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 \
  94         (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  95
  96
  97 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  98 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  99
 100 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 101
 102 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 103
 104 #define PT64_LEVEL_BITS 9
 105
 106 #define PT64_LEVEL_SHIFT(level) \
 107                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
 108
 109 #define PT64_LEVEL_MASK(level) \
 110                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 111
 112 #define PT64_INDEX(address, level)\
 113         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 114
 115
 116 #define PT32_LEVEL_BITS 10
 117
 118 #define PT32_LEVEL_SHIFT(level) \
 119                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 120
 121 #define PT32_LEVEL_MASK(level) \
 122                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 123
 124 #define PT32_INDEX(address, level)\
 125         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 126
 127
 128 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 129 #define PT64_DIR_BASE_ADDR_MASK \
 130         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 131
 132 #define PT32_BASE_ADDR_MASK PAGE_MASK
 133 #define PT32_DIR_BASE_ADDR_MASK \
 134         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 135
 136
 137 #define PFERR_PRESENT_MASK (1U << 0)
 138 #define PFERR_WRITE_MASK (1U << 1)
 139 #define PFERR_USER_MASK (1U << 2)
 140 #define PFERR_FETCH_MASK (1U << 4)
 141
 142 #define PT64_ROOT_LEVEL 4
 143 #define PT32_ROOT_LEVEL 2
 144 #define PT32E_ROOT_LEVEL 3
 145
 146 #define PT_DIRECTORY_LEVEL 2
 147 #define PT_PAGE_TABLE_LEVEL 1
 148
 149 #define RMAP_EXT 4
 150
 151 struct kvm_rmap_desc {
 152         u64 *shadow_ptes[RMAP_EXT];
 153         struct kvm_rmap_desc *more;
 154 };
 155
 156 static struct kmem_cache *pte_chain_cache;
 157 static struct kmem_cache *rmap_desc_cache;
 158 static struct kmem_cache *mmu_page_header_cache;
 159
 160 static u64 __read_mostly shadow_trap_nonpresent_pte;
 161 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 162
 163 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 164 {
 165         shadow_trap_nonpresent_pte = trap_pte;
 166         shadow_notrap_nonpresent_pte = notrap_pte;
 167 }
 168 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 169
 170 static int is_write_protection(struct kvm_vcpu *vcpu)
 171 {
 172         return vcpu->cr0 & X86_CR0_WP;
 173 }
 174
 175 static int is_cpuid_PSE36(void)
 176 {
 177         return 1;
 178 }
 179
 180 static int is_nx(struct kvm_vcpu *vcpu)
 181 {
 182         return vcpu->shadow_efer & EFER_NX;
 183 }
 184
 185 static int is_present_pte(unsigned long pte)
 186 {
 187         return pte & PT_PRESENT_MASK;
 188 }
 189
 190 static int is_shadow_present_pte(u64 pte)
 191 {
 192         pte &= ~PT_SHADOW_IO_MARK;
 193         return pte != shadow_trap_nonpresent_pte
 194                 && pte != shadow_notrap_nonpresent_pte;
 195 }
 196
 197 static int is_writeble_pte(unsigned long pte)
 198 {
 199         return pte & PT_WRITABLE_MASK;
 200 }
 201
 202 static int is_dirty_pte(unsigned long pte)
 203 {
 204         return pte & PT_DIRTY_MASK;
 205 }
 206
 207 static int is_io_pte(unsigned long pte)
 208 {
 209         return pte & PT_SHADOW_IO_MARK;
 210 }
 211
 212 static int is_rmap_pte(u64 pte)
 213 {
 214         return (pte & (PT_WRITABLE_MASK | PT_PRESENT_MASK))
 215                 == (PT_WRITABLE_MASK | PT_PRESENT_MASK);
 216 }
 217
 218 static void set_shadow_pte(u64 *sptep, u64 spte)
 219 {
 220 #ifdef CONFIG_X86_64
 221         set_64bit((unsigned long *)sptep, spte);
 222 #else
 223         set_64bit((unsigned long long *)sptep, spte);
 224 #endif
 225 }
 226
 227 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 228                                   struct kmem_cache *base_cache, int min)
 229 {
 230         void *obj;
 231
 232         if (cache->nobjs >= min)
 233                 return 0;
 234         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 235                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 236                 if (!obj)
 237                         return -ENOMEM;
 238                 cache->objects[cache->nobjs++] = obj;
 239         }
 240         return 0;
 241 }
 242
 243 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 244 {
 245         while (mc->nobjs)
 246                 kfree(mc->objects[--mc->nobjs]);
 247 }
 248
 249 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 250                                        int min)
 251 {
 252         struct page *page;
 253
 254         if (cache->nobjs >= min)
 255                 return 0;
 256         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 257                 page = alloc_page(GFP_KERNEL);
 258                 if (!page)
 259                         return -ENOMEM;
 260                 set_page_private(page, 0);
 261                 cache->objects[cache->nobjs++] = page_address(page);
 262         }
 263         return 0;
 264 }
 265
 266 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 267 {
 268         while (mc->nobjs)
 269                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 270 }
 271
 272 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 273 {
 274         int r;
 275
 276         kvm_mmu_free_some_pages(vcpu);
 277         r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
 278                                    pte_chain_cache, 4);
 279         if (r)
 280                 goto out;
 281         r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
 282                                    rmap_desc_cache, 1);
 283         if (r)
 284                 goto out;
 285         r = mmu_topup_memory_cache_page(&vcpu->mmu_page_cache, 8);
 286         if (r)
 287                 goto out;
 288         r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
 289                                    mmu_page_header_cache, 4);
 290 out:
 291         return r;
 292 }
 293
 294 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 295 {
 296         mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
 297         mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
 298         mmu_free_memory_cache_page(&vcpu->mmu_page_cache);
 299         mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
 300 }
 301
 302 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 303                                     size_t size)
 304 {
 305         void *p;
 306
 307         BUG_ON(!mc->nobjs);
 308         p = mc->objects[--mc->nobjs];
 309         memset(p, 0, size);
 310         return p;
 311 }
 312
 313 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 314 {
 315         return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
 316                                       sizeof(struct kvm_pte_chain));
 317 }
 318
 319 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 320 {
 321         kfree(pc);
 322 }
 323
 324 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 325 {
 326         return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
 327                                       sizeof(struct kvm_rmap_desc));
 328 }
 329
 330 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 331 {
 332         kfree(rd);
 333 }
 334
 335 /*
 336  * Take gfn and return the reverse mapping to it.
 337  * Note: gfn must be unaliased before this function get called
 338  */
 339
 340 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
 341 {
 342         struct kvm_memory_slot *slot;
 343
 344         slot = gfn_to_memslot(kvm, gfn);
 345         return &slot->rmap[gfn - slot->base_gfn];
 346 }
 347
 348 /*
 349  * Reverse mapping data structures:
 350  *
 351  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 352  * that points to page_address(page).
 353  *
 354  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 355  * containing more mappings.
 356  */
 357 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 358 {
 359         struct kvm_mmu_page *page;
 360         struct kvm_rmap_desc *desc;
 361         unsigned long *rmapp;
 362         int i;
 363
 364         if (!is_rmap_pte(*spte))
 365                 return;
 366         gfn = unalias_gfn(vcpu->kvm, gfn);
 367         page = page_header(__pa(spte));
 368         page->gfns[spte - page->spt] = gfn;
 369         rmapp = gfn_to_rmap(vcpu->kvm, gfn);
 370         if (!*rmapp) {
 371                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 372                 *rmapp = (unsigned long)spte;
 373         } else if (!(*rmapp & 1)) {
 374                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 375                 desc = mmu_alloc_rmap_desc(vcpu);
 376                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 377                 desc->shadow_ptes[1] = spte;
 378                 *rmapp = (unsigned long)desc | 1;
 379         } else {
 380                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 381                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 382                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 383                         desc = desc->more;
 384                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 385                         desc->more = mmu_alloc_rmap_desc(vcpu);
 386                         desc = desc->more;
 387                 }
 388                 for (i = 0; desc->shadow_ptes[i]; ++i)
 389                         ;
 390                 desc->shadow_ptes[i] = spte;
 391         }
 392 }
 393
 394 static void rmap_desc_remove_entry(unsigned long *rmapp,
 395                                    struct kvm_rmap_desc *desc,
 396                                    int i,
 397                                    struct kvm_rmap_desc *prev_desc)
 398 {
 399         int j;
 400
 401         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 402                 ;
 403         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 404         desc->shadow_ptes[j] = NULL;
 405         if (j != 0)
 406                 return;
 407         if (!prev_desc && !desc->more)
 408                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 409         else
 410                 if (prev_desc)
 411                         prev_desc->more = desc->more;
 412                 else
 413                         *rmapp = (unsigned long)desc->more | 1;
 414         mmu_free_rmap_desc(desc);
 415 }
 416
 417 static void rmap_remove(struct kvm *kvm, u64 *spte)
 418 {
 419         struct kvm_rmap_desc *desc;
 420         struct kvm_rmap_desc *prev_desc;
 421         struct kvm_mmu_page *page;
 422         unsigned long *rmapp;
 423         int i;
 424
 425         if (!is_rmap_pte(*spte))
 426                 return;
 427         page = page_header(__pa(spte));
 428         rmapp = gfn_to_rmap(kvm, page->gfns[spte - page->spt]);
 429         if (!*rmapp) {
 430                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 431                 BUG();
 432         } else if (!(*rmapp & 1)) {
 433                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 434                 if ((u64 *)*rmapp != spte) {
 435                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 436                                spte, *spte);
 437                         BUG();
 438                 }
 439                 *rmapp = 0;
 440         } else {
 441                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 442                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 443                 prev_desc = NULL;
 444                 while (desc) {
 445                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 446                                 if (desc->shadow_ptes[i] == spte) {
 447                                         rmap_desc_remove_entry(rmapp,
 448                                                                desc, i,
 449                                                                prev_desc);
 450                                         return;
 451                                 }
 452                         prev_desc = desc;
 453                         desc = desc->more;
 454                 }
 455                 BUG();
 456         }
 457 }
 458
 459 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 460 {
 461         struct kvm_rmap_desc *desc;
 462         unsigned long *rmapp;
 463         u64 *spte;
 464
 465         gfn = unalias_gfn(kvm, gfn);
 466         rmapp = gfn_to_rmap(kvm, gfn);
 467
 468         while (*rmapp) {
 469                 if (!(*rmapp & 1))
 470                         spte = (u64 *)*rmapp;
 471                 else {
 472                         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 473                         spte = desc->shadow_ptes[0];
 474                 }
 475                 BUG_ON(!spte);
 476                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 477                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 478                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 479                 rmap_remove(kvm, spte);
 480                 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 481                 kvm_flush_remote_tlbs(kvm);
 482         }
 483 }
 484
 485 #ifdef MMU_DEBUG
 486 static int is_empty_shadow_page(u64 *spt)
 487 {
 488         u64 *pos;
 489         u64 *end;
 490
 491         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 492                 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
 493                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 494                                pos, *pos);
 495                         return 0;
 496                 }
 497         return 1;
 498 }
 499 #endif
 500
 501 static void kvm_mmu_free_page(struct kvm *kvm,
 502                               struct kvm_mmu_page *page_head)
 503 {
 504         ASSERT(is_empty_shadow_page(page_head->spt));
 505         list_del(&page_head->link);
 506         __free_page(virt_to_page(page_head->spt));
 507         __free_page(virt_to_page(page_head->gfns));
 508         kfree(page_head);
 509         ++kvm->n_free_mmu_pages;
 510 }
 511
 512 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 513 {
 514         return gfn;
 515 }
 516
 517 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 518                                                u64 *parent_pte)
 519 {
 520         struct kvm_mmu_page *page;
 521
 522         if (!vcpu->kvm->n_free_mmu_pages)
 523                 return NULL;
 524
 525         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 526                                       sizeof *page);
 527         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 528         page->gfns = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 529         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 530         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 531         ASSERT(is_empty_shadow_page(page->spt));
 532         page->slot_bitmap = 0;
 533         page->multimapped = 0;
 534         page->parent_pte = parent_pte;
 535         --vcpu->kvm->n_free_mmu_pages;
 536         return page;
 537 }
 538
 539 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 540                                     struct kvm_mmu_page *page, u64 *parent_pte)
 541 {
 542         struct kvm_pte_chain *pte_chain;
 543         struct hlist_node *node;
 544         int i;
 545
 546         if (!parent_pte)
 547                 return;
 548         if (!page->multimapped) {
 549                 u64 *old = page->parent_pte;
 550
 551                 if (!old) {
 552                         page->parent_pte = parent_pte;
 553                         return;
 554                 }
 555                 page->multimapped = 1;
 556                 pte_chain = mmu_alloc_pte_chain(vcpu);
 557                 INIT_HLIST_HEAD(&page->parent_ptes);
 558                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 559                 pte_chain->parent_ptes[0] = old;
 560         }
 561         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 562                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 563                         continue;
 564                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 565                         if (!pte_chain->parent_ptes[i]) {
 566                                 pte_chain->parent_ptes[i] = parent_pte;
 567                                 return;
 568                         }
 569         }
 570         pte_chain = mmu_alloc_pte_chain(vcpu);
 571         BUG_ON(!pte_chain);
 572         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 573         pte_chain->parent_ptes[0] = parent_pte;
 574 }
 575
 576 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
 577                                        u64 *parent_pte)
 578 {
 579         struct kvm_pte_chain *pte_chain;
 580         struct hlist_node *node;
 581         int i;
 582
 583         if (!page->multimapped) {
 584                 BUG_ON(page->parent_pte != parent_pte);
 585                 page->parent_pte = NULL;
 586                 return;
 587         }
 588         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 589                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 590                         if (!pte_chain->parent_ptes[i])
 591                                 break;
 592                         if (pte_chain->parent_ptes[i] != parent_pte)
 593                                 continue;
 594                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 595                                 && pte_chain->parent_ptes[i + 1]) {
 596                                 pte_chain->parent_ptes[i]
 597                                         = pte_chain->parent_ptes[i + 1];
 598                                 ++i;
 599                         }
 600                         pte_chain->parent_ptes[i] = NULL;
 601                         if (i == 0) {
 602                                 hlist_del(&pte_chain->link);
 603                                 mmu_free_pte_chain(pte_chain);
 604                                 if (hlist_empty(&page->parent_ptes)) {
 605                                         page->multimapped = 0;
 606                                         page->parent_pte = NULL;
 607                                 }
 608                         }
 609                         return;
 610                 }
 611         BUG();
 612 }
 613
 614 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm,
 615                                                 gfn_t gfn)
 616 {
 617         unsigned index;
 618         struct hlist_head *bucket;
 619         struct kvm_mmu_page *page;
 620         struct hlist_node *node;
 621
 622         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 623         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 624         bucket = &kvm->mmu_page_hash[index];
 625         hlist_for_each_entry(page, node, bucket, hash_link)
 626                 if (page->gfn == gfn && !page->role.metaphysical) {
 627                         pgprintk("%s: found role %x\n",
 628                                  __FUNCTION__, page->role.word);
 629                         return page;
 630                 }
 631         return NULL;
 632 }
 633
 634 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 635                                              gfn_t gfn,
 636                                              gva_t gaddr,
 637                                              unsigned level,
 638                                              int metaphysical,
 639                                              unsigned hugepage_access,
 640                                              u64 *parent_pte)
 641 {
 642         union kvm_mmu_page_role role;
 643         unsigned index;
 644         unsigned quadrant;
 645         struct hlist_head *bucket;
 646         struct kvm_mmu_page *page;
 647         struct hlist_node *node;
 648
 649         role.word = 0;
 650         role.glevels = vcpu->mmu.root_level;
 651         role.level = level;
 652         role.metaphysical = metaphysical;
 653         role.hugepage_access = hugepage_access;
 654         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 655                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 656                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 657                 role.quadrant = quadrant;
 658         }
 659         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 660                  gfn, role.word);
 661         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 662         bucket = &vcpu->kvm->mmu_page_hash[index];
 663         hlist_for_each_entry(page, node, bucket, hash_link)
 664                 if (page->gfn == gfn && page->role.word == role.word) {
 665                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 666                         pgprintk("%s: found\n", __FUNCTION__);
 667                         return page;
 668                 }
 669         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 670         if (!page)
 671                 return page;
 672         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 673         page->gfn = gfn;
 674         page->role = role;
 675         hlist_add_head(&page->hash_link, bucket);
 676         vcpu->mmu.prefetch_page(vcpu, page);
 677         if (!metaphysical)
 678                 rmap_write_protect(vcpu->kvm, gfn);
 679         return page;
 680 }
 681
 682 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 683                                          struct kvm_mmu_page *page)
 684 {
 685         unsigned i;
 686         u64 *pt;
 687         u64 ent;
 688
 689         pt = page->spt;
 690
 691         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 692                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 693                         if (is_shadow_present_pte(pt[i]))
 694                                 rmap_remove(kvm, &pt[i]);
 695                         pt[i] = shadow_trap_nonpresent_pte;
 696                 }
 697                 kvm_flush_remote_tlbs(kvm);
 698                 return;
 699         }
 700
 701         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 702                 ent = pt[i];
 703
 704                 pt[i] = shadow_trap_nonpresent_pte;
 705                 if (!is_shadow_present_pte(ent))
 706                         continue;
 707                 ent &= PT64_BASE_ADDR_MASK;
 708                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 709         }
 710         kvm_flush_remote_tlbs(kvm);
 711 }
 712
 713 static void kvm_mmu_put_page(struct kvm_mmu_page *page,
 714                              u64 *parent_pte)
 715 {
 716         mmu_page_remove_parent_pte(page, parent_pte);
 717 }
 718
 719 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 720 {
 721         int i;
 722
 723         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 724                 if (kvm->vcpus[i])
 725                         kvm->vcpus[i]->last_pte_updated = NULL;
 726 }
 727
 728 static void kvm_mmu_zap_page(struct kvm *kvm,
 729                              struct kvm_mmu_page *page)
 730 {
 731         u64 *parent_pte;
 732
 733         while (page->multimapped || page->parent_pte) {
 734                 if (!page->multimapped)
 735                         parent_pte = page->parent_pte;
 736                 else {
 737                         struct kvm_pte_chain *chain;
 738
 739                         chain = container_of(page->parent_ptes.first,
 740                                              struct kvm_pte_chain, link);
 741                         parent_pte = chain->parent_ptes[0];
 742                 }
 743                 BUG_ON(!parent_pte);
 744                 kvm_mmu_put_page(page, parent_pte);
 745                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 746         }
 747         kvm_mmu_page_unlink_children(kvm, page);
 748         if (!page->root_count) {
 749                 hlist_del(&page->hash_link);
 750                 kvm_mmu_free_page(kvm, page);
 751         } else
 752                 list_move(&page->link, &kvm->active_mmu_pages);
 753         kvm_mmu_reset_last_pte_updated(kvm);
 754 }
 755
 756 /*
 757  * Changing the number of mmu pages allocated to the vm
 758  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 759  */
 760 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 761 {
 762         /*
 763          * If we set the number of mmu pages to be smaller be than the
 764          * number of actived pages , we must to free some mmu pages before we
 765          * change the value
 766          */
 767
 768         if ((kvm->n_alloc_mmu_pages - kvm->n_free_mmu_pages) >
 769             kvm_nr_mmu_pages) {
 770                 int n_used_mmu_pages = kvm->n_alloc_mmu_pages
 771                                        - kvm->n_free_mmu_pages;
 772
 773                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 774                         struct kvm_mmu_page *page;
 775
 776                         page = container_of(kvm->active_mmu_pages.prev,
 777                                             struct kvm_mmu_page, link);
 778                         kvm_mmu_zap_page(kvm, page);
 779                         n_used_mmu_pages--;
 780                 }
 781                 kvm->n_free_mmu_pages = 0;
 782         }
 783         else
 784                 kvm->n_free_mmu_pages += kvm_nr_mmu_pages
 785                                          - kvm->n_alloc_mmu_pages;
 786
 787         kvm->n_alloc_mmu_pages = kvm_nr_mmu_pages;
 788 }
 789
 790 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 791 {
 792         unsigned index;
 793         struct hlist_head *bucket;
 794         struct kvm_mmu_page *page;
 795         struct hlist_node *node, *n;
 796         int r;
 797
 798         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 799         r = 0;
 800         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 801         bucket = &kvm->mmu_page_hash[index];
 802         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 803                 if (page->gfn == gfn && !page->role.metaphysical) {
 804                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 805                                  page->role.word);
 806                         kvm_mmu_zap_page(kvm, page);
 807                         r = 1;
 808                 }
 809         return r;
 810 }
 811
 812 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
 813 {
 814         struct kvm_mmu_page *page;
 815
 816         while ((page = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
 817                 pgprintk("%s: zap %lx %x\n",
 818                          __FUNCTION__, gfn, page->role.word);
 819                 kvm_mmu_zap_page(kvm, page);
 820         }
 821 }
 822
 823 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 824 {
 825         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 826         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 827
 828         __set_bit(slot, &page_head->slot_bitmap);
 829 }
 830
 831 hpa_t safe_gpa_to_hpa(struct kvm *kvm, gpa_t gpa)
 832 {
 833         hpa_t hpa = gpa_to_hpa(kvm, gpa);
 834
 835         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 836 }
 837
 838 hpa_t gpa_to_hpa(struct kvm *kvm, gpa_t gpa)
 839 {
 840         struct page *page;
 841
 842         ASSERT((gpa & HPA_ERR_MASK) == 0);
 843         page = gfn_to_page(kvm, gpa >> PAGE_SHIFT);
 844         if (!page)
 845                 return gpa | HPA_ERR_MASK;
 846         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 847                 | (gpa & (PAGE_SIZE-1));
 848 }
 849
 850 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 851 {
 852         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 853
 854         if (gpa == UNMAPPED_GVA)
 855                 return UNMAPPED_GVA;
 856         return gpa_to_hpa(vcpu->kvm, gpa);
 857 }
 858
 859 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 860 {
 861         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 862
 863         if (gpa == UNMAPPED_GVA)
 864                 return NULL;
 865         return pfn_to_page(gpa_to_hpa(vcpu->kvm, gpa) >> PAGE_SHIFT);
 866 }
 867
 868 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 869 {
 870 }
 871
 872 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 873 {
 874         int level = PT32E_ROOT_LEVEL;
 875         hpa_t table_addr = vcpu->mmu.root_hpa;
 876
 877         for (; ; level--) {
 878                 u32 index = PT64_INDEX(v, level);
 879                 u64 *table;
 880                 u64 pte;
 881
 882                 ASSERT(VALID_PAGE(table_addr));
 883                 table = __va(table_addr);
 884
 885                 if (level == 1) {
 886                         pte = table[index];
 887                         if (is_shadow_present_pte(pte) && is_writeble_pte(pte))
 888                                 return 0;
 889                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 890                         page_header_update_slot(vcpu->kvm, table, v);
 891                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 892                                                                 PT_USER_MASK;
 893                         rmap_add(vcpu, &table[index], v >> PAGE_SHIFT);
 894                         return 0;
 895                 }
 896
 897                 if (table[index] == shadow_trap_nonpresent_pte) {
 898                         struct kvm_mmu_page *new_table;
 899                         gfn_t pseudo_gfn;
 900
 901                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 902                                 >> PAGE_SHIFT;
 903                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 904                                                      v, level - 1,
 905                                                      1, 3, &table[index]);
 906                         if (!new_table) {
 907                                 pgprintk("nonpaging_map: ENOMEM\n");
 908                                 return -ENOMEM;
 909                         }
 910
 911                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 912                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 913                 }
 914                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 915         }
 916 }
 917
 918 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 919                                     struct kvm_mmu_page *sp)
 920 {
 921         int i;
 922
 923         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 924                 sp->spt[i] = shadow_trap_nonpresent_pte;
 925 }
 926
 927 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 928 {
 929         int i;
 930         struct kvm_mmu_page *page;
 931
 932         if (!VALID_PAGE(vcpu->mmu.root_hpa))
 933                 return;
 934 #ifdef CONFIG_X86_64
 935         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 936                 hpa_t root = vcpu->mmu.root_hpa;
 937
 938                 page = page_header(root);
 939                 --page->root_count;
 940                 vcpu->mmu.root_hpa = INVALID_PAGE;
 941                 return;
 942         }
 943 #endif
 944         for (i = 0; i < 4; ++i) {
 945                 hpa_t root = vcpu->mmu.pae_root[i];
 946
 947                 if (root) {
 948                         root &= PT64_BASE_ADDR_MASK;
 949                         page = page_header(root);
 950                         --page->root_count;
 951                 }
 952                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 953         }
 954         vcpu->mmu.root_hpa = INVALID_PAGE;
 955 }
 956
 957 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 958 {
 959         int i;
 960         gfn_t root_gfn;
 961         struct kvm_mmu_page *page;
 962
 963         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 964
 965 #ifdef CONFIG_X86_64
 966         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 967                 hpa_t root = vcpu->mmu.root_hpa;
 968
 969                 ASSERT(!VALID_PAGE(root));
 970                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 971                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 972                 root = __pa(page->spt);
 973                 ++page->root_count;
 974                 vcpu->mmu.root_hpa = root;
 975                 return;
 976         }
 977 #endif
 978         for (i = 0; i < 4; ++i) {
 979                 hpa_t root = vcpu->mmu.pae_root[i];
 980
 981                 ASSERT(!VALID_PAGE(root));
 982                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 983                         if (!is_present_pte(vcpu->pdptrs[i])) {
 984                                 vcpu->mmu.pae_root[i] = 0;
 985                                 continue;
 986                         }
 987                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 988                 } else if (vcpu->mmu.root_level == 0)
 989                         root_gfn = 0;
 990                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 991                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 992                                         0, NULL);
 993                 root = __pa(page->spt);
 994                 ++page->root_count;
 995                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 996         }
 997         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 998 }
 999
1000 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1001 {
1002         return vaddr;
1003 }
1004
1005 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1006                                u32 error_code)
1007 {
1008         gpa_t addr = gva;
1009         hpa_t paddr;
1010         int r;
1011
1012         r = mmu_topup_memory_caches(vcpu);
1013         if (r)
1014                 return r;
1015
1016         ASSERT(vcpu);
1017         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
1018
1019
1020         paddr = gpa_to_hpa(vcpu->kvm, addr & PT64_BASE_ADDR_MASK);
1021
1022         if (is_error_hpa(paddr))
1023                 return 1;
1024
1025         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
1026 }
1027
1028 static void nonpaging_free(struct kvm_vcpu *vcpu)
1029 {
1030         mmu_free_roots(vcpu);
1031 }
1032
1033 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1034 {
1035         struct kvm_mmu *context = &vcpu->mmu;
1036
1037         context->new_cr3 = nonpaging_new_cr3;
1038         context->page_fault = nonpaging_page_fault;
1039         context->gva_to_gpa = nonpaging_gva_to_gpa;
1040         context->free = nonpaging_free;
1041         context->prefetch_page = nonpaging_prefetch_page;
1042         context->root_level = 0;
1043         context->shadow_root_level = PT32E_ROOT_LEVEL;
1044         context->root_hpa = INVALID_PAGE;
1045         return 0;
1046 }
1047
1048 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1049 {
1050         ++vcpu->stat.tlb_flush;
1051         kvm_x86_ops->tlb_flush(vcpu);
1052 }
1053
1054 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1055 {
1056         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
1057         mmu_free_roots(vcpu);
1058 }
1059
1060 static void inject_page_fault(struct kvm_vcpu *vcpu,
1061                               u64 addr,
1062                               u32 err_code)
1063 {
1064         kvm_x86_ops->inject_page_fault(vcpu, addr, err_code);
1065 }
1066
1067 static void paging_free(struct kvm_vcpu *vcpu)
1068 {
1069         nonpaging_free(vcpu);
1070 }
1071
1072 #define PTTYPE 64
1073 #include "paging_tmpl.h"
1074 #undef PTTYPE
1075
1076 #define PTTYPE 32
1077 #include "paging_tmpl.h"
1078 #undef PTTYPE
1079
1080 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1081 {
1082         struct kvm_mmu *context = &vcpu->mmu;
1083
1084         ASSERT(is_pae(vcpu));
1085         context->new_cr3 = paging_new_cr3;
1086         context->page_fault = paging64_page_fault;
1087         context->gva_to_gpa = paging64_gva_to_gpa;
1088         context->prefetch_page = paging64_prefetch_page;
1089         context->free = paging_free;
1090         context->root_level = level;
1091         context->shadow_root_level = level;
1092         context->root_hpa = INVALID_PAGE;
1093         return 0;
1094 }
1095
1096 static int paging64_init_context(struct kvm_vcpu *vcpu)
1097 {
1098         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1099 }
1100
1101 static int paging32_init_context(struct kvm_vcpu *vcpu)
1102 {
1103         struct kvm_mmu *context = &vcpu->mmu;
1104
1105         context->new_cr3 = paging_new_cr3;
1106         context->page_fault = paging32_page_fault;
1107         context->gva_to_gpa = paging32_gva_to_gpa;
1108         context->free = paging_free;
1109         context->prefetch_page = paging32_prefetch_page;
1110         context->root_level = PT32_ROOT_LEVEL;
1111         context->shadow_root_level = PT32E_ROOT_LEVEL;
1112         context->root_hpa = INVALID_PAGE;
1113         return 0;
1114 }
1115
1116 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1117 {
1118         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1119 }
1120
1121 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1122 {
1123         ASSERT(vcpu);
1124         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1125
1126         if (!is_paging(vcpu))
1127                 return nonpaging_init_context(vcpu);
1128         else if (is_long_mode(vcpu))
1129                 return paging64_init_context(vcpu);
1130         else if (is_pae(vcpu))
1131                 return paging32E_init_context(vcpu);
1132         else
1133                 return paging32_init_context(vcpu);
1134 }
1135
1136 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1137 {
1138         ASSERT(vcpu);
1139         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1140                 vcpu->mmu.free(vcpu);
1141                 vcpu->mmu.root_hpa = INVALID_PAGE;
1142         }
1143 }
1144
1145 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1146 {
1147         destroy_kvm_mmu(vcpu);
1148         return init_kvm_mmu(vcpu);
1149 }
1150 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1151
1152 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1153 {
1154         int r;
1155
1156         mutex_lock(&vcpu->kvm->lock);
1157         r = mmu_topup_memory_caches(vcpu);
1158         if (r)
1159                 goto out;
1160         mmu_alloc_roots(vcpu);
1161         kvm_x86_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1162         kvm_mmu_flush_tlb(vcpu);
1163 out:
1164         mutex_unlock(&vcpu->kvm->lock);
1165         return r;
1166 }
1167 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1168
1169 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1170 {
1171         mmu_free_roots(vcpu);
1172 }
1173
1174 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1175                                   struct kvm_mmu_page *page,
1176                                   u64 *spte)
1177 {
1178         u64 pte;
1179         struct kvm_mmu_page *child;
1180
1181         pte = *spte;
1182         if (is_shadow_present_pte(pte)) {
1183                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1184                         rmap_remove(vcpu->kvm, spte);
1185                 else {
1186                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1187                         mmu_page_remove_parent_pte(child, spte);
1188                 }
1189         }
1190         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1191         kvm_flush_remote_tlbs(vcpu->kvm);
1192 }
1193
1194 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1195                                   struct kvm_mmu_page *page,
1196                                   u64 *spte,
1197                                   const void *new, int bytes,
1198                                   int offset_in_pte)
1199 {
1200         if (page->role.level != PT_PAGE_TABLE_LEVEL)
1201                 return;
1202
1203         if (page->role.glevels == PT32_ROOT_LEVEL)
1204                 paging32_update_pte(vcpu, page, spte, new, bytes,
1205                                     offset_in_pte);
1206         else
1207                 paging64_update_pte(vcpu, page, spte, new, bytes,
1208                                     offset_in_pte);
1209 }
1210
1211 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1212 {
1213         u64 *spte = vcpu->last_pte_updated;
1214
1215         return !!(spte && (*spte & PT_ACCESSED_MASK));
1216 }
1217
1218 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1219                        const u8 *new, int bytes)
1220 {
1221         gfn_t gfn = gpa >> PAGE_SHIFT;
1222         struct kvm_mmu_page *page;
1223         struct hlist_node *node, *n;
1224         struct hlist_head *bucket;
1225         unsigned index;
1226         u64 *spte;
1227         unsigned offset = offset_in_page(gpa);
1228         unsigned pte_size;
1229         unsigned page_offset;
1230         unsigned misaligned;
1231         unsigned quadrant;
1232         int level;
1233         int flooded = 0;
1234         int npte;
1235
1236         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1237         kvm_mmu_audit(vcpu, "pre pte write");
1238         if (gfn == vcpu->last_pt_write_gfn
1239             && !last_updated_pte_accessed(vcpu)) {
1240                 ++vcpu->last_pt_write_count;
1241                 if (vcpu->last_pt_write_count >= 3)
1242                         flooded = 1;
1243         } else {
1244                 vcpu->last_pt_write_gfn = gfn;
1245                 vcpu->last_pt_write_count = 1;
1246                 vcpu->last_pte_updated = NULL;
1247         }
1248         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1249         bucket = &vcpu->kvm->mmu_page_hash[index];
1250         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1251                 if (page->gfn != gfn || page->role.metaphysical)
1252                         continue;
1253                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1254                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1255                 misaligned |= bytes < 4;
1256                 if (misaligned || flooded) {
1257                         /*
1258                          * Misaligned accesses are too much trouble to fix
1259                          * up; also, they usually indicate a page is not used
1260                          * as a page table.
1261                          *
1262                          * If we're seeing too many writes to a page,
1263                          * it may no longer be a page table, or we may be
1264                          * forking, in which case it is better to unmap the
1265                          * page.
1266                          */
1267                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1268                                  gpa, bytes, page->role.word);
1269                         kvm_mmu_zap_page(vcpu->kvm, page);
1270                         continue;
1271                 }
1272                 page_offset = offset;
1273                 level = page->role.level;
1274                 npte = 1;
1275                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1276                         page_offset <<= 1;      /* 32->64 */
1277                         /*
1278                          * A 32-bit pde maps 4MB while the shadow pdes map
1279                          * only 2MB.  So we need to double the offset again
1280                          * and zap two pdes instead of one.
1281                          */
1282                         if (level == PT32_ROOT_LEVEL) {
1283                                 page_offset &= ~7; /* kill rounding error */
1284                                 page_offset <<= 1;
1285                                 npte = 2;
1286                         }
1287                         quadrant = page_offset >> PAGE_SHIFT;
1288                         page_offset &= ~PAGE_MASK;
1289                         if (quadrant != page->role.quadrant)
1290                                 continue;
1291                 }
1292                 spte = &page->spt[page_offset / sizeof(*spte)];
1293                 while (npte--) {
1294                         mmu_pte_write_zap_pte(vcpu, page, spte);
1295                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes,
1296                                               page_offset & (pte_size - 1));
1297                         ++spte;
1298                 }
1299         }
1300         kvm_mmu_audit(vcpu, "post pte write");
1301 }
1302
1303 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1304 {
1305         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1306
1307         return kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1308 }
1309
1310 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1311 {
1312         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1313                 struct kvm_mmu_page *page;
1314
1315                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1316                                     struct kvm_mmu_page, link);
1317                 kvm_mmu_zap_page(vcpu->kvm, page);
1318         }
1319 }
1320
1321 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1322 {
1323         struct kvm_mmu_page *page;
1324
1325         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1326                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1327                                     struct kvm_mmu_page, link);
1328                 kvm_mmu_zap_page(vcpu->kvm, page);
1329         }
1330         free_page((unsigned long)vcpu->mmu.pae_root);
1331 }
1332
1333 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1334 {
1335         struct page *page;
1336         int i;
1337
1338         ASSERT(vcpu);
1339
1340         if (vcpu->kvm->n_requested_mmu_pages)
1341                 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_requested_mmu_pages;
1342         else
1343                 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_alloc_mmu_pages;
1344         /*
1345          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1346          * Therefore we need to allocate shadow page tables in the first
1347          * 4GB of memory, which happens to fit the DMA32 zone.
1348          */
1349         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1350         if (!page)
1351                 goto error_1;
1352         vcpu->mmu.pae_root = page_address(page);
1353         for (i = 0; i < 4; ++i)
1354                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1355
1356         return 0;
1357
1358 error_1:
1359         free_mmu_pages(vcpu);
1360         return -ENOMEM;
1361 }
1362
1363 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1364 {
1365         ASSERT(vcpu);
1366         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1367
1368         return alloc_mmu_pages(vcpu);
1369 }
1370
1371 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1372 {
1373         ASSERT(vcpu);
1374         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1375
1376         return init_kvm_mmu(vcpu);
1377 }
1378
1379 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1380 {
1381         ASSERT(vcpu);
1382
1383         destroy_kvm_mmu(vcpu);
1384         free_mmu_pages(vcpu);
1385         mmu_free_memory_caches(vcpu);
1386 }
1387
1388 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1389 {
1390         struct kvm_mmu_page *page;
1391
1392         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1393                 int i;
1394                 u64 *pt;
1395
1396                 if (!test_bit(slot, &page->slot_bitmap))
1397                         continue;
1398
1399                 pt = page->spt;
1400                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1401                         /* avoid RMW */
1402                         if (pt[i] & PT_WRITABLE_MASK) {
1403                                 rmap_remove(kvm, &pt[i]);
1404                                 pt[i] &= ~PT_WRITABLE_MASK;
1405                         }
1406         }
1407 }
1408
1409 void kvm_mmu_zap_all(struct kvm *kvm)
1410 {
1411         struct kvm_mmu_page *page, *node;
1412
1413         list_for_each_entry_safe(page, node, &kvm->active_mmu_pages, link)
1414                 kvm_mmu_zap_page(kvm, page);
1415
1416         kvm_flush_remote_tlbs(kvm);
1417 }
1418
1419 void kvm_mmu_module_exit(void)
1420 {
1421         if (pte_chain_cache)
1422                 kmem_cache_destroy(pte_chain_cache);
1423         if (rmap_desc_cache)
1424                 kmem_cache_destroy(rmap_desc_cache);
1425         if (mmu_page_header_cache)
1426                 kmem_cache_destroy(mmu_page_header_cache);
1427 }
1428
1429 int kvm_mmu_module_init(void)
1430 {
1431         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1432                                             sizeof(struct kvm_pte_chain),
1433                                             0, 0, NULL);
1434         if (!pte_chain_cache)
1435                 goto nomem;
1436         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1437                                             sizeof(struct kvm_rmap_desc),
1438                                             0, 0, NULL);
1439         if (!rmap_desc_cache)
1440                 goto nomem;
1441
1442         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1443                                                   sizeof(struct kvm_mmu_page),
1444                                                   0, 0, NULL);
1445         if (!mmu_page_header_cache)
1446                 goto nomem;
1447
1448         return 0;
1449
1450 nomem:
1451         kvm_mmu_module_exit();
1452         return -ENOMEM;
1453 }
1454
1455 #ifdef AUDIT
1456
1457 static const char *audit_msg;
1458
1459 static gva_t canonicalize(gva_t gva)
1460 {
1461 #ifdef CONFIG_X86_64
1462         gva = (long long)(gva << 16) >> 16;
1463 #endif
1464         return gva;
1465 }
1466
1467 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1468                                 gva_t va, int level)
1469 {
1470         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1471         int i;
1472         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1473
1474         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1475                 u64 ent = pt[i];
1476
1477                 if (ent == shadow_trap_nonpresent_pte)
1478                         continue;
1479
1480                 va = canonicalize(va);
1481                 if (level > 1) {
1482                         if (ent == shadow_notrap_nonpresent_pte)
1483                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1484                                        " in nonleaf level: levels %d gva %lx"
1485                                        " level %d pte %llx\n", audit_msg,
1486                                        vcpu->mmu.root_level, va, level, ent);
1487
1488                         audit_mappings_page(vcpu, ent, va, level - 1);
1489                 } else {
1490                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1491                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1492
1493                         if (is_shadow_present_pte(ent)
1494                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1495                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1496                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1497                                        audit_msg, vcpu->mmu.root_level,
1498                                        va, gpa, hpa, ent,
1499                                        is_shadow_present_pte(ent));
1500                         else if (ent == shadow_notrap_nonpresent_pte
1501                                  && !is_error_hpa(hpa))
1502                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1503                                        " valid guest gva %lx\n", audit_msg, va);
1504
1505                 }
1506         }
1507 }
1508
1509 static void audit_mappings(struct kvm_vcpu *vcpu)
1510 {
1511         unsigned i;
1512
1513         if (vcpu->mmu.root_level == 4)
1514                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1515         else
1516                 for (i = 0; i < 4; ++i)
1517                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1518                                 audit_mappings_page(vcpu,
1519                                                     vcpu->mmu.pae_root[i],
1520                                                     i << 30,
1521                                                     2);
1522 }
1523
1524 static int count_rmaps(struct kvm_vcpu *vcpu)
1525 {
1526         int nmaps = 0;
1527         int i, j, k;
1528
1529         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1530                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1531                 struct kvm_rmap_desc *d;
1532
1533                 for (j = 0; j < m->npages; ++j) {
1534                         unsigned long *rmapp = &m->rmap[j];
1535
1536                         if (!*rmapp)
1537                                 continue;
1538                         if (!(*rmapp & 1)) {
1539                                 ++nmaps;
1540                                 continue;
1541                         }
1542                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1543                         while (d) {
1544                                 for (k = 0; k < RMAP_EXT; ++k)
1545                                         if (d->shadow_ptes[k])
1546                                                 ++nmaps;
1547                                         else
1548                                                 break;
1549                                 d = d->more;
1550                         }
1551                 }
1552         }
1553         return nmaps;
1554 }
1555
1556 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1557 {
1558         int nmaps = 0;
1559         struct kvm_mmu_page *page;
1560         int i;
1561
1562         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1563                 u64 *pt = page->spt;
1564
1565                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1566                         continue;
1567
1568                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1569                         u64 ent = pt[i];
1570
1571                         if (!(ent & PT_PRESENT_MASK))
1572                                 continue;
1573                         if (!(ent & PT_WRITABLE_MASK))
1574                                 continue;
1575                         ++nmaps;
1576                 }
1577         }
1578         return nmaps;
1579 }
1580
1581 static void audit_rmap(struct kvm_vcpu *vcpu)
1582 {
1583         int n_rmap = count_rmaps(vcpu);
1584         int n_actual = count_writable_mappings(vcpu);
1585
1586         if (n_rmap != n_actual)
1587                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1588                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1589 }
1590
1591 static void audit_write_protection(struct kvm_vcpu *vcpu)
1592 {
1593         struct kvm_mmu_page *page;
1594         struct kvm_memory_slot *slot;
1595         unsigned long *rmapp;
1596         gfn_t gfn;
1597
1598         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1599                 if (page->role.metaphysical)
1600                         continue;
1601
1602                 slot = gfn_to_memslot(vcpu->kvm, page->gfn);
1603                 gfn = unalias_gfn(vcpu->kvm, page->gfn);
1604                 rmapp = &slot->rmap[gfn - slot->base_gfn];
1605                 if (*rmapp)
1606                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1607                                " mappings: gfn %lx role %x\n",
1608                                __FUNCTION__, audit_msg, page->gfn,
1609                                page->role.word);
1610         }
1611 }
1612
1613 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1614 {
1615         int olddbg = dbg;
1616
1617         dbg = 0;
1618         audit_msg = msg;
1619         audit_rmap(vcpu);
1620         audit_write_protection(vcpu);
1621         audit_mappings(vcpu);
1622         dbg = olddbg;
1623 }
1624
1625 #endif