git.oblomov.eu Git - linux-2.6/blob - drivers/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "vmx.h"
  21 #include "kvm.h"
  22
  23 #include <linux/types.h>
  24 #include <linux/string.h>
  25 #include <linux/mm.h>
  26 #include <linux/highmem.h>
  27 #include <linux/module.h>
  28
  29 #include <asm/page.h>
  30 #include <asm/cmpxchg.h>
  31
  32 #undef MMU_DEBUG
  33
  34 #undef AUDIT
  35
  36 #ifdef AUDIT
  37 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  38 #else
  39 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  40 #endif
  41
  42 #ifdef MMU_DEBUG
  43
  44 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  45 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  46
  47 #else
  48
  49 #define pgprintk(x...) do { } while (0)
  50 #define rmap_printk(x...) do { } while (0)
  51
  52 #endif
  53
  54 #if defined(MMU_DEBUG) || defined(AUDIT)
  55 static int dbg = 1;
  56 #endif
  57
  58 #ifndef MMU_DEBUG
  59 #define ASSERT(x) do { } while (0)
  60 #else
  61 #define ASSERT(x)                                                       \
  62         if (!(x)) {                                                     \
  63                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  64                        __FILE__, __LINE__, #x);                         \
  65         }
  66 #endif
  67
  68 #define PT64_PT_BITS 9
  69 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  70 #define PT32_PT_BITS 10
  71 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  72
  73 #define PT_WRITABLE_SHIFT 1
  74
  75 #define PT_PRESENT_MASK (1ULL << 0)
  76 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  77 #define PT_USER_MASK (1ULL << 2)
  78 #define PT_PWT_MASK (1ULL << 3)
  79 #define PT_PCD_MASK (1ULL << 4)
  80 #define PT_ACCESSED_MASK (1ULL << 5)
  81 #define PT_DIRTY_MASK (1ULL << 6)
  82 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  83 #define PT_PAT_MASK (1ULL << 7)
  84 #define PT_GLOBAL_MASK (1ULL << 8)
  85 #define PT64_NX_MASK (1ULL << 63)
  86
  87 #define PT_PAT_SHIFT 7
  88 #define PT_DIR_PAT_SHIFT 12
  89 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  90
  91 #define PT32_DIR_PSE36_SIZE 4
  92 #define PT32_DIR_PSE36_SHIFT 13
  93 #define PT32_DIR_PSE36_MASK (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  94
  95
  96 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  97 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  98
  99 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 100
 101 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 102
 103 #define PT64_LEVEL_BITS 9
 104
 105 #define PT64_LEVEL_SHIFT(level) \
 106                 ( PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS )
 107
 108 #define PT64_LEVEL_MASK(level) \
 109                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 110
 111 #define PT64_INDEX(address, level)\
 112         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 113
 114
 115 #define PT32_LEVEL_BITS 10
 116
 117 #define PT32_LEVEL_SHIFT(level) \
 118                 ( PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS )
 119
 120 #define PT32_LEVEL_MASK(level) \
 121                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 122
 123 #define PT32_INDEX(address, level)\
 124         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 125
 126
 127 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 128 #define PT64_DIR_BASE_ADDR_MASK \
 129         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 130
 131 #define PT32_BASE_ADDR_MASK PAGE_MASK
 132 #define PT32_DIR_BASE_ADDR_MASK \
 133         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 134
 135
 136 #define PFERR_PRESENT_MASK (1U << 0)
 137 #define PFERR_WRITE_MASK (1U << 1)
 138 #define PFERR_USER_MASK (1U << 2)
 139 #define PFERR_FETCH_MASK (1U << 4)
 140
 141 #define PT64_ROOT_LEVEL 4
 142 #define PT32_ROOT_LEVEL 2
 143 #define PT32E_ROOT_LEVEL 3
 144
 145 #define PT_DIRECTORY_LEVEL 2
 146 #define PT_PAGE_TABLE_LEVEL 1
 147
 148 #define RMAP_EXT 4
 149
 150 struct kvm_rmap_desc {
 151         u64 *shadow_ptes[RMAP_EXT];
 152         struct kvm_rmap_desc *more;
 153 };
 154
 155 static struct kmem_cache *pte_chain_cache;
 156 static struct kmem_cache *rmap_desc_cache;
 157 static struct kmem_cache *mmu_page_header_cache;
 158
 159 static u64 __read_mostly shadow_trap_nonpresent_pte;
 160 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 161
 162 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 163 {
 164         shadow_trap_nonpresent_pte = trap_pte;
 165         shadow_notrap_nonpresent_pte = notrap_pte;
 166 }
 167 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 168
 169 static int is_write_protection(struct kvm_vcpu *vcpu)
 170 {
 171         return vcpu->cr0 & X86_CR0_WP;
 172 }
 173
 174 static int is_cpuid_PSE36(void)
 175 {
 176         return 1;
 177 }
 178
 179 static int is_nx(struct kvm_vcpu *vcpu)
 180 {
 181         return vcpu->shadow_efer & EFER_NX;
 182 }
 183
 184 static int is_present_pte(unsigned long pte)
 185 {
 186         return pte & PT_PRESENT_MASK;
 187 }
 188
 189 static int is_shadow_present_pte(u64 pte)
 190 {
 191         pte &= ~PT_SHADOW_IO_MARK;
 192         return pte != shadow_trap_nonpresent_pte
 193                 && pte != shadow_notrap_nonpresent_pte;
 194 }
 195
 196 static int is_writeble_pte(unsigned long pte)
 197 {
 198         return pte & PT_WRITABLE_MASK;
 199 }
 200
 201 static int is_io_pte(unsigned long pte)
 202 {
 203         return pte & PT_SHADOW_IO_MARK;
 204 }
 205
 206 static int is_rmap_pte(u64 pte)
 207 {
 208         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 209                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 210 }
 211
 212 static void set_shadow_pte(u64 *sptep, u64 spte)
 213 {
 214 #ifdef CONFIG_X86_64
 215         set_64bit((unsigned long *)sptep, spte);
 216 #else
 217         set_64bit((unsigned long long *)sptep, spte);
 218 #endif
 219 }
 220
 221 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 222                                   struct kmem_cache *base_cache, int min)
 223 {
 224         void *obj;
 225
 226         if (cache->nobjs >= min)
 227                 return 0;
 228         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 229                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 230                 if (!obj)
 231                         return -ENOMEM;
 232                 cache->objects[cache->nobjs++] = obj;
 233         }
 234         return 0;
 235 }
 236
 237 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 238 {
 239         while (mc->nobjs)
 240                 kfree(mc->objects[--mc->nobjs]);
 241 }
 242
 243 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 244                                        int min)
 245 {
 246         struct page *page;
 247
 248         if (cache->nobjs >= min)
 249                 return 0;
 250         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 251                 page = alloc_page(GFP_KERNEL);
 252                 if (!page)
 253                         return -ENOMEM;
 254                 set_page_private(page, 0);
 255                 cache->objects[cache->nobjs++] = page_address(page);
 256         }
 257         return 0;
 258 }
 259
 260 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 261 {
 262         while (mc->nobjs)
 263                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 264 }
 265
 266 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 267 {
 268         int r;
 269
 270         kvm_mmu_free_some_pages(vcpu);
 271         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 272                                    pte_chain_cache, 4);
 273         if (r)
 274                 goto out;
 275         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 276                                    rmap_desc_cache, 1);
 277         if (r)
 278                 goto out;
 279         r = mmu_topup_memory_cache_page(&vcpu->mmu_page_cache, 4);
 280         if (r)
 281                 goto out;
 282         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
 283                                    mmu_page_header_cache, 4);
 284 out:
 285         return r;
 286 }
 287
 288 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 289 {
 290         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 291         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 292         mmu_free_memory_cache_page(&vcpu->mmu_page_cache);
 293         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
 294 }
 295
 296 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 297                                     size_t size)
 298 {
 299         void *p;
 300
 301         BUG_ON(!mc->nobjs);
 302         p = mc->objects[--mc->nobjs];
 303         memset(p, 0, size);
 304         return p;
 305 }
 306
 307 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 308 {
 309         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 310                                       sizeof(struct kvm_pte_chain));
 311 }
 312
 313 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 314 {
 315         kfree(pc);
 316 }
 317
 318 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 319 {
 320         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 321                                       sizeof(struct kvm_rmap_desc));
 322 }
 323
 324 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 325 {
 326         kfree(rd);
 327 }
 328
 329 /*
 330  * Reverse mapping data structures:
 331  *
 332  * If page->private bit zero is zero, then page->private points to the
 333  * shadow page table entry that points to page_address(page).
 334  *
 335  * If page->private bit zero is one, (then page->private & ~1) points
 336  * to a struct kvm_rmap_desc containing more mappings.
 337  */
 338 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte)
 339 {
 340         struct page *page;
 341         struct kvm_rmap_desc *desc;
 342         int i;
 343
 344         if (!is_rmap_pte(*spte))
 345                 return;
 346         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 347         if (!page_private(page)) {
 348                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 349                 set_page_private(page,(unsigned long)spte);
 350         } else if (!(page_private(page) & 1)) {
 351                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 352                 desc = mmu_alloc_rmap_desc(vcpu);
 353                 desc->shadow_ptes[0] = (u64 *)page_private(page);
 354                 desc->shadow_ptes[1] = spte;
 355                 set_page_private(page,(unsigned long)desc | 1);
 356         } else {
 357                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 358                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 359                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 360                         desc = desc->more;
 361                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 362                         desc->more = mmu_alloc_rmap_desc(vcpu);
 363                         desc = desc->more;
 364                 }
 365                 for (i = 0; desc->shadow_ptes[i]; ++i)
 366                         ;
 367                 desc->shadow_ptes[i] = spte;
 368         }
 369 }
 370
 371 static void rmap_desc_remove_entry(struct page *page,
 372                                    struct kvm_rmap_desc *desc,
 373                                    int i,
 374                                    struct kvm_rmap_desc *prev_desc)
 375 {
 376         int j;
 377
 378         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 379                 ;
 380         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 381         desc->shadow_ptes[j] = NULL;
 382         if (j != 0)
 383                 return;
 384         if (!prev_desc && !desc->more)
 385                 set_page_private(page,(unsigned long)desc->shadow_ptes[0]);
 386         else
 387                 if (prev_desc)
 388                         prev_desc->more = desc->more;
 389                 else
 390                         set_page_private(page,(unsigned long)desc->more | 1);
 391         mmu_free_rmap_desc(desc);
 392 }
 393
 394 static void rmap_remove(u64 *spte)
 395 {
 396         struct page *page;
 397         struct kvm_rmap_desc *desc;
 398         struct kvm_rmap_desc *prev_desc;
 399         int i;
 400
 401         if (!is_rmap_pte(*spte))
 402                 return;
 403         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 404         if (!page_private(page)) {
 405                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 406                 BUG();
 407         } else if (!(page_private(page) & 1)) {
 408                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 409                 if ((u64 *)page_private(page) != spte) {
 410                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 411                                spte, *spte);
 412                         BUG();
 413                 }
 414                 set_page_private(page,0);
 415         } else {
 416                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 417                 desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 418                 prev_desc = NULL;
 419                 while (desc) {
 420                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 421                                 if (desc->shadow_ptes[i] == spte) {
 422                                         rmap_desc_remove_entry(page,
 423                                                                desc, i,
 424                                                                prev_desc);
 425                                         return;
 426                                 }
 427                         prev_desc = desc;
 428                         desc = desc->more;
 429                 }
 430                 BUG();
 431         }
 432 }
 433
 434 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 435 {
 436         struct kvm *kvm = vcpu->kvm;
 437         struct page *page;
 438         struct kvm_rmap_desc *desc;
 439         u64 *spte;
 440
 441         page = gfn_to_page(kvm, gfn);
 442         BUG_ON(!page);
 443
 444         while (page_private(page)) {
 445                 if (!(page_private(page) & 1))
 446                         spte = (u64 *)page_private(page);
 447                 else {
 448                         desc = (struct kvm_rmap_desc *)(page_private(page) & ~1ul);
 449                         spte = desc->shadow_ptes[0];
 450                 }
 451                 BUG_ON(!spte);
 452                 BUG_ON((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT
 453                        != page_to_pfn(page));
 454                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 455                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 456                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 457                 rmap_remove(spte);
 458                 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 459                 kvm_flush_remote_tlbs(vcpu->kvm);
 460         }
 461 }
 462
 463 #ifdef MMU_DEBUG
 464 static int is_empty_shadow_page(u64 *spt)
 465 {
 466         u64 *pos;
 467         u64 *end;
 468
 469         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 470                 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
 471                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 472                                pos, *pos);
 473                         return 0;
 474                 }
 475         return 1;
 476 }
 477 #endif
 478
 479 static void kvm_mmu_free_page(struct kvm *kvm,
 480                               struct kvm_mmu_page *page_head)
 481 {
 482         ASSERT(is_empty_shadow_page(page_head->spt));
 483         list_del(&page_head->link);
 484         __free_page(virt_to_page(page_head->spt));
 485         kfree(page_head);
 486         ++kvm->n_free_mmu_pages;
 487 }
 488
 489 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 490 {
 491         return gfn;
 492 }
 493
 494 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 495                                                u64 *parent_pte)
 496 {
 497         struct kvm_mmu_page *page;
 498
 499         if (!vcpu->kvm->n_free_mmu_pages)
 500                 return NULL;
 501
 502         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 503                                       sizeof *page);
 504         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 505         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 506         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 507         ASSERT(is_empty_shadow_page(page->spt));
 508         page->slot_bitmap = 0;
 509         page->multimapped = 0;
 510         page->parent_pte = parent_pte;
 511         --vcpu->kvm->n_free_mmu_pages;
 512         return page;
 513 }
 514
 515 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 516                                     struct kvm_mmu_page *page, u64 *parent_pte)
 517 {
 518         struct kvm_pte_chain *pte_chain;
 519         struct hlist_node *node;
 520         int i;
 521
 522         if (!parent_pte)
 523                 return;
 524         if (!page->multimapped) {
 525                 u64 *old = page->parent_pte;
 526
 527                 if (!old) {
 528                         page->parent_pte = parent_pte;
 529                         return;
 530                 }
 531                 page->multimapped = 1;
 532                 pte_chain = mmu_alloc_pte_chain(vcpu);
 533                 INIT_HLIST_HEAD(&page->parent_ptes);
 534                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 535                 pte_chain->parent_ptes[0] = old;
 536         }
 537         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 538                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 539                         continue;
 540                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 541                         if (!pte_chain->parent_ptes[i]) {
 542                                 pte_chain->parent_ptes[i] = parent_pte;
 543                                 return;
 544                         }
 545         }
 546         pte_chain = mmu_alloc_pte_chain(vcpu);
 547         BUG_ON(!pte_chain);
 548         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 549         pte_chain->parent_ptes[0] = parent_pte;
 550 }
 551
 552 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
 553                                        u64 *parent_pte)
 554 {
 555         struct kvm_pte_chain *pte_chain;
 556         struct hlist_node *node;
 557         int i;
 558
 559         if (!page->multimapped) {
 560                 BUG_ON(page->parent_pte != parent_pte);
 561                 page->parent_pte = NULL;
 562                 return;
 563         }
 564         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 565                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 566                         if (!pte_chain->parent_ptes[i])
 567                                 break;
 568                         if (pte_chain->parent_ptes[i] != parent_pte)
 569                                 continue;
 570                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 571                                 && pte_chain->parent_ptes[i + 1]) {
 572                                 pte_chain->parent_ptes[i]
 573                                         = pte_chain->parent_ptes[i + 1];
 574                                 ++i;
 575                         }
 576                         pte_chain->parent_ptes[i] = NULL;
 577                         if (i == 0) {
 578                                 hlist_del(&pte_chain->link);
 579                                 mmu_free_pte_chain(pte_chain);
 580                                 if (hlist_empty(&page->parent_ptes)) {
 581                                         page->multimapped = 0;
 582                                         page->parent_pte = NULL;
 583                                 }
 584                         }
 585                         return;
 586                 }
 587         BUG();
 588 }
 589
 590 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 591                                                 gfn_t gfn)
 592 {
 593         unsigned index;
 594         struct hlist_head *bucket;
 595         struct kvm_mmu_page *page;
 596         struct hlist_node *node;
 597
 598         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 599         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 600         bucket = &vcpu->kvm->mmu_page_hash[index];
 601         hlist_for_each_entry(page, node, bucket, hash_link)
 602                 if (page->gfn == gfn && !page->role.metaphysical) {
 603                         pgprintk("%s: found role %x\n",
 604                                  __FUNCTION__, page->role.word);
 605                         return page;
 606                 }
 607         return NULL;
 608 }
 609
 610 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 611                                              gfn_t gfn,
 612                                              gva_t gaddr,
 613                                              unsigned level,
 614                                              int metaphysical,
 615                                              unsigned hugepage_access,
 616                                              u64 *parent_pte)
 617 {
 618         union kvm_mmu_page_role role;
 619         unsigned index;
 620         unsigned quadrant;
 621         struct hlist_head *bucket;
 622         struct kvm_mmu_page *page;
 623         struct hlist_node *node;
 624
 625         role.word = 0;
 626         role.glevels = vcpu->mmu.root_level;
 627         role.level = level;
 628         role.metaphysical = metaphysical;
 629         role.hugepage_access = hugepage_access;
 630         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 631                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 632                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 633                 role.quadrant = quadrant;
 634         }
 635         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 636                  gfn, role.word);
 637         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 638         bucket = &vcpu->kvm->mmu_page_hash[index];
 639         hlist_for_each_entry(page, node, bucket, hash_link)
 640                 if (page->gfn == gfn && page->role.word == role.word) {
 641                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 642                         pgprintk("%s: found\n", __FUNCTION__);
 643                         return page;
 644                 }
 645         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 646         if (!page)
 647                 return page;
 648         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 649         page->gfn = gfn;
 650         page->role = role;
 651         hlist_add_head(&page->hash_link, bucket);
 652         vcpu->mmu.prefetch_page(vcpu, page);
 653         if (!metaphysical)
 654                 rmap_write_protect(vcpu, gfn);
 655         return page;
 656 }
 657
 658 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 659                                          struct kvm_mmu_page *page)
 660 {
 661         unsigned i;
 662         u64 *pt;
 663         u64 ent;
 664
 665         pt = page->spt;
 666
 667         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 668                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 669                         if (is_shadow_present_pte(pt[i]))
 670                                 rmap_remove(&pt[i]);
 671                         pt[i] = shadow_trap_nonpresent_pte;
 672                 }
 673                 kvm_flush_remote_tlbs(kvm);
 674                 return;
 675         }
 676
 677         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 678                 ent = pt[i];
 679
 680                 pt[i] = shadow_trap_nonpresent_pte;
 681                 if (!is_shadow_present_pte(ent))
 682                         continue;
 683                 ent &= PT64_BASE_ADDR_MASK;
 684                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 685         }
 686         kvm_flush_remote_tlbs(kvm);
 687 }
 688
 689 static void kvm_mmu_put_page(struct kvm_mmu_page *page,
 690                              u64 *parent_pte)
 691 {
 692         mmu_page_remove_parent_pte(page, parent_pte);
 693 }
 694
 695 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 696 {
 697         int i;
 698
 699         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 700                 if (kvm->vcpus[i])
 701                         kvm->vcpus[i]->last_pte_updated = NULL;
 702 }
 703
 704 static void kvm_mmu_zap_page(struct kvm *kvm,
 705                              struct kvm_mmu_page *page)
 706 {
 707         u64 *parent_pte;
 708
 709         while (page->multimapped || page->parent_pte) {
 710                 if (!page->multimapped)
 711                         parent_pte = page->parent_pte;
 712                 else {
 713                         struct kvm_pte_chain *chain;
 714
 715                         chain = container_of(page->parent_ptes.first,
 716                                              struct kvm_pte_chain, link);
 717                         parent_pte = chain->parent_ptes[0];
 718                 }
 719                 BUG_ON(!parent_pte);
 720                 kvm_mmu_put_page(page, parent_pte);
 721                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 722         }
 723         kvm_mmu_page_unlink_children(kvm, page);
 724         if (!page->root_count) {
 725                 hlist_del(&page->hash_link);
 726                 kvm_mmu_free_page(kvm, page);
 727         } else
 728                 list_move(&page->link, &kvm->active_mmu_pages);
 729         kvm_mmu_reset_last_pte_updated(kvm);
 730 }
 731
 732 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 733 {
 734         unsigned index;
 735         struct hlist_head *bucket;
 736         struct kvm_mmu_page *page;
 737         struct hlist_node *node, *n;
 738         int r;
 739
 740         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 741         r = 0;
 742         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 743         bucket = &vcpu->kvm->mmu_page_hash[index];
 744         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 745                 if (page->gfn == gfn && !page->role.metaphysical) {
 746                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 747                                  page->role.word);
 748                         kvm_mmu_zap_page(vcpu->kvm, page);
 749                         r = 1;
 750                 }
 751         return r;
 752 }
 753
 754 static void mmu_unshadow(struct kvm_vcpu *vcpu, gfn_t gfn)
 755 {
 756         struct kvm_mmu_page *page;
 757
 758         while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
 759                 pgprintk("%s: zap %lx %x\n",
 760                          __FUNCTION__, gfn, page->role.word);
 761                 kvm_mmu_zap_page(vcpu->kvm, page);
 762         }
 763 }
 764
 765 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 766 {
 767         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 768         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 769
 770         __set_bit(slot, &page_head->slot_bitmap);
 771 }
 772
 773 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 774 {
 775         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 776
 777         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 778 }
 779
 780 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 781 {
 782         struct page *page;
 783
 784         ASSERT((gpa & HPA_ERR_MASK) == 0);
 785         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 786         if (!page)
 787                 return gpa | HPA_ERR_MASK;
 788         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 789                 | (gpa & (PAGE_SIZE-1));
 790 }
 791
 792 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 793 {
 794         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 795
 796         if (gpa == UNMAPPED_GVA)
 797                 return UNMAPPED_GVA;
 798         return gpa_to_hpa(vcpu, gpa);
 799 }
 800
 801 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 802 {
 803         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 804
 805         if (gpa == UNMAPPED_GVA)
 806                 return NULL;
 807         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 808 }
 809
 810 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 811 {
 812 }
 813
 814 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 815 {
 816         int level = PT32E_ROOT_LEVEL;
 817         hpa_t table_addr = vcpu->mmu.root_hpa;
 818
 819         for (; ; level--) {
 820                 u32 index = PT64_INDEX(v, level);
 821                 u64 *table;
 822                 u64 pte;
 823
 824                 ASSERT(VALID_PAGE(table_addr));
 825                 table = __va(table_addr);
 826
 827                 if (level == 1) {
 828                         pte = table[index];
 829                         if (is_shadow_present_pte(pte) && is_writeble_pte(pte))
 830                                 return 0;
 831                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 832                         page_header_update_slot(vcpu->kvm, table, v);
 833                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 834                                                                 PT_USER_MASK;
 835                         rmap_add(vcpu, &table[index]);
 836                         return 0;
 837                 }
 838
 839                 if (table[index] == shadow_trap_nonpresent_pte) {
 840                         struct kvm_mmu_page *new_table;
 841                         gfn_t pseudo_gfn;
 842
 843                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 844                                 >> PAGE_SHIFT;
 845                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 846                                                      v, level - 1,
 847                                                      1, 0, &table[index]);
 848                         if (!new_table) {
 849                                 pgprintk("nonpaging_map: ENOMEM\n");
 850                                 return -ENOMEM;
 851                         }
 852
 853                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 854                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 855                 }
 856                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 857         }
 858 }
 859
 860 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 861                                     struct kvm_mmu_page *sp)
 862 {
 863         int i;
 864
 865         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 866                 sp->spt[i] = shadow_trap_nonpresent_pte;
 867 }
 868
 869 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 870 {
 871         int i;
 872         struct kvm_mmu_page *page;
 873
 874         if (!VALID_PAGE(vcpu->mmu.root_hpa))
 875                 return;
 876 #ifdef CONFIG_X86_64
 877         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 878                 hpa_t root = vcpu->mmu.root_hpa;
 879
 880                 page = page_header(root);
 881                 --page->root_count;
 882                 vcpu->mmu.root_hpa = INVALID_PAGE;
 883                 return;
 884         }
 885 #endif
 886         for (i = 0; i < 4; ++i) {
 887                 hpa_t root = vcpu->mmu.pae_root[i];
 888
 889                 if (root) {
 890                         root &= PT64_BASE_ADDR_MASK;
 891                         page = page_header(root);
 892                         --page->root_count;
 893                 }
 894                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 895         }
 896         vcpu->mmu.root_hpa = INVALID_PAGE;
 897 }
 898
 899 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 900 {
 901         int i;
 902         gfn_t root_gfn;
 903         struct kvm_mmu_page *page;
 904
 905         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 906
 907 #ifdef CONFIG_X86_64
 908         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 909                 hpa_t root = vcpu->mmu.root_hpa;
 910
 911                 ASSERT(!VALID_PAGE(root));
 912                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 913                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 914                 root = __pa(page->spt);
 915                 ++page->root_count;
 916                 vcpu->mmu.root_hpa = root;
 917                 return;
 918         }
 919 #endif
 920         for (i = 0; i < 4; ++i) {
 921                 hpa_t root = vcpu->mmu.pae_root[i];
 922
 923                 ASSERT(!VALID_PAGE(root));
 924                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 925                         if (!is_present_pte(vcpu->pdptrs[i])) {
 926                                 vcpu->mmu.pae_root[i] = 0;
 927                                 continue;
 928                         }
 929                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 930                 } else if (vcpu->mmu.root_level == 0)
 931                         root_gfn = 0;
 932                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 933                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 934                                         0, NULL);
 935                 root = __pa(page->spt);
 936                 ++page->root_count;
 937                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 938         }
 939         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 940 }
 941
 942 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 943 {
 944         return vaddr;
 945 }
 946
 947 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 948                                u32 error_code)
 949 {
 950         gpa_t addr = gva;
 951         hpa_t paddr;
 952         int r;
 953
 954         r = mmu_topup_memory_caches(vcpu);
 955         if (r)
 956                 return r;
 957
 958         ASSERT(vcpu);
 959         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
 960
 961
 962         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
 963
 964         if (is_error_hpa(paddr))
 965                 return 1;
 966
 967         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
 968 }
 969
 970 static void nonpaging_free(struct kvm_vcpu *vcpu)
 971 {
 972         mmu_free_roots(vcpu);
 973 }
 974
 975 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
 976 {
 977         struct kvm_mmu *context = &vcpu->mmu;
 978
 979         context->new_cr3 = nonpaging_new_cr3;
 980         context->page_fault = nonpaging_page_fault;
 981         context->gva_to_gpa = nonpaging_gva_to_gpa;
 982         context->free = nonpaging_free;
 983         context->prefetch_page = nonpaging_prefetch_page;
 984         context->root_level = 0;
 985         context->shadow_root_level = PT32E_ROOT_LEVEL;
 986         context->root_hpa = INVALID_PAGE;
 987         return 0;
 988 }
 989
 990 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
 991 {
 992         ++vcpu->stat.tlb_flush;
 993         kvm_x86_ops->tlb_flush(vcpu);
 994 }
 995
 996 static void paging_new_cr3(struct kvm_vcpu *vcpu)
 997 {
 998         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
 999         mmu_free_roots(vcpu);
1000 }
1001
1002 static void inject_page_fault(struct kvm_vcpu *vcpu,
1003                               u64 addr,
1004                               u32 err_code)
1005 {
1006         kvm_x86_ops->inject_page_fault(vcpu, addr, err_code);
1007 }
1008
1009 static void paging_free(struct kvm_vcpu *vcpu)
1010 {
1011         nonpaging_free(vcpu);
1012 }
1013
1014 #define PTTYPE 64
1015 #include "paging_tmpl.h"
1016 #undef PTTYPE
1017
1018 #define PTTYPE 32
1019 #include "paging_tmpl.h"
1020 #undef PTTYPE
1021
1022 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1023 {
1024         struct kvm_mmu *context = &vcpu->mmu;
1025
1026         ASSERT(is_pae(vcpu));
1027         context->new_cr3 = paging_new_cr3;
1028         context->page_fault = paging64_page_fault;
1029         context->gva_to_gpa = paging64_gva_to_gpa;
1030         context->prefetch_page = paging64_prefetch_page;
1031         context->free = paging_free;
1032         context->root_level = level;
1033         context->shadow_root_level = level;
1034         context->root_hpa = INVALID_PAGE;
1035         return 0;
1036 }
1037
1038 static int paging64_init_context(struct kvm_vcpu *vcpu)
1039 {
1040         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1041 }
1042
1043 static int paging32_init_context(struct kvm_vcpu *vcpu)
1044 {
1045         struct kvm_mmu *context = &vcpu->mmu;
1046
1047         context->new_cr3 = paging_new_cr3;
1048         context->page_fault = paging32_page_fault;
1049         context->gva_to_gpa = paging32_gva_to_gpa;
1050         context->free = paging_free;
1051         context->prefetch_page = paging32_prefetch_page;
1052         context->root_level = PT32_ROOT_LEVEL;
1053         context->shadow_root_level = PT32E_ROOT_LEVEL;
1054         context->root_hpa = INVALID_PAGE;
1055         return 0;
1056 }
1057
1058 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1059 {
1060         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1061 }
1062
1063 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1064 {
1065         ASSERT(vcpu);
1066         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1067
1068         if (!is_paging(vcpu))
1069                 return nonpaging_init_context(vcpu);
1070         else if (is_long_mode(vcpu))
1071                 return paging64_init_context(vcpu);
1072         else if (is_pae(vcpu))
1073                 return paging32E_init_context(vcpu);
1074         else
1075                 return paging32_init_context(vcpu);
1076 }
1077
1078 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1079 {
1080         ASSERT(vcpu);
1081         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1082                 vcpu->mmu.free(vcpu);
1083                 vcpu->mmu.root_hpa = INVALID_PAGE;
1084         }
1085 }
1086
1087 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1088 {
1089         destroy_kvm_mmu(vcpu);
1090         return init_kvm_mmu(vcpu);
1091 }
1092 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1093
1094 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1095 {
1096         int r;
1097
1098         mutex_lock(&vcpu->kvm->lock);
1099         r = mmu_topup_memory_caches(vcpu);
1100         if (r)
1101                 goto out;
1102         mmu_alloc_roots(vcpu);
1103         kvm_x86_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1104         kvm_mmu_flush_tlb(vcpu);
1105 out:
1106         mutex_unlock(&vcpu->kvm->lock);
1107         return r;
1108 }
1109 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1110
1111 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1112 {
1113         mmu_free_roots(vcpu);
1114 }
1115
1116 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1117                                   struct kvm_mmu_page *page,
1118                                   u64 *spte)
1119 {
1120         u64 pte;
1121         struct kvm_mmu_page *child;
1122
1123         pte = *spte;
1124         if (is_shadow_present_pte(pte)) {
1125                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1126                         rmap_remove(spte);
1127                 else {
1128                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1129                         mmu_page_remove_parent_pte(child, spte);
1130                 }
1131         }
1132         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1133         kvm_flush_remote_tlbs(vcpu->kvm);
1134 }
1135
1136 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1137                                   struct kvm_mmu_page *page,
1138                                   u64 *spte,
1139                                   const void *new, int bytes,
1140                                   int offset_in_pte)
1141 {
1142         if (page->role.level != PT_PAGE_TABLE_LEVEL)
1143                 return;
1144
1145         if (page->role.glevels == PT32_ROOT_LEVEL)
1146                 paging32_update_pte(vcpu, page, spte, new, bytes,
1147                                     offset_in_pte);
1148         else
1149                 paging64_update_pte(vcpu, page, spte, new, bytes,
1150                                     offset_in_pte);
1151 }
1152
1153 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1154 {
1155         u64 *spte = vcpu->last_pte_updated;
1156
1157         return !!(spte && (*spte & PT_ACCESSED_MASK));
1158 }
1159
1160 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1161                        const u8 *new, int bytes)
1162 {
1163         gfn_t gfn = gpa >> PAGE_SHIFT;
1164         struct kvm_mmu_page *page;
1165         struct hlist_node *node, *n;
1166         struct hlist_head *bucket;
1167         unsigned index;
1168         u64 *spte;
1169         unsigned offset = offset_in_page(gpa);
1170         unsigned pte_size;
1171         unsigned page_offset;
1172         unsigned misaligned;
1173         unsigned quadrant;
1174         int level;
1175         int flooded = 0;
1176         int npte;
1177
1178         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1179         kvm_mmu_audit(vcpu, "pre pte write");
1180         if (gfn == vcpu->last_pt_write_gfn
1181             && !last_updated_pte_accessed(vcpu)) {
1182                 ++vcpu->last_pt_write_count;
1183                 if (vcpu->last_pt_write_count >= 3)
1184                         flooded = 1;
1185         } else {
1186                 vcpu->last_pt_write_gfn = gfn;
1187                 vcpu->last_pt_write_count = 1;
1188                 vcpu->last_pte_updated = NULL;
1189         }
1190         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1191         bucket = &vcpu->kvm->mmu_page_hash[index];
1192         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1193                 if (page->gfn != gfn || page->role.metaphysical)
1194                         continue;
1195                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1196                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1197                 misaligned |= bytes < 4;
1198                 if (misaligned || flooded) {
1199                         /*
1200                          * Misaligned accesses are too much trouble to fix
1201                          * up; also, they usually indicate a page is not used
1202                          * as a page table.
1203                          *
1204                          * If we're seeing too many writes to a page,
1205                          * it may no longer be a page table, or we may be
1206                          * forking, in which case it is better to unmap the
1207                          * page.
1208                          */
1209                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1210                                  gpa, bytes, page->role.word);
1211                         kvm_mmu_zap_page(vcpu->kvm, page);
1212                         continue;
1213                 }
1214                 page_offset = offset;
1215                 level = page->role.level;
1216                 npte = 1;
1217                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1218                         page_offset <<= 1;      /* 32->64 */
1219                         /*
1220                          * A 32-bit pde maps 4MB while the shadow pdes map
1221                          * only 2MB.  So we need to double the offset again
1222                          * and zap two pdes instead of one.
1223                          */
1224                         if (level == PT32_ROOT_LEVEL) {
1225                                 page_offset &= ~7; /* kill rounding error */
1226                                 page_offset <<= 1;
1227                                 npte = 2;
1228                         }
1229                         quadrant = page_offset >> PAGE_SHIFT;
1230                         page_offset &= ~PAGE_MASK;
1231                         if (quadrant != page->role.quadrant)
1232                                 continue;
1233                 }
1234                 spte = &page->spt[page_offset / sizeof(*spte)];
1235                 while (npte--) {
1236                         mmu_pte_write_zap_pte(vcpu, page, spte);
1237                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes,
1238                                               page_offset & (pte_size - 1));
1239                         ++spte;
1240                 }
1241         }
1242         kvm_mmu_audit(vcpu, "post pte write");
1243 }
1244
1245 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1246 {
1247         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1248
1249         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1250 }
1251
1252 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1253 {
1254         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1255                 struct kvm_mmu_page *page;
1256
1257                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1258                                     struct kvm_mmu_page, link);
1259                 kvm_mmu_zap_page(vcpu->kvm, page);
1260         }
1261 }
1262
1263 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1264 {
1265         struct kvm_mmu_page *page;
1266
1267         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1268                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1269                                     struct kvm_mmu_page, link);
1270                 kvm_mmu_zap_page(vcpu->kvm, page);
1271         }
1272         free_page((unsigned long)vcpu->mmu.pae_root);
1273 }
1274
1275 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1276 {
1277         struct page *page;
1278         int i;
1279
1280         ASSERT(vcpu);
1281
1282         vcpu->kvm->n_free_mmu_pages = KVM_NUM_MMU_PAGES;
1283
1284         /*
1285          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1286          * Therefore we need to allocate shadow page tables in the first
1287          * 4GB of memory, which happens to fit the DMA32 zone.
1288          */
1289         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1290         if (!page)
1291                 goto error_1;
1292         vcpu->mmu.pae_root = page_address(page);
1293         for (i = 0; i < 4; ++i)
1294                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1295
1296         return 0;
1297
1298 error_1:
1299         free_mmu_pages(vcpu);
1300         return -ENOMEM;
1301 }
1302
1303 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1304 {
1305         ASSERT(vcpu);
1306         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1307
1308         return alloc_mmu_pages(vcpu);
1309 }
1310
1311 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1312 {
1313         ASSERT(vcpu);
1314         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1315
1316         return init_kvm_mmu(vcpu);
1317 }
1318
1319 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1320 {
1321         ASSERT(vcpu);
1322
1323         destroy_kvm_mmu(vcpu);
1324         free_mmu_pages(vcpu);
1325         mmu_free_memory_caches(vcpu);
1326 }
1327
1328 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1329 {
1330         struct kvm_mmu_page *page;
1331
1332         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1333                 int i;
1334                 u64 *pt;
1335
1336                 if (!test_bit(slot, &page->slot_bitmap))
1337                         continue;
1338
1339                 pt = page->spt;
1340                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1341                         /* avoid RMW */
1342                         if (pt[i] & PT_WRITABLE_MASK) {
1343                                 rmap_remove(&pt[i]);
1344                                 pt[i] &= ~PT_WRITABLE_MASK;
1345                         }
1346         }
1347 }
1348
1349 void kvm_mmu_zap_all(struct kvm *kvm)
1350 {
1351         struct kvm_mmu_page *page, *node;
1352
1353         list_for_each_entry_safe(page, node, &kvm->active_mmu_pages, link)
1354                 kvm_mmu_zap_page(kvm, page);
1355
1356         kvm_flush_remote_tlbs(kvm);
1357 }
1358
1359 void kvm_mmu_module_exit(void)
1360 {
1361         if (pte_chain_cache)
1362                 kmem_cache_destroy(pte_chain_cache);
1363         if (rmap_desc_cache)
1364                 kmem_cache_destroy(rmap_desc_cache);
1365         if (mmu_page_header_cache)
1366                 kmem_cache_destroy(mmu_page_header_cache);
1367 }
1368
1369 int kvm_mmu_module_init(void)
1370 {
1371         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1372                                             sizeof(struct kvm_pte_chain),
1373                                             0, 0, NULL);
1374         if (!pte_chain_cache)
1375                 goto nomem;
1376         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1377                                             sizeof(struct kvm_rmap_desc),
1378                                             0, 0, NULL);
1379         if (!rmap_desc_cache)
1380                 goto nomem;
1381
1382         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1383                                                   sizeof(struct kvm_mmu_page),
1384                                                   0, 0, NULL);
1385         if (!mmu_page_header_cache)
1386                 goto nomem;
1387
1388         return 0;
1389
1390 nomem:
1391         kvm_mmu_module_exit();
1392         return -ENOMEM;
1393 }
1394
1395 #ifdef AUDIT
1396
1397 static const char *audit_msg;
1398
1399 static gva_t canonicalize(gva_t gva)
1400 {
1401 #ifdef CONFIG_X86_64
1402         gva = (long long)(gva << 16) >> 16;
1403 #endif
1404         return gva;
1405 }
1406
1407 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1408                                 gva_t va, int level)
1409 {
1410         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1411         int i;
1412         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1413
1414         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1415                 u64 ent = pt[i];
1416
1417                 if (ent == shadow_trap_nonpresent_pte)
1418                         continue;
1419
1420                 va = canonicalize(va);
1421                 if (level > 1) {
1422                         if (ent == shadow_notrap_nonpresent_pte)
1423                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1424                                        " in nonleaf level: levels %d gva %lx"
1425                                        " level %d pte %llx\n", audit_msg,
1426                                        vcpu->mmu.root_level, va, level, ent);
1427
1428                         audit_mappings_page(vcpu, ent, va, level - 1);
1429                 } else {
1430                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1431                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1432
1433                         if (is_shadow_present_pte(ent)
1434                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1435                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1436                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1437                                        audit_msg, vcpu->mmu.root_level,
1438                                        va, gpa, hpa, ent, is_shadow_present_pte(ent));
1439                         else if (ent == shadow_notrap_nonpresent_pte
1440                                  && !is_error_hpa(hpa))
1441                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1442                                        " valid guest gva %lx\n", audit_msg, va);
1443
1444                 }
1445         }
1446 }
1447
1448 static void audit_mappings(struct kvm_vcpu *vcpu)
1449 {
1450         unsigned i;
1451
1452         if (vcpu->mmu.root_level == 4)
1453                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1454         else
1455                 for (i = 0; i < 4; ++i)
1456                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1457                                 audit_mappings_page(vcpu,
1458                                                     vcpu->mmu.pae_root[i],
1459                                                     i << 30,
1460                                                     2);
1461 }
1462
1463 static int count_rmaps(struct kvm_vcpu *vcpu)
1464 {
1465         int nmaps = 0;
1466         int i, j, k;
1467
1468         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1469                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1470                 struct kvm_rmap_desc *d;
1471
1472                 for (j = 0; j < m->npages; ++j) {
1473                         struct page *page = m->phys_mem[j];
1474
1475                         if (!page->private)
1476                                 continue;
1477                         if (!(page->private & 1)) {
1478                                 ++nmaps;
1479                                 continue;
1480                         }
1481                         d = (struct kvm_rmap_desc *)(page->private & ~1ul);
1482                         while (d) {
1483                                 for (k = 0; k < RMAP_EXT; ++k)
1484                                         if (d->shadow_ptes[k])
1485                                                 ++nmaps;
1486                                         else
1487                                                 break;
1488                                 d = d->more;
1489                         }
1490                 }
1491         }
1492         return nmaps;
1493 }
1494
1495 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1496 {
1497         int nmaps = 0;
1498         struct kvm_mmu_page *page;
1499         int i;
1500
1501         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1502                 u64 *pt = page->spt;
1503
1504                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1505                         continue;
1506
1507                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1508                         u64 ent = pt[i];
1509
1510                         if (!(ent & PT_PRESENT_MASK))
1511                                 continue;
1512                         if (!(ent & PT_WRITABLE_MASK))
1513                                 continue;
1514                         ++nmaps;
1515                 }
1516         }
1517         return nmaps;
1518 }
1519
1520 static void audit_rmap(struct kvm_vcpu *vcpu)
1521 {
1522         int n_rmap = count_rmaps(vcpu);
1523         int n_actual = count_writable_mappings(vcpu);
1524
1525         if (n_rmap != n_actual)
1526                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1527                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1528 }
1529
1530 static void audit_write_protection(struct kvm_vcpu *vcpu)
1531 {
1532         struct kvm_mmu_page *page;
1533
1534         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1535                 hfn_t hfn;
1536                 struct page *pg;
1537
1538                 if (page->role.metaphysical)
1539                         continue;
1540
1541                 hfn = gpa_to_hpa(vcpu, (gpa_t)page->gfn << PAGE_SHIFT)
1542                         >> PAGE_SHIFT;
1543                 pg = pfn_to_page(hfn);
1544                 if (pg->private)
1545                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1546                                " mappings: gfn %lx role %x\n",
1547                                __FUNCTION__, audit_msg, page->gfn,
1548                                page->role.word);
1549         }
1550 }
1551
1552 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1553 {
1554         int olddbg = dbg;
1555
1556         dbg = 0;
1557         audit_msg = msg;
1558         audit_rmap(vcpu);
1559         audit_write_protection(vcpu);
1560         audit_mappings(vcpu);
1561         dbg = olddbg;
1562 }
1563
1564 #endif