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

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  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, 8);
 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  * Take gfn and return the reverse mapping to it.
 331  * Note: gfn must be unaliased before this function get called
 332  */
 333
 334 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
 335 {
 336         struct kvm_memory_slot *slot;
 337
 338         slot = gfn_to_memslot(kvm, gfn);
 339         return &slot->rmap[gfn - slot->base_gfn];
 340 }
 341
 342 /*
 343  * Reverse mapping data structures:
 344  *
 345  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 346  * that points to page_address(page).
 347  *
 348  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 349  * containing more mappings.
 350  */
 351 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 352 {
 353         struct kvm_mmu_page *page;
 354         struct kvm_rmap_desc *desc;
 355         unsigned long *rmapp;
 356         int i;
 357
 358         if (!is_rmap_pte(*spte))
 359                 return;
 360         gfn = unalias_gfn(vcpu->kvm, gfn);
 361         page = page_header(__pa(spte));
 362         page->gfns[spte - page->spt] = gfn;
 363         rmapp = gfn_to_rmap(vcpu->kvm, gfn);
 364         if (!*rmapp) {
 365                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 366                 *rmapp = (unsigned long)spte;
 367         } else if (!(*rmapp & 1)) {
 368                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 369                 desc = mmu_alloc_rmap_desc(vcpu);
 370                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 371                 desc->shadow_ptes[1] = spte;
 372                 *rmapp = (unsigned long)desc | 1;
 373         } else {
 374                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 375                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 376                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 377                         desc = desc->more;
 378                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 379                         desc->more = mmu_alloc_rmap_desc(vcpu);
 380                         desc = desc->more;
 381                 }
 382                 for (i = 0; desc->shadow_ptes[i]; ++i)
 383                         ;
 384                 desc->shadow_ptes[i] = spte;
 385         }
 386 }
 387
 388 static void rmap_desc_remove_entry(unsigned long *rmapp,
 389                                    struct kvm_rmap_desc *desc,
 390                                    int i,
 391                                    struct kvm_rmap_desc *prev_desc)
 392 {
 393         int j;
 394
 395         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 396                 ;
 397         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 398         desc->shadow_ptes[j] = NULL;
 399         if (j != 0)
 400                 return;
 401         if (!prev_desc && !desc->more)
 402                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 403         else
 404                 if (prev_desc)
 405                         prev_desc->more = desc->more;
 406                 else
 407                         *rmapp = (unsigned long)desc->more | 1;
 408         mmu_free_rmap_desc(desc);
 409 }
 410
 411 static void rmap_remove(struct kvm *kvm, u64 *spte)
 412 {
 413         struct kvm_rmap_desc *desc;
 414         struct kvm_rmap_desc *prev_desc;
 415         struct kvm_mmu_page *page;
 416         unsigned long *rmapp;
 417         int i;
 418
 419         if (!is_rmap_pte(*spte))
 420                 return;
 421         page = page_header(__pa(spte));
 422         rmapp = gfn_to_rmap(kvm, page->gfns[spte - page->spt]);
 423         if (!*rmapp) {
 424                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 425                 BUG();
 426         } else if (!(*rmapp & 1)) {
 427                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 428                 if ((u64 *)*rmapp != spte) {
 429                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 430                                spte, *spte);
 431                         BUG();
 432                 }
 433                 *rmapp = 0;
 434         } else {
 435                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 436                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 437                 prev_desc = NULL;
 438                 while (desc) {
 439                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 440                                 if (desc->shadow_ptes[i] == spte) {
 441                                         rmap_desc_remove_entry(rmapp,
 442                                                                desc, i,
 443                                                                prev_desc);
 444                                         return;
 445                                 }
 446                         prev_desc = desc;
 447                         desc = desc->more;
 448                 }
 449                 BUG();
 450         }
 451 }
 452
 453 static void rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
 454 {
 455         struct kvm_rmap_desc *desc;
 456         unsigned long *rmapp;
 457         u64 *spte;
 458
 459         gfn = unalias_gfn(vcpu->kvm, gfn);
 460         rmapp = gfn_to_rmap(vcpu->kvm, gfn);
 461
 462         while (*rmapp) {
 463                 if (!(*rmapp & 1))
 464                         spte = (u64 *)*rmapp;
 465                 else {
 466                         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 467                         spte = desc->shadow_ptes[0];
 468                 }
 469                 BUG_ON(!spte);
 470                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 471                 BUG_ON(!(*spte & PT_WRITABLE_MASK));
 472                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 473                 rmap_remove(vcpu->kvm, spte);
 474                 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 475                 kvm_flush_remote_tlbs(vcpu->kvm);
 476         }
 477 }
 478
 479 #ifdef MMU_DEBUG
 480 static int is_empty_shadow_page(u64 *spt)
 481 {
 482         u64 *pos;
 483         u64 *end;
 484
 485         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 486                 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
 487                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 488                                pos, *pos);
 489                         return 0;
 490                 }
 491         return 1;
 492 }
 493 #endif
 494
 495 static void kvm_mmu_free_page(struct kvm *kvm,
 496                               struct kvm_mmu_page *page_head)
 497 {
 498         ASSERT(is_empty_shadow_page(page_head->spt));
 499         list_del(&page_head->link);
 500         __free_page(virt_to_page(page_head->spt));
 501         __free_page(virt_to_page(page_head->gfns));
 502         kfree(page_head);
 503         ++kvm->n_free_mmu_pages;
 504 }
 505
 506 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 507 {
 508         return gfn;
 509 }
 510
 511 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 512                                                u64 *parent_pte)
 513 {
 514         struct kvm_mmu_page *page;
 515
 516         if (!vcpu->kvm->n_free_mmu_pages)
 517                 return NULL;
 518
 519         page = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache,
 520                                       sizeof *page);
 521         page->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 522         page->gfns = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
 523         set_page_private(virt_to_page(page->spt), (unsigned long)page);
 524         list_add(&page->link, &vcpu->kvm->active_mmu_pages);
 525         ASSERT(is_empty_shadow_page(page->spt));
 526         page->slot_bitmap = 0;
 527         page->multimapped = 0;
 528         page->parent_pte = parent_pte;
 529         --vcpu->kvm->n_free_mmu_pages;
 530         return page;
 531 }
 532
 533 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 534                                     struct kvm_mmu_page *page, u64 *parent_pte)
 535 {
 536         struct kvm_pte_chain *pte_chain;
 537         struct hlist_node *node;
 538         int i;
 539
 540         if (!parent_pte)
 541                 return;
 542         if (!page->multimapped) {
 543                 u64 *old = page->parent_pte;
 544
 545                 if (!old) {
 546                         page->parent_pte = parent_pte;
 547                         return;
 548                 }
 549                 page->multimapped = 1;
 550                 pte_chain = mmu_alloc_pte_chain(vcpu);
 551                 INIT_HLIST_HEAD(&page->parent_ptes);
 552                 hlist_add_head(&pte_chain->link, &page->parent_ptes);
 553                 pte_chain->parent_ptes[0] = old;
 554         }
 555         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link) {
 556                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 557                         continue;
 558                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 559                         if (!pte_chain->parent_ptes[i]) {
 560                                 pte_chain->parent_ptes[i] = parent_pte;
 561                                 return;
 562                         }
 563         }
 564         pte_chain = mmu_alloc_pte_chain(vcpu);
 565         BUG_ON(!pte_chain);
 566         hlist_add_head(&pte_chain->link, &page->parent_ptes);
 567         pte_chain->parent_ptes[0] = parent_pte;
 568 }
 569
 570 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *page,
 571                                        u64 *parent_pte)
 572 {
 573         struct kvm_pte_chain *pte_chain;
 574         struct hlist_node *node;
 575         int i;
 576
 577         if (!page->multimapped) {
 578                 BUG_ON(page->parent_pte != parent_pte);
 579                 page->parent_pte = NULL;
 580                 return;
 581         }
 582         hlist_for_each_entry(pte_chain, node, &page->parent_ptes, link)
 583                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 584                         if (!pte_chain->parent_ptes[i])
 585                                 break;
 586                         if (pte_chain->parent_ptes[i] != parent_pte)
 587                                 continue;
 588                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 589                                 && pte_chain->parent_ptes[i + 1]) {
 590                                 pte_chain->parent_ptes[i]
 591                                         = pte_chain->parent_ptes[i + 1];
 592                                 ++i;
 593                         }
 594                         pte_chain->parent_ptes[i] = NULL;
 595                         if (i == 0) {
 596                                 hlist_del(&pte_chain->link);
 597                                 mmu_free_pte_chain(pte_chain);
 598                                 if (hlist_empty(&page->parent_ptes)) {
 599                                         page->multimapped = 0;
 600                                         page->parent_pte = NULL;
 601                                 }
 602                         }
 603                         return;
 604                 }
 605         BUG();
 606 }
 607
 608 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm_vcpu *vcpu,
 609                                                 gfn_t gfn)
 610 {
 611         unsigned index;
 612         struct hlist_head *bucket;
 613         struct kvm_mmu_page *page;
 614         struct hlist_node *node;
 615
 616         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 617         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 618         bucket = &vcpu->kvm->mmu_page_hash[index];
 619         hlist_for_each_entry(page, node, bucket, hash_link)
 620                 if (page->gfn == gfn && !page->role.metaphysical) {
 621                         pgprintk("%s: found role %x\n",
 622                                  __FUNCTION__, page->role.word);
 623                         return page;
 624                 }
 625         return NULL;
 626 }
 627
 628 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 629                                              gfn_t gfn,
 630                                              gva_t gaddr,
 631                                              unsigned level,
 632                                              int metaphysical,
 633                                              unsigned hugepage_access,
 634                                              u64 *parent_pte)
 635 {
 636         union kvm_mmu_page_role role;
 637         unsigned index;
 638         unsigned quadrant;
 639         struct hlist_head *bucket;
 640         struct kvm_mmu_page *page;
 641         struct hlist_node *node;
 642
 643         role.word = 0;
 644         role.glevels = vcpu->mmu.root_level;
 645         role.level = level;
 646         role.metaphysical = metaphysical;
 647         role.hugepage_access = hugepage_access;
 648         if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
 649                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 650                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 651                 role.quadrant = quadrant;
 652         }
 653         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 654                  gfn, role.word);
 655         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 656         bucket = &vcpu->kvm->mmu_page_hash[index];
 657         hlist_for_each_entry(page, node, bucket, hash_link)
 658                 if (page->gfn == gfn && page->role.word == role.word) {
 659                         mmu_page_add_parent_pte(vcpu, page, parent_pte);
 660                         pgprintk("%s: found\n", __FUNCTION__);
 661                         return page;
 662                 }
 663         page = kvm_mmu_alloc_page(vcpu, parent_pte);
 664         if (!page)
 665                 return page;
 666         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 667         page->gfn = gfn;
 668         page->role = role;
 669         hlist_add_head(&page->hash_link, bucket);
 670         vcpu->mmu.prefetch_page(vcpu, page);
 671         if (!metaphysical)
 672                 rmap_write_protect(vcpu, gfn);
 673         return page;
 674 }
 675
 676 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 677                                          struct kvm_mmu_page *page)
 678 {
 679         unsigned i;
 680         u64 *pt;
 681         u64 ent;
 682
 683         pt = page->spt;
 684
 685         if (page->role.level == PT_PAGE_TABLE_LEVEL) {
 686                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 687                         if (is_shadow_present_pte(pt[i]))
 688                                 rmap_remove(kvm, &pt[i]);
 689                         pt[i] = shadow_trap_nonpresent_pte;
 690                 }
 691                 kvm_flush_remote_tlbs(kvm);
 692                 return;
 693         }
 694
 695         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 696                 ent = pt[i];
 697
 698                 pt[i] = shadow_trap_nonpresent_pte;
 699                 if (!is_shadow_present_pte(ent))
 700                         continue;
 701                 ent &= PT64_BASE_ADDR_MASK;
 702                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 703         }
 704         kvm_flush_remote_tlbs(kvm);
 705 }
 706
 707 static void kvm_mmu_put_page(struct kvm_mmu_page *page,
 708                              u64 *parent_pte)
 709 {
 710         mmu_page_remove_parent_pte(page, parent_pte);
 711 }
 712
 713 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 714 {
 715         int i;
 716
 717         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 718                 if (kvm->vcpus[i])
 719                         kvm->vcpus[i]->last_pte_updated = NULL;
 720 }
 721
 722 static void kvm_mmu_zap_page(struct kvm *kvm,
 723                              struct kvm_mmu_page *page)
 724 {
 725         u64 *parent_pte;
 726
 727         while (page->multimapped || page->parent_pte) {
 728                 if (!page->multimapped)
 729                         parent_pte = page->parent_pte;
 730                 else {
 731                         struct kvm_pte_chain *chain;
 732
 733                         chain = container_of(page->parent_ptes.first,
 734                                              struct kvm_pte_chain, link);
 735                         parent_pte = chain->parent_ptes[0];
 736                 }
 737                 BUG_ON(!parent_pte);
 738                 kvm_mmu_put_page(page, parent_pte);
 739                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 740         }
 741         kvm_mmu_page_unlink_children(kvm, page);
 742         if (!page->root_count) {
 743                 hlist_del(&page->hash_link);
 744                 kvm_mmu_free_page(kvm, page);
 745         } else
 746                 list_move(&page->link, &kvm->active_mmu_pages);
 747         kvm_mmu_reset_last_pte_updated(kvm);
 748 }
 749
 750 /*
 751  * Changing the number of mmu pages allocated to the vm
 752  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 753  */
 754 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 755 {
 756         /*
 757          * If we set the number of mmu pages to be smaller be than the
 758          * number of actived pages , we must to free some mmu pages before we
 759          * change the value
 760          */
 761
 762         if ((kvm->n_alloc_mmu_pages - kvm->n_free_mmu_pages) >
 763             kvm_nr_mmu_pages) {
 764                 int n_used_mmu_pages = kvm->n_alloc_mmu_pages
 765                                        - kvm->n_free_mmu_pages;
 766
 767                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 768                         struct kvm_mmu_page *page;
 769
 770                         page = container_of(kvm->active_mmu_pages.prev,
 771                                             struct kvm_mmu_page, link);
 772                         kvm_mmu_zap_page(kvm, page);
 773                         n_used_mmu_pages--;
 774                 }
 775                 kvm->n_free_mmu_pages = 0;
 776         }
 777         else
 778                 kvm->n_free_mmu_pages += kvm_nr_mmu_pages
 779                                          - kvm->n_alloc_mmu_pages;
 780
 781         kvm->n_alloc_mmu_pages = kvm_nr_mmu_pages;
 782 }
 783
 784 static int kvm_mmu_unprotect_page(struct kvm_vcpu *vcpu, gfn_t gfn)
 785 {
 786         unsigned index;
 787         struct hlist_head *bucket;
 788         struct kvm_mmu_page *page;
 789         struct hlist_node *node, *n;
 790         int r;
 791
 792         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 793         r = 0;
 794         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 795         bucket = &vcpu->kvm->mmu_page_hash[index];
 796         hlist_for_each_entry_safe(page, node, n, bucket, hash_link)
 797                 if (page->gfn == gfn && !page->role.metaphysical) {
 798                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 799                                  page->role.word);
 800                         kvm_mmu_zap_page(vcpu->kvm, page);
 801                         r = 1;
 802                 }
 803         return r;
 804 }
 805
 806 static void mmu_unshadow(struct kvm_vcpu *vcpu, gfn_t gfn)
 807 {
 808         struct kvm_mmu_page *page;
 809
 810         while ((page = kvm_mmu_lookup_page(vcpu, gfn)) != NULL) {
 811                 pgprintk("%s: zap %lx %x\n",
 812                          __FUNCTION__, gfn, page->role.word);
 813                 kvm_mmu_zap_page(vcpu->kvm, page);
 814         }
 815 }
 816
 817 static void page_header_update_slot(struct kvm *kvm, void *pte, gpa_t gpa)
 818 {
 819         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gpa >> PAGE_SHIFT));
 820         struct kvm_mmu_page *page_head = page_header(__pa(pte));
 821
 822         __set_bit(slot, &page_head->slot_bitmap);
 823 }
 824
 825 hpa_t safe_gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 826 {
 827         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
 828
 829         return is_error_hpa(hpa) ? bad_page_address | (gpa & ~PAGE_MASK): hpa;
 830 }
 831
 832 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa)
 833 {
 834         struct page *page;
 835
 836         ASSERT((gpa & HPA_ERR_MASK) == 0);
 837         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 838         if (!page)
 839                 return gpa | HPA_ERR_MASK;
 840         return ((hpa_t)page_to_pfn(page) << PAGE_SHIFT)
 841                 | (gpa & (PAGE_SIZE-1));
 842 }
 843
 844 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva)
 845 {
 846         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 847
 848         if (gpa == UNMAPPED_GVA)
 849                 return UNMAPPED_GVA;
 850         return gpa_to_hpa(vcpu, gpa);
 851 }
 852
 853 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 854 {
 855         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
 856
 857         if (gpa == UNMAPPED_GVA)
 858                 return NULL;
 859         return pfn_to_page(gpa_to_hpa(vcpu, gpa) >> PAGE_SHIFT);
 860 }
 861
 862 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 863 {
 864 }
 865
 866 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, hpa_t p)
 867 {
 868         int level = PT32E_ROOT_LEVEL;
 869         hpa_t table_addr = vcpu->mmu.root_hpa;
 870
 871         for (; ; level--) {
 872                 u32 index = PT64_INDEX(v, level);
 873                 u64 *table;
 874                 u64 pte;
 875
 876                 ASSERT(VALID_PAGE(table_addr));
 877                 table = __va(table_addr);
 878
 879                 if (level == 1) {
 880                         pte = table[index];
 881                         if (is_shadow_present_pte(pte) && is_writeble_pte(pte))
 882                                 return 0;
 883                         mark_page_dirty(vcpu->kvm, v >> PAGE_SHIFT);
 884                         page_header_update_slot(vcpu->kvm, table, v);
 885                         table[index] = p | PT_PRESENT_MASK | PT_WRITABLE_MASK |
 886                                                                 PT_USER_MASK;
 887                         rmap_add(vcpu, &table[index], v >> PAGE_SHIFT);
 888                         return 0;
 889                 }
 890
 891                 if (table[index] == shadow_trap_nonpresent_pte) {
 892                         struct kvm_mmu_page *new_table;
 893                         gfn_t pseudo_gfn;
 894
 895                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
 896                                 >> PAGE_SHIFT;
 897                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
 898                                                      v, level - 1,
 899                                                      1, 0, &table[index]);
 900                         if (!new_table) {
 901                                 pgprintk("nonpaging_map: ENOMEM\n");
 902                                 return -ENOMEM;
 903                         }
 904
 905                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
 906                                 | PT_WRITABLE_MASK | PT_USER_MASK;
 907                 }
 908                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
 909         }
 910 }
 911
 912 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 913                                     struct kvm_mmu_page *sp)
 914 {
 915         int i;
 916
 917         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 918                 sp->spt[i] = shadow_trap_nonpresent_pte;
 919 }
 920
 921 static void mmu_free_roots(struct kvm_vcpu *vcpu)
 922 {
 923         int i;
 924         struct kvm_mmu_page *page;
 925
 926         if (!VALID_PAGE(vcpu->mmu.root_hpa))
 927                 return;
 928 #ifdef CONFIG_X86_64
 929         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 930                 hpa_t root = vcpu->mmu.root_hpa;
 931
 932                 page = page_header(root);
 933                 --page->root_count;
 934                 vcpu->mmu.root_hpa = INVALID_PAGE;
 935                 return;
 936         }
 937 #endif
 938         for (i = 0; i < 4; ++i) {
 939                 hpa_t root = vcpu->mmu.pae_root[i];
 940
 941                 if (root) {
 942                         root &= PT64_BASE_ADDR_MASK;
 943                         page = page_header(root);
 944                         --page->root_count;
 945                 }
 946                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
 947         }
 948         vcpu->mmu.root_hpa = INVALID_PAGE;
 949 }
 950
 951 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 952 {
 953         int i;
 954         gfn_t root_gfn;
 955         struct kvm_mmu_page *page;
 956
 957         root_gfn = vcpu->cr3 >> PAGE_SHIFT;
 958
 959 #ifdef CONFIG_X86_64
 960         if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
 961                 hpa_t root = vcpu->mmu.root_hpa;
 962
 963                 ASSERT(!VALID_PAGE(root));
 964                 page = kvm_mmu_get_page(vcpu, root_gfn, 0,
 965                                         PT64_ROOT_LEVEL, 0, 0, NULL);
 966                 root = __pa(page->spt);
 967                 ++page->root_count;
 968                 vcpu->mmu.root_hpa = root;
 969                 return;
 970         }
 971 #endif
 972         for (i = 0; i < 4; ++i) {
 973                 hpa_t root = vcpu->mmu.pae_root[i];
 974
 975                 ASSERT(!VALID_PAGE(root));
 976                 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
 977                         if (!is_present_pte(vcpu->pdptrs[i])) {
 978                                 vcpu->mmu.pae_root[i] = 0;
 979                                 continue;
 980                         }
 981                         root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
 982                 } else if (vcpu->mmu.root_level == 0)
 983                         root_gfn = 0;
 984                 page = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 985                                         PT32_ROOT_LEVEL, !is_paging(vcpu),
 986                                         0, NULL);
 987                 root = __pa(page->spt);
 988                 ++page->root_count;
 989                 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
 990         }
 991         vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
 992 }
 993
 994 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
 995 {
 996         return vaddr;
 997 }
 998
 999 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1000                                u32 error_code)
1001 {
1002         gpa_t addr = gva;
1003         hpa_t paddr;
1004         int r;
1005
1006         r = mmu_topup_memory_caches(vcpu);
1007         if (r)
1008                 return r;
1009
1010         ASSERT(vcpu);
1011         ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
1012
1013
1014         paddr = gpa_to_hpa(vcpu , addr & PT64_BASE_ADDR_MASK);
1015
1016         if (is_error_hpa(paddr))
1017                 return 1;
1018
1019         return nonpaging_map(vcpu, addr & PAGE_MASK, paddr);
1020 }
1021
1022 static void nonpaging_free(struct kvm_vcpu *vcpu)
1023 {
1024         mmu_free_roots(vcpu);
1025 }
1026
1027 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1028 {
1029         struct kvm_mmu *context = &vcpu->mmu;
1030
1031         context->new_cr3 = nonpaging_new_cr3;
1032         context->page_fault = nonpaging_page_fault;
1033         context->gva_to_gpa = nonpaging_gva_to_gpa;
1034         context->free = nonpaging_free;
1035         context->prefetch_page = nonpaging_prefetch_page;
1036         context->root_level = 0;
1037         context->shadow_root_level = PT32E_ROOT_LEVEL;
1038         context->root_hpa = INVALID_PAGE;
1039         return 0;
1040 }
1041
1042 static void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1043 {
1044         ++vcpu->stat.tlb_flush;
1045         kvm_x86_ops->tlb_flush(vcpu);
1046 }
1047
1048 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1049 {
1050         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
1051         mmu_free_roots(vcpu);
1052 }
1053
1054 static void inject_page_fault(struct kvm_vcpu *vcpu,
1055                               u64 addr,
1056                               u32 err_code)
1057 {
1058         kvm_x86_ops->inject_page_fault(vcpu, addr, err_code);
1059 }
1060
1061 static void paging_free(struct kvm_vcpu *vcpu)
1062 {
1063         nonpaging_free(vcpu);
1064 }
1065
1066 #define PTTYPE 64
1067 #include "paging_tmpl.h"
1068 #undef PTTYPE
1069
1070 #define PTTYPE 32
1071 #include "paging_tmpl.h"
1072 #undef PTTYPE
1073
1074 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1075 {
1076         struct kvm_mmu *context = &vcpu->mmu;
1077
1078         ASSERT(is_pae(vcpu));
1079         context->new_cr3 = paging_new_cr3;
1080         context->page_fault = paging64_page_fault;
1081         context->gva_to_gpa = paging64_gva_to_gpa;
1082         context->prefetch_page = paging64_prefetch_page;
1083         context->free = paging_free;
1084         context->root_level = level;
1085         context->shadow_root_level = level;
1086         context->root_hpa = INVALID_PAGE;
1087         return 0;
1088 }
1089
1090 static int paging64_init_context(struct kvm_vcpu *vcpu)
1091 {
1092         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1093 }
1094
1095 static int paging32_init_context(struct kvm_vcpu *vcpu)
1096 {
1097         struct kvm_mmu *context = &vcpu->mmu;
1098
1099         context->new_cr3 = paging_new_cr3;
1100         context->page_fault = paging32_page_fault;
1101         context->gva_to_gpa = paging32_gva_to_gpa;
1102         context->free = paging_free;
1103         context->prefetch_page = paging32_prefetch_page;
1104         context->root_level = PT32_ROOT_LEVEL;
1105         context->shadow_root_level = PT32E_ROOT_LEVEL;
1106         context->root_hpa = INVALID_PAGE;
1107         return 0;
1108 }
1109
1110 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1111 {
1112         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1113 }
1114
1115 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1116 {
1117         ASSERT(vcpu);
1118         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1119
1120         if (!is_paging(vcpu))
1121                 return nonpaging_init_context(vcpu);
1122         else if (is_long_mode(vcpu))
1123                 return paging64_init_context(vcpu);
1124         else if (is_pae(vcpu))
1125                 return paging32E_init_context(vcpu);
1126         else
1127                 return paging32_init_context(vcpu);
1128 }
1129
1130 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1131 {
1132         ASSERT(vcpu);
1133         if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1134                 vcpu->mmu.free(vcpu);
1135                 vcpu->mmu.root_hpa = INVALID_PAGE;
1136         }
1137 }
1138
1139 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1140 {
1141         destroy_kvm_mmu(vcpu);
1142         return init_kvm_mmu(vcpu);
1143 }
1144 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1145
1146 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1147 {
1148         int r;
1149
1150         mutex_lock(&vcpu->kvm->lock);
1151         r = mmu_topup_memory_caches(vcpu);
1152         if (r)
1153                 goto out;
1154         mmu_alloc_roots(vcpu);
1155         kvm_x86_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1156         kvm_mmu_flush_tlb(vcpu);
1157 out:
1158         mutex_unlock(&vcpu->kvm->lock);
1159         return r;
1160 }
1161 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1162
1163 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1164 {
1165         mmu_free_roots(vcpu);
1166 }
1167
1168 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1169                                   struct kvm_mmu_page *page,
1170                                   u64 *spte)
1171 {
1172         u64 pte;
1173         struct kvm_mmu_page *child;
1174
1175         pte = *spte;
1176         if (is_shadow_present_pte(pte)) {
1177                 if (page->role.level == PT_PAGE_TABLE_LEVEL)
1178                         rmap_remove(vcpu->kvm, spte);
1179                 else {
1180                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1181                         mmu_page_remove_parent_pte(child, spte);
1182                 }
1183         }
1184         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1185         kvm_flush_remote_tlbs(vcpu->kvm);
1186 }
1187
1188 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1189                                   struct kvm_mmu_page *page,
1190                                   u64 *spte,
1191                                   const void *new, int bytes,
1192                                   int offset_in_pte)
1193 {
1194         if (page->role.level != PT_PAGE_TABLE_LEVEL)
1195                 return;
1196
1197         if (page->role.glevels == PT32_ROOT_LEVEL)
1198                 paging32_update_pte(vcpu, page, spte, new, bytes,
1199                                     offset_in_pte);
1200         else
1201                 paging64_update_pte(vcpu, page, spte, new, bytes,
1202                                     offset_in_pte);
1203 }
1204
1205 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1206 {
1207         u64 *spte = vcpu->last_pte_updated;
1208
1209         return !!(spte && (*spte & PT_ACCESSED_MASK));
1210 }
1211
1212 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1213                        const u8 *new, int bytes)
1214 {
1215         gfn_t gfn = gpa >> PAGE_SHIFT;
1216         struct kvm_mmu_page *page;
1217         struct hlist_node *node, *n;
1218         struct hlist_head *bucket;
1219         unsigned index;
1220         u64 *spte;
1221         unsigned offset = offset_in_page(gpa);
1222         unsigned pte_size;
1223         unsigned page_offset;
1224         unsigned misaligned;
1225         unsigned quadrant;
1226         int level;
1227         int flooded = 0;
1228         int npte;
1229
1230         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1231         kvm_mmu_audit(vcpu, "pre pte write");
1232         if (gfn == vcpu->last_pt_write_gfn
1233             && !last_updated_pte_accessed(vcpu)) {
1234                 ++vcpu->last_pt_write_count;
1235                 if (vcpu->last_pt_write_count >= 3)
1236                         flooded = 1;
1237         } else {
1238                 vcpu->last_pt_write_gfn = gfn;
1239                 vcpu->last_pt_write_count = 1;
1240                 vcpu->last_pte_updated = NULL;
1241         }
1242         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1243         bucket = &vcpu->kvm->mmu_page_hash[index];
1244         hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
1245                 if (page->gfn != gfn || page->role.metaphysical)
1246                         continue;
1247                 pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1248                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1249                 misaligned |= bytes < 4;
1250                 if (misaligned || flooded) {
1251                         /*
1252                          * Misaligned accesses are too much trouble to fix
1253                          * up; also, they usually indicate a page is not used
1254                          * as a page table.
1255                          *
1256                          * If we're seeing too many writes to a page,
1257                          * it may no longer be a page table, or we may be
1258                          * forking, in which case it is better to unmap the
1259                          * page.
1260                          */
1261                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1262                                  gpa, bytes, page->role.word);
1263                         kvm_mmu_zap_page(vcpu->kvm, page);
1264                         continue;
1265                 }
1266                 page_offset = offset;
1267                 level = page->role.level;
1268                 npte = 1;
1269                 if (page->role.glevels == PT32_ROOT_LEVEL) {
1270                         page_offset <<= 1;      /* 32->64 */
1271                         /*
1272                          * A 32-bit pde maps 4MB while the shadow pdes map
1273                          * only 2MB.  So we need to double the offset again
1274                          * and zap two pdes instead of one.
1275                          */
1276                         if (level == PT32_ROOT_LEVEL) {
1277                                 page_offset &= ~7; /* kill rounding error */
1278                                 page_offset <<= 1;
1279                                 npte = 2;
1280                         }
1281                         quadrant = page_offset >> PAGE_SHIFT;
1282                         page_offset &= ~PAGE_MASK;
1283                         if (quadrant != page->role.quadrant)
1284                                 continue;
1285                 }
1286                 spte = &page->spt[page_offset / sizeof(*spte)];
1287                 while (npte--) {
1288                         mmu_pte_write_zap_pte(vcpu, page, spte);
1289                         mmu_pte_write_new_pte(vcpu, page, spte, new, bytes,
1290                                               page_offset & (pte_size - 1));
1291                         ++spte;
1292                 }
1293         }
1294         kvm_mmu_audit(vcpu, "post pte write");
1295 }
1296
1297 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1298 {
1299         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1300
1301         return kvm_mmu_unprotect_page(vcpu, gpa >> PAGE_SHIFT);
1302 }
1303
1304 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1305 {
1306         while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1307                 struct kvm_mmu_page *page;
1308
1309                 page = container_of(vcpu->kvm->active_mmu_pages.prev,
1310                                     struct kvm_mmu_page, link);
1311                 kvm_mmu_zap_page(vcpu->kvm, page);
1312         }
1313 }
1314
1315 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1316 {
1317         struct kvm_mmu_page *page;
1318
1319         while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1320                 page = container_of(vcpu->kvm->active_mmu_pages.next,
1321                                     struct kvm_mmu_page, link);
1322                 kvm_mmu_zap_page(vcpu->kvm, page);
1323         }
1324         free_page((unsigned long)vcpu->mmu.pae_root);
1325 }
1326
1327 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1328 {
1329         struct page *page;
1330         int i;
1331
1332         ASSERT(vcpu);
1333
1334         if (vcpu->kvm->n_requested_mmu_pages)
1335                 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_requested_mmu_pages;
1336         else
1337                 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_alloc_mmu_pages;
1338         /*
1339          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1340          * Therefore we need to allocate shadow page tables in the first
1341          * 4GB of memory, which happens to fit the DMA32 zone.
1342          */
1343         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1344         if (!page)
1345                 goto error_1;
1346         vcpu->mmu.pae_root = page_address(page);
1347         for (i = 0; i < 4; ++i)
1348                 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1349
1350         return 0;
1351
1352 error_1:
1353         free_mmu_pages(vcpu);
1354         return -ENOMEM;
1355 }
1356
1357 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1358 {
1359         ASSERT(vcpu);
1360         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1361
1362         return alloc_mmu_pages(vcpu);
1363 }
1364
1365 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1366 {
1367         ASSERT(vcpu);
1368         ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1369
1370         return init_kvm_mmu(vcpu);
1371 }
1372
1373 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1374 {
1375         ASSERT(vcpu);
1376
1377         destroy_kvm_mmu(vcpu);
1378         free_mmu_pages(vcpu);
1379         mmu_free_memory_caches(vcpu);
1380 }
1381
1382 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1383 {
1384         struct kvm_mmu_page *page;
1385
1386         list_for_each_entry(page, &kvm->active_mmu_pages, link) {
1387                 int i;
1388                 u64 *pt;
1389
1390                 if (!test_bit(slot, &page->slot_bitmap))
1391                         continue;
1392
1393                 pt = page->spt;
1394                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1395                         /* avoid RMW */
1396                         if (pt[i] & PT_WRITABLE_MASK) {
1397                                 rmap_remove(kvm, &pt[i]);
1398                                 pt[i] &= ~PT_WRITABLE_MASK;
1399                         }
1400         }
1401 }
1402
1403 void kvm_mmu_zap_all(struct kvm *kvm)
1404 {
1405         struct kvm_mmu_page *page, *node;
1406
1407         list_for_each_entry_safe(page, node, &kvm->active_mmu_pages, link)
1408                 kvm_mmu_zap_page(kvm, page);
1409
1410         kvm_flush_remote_tlbs(kvm);
1411 }
1412
1413 void kvm_mmu_module_exit(void)
1414 {
1415         if (pte_chain_cache)
1416                 kmem_cache_destroy(pte_chain_cache);
1417         if (rmap_desc_cache)
1418                 kmem_cache_destroy(rmap_desc_cache);
1419         if (mmu_page_header_cache)
1420                 kmem_cache_destroy(mmu_page_header_cache);
1421 }
1422
1423 int kvm_mmu_module_init(void)
1424 {
1425         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1426                                             sizeof(struct kvm_pte_chain),
1427                                             0, 0, NULL);
1428         if (!pte_chain_cache)
1429                 goto nomem;
1430         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1431                                             sizeof(struct kvm_rmap_desc),
1432                                             0, 0, NULL);
1433         if (!rmap_desc_cache)
1434                 goto nomem;
1435
1436         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1437                                                   sizeof(struct kvm_mmu_page),
1438                                                   0, 0, NULL);
1439         if (!mmu_page_header_cache)
1440                 goto nomem;
1441
1442         return 0;
1443
1444 nomem:
1445         kvm_mmu_module_exit();
1446         return -ENOMEM;
1447 }
1448
1449 #ifdef AUDIT
1450
1451 static const char *audit_msg;
1452
1453 static gva_t canonicalize(gva_t gva)
1454 {
1455 #ifdef CONFIG_X86_64
1456         gva = (long long)(gva << 16) >> 16;
1457 #endif
1458         return gva;
1459 }
1460
1461 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1462                                 gva_t va, int level)
1463 {
1464         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1465         int i;
1466         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1467
1468         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1469                 u64 ent = pt[i];
1470
1471                 if (ent == shadow_trap_nonpresent_pte)
1472                         continue;
1473
1474                 va = canonicalize(va);
1475                 if (level > 1) {
1476                         if (ent == shadow_notrap_nonpresent_pte)
1477                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1478                                        " in nonleaf level: levels %d gva %lx"
1479                                        " level %d pte %llx\n", audit_msg,
1480                                        vcpu->mmu.root_level, va, level, ent);
1481
1482                         audit_mappings_page(vcpu, ent, va, level - 1);
1483                 } else {
1484                         gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1485                         hpa_t hpa = gpa_to_hpa(vcpu, gpa);
1486
1487                         if (is_shadow_present_pte(ent)
1488                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1489                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1490                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1491                                        audit_msg, vcpu->mmu.root_level,
1492                                        va, gpa, hpa, ent, is_shadow_present_pte(ent));
1493                         else if (ent == shadow_notrap_nonpresent_pte
1494                                  && !is_error_hpa(hpa))
1495                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1496                                        " valid guest gva %lx\n", audit_msg, va);
1497
1498                 }
1499         }
1500 }
1501
1502 static void audit_mappings(struct kvm_vcpu *vcpu)
1503 {
1504         unsigned i;
1505
1506         if (vcpu->mmu.root_level == 4)
1507                 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1508         else
1509                 for (i = 0; i < 4; ++i)
1510                         if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1511                                 audit_mappings_page(vcpu,
1512                                                     vcpu->mmu.pae_root[i],
1513                                                     i << 30,
1514                                                     2);
1515 }
1516
1517 static int count_rmaps(struct kvm_vcpu *vcpu)
1518 {
1519         int nmaps = 0;
1520         int i, j, k;
1521
1522         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1523                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1524                 struct kvm_rmap_desc *d;
1525
1526                 for (j = 0; j < m->npages; ++j) {
1527                         unsigned long *rmapp = &m->rmap[j];
1528
1529                         if (!*rmapp)
1530                                 continue;
1531                         if (!(*rmapp & 1)) {
1532                                 ++nmaps;
1533                                 continue;
1534                         }
1535                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1536                         while (d) {
1537                                 for (k = 0; k < RMAP_EXT; ++k)
1538                                         if (d->shadow_ptes[k])
1539                                                 ++nmaps;
1540                                         else
1541                                                 break;
1542                                 d = d->more;
1543                         }
1544                 }
1545         }
1546         return nmaps;
1547 }
1548
1549 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1550 {
1551         int nmaps = 0;
1552         struct kvm_mmu_page *page;
1553         int i;
1554
1555         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1556                 u64 *pt = page->spt;
1557
1558                 if (page->role.level != PT_PAGE_TABLE_LEVEL)
1559                         continue;
1560
1561                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1562                         u64 ent = pt[i];
1563
1564                         if (!(ent & PT_PRESENT_MASK))
1565                                 continue;
1566                         if (!(ent & PT_WRITABLE_MASK))
1567                                 continue;
1568                         ++nmaps;
1569                 }
1570         }
1571         return nmaps;
1572 }
1573
1574 static void audit_rmap(struct kvm_vcpu *vcpu)
1575 {
1576         int n_rmap = count_rmaps(vcpu);
1577         int n_actual = count_writable_mappings(vcpu);
1578
1579         if (n_rmap != n_actual)
1580                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1581                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1582 }
1583
1584 static void audit_write_protection(struct kvm_vcpu *vcpu)
1585 {
1586         struct kvm_mmu_page *page;
1587         struct kvm_memory_slot *slot;
1588         unsigned long *rmapp;
1589         gfn_t gfn;
1590
1591         list_for_each_entry(page, &vcpu->kvm->active_mmu_pages, link) {
1592                 if (page->role.metaphysical)
1593                         continue;
1594
1595                 slot = gfn_to_memslot(vcpu->kvm, page->gfn);
1596                 gfn = unalias_gfn(vcpu->kvm, page->gfn);
1597                 rmapp = &slot->rmap[gfn - slot->base_gfn];
1598                 if (*rmapp)
1599                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1600                                " mappings: gfn %lx role %x\n",
1601                                __FUNCTION__, audit_msg, page->gfn,
1602                                page->role.word);
1603         }
1604 }
1605
1606 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1607 {
1608         int olddbg = dbg;
1609
1610         dbg = 0;
1611         audit_msg = msg;
1612         audit_rmap(vcpu);
1613         audit_write_protection(vcpu);
1614         audit_mappings(vcpu);
1615         dbg = olddbg;
1616 }
1617
1618 #endif