                 d = d->more;
1400                         }
1401                 }
1402         }
1403         return nmaps;
1404 }
1405
1406 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1407 {
1408         int nmaps = 0;
1409         struct kvm_mmu_page *page;
1410         int i;
1411
1412         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1413                 u64 *pt = __va(page->page_hpa);
1414
1415                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1416                         continue;
1417
1418                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1419                         u64 ent = pt[i];
1420
1421                         if (!(ent & PT_PRESENT_MASK))
1422                                 continue;
1423                         if (!(ent & PT_WRITABLE_MASK))
1424                                 continue;
1425                         ++nmaps;
1426                 }
1427         }
1428         return nmaps;
1429 }
1430
1431 static void audit_rmap(struct kvm_vcpu *vcpu)
1432 {
1433         int n_rmap = count_rmaps(vcpu);
1434         int n_actual = count_writable_mappings(vcpu);
1435
1436         if (n_rmap != n_actual)
1437                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1438                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1439 }
1440
1441 static void audit_write_protection(struct kvm_vcpu *vcpu)
1442 {
1443         struct kvm_mmu_page *page;
1444
1445         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1446                 hfn_t hfn;
1447                 struct page *pg;
1448
1449                 if (page->role.metaphysical)
1450                         continue;
1451
1452                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1453                         >> PAGE_SHIFT;
1454                 pg = pfn_to_page(hfn);
1455                 if (pg->private)
1456                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1457                                " mappings: gfn %lx role %x\n",
1458                                __FUNCTION__, audit_msg, page->gfn,
1459                                page->role.word);
1460         }
1461 }
1462
1463 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1464 {
1465         int olddbg = dbg;
1466
1467         dbg = 0;
1468         audit_msg = msg;
1469         audit_rmap(vcpu);
1470         audit_write_protection(vcpu);
1471         audit_mappings(vcpu);
1472         dbg = olddbg;
1473 }
1474
1475 #endif