err.no Git - linux-2.6/blob - arch/x86/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 "mmu.h"
  22
  23 #include <linux/kvm_host.h>
  24 #include <linux/types.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/highmem.h>
  28 #include <linux/module.h>
  29 #include <linux/swap.h>
  30
  31 #include <asm/page.h>
  32 #include <asm/cmpxchg.h>
  33 #include <asm/io.h>
  34
  35 #undef MMU_DEBUG
  36
  37 #undef AUDIT
  38
  39 #ifdef AUDIT
  40 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  41 #else
  42 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  43 #endif
  44
  45 #ifdef MMU_DEBUG
  46
  47 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  48 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  49
  50 #else
  51
  52 #define pgprintk(x...) do { } while (0)
  53 #define rmap_printk(x...) do { } while (0)
  54
  55 #endif
  56
  57 #if defined(MMU_DEBUG) || defined(AUDIT)
  58 static int dbg = 1;
  59 #endif
  60
  61 #ifndef MMU_DEBUG
  62 #define ASSERT(x) do { } while (0)
  63 #else
  64 #define ASSERT(x)                                                       \
  65         if (!(x)) {                                                     \
  66                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  67                        __FILE__, __LINE__, #x);                         \
  68         }
  69 #endif
  70
  71 #define PT64_PT_BITS 9
  72 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  73 #define PT32_PT_BITS 10
  74 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  75
  76 #define PT_WRITABLE_SHIFT 1
  77
  78 #define PT_PRESENT_MASK (1ULL << 0)
  79 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  80 #define PT_USER_MASK (1ULL << 2)
  81 #define PT_PWT_MASK (1ULL << 3)
  82 #define PT_PCD_MASK (1ULL << 4)
  83 #define PT_ACCESSED_MASK (1ULL << 5)
  84 #define PT_DIRTY_MASK (1ULL << 6)
  85 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  86 #define PT_PAT_MASK (1ULL << 7)
  87 #define PT_GLOBAL_MASK (1ULL << 8)
  88 #define PT64_NX_SHIFT 63
  89 #define PT64_NX_MASK (1ULL << PT64_NX_SHIFT)
  90
  91 #define PT_PAT_SHIFT 7
  92 #define PT_DIR_PAT_SHIFT 12
  93 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  94
  95 #define PT32_DIR_PSE36_SIZE 4
  96 #define PT32_DIR_PSE36_SHIFT 13
  97 #define PT32_DIR_PSE36_MASK \
  98         (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  99
 100
 101 #define PT_FIRST_AVAIL_BITS_SHIFT 9
 102 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
 103
 104 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 105
 106 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 107
 108 #define PT64_LEVEL_BITS 9
 109
 110 #define PT64_LEVEL_SHIFT(level) \
 111                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
 112
 113 #define PT64_LEVEL_MASK(level) \
 114                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 115
 116 #define PT64_INDEX(address, level)\
 117         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 118
 119
 120 #define PT32_LEVEL_BITS 10
 121
 122 #define PT32_LEVEL_SHIFT(level) \
 123                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 124
 125 #define PT32_LEVEL_MASK(level) \
 126                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 127
 128 #define PT32_INDEX(address, level)\
 129         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 130
 131
 132 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 133 #define PT64_DIR_BASE_ADDR_MASK \
 134         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 135
 136 #define PT32_BASE_ADDR_MASK PAGE_MASK
 137 #define PT32_DIR_BASE_ADDR_MASK \
 138         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 139
 140 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 141                         | PT64_NX_MASK)
 142
 143 #define PFERR_PRESENT_MASK (1U << 0)
 144 #define PFERR_WRITE_MASK (1U << 1)
 145 #define PFERR_USER_MASK (1U << 2)
 146 #define PFERR_FETCH_MASK (1U << 4)
 147
 148 #define PT64_ROOT_LEVEL 4
 149 #define PT32_ROOT_LEVEL 2
 150 #define PT32E_ROOT_LEVEL 3
 151
 152 #define PT_DIRECTORY_LEVEL 2
 153 #define PT_PAGE_TABLE_LEVEL 1
 154
 155 #define RMAP_EXT 4
 156
 157 #define ACC_EXEC_MASK    1
 158 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 159 #define ACC_USER_MASK    PT_USER_MASK
 160 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 161
 162 struct kvm_rmap_desc {
 163         u64 *shadow_ptes[RMAP_EXT];
 164         struct kvm_rmap_desc *more;
 165 };
 166
 167 static struct kmem_cache *pte_chain_cache;
 168 static struct kmem_cache *rmap_desc_cache;
 169 static struct kmem_cache *mmu_page_header_cache;
 170
 171 static u64 __read_mostly shadow_trap_nonpresent_pte;
 172 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 173
 174 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 175 {
 176         shadow_trap_nonpresent_pte = trap_pte;
 177         shadow_notrap_nonpresent_pte = notrap_pte;
 178 }
 179 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 180
 181 static int is_write_protection(struct kvm_vcpu *vcpu)
 182 {
 183         return vcpu->arch.cr0 & X86_CR0_WP;
 184 }
 185
 186 static int is_cpuid_PSE36(void)
 187 {
 188         return 1;
 189 }
 190
 191 static int is_nx(struct kvm_vcpu *vcpu)
 192 {
 193         return vcpu->arch.shadow_efer & EFER_NX;
 194 }
 195
 196 static int is_present_pte(unsigned long pte)
 197 {
 198         return pte & PT_PRESENT_MASK;
 199 }
 200
 201 static int is_shadow_present_pte(u64 pte)
 202 {
 203         pte &= ~PT_SHADOW_IO_MARK;
 204         return pte != shadow_trap_nonpresent_pte
 205                 && pte != shadow_notrap_nonpresent_pte;
 206 }
 207
 208 static int is_writeble_pte(unsigned long pte)
 209 {
 210         return pte & PT_WRITABLE_MASK;
 211 }
 212
 213 static int is_dirty_pte(unsigned long pte)
 214 {
 215         return pte & PT_DIRTY_MASK;
 216 }
 217
 218 static int is_io_pte(unsigned long pte)
 219 {
 220         return pte & PT_SHADOW_IO_MARK;
 221 }
 222
 223 static int is_rmap_pte(u64 pte)
 224 {
 225         return is_shadow_present_pte(pte);
 226 }
 227
 228 static gfn_t pse36_gfn_delta(u32 gpte)
 229 {
 230         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 231
 232         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 233 }
 234
 235 static void set_shadow_pte(u64 *sptep, u64 spte)
 236 {
 237 #ifdef CONFIG_X86_64
 238         set_64bit((unsigned long *)sptep, spte);
 239 #else
 240         set_64bit((unsigned long long *)sptep, spte);
 241 #endif
 242 }
 243
 244 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 245                                   struct kmem_cache *base_cache, int min)
 246 {
 247         void *obj;
 248
 249         if (cache->nobjs >= min)
 250                 return 0;
 251         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 252                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 253                 if (!obj)
 254                         return -ENOMEM;
 255                 cache->objects[cache->nobjs++] = obj;
 256         }
 257         return 0;
 258 }
 259
 260 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 261 {
 262         while (mc->nobjs)
 263                 kfree(mc->objects[--mc->nobjs]);
 264 }
 265
 266 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 267                                        int min)
 268 {
 269         struct page *page;
 270
 271         if (cache->nobjs >= min)
 272                 return 0;
 273         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 274                 page = alloc_page(GFP_KERNEL);
 275                 if (!page)
 276                         return -ENOMEM;
 277                 set_page_private(page, 0);
 278                 cache->objects[cache->nobjs++] = page_address(page);
 279         }
 280         return 0;
 281 }
 282
 283 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 284 {
 285         while (mc->nobjs)
 286                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 287 }
 288
 289 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 290 {
 291         int r;
 292
 293         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 294                                    pte_chain_cache, 4);
 295         if (r)
 296                 goto out;
 297         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 298                                    rmap_desc_cache, 1);
 299         if (r)
 300                 goto out;
 301         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 302         if (r)
 303                 goto out;
 304         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 305                                    mmu_page_header_cache, 4);
 306 out:
 307         return r;
 308 }
 309
 310 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 311 {
 312         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 313         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 314         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 315         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 316 }
 317
 318 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 319                                     size_t size)
 320 {
 321         void *p;
 322
 323         BUG_ON(!mc->nobjs);
 324         p = mc->objects[--mc->nobjs];
 325         memset(p, 0, size);
 326         return p;
 327 }
 328
 329 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 330 {
 331         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 332                                       sizeof(struct kvm_pte_chain));
 333 }
 334
 335 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 336 {
 337         kfree(pc);
 338 }
 339
 340 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 341 {
 342         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 343                                       sizeof(struct kvm_rmap_desc));
 344 }
 345
 346 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 347 {
 348         kfree(rd);
 349 }
 350
 351 /*
 352  * Take gfn and return the reverse mapping to it.
 353  * Note: gfn must be unaliased before this function get called
 354  */
 355
 356 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
 357 {
 358         struct kvm_memory_slot *slot;
 359
 360         slot = gfn_to_memslot(kvm, gfn);
 361         return &slot->rmap[gfn - slot->base_gfn];
 362 }
 363
 364 /*
 365  * Reverse mapping data structures:
 366  *
 367  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 368  * that points to page_address(page).
 369  *
 370  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 371  * containing more mappings.
 372  */
 373 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 374 {
 375         struct kvm_mmu_page *sp;
 376         struct kvm_rmap_desc *desc;
 377         unsigned long *rmapp;
 378         int i;
 379
 380         if (!is_rmap_pte(*spte))
 381                 return;
 382         gfn = unalias_gfn(vcpu->kvm, gfn);
 383         sp = page_header(__pa(spte));
 384         sp->gfns[spte - sp->spt] = gfn;
 385         rmapp = gfn_to_rmap(vcpu->kvm, gfn);
 386         if (!*rmapp) {
 387                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 388                 *rmapp = (unsigned long)spte;
 389         } else if (!(*rmapp & 1)) {
 390                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 391                 desc = mmu_alloc_rmap_desc(vcpu);
 392                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 393                 desc->shadow_ptes[1] = spte;
 394                 *rmapp = (unsigned long)desc | 1;
 395         } else {
 396                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 397                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 398                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 399                         desc = desc->more;
 400                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 401                         desc->more = mmu_alloc_rmap_desc(vcpu);
 402                         desc = desc->more;
 403                 }
 404                 for (i = 0; desc->shadow_ptes[i]; ++i)
 405                         ;
 406                 desc->shadow_ptes[i] = spte;
 407         }
 408 }
 409
 410 static void rmap_desc_remove_entry(unsigned long *rmapp,
 411                                    struct kvm_rmap_desc *desc,
 412                                    int i,
 413                                    struct kvm_rmap_desc *prev_desc)
 414 {
 415         int j;
 416
 417         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 418                 ;
 419         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 420         desc->shadow_ptes[j] = NULL;
 421         if (j != 0)
 422                 return;
 423         if (!prev_desc && !desc->more)
 424                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 425         else
 426                 if (prev_desc)
 427                         prev_desc->more = desc->more;
 428                 else
 429                         *rmapp = (unsigned long)desc->more | 1;
 430         mmu_free_rmap_desc(desc);
 431 }
 432
 433 static void rmap_remove(struct kvm *kvm, u64 *spte)
 434 {
 435         struct kvm_rmap_desc *desc;
 436         struct kvm_rmap_desc *prev_desc;
 437         struct kvm_mmu_page *sp;
 438         struct page *page;
 439         unsigned long *rmapp;
 440         int i;
 441
 442         if (!is_rmap_pte(*spte))
 443                 return;
 444         sp = page_header(__pa(spte));
 445         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 446         mark_page_accessed(page);
 447         if (is_writeble_pte(*spte))
 448                 kvm_release_page_dirty(page);
 449         else
 450                 kvm_release_page_clean(page);
 451         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt]);
 452         if (!*rmapp) {
 453                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 454                 BUG();
 455         } else if (!(*rmapp & 1)) {
 456                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 457                 if ((u64 *)*rmapp != spte) {
 458                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 459                                spte, *spte);
 460                         BUG();
 461                 }
 462                 *rmapp = 0;
 463         } else {
 464                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 465                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 466                 prev_desc = NULL;
 467                 while (desc) {
 468                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 469                                 if (desc->shadow_ptes[i] == spte) {
 470                                         rmap_desc_remove_entry(rmapp,
 471                                                                desc, i,
 472                                                                prev_desc);
 473                                         return;
 474                                 }
 475                         prev_desc = desc;
 476                         desc = desc->more;
 477                 }
 478                 BUG();
 479         }
 480 }
 481
 482 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 483 {
 484         struct kvm_rmap_desc *desc;
 485         struct kvm_rmap_desc *prev_desc;
 486         u64 *prev_spte;
 487         int i;
 488
 489         if (!*rmapp)
 490                 return NULL;
 491         else if (!(*rmapp & 1)) {
 492                 if (!spte)
 493                         return (u64 *)*rmapp;
 494                 return NULL;
 495         }
 496         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 497         prev_desc = NULL;
 498         prev_spte = NULL;
 499         while (desc) {
 500                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 501                         if (prev_spte == spte)
 502                                 return desc->shadow_ptes[i];
 503                         prev_spte = desc->shadow_ptes[i];
 504                 }
 505                 desc = desc->more;
 506         }
 507         return NULL;
 508 }
 509
 510 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 511 {
 512         unsigned long *rmapp;
 513         u64 *spte;
 514         int write_protected = 0;
 515
 516         gfn = unalias_gfn(kvm, gfn);
 517         rmapp = gfn_to_rmap(kvm, gfn);
 518
 519         spte = rmap_next(kvm, rmapp, NULL);
 520         while (spte) {
 521                 BUG_ON(!spte);
 522                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 523                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 524                 if (is_writeble_pte(*spte)) {
 525                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 526                         write_protected = 1;
 527                 }
 528                 spte = rmap_next(kvm, rmapp, spte);
 529         }
 530         if (write_protected)
 531                 kvm_flush_remote_tlbs(kvm);
 532 }
 533
 534 #ifdef MMU_DEBUG
 535 static int is_empty_shadow_page(u64 *spt)
 536 {
 537         u64 *pos;
 538         u64 *end;
 539
 540         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 541                 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
 542                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 543                                pos, *pos);
 544                         return 0;
 545                 }
 546         return 1;
 547 }
 548 #endif
 549
 550 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 551 {
 552         ASSERT(is_empty_shadow_page(sp->spt));
 553         list_del(&sp->link);
 554         __free_page(virt_to_page(sp->spt));
 555         __free_page(virt_to_page(sp->gfns));
 556         kfree(sp);
 557         ++kvm->arch.n_free_mmu_pages;
 558 }
 559
 560 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 561 {
 562         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 563 }
 564
 565 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 566                                                u64 *parent_pte)
 567 {
 568         struct kvm_mmu_page *sp;
 569
 570         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 571         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 572         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 573         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 574         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 575         ASSERT(is_empty_shadow_page(sp->spt));
 576         sp->slot_bitmap = 0;
 577         sp->multimapped = 0;
 578         sp->parent_pte = parent_pte;
 579         --vcpu->kvm->arch.n_free_mmu_pages;
 580         return sp;
 581 }
 582
 583 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 584                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 585 {
 586         struct kvm_pte_chain *pte_chain;
 587         struct hlist_node *node;
 588         int i;
 589
 590         if (!parent_pte)
 591                 return;
 592         if (!sp->multimapped) {
 593                 u64 *old = sp->parent_pte;
 594
 595                 if (!old) {
 596                         sp->parent_pte = parent_pte;
 597                         return;
 598                 }
 599                 sp->multimapped = 1;
 600                 pte_chain = mmu_alloc_pte_chain(vcpu);
 601                 INIT_HLIST_HEAD(&sp->parent_ptes);
 602                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 603                 pte_chain->parent_ptes[0] = old;
 604         }
 605         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 606                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 607                         continue;
 608                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 609                         if (!pte_chain->parent_ptes[i]) {
 610                                 pte_chain->parent_ptes[i] = parent_pte;
 611                                 return;
 612                         }
 613         }
 614         pte_chain = mmu_alloc_pte_chain(vcpu);
 615         BUG_ON(!pte_chain);
 616         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 617         pte_chain->parent_ptes[0] = parent_pte;
 618 }
 619
 620 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 621                                        u64 *parent_pte)
 622 {
 623         struct kvm_pte_chain *pte_chain;
 624         struct hlist_node *node;
 625         int i;
 626
 627         if (!sp->multimapped) {
 628                 BUG_ON(sp->parent_pte != parent_pte);
 629                 sp->parent_pte = NULL;
 630                 return;
 631         }
 632         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 633                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 634                         if (!pte_chain->parent_ptes[i])
 635                                 break;
 636                         if (pte_chain->parent_ptes[i] != parent_pte)
 637                                 continue;
 638                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 639                                 && pte_chain->parent_ptes[i + 1]) {
 640                                 pte_chain->parent_ptes[i]
 641                                         = pte_chain->parent_ptes[i + 1];
 642                                 ++i;
 643                         }
 644                         pte_chain->parent_ptes[i] = NULL;
 645                         if (i == 0) {
 646                                 hlist_del(&pte_chain->link);
 647                                 mmu_free_pte_chain(pte_chain);
 648                                 if (hlist_empty(&sp->parent_ptes)) {
 649                                         sp->multimapped = 0;
 650                                         sp->parent_pte = NULL;
 651                                 }
 652                         }
 653                         return;
 654                 }
 655         BUG();
 656 }
 657
 658 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
 659 {
 660         unsigned index;
 661         struct hlist_head *bucket;
 662         struct kvm_mmu_page *sp;
 663         struct hlist_node *node;
 664
 665         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 666         index = kvm_page_table_hashfn(gfn);
 667         bucket = &kvm->arch.mmu_page_hash[index];
 668         hlist_for_each_entry(sp, node, bucket, hash_link)
 669                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 670                         pgprintk("%s: found role %x\n",
 671                                  __FUNCTION__, sp->role.word);
 672                         return sp;
 673                 }
 674         return NULL;
 675 }
 676
 677 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 678                                              gfn_t gfn,
 679                                              gva_t gaddr,
 680                                              unsigned level,
 681                                              int metaphysical,
 682                                              unsigned access,
 683                                              u64 *parent_pte)
 684 {
 685         union kvm_mmu_page_role role;
 686         unsigned index;
 687         unsigned quadrant;
 688         struct hlist_head *bucket;
 689         struct kvm_mmu_page *sp;
 690         struct hlist_node *node;
 691
 692         role.word = 0;
 693         role.glevels = vcpu->arch.mmu.root_level;
 694         role.level = level;
 695         role.metaphysical = metaphysical;
 696         role.access = access;
 697         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
 698                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 699                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 700                 role.quadrant = quadrant;
 701         }
 702         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 703                  gfn, role.word);
 704         index = kvm_page_table_hashfn(gfn);
 705         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
 706         hlist_for_each_entry(sp, node, bucket, hash_link)
 707                 if (sp->gfn == gfn && sp->role.word == role.word) {
 708                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
 709                         pgprintk("%s: found\n", __FUNCTION__);
 710                         return sp;
 711                 }
 712         ++vcpu->kvm->stat.mmu_cache_miss;
 713         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
 714         if (!sp)
 715                 return sp;
 716         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 717         sp->gfn = gfn;
 718         sp->role = role;
 719         hlist_add_head(&sp->hash_link, bucket);
 720         vcpu->arch.mmu.prefetch_page(vcpu, sp);
 721         if (!metaphysical)
 722                 rmap_write_protect(vcpu->kvm, gfn);
 723         return sp;
 724 }
 725
 726 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 727                                          struct kvm_mmu_page *sp)
 728 {
 729         unsigned i;
 730         u64 *pt;
 731         u64 ent;
 732
 733         pt = sp->spt;
 734
 735         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
 736                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 737                         if (is_shadow_present_pte(pt[i]))
 738                                 rmap_remove(kvm, &pt[i]);
 739                         pt[i] = shadow_trap_nonpresent_pte;
 740                 }
 741                 kvm_flush_remote_tlbs(kvm);
 742                 return;
 743         }
 744
 745         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 746                 ent = pt[i];
 747
 748                 pt[i] = shadow_trap_nonpresent_pte;
 749                 if (!is_shadow_present_pte(ent))
 750                         continue;
 751                 ent &= PT64_BASE_ADDR_MASK;
 752                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 753         }
 754         kvm_flush_remote_tlbs(kvm);
 755 }
 756
 757 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
 758 {
 759         mmu_page_remove_parent_pte(sp, parent_pte);
 760 }
 761
 762 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 763 {
 764         int i;
 765
 766         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 767                 if (kvm->vcpus[i])
 768                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
 769 }
 770
 771 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 772 {
 773         u64 *parent_pte;
 774
 775         ++kvm->stat.mmu_shadow_zapped;
 776         while (sp->multimapped || sp->parent_pte) {
 777                 if (!sp->multimapped)
 778                         parent_pte = sp->parent_pte;
 779                 else {
 780                         struct kvm_pte_chain *chain;
 781
 782                         chain = container_of(sp->parent_ptes.first,
 783                                              struct kvm_pte_chain, link);
 784                         parent_pte = chain->parent_ptes[0];
 785                 }
 786                 BUG_ON(!parent_pte);
 787                 kvm_mmu_put_page(sp, parent_pte);
 788                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 789         }
 790         kvm_mmu_page_unlink_children(kvm, sp);
 791         if (!sp->root_count) {
 792                 hlist_del(&sp->hash_link);
 793                 kvm_mmu_free_page(kvm, sp);
 794         } else
 795                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
 796         kvm_mmu_reset_last_pte_updated(kvm);
 797 }
 798
 799 /*
 800  * Changing the number of mmu pages allocated to the vm
 801  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 802  */
 803 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 804 {
 805         /*
 806          * If we set the number of mmu pages to be smaller be than the
 807          * number of actived pages , we must to free some mmu pages before we
 808          * change the value
 809          */
 810
 811         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
 812             kvm_nr_mmu_pages) {
 813                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
 814                                        - kvm->arch.n_free_mmu_pages;
 815
 816                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 817                         struct kvm_mmu_page *page;
 818
 819                         page = container_of(kvm->arch.active_mmu_pages.prev,
 820                                             struct kvm_mmu_page, link);
 821                         kvm_mmu_zap_page(kvm, page);
 822                         n_used_mmu_pages--;
 823                 }
 824                 kvm->arch.n_free_mmu_pages = 0;
 825         }
 826         else
 827                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
 828                                          - kvm->arch.n_alloc_mmu_pages;
 829
 830         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
 831 }
 832
 833 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 834 {
 835         unsigned index;
 836         struct hlist_head *bucket;
 837         struct kvm_mmu_page *sp;
 838         struct hlist_node *node, *n;
 839         int r;
 840
 841         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 842         r = 0;
 843         index = kvm_page_table_hashfn(gfn);
 844         bucket = &kvm->arch.mmu_page_hash[index];
 845         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
 846                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 847                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 848                                  sp->role.word);
 849                         kvm_mmu_zap_page(kvm, sp);
 850                         r = 1;
 851                 }
 852         return r;
 853 }
 854
 855 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
 856 {
 857         struct kvm_mmu_page *sp;
 858
 859         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
 860                 pgprintk("%s: zap %lx %x\n", __FUNCTION__, gfn, sp->role.word);
 861                 kvm_mmu_zap_page(kvm, sp);
 862         }
 863 }
 864
 865 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
 866 {
 867         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
 868         struct kvm_mmu_page *sp = page_header(__pa(pte));
 869
 870         __set_bit(slot, &sp->slot_bitmap);
 871 }
 872
 873 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 874 {
 875         struct page *page;
 876
 877         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
 878
 879         if (gpa == UNMAPPED_GVA)
 880                 return NULL;
 881
 882         down_read(&current->mm->mmap_sem);
 883         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 884         up_read(&current->mm->mmap_sem);
 885
 886         return page;
 887 }
 888
 889 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
 890                          unsigned pt_access, unsigned pte_access,
 891                          int user_fault, int write_fault, int dirty,
 892                          int *ptwrite, gfn_t gfn, struct page *page)
 893 {
 894         u64 spte;
 895         int was_rmapped = 0;
 896         int was_writeble = is_writeble_pte(*shadow_pte);
 897         hfn_t host_pfn = (*shadow_pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 898
 899         pgprintk("%s: spte %llx access %x write_fault %d"
 900                  " user_fault %d gfn %lx\n",
 901                  __FUNCTION__, *shadow_pte, pt_access,
 902                  write_fault, user_fault, gfn);
 903
 904         if (is_rmap_pte(*shadow_pte)) {
 905                 if (host_pfn != page_to_pfn(page)) {
 906                         pgprintk("hfn old %lx new %lx\n",
 907                                  host_pfn, page_to_pfn(page));
 908                         rmap_remove(vcpu->kvm, shadow_pte);
 909                 }
 910                 else
 911                         was_rmapped = 1;
 912         }
 913
 914         /*
 915          * We don't set the accessed bit, since we sometimes want to see
 916          * whether the guest actually used the pte (in order to detect
 917          * demand paging).
 918          */
 919         spte = PT_PRESENT_MASK | PT_DIRTY_MASK;
 920         if (!dirty)
 921                 pte_access &= ~ACC_WRITE_MASK;
 922         if (!(pte_access & ACC_EXEC_MASK))
 923                 spte |= PT64_NX_MASK;
 924
 925         spte |= PT_PRESENT_MASK;
 926         if (pte_access & ACC_USER_MASK)
 927                 spte |= PT_USER_MASK;
 928
 929         if (is_error_page(page)) {
 930                 set_shadow_pte(shadow_pte,
 931                                shadow_trap_nonpresent_pte | PT_SHADOW_IO_MARK);
 932                 kvm_release_page_clean(page);
 933                 return;
 934         }
 935
 936         spte |= page_to_phys(page);
 937
 938         if ((pte_access & ACC_WRITE_MASK)
 939             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
 940                 struct kvm_mmu_page *shadow;
 941
 942                 spte |= PT_WRITABLE_MASK;
 943                 if (user_fault) {
 944                         mmu_unshadow(vcpu->kvm, gfn);
 945                         goto unshadowed;
 946                 }
 947
 948                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
 949                 if (shadow) {
 950                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 951                                  __FUNCTION__, gfn);
 952                         pte_access &= ~ACC_WRITE_MASK;
 953                         if (is_writeble_pte(spte)) {
 954                                 spte &= ~PT_WRITABLE_MASK;
 955                                 kvm_x86_ops->tlb_flush(vcpu);
 956                         }
 957                         if (write_fault)
 958                                 *ptwrite = 1;
 959                 }
 960         }
 961
 962 unshadowed:
 963
 964         if (pte_access & ACC_WRITE_MASK)
 965                 mark_page_dirty(vcpu->kvm, gfn);
 966
 967         pgprintk("%s: setting spte %llx\n", __FUNCTION__, spte);
 968         set_shadow_pte(shadow_pte, spte);
 969         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
 970         if (!was_rmapped) {
 971                 rmap_add(vcpu, shadow_pte, gfn);
 972                 if (!is_rmap_pte(*shadow_pte))
 973                         kvm_release_page_clean(page);
 974         } else {
 975                 if (was_writeble)
 976                         kvm_release_page_dirty(page);
 977                 else
 978                         kvm_release_page_clean(page);
 979         }
 980         if (!ptwrite || !*ptwrite)
 981                 vcpu->arch.last_pte_updated = shadow_pte;
 982 }
 983
 984 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 985 {
 986 }
 987
 988 static int __nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write,
 989                            gfn_t gfn, struct page *page)
 990 {
 991         int level = PT32E_ROOT_LEVEL;
 992         hpa_t table_addr = vcpu->arch.mmu.root_hpa;
 993         int pt_write = 0;
 994
 995         for (; ; level--) {
 996                 u32 index = PT64_INDEX(v, level);
 997                 u64 *table;
 998
 999                 ASSERT(VALID_PAGE(table_addr));
1000                 table = __va(table_addr);
1001
1002                 if (level == 1) {
1003                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1004                                      0, write, 1, &pt_write, gfn, page);
1005                         return pt_write || is_io_pte(table[index]);
1006                 }
1007
1008                 if (table[index] == shadow_trap_nonpresent_pte) {
1009                         struct kvm_mmu_page *new_table;
1010                         gfn_t pseudo_gfn;
1011
1012                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1013                                 >> PAGE_SHIFT;
1014                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1015                                                      v, level - 1,
1016                                                      1, ACC_ALL, &table[index]);
1017                         if (!new_table) {
1018                                 pgprintk("nonpaging_map: ENOMEM\n");
1019                                 kvm_release_page_clean(page);
1020                                 return -ENOMEM;
1021                         }
1022
1023                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
1024                                 | PT_WRITABLE_MASK | PT_USER_MASK;
1025                 }
1026                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1027         }
1028 }
1029
1030 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1031 {
1032         int r;
1033
1034         struct page *page;
1035
1036         down_read(&vcpu->kvm->slots_lock);
1037
1038         down_read(&current->mm->mmap_sem);
1039         page = gfn_to_page(vcpu->kvm, gfn);
1040         up_read(&current->mm->mmap_sem);
1041
1042         spin_lock(&vcpu->kvm->mmu_lock);
1043         kvm_mmu_free_some_pages(vcpu);
1044         r = __nonpaging_map(vcpu, v, write, gfn, page);
1045         spin_unlock(&vcpu->kvm->mmu_lock);
1046
1047         up_read(&vcpu->kvm->slots_lock);
1048
1049         return r;
1050 }
1051
1052
1053 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1054                                     struct kvm_mmu_page *sp)
1055 {
1056         int i;
1057
1058         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1059                 sp->spt[i] = shadow_trap_nonpresent_pte;
1060 }
1061
1062 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1063 {
1064         int i;
1065         struct kvm_mmu_page *sp;
1066
1067         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1068                 return;
1069         spin_lock(&vcpu->kvm->mmu_lock);
1070 #ifdef CONFIG_X86_64
1071         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1072                 hpa_t root = vcpu->arch.mmu.root_hpa;
1073
1074                 sp = page_header(root);
1075                 --sp->root_count;
1076                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1077                 spin_unlock(&vcpu->kvm->mmu_lock);
1078                 return;
1079         }
1080 #endif
1081         for (i = 0; i < 4; ++i) {
1082                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1083
1084                 if (root) {
1085                         root &= PT64_BASE_ADDR_MASK;
1086                         sp = page_header(root);
1087                         --sp->root_count;
1088                 }
1089                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1090         }
1091         spin_unlock(&vcpu->kvm->mmu_lock);
1092         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1093 }
1094
1095 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1096 {
1097         int i;
1098         gfn_t root_gfn;
1099         struct kvm_mmu_page *sp;
1100
1101         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1102
1103 #ifdef CONFIG_X86_64
1104         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1105                 hpa_t root = vcpu->arch.mmu.root_hpa;
1106
1107                 ASSERT(!VALID_PAGE(root));
1108                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1109                                       PT64_ROOT_LEVEL, 0, ACC_ALL, NULL);
1110                 root = __pa(sp->spt);
1111                 ++sp->root_count;
1112                 vcpu->arch.mmu.root_hpa = root;
1113                 return;
1114         }
1115 #endif
1116         for (i = 0; i < 4; ++i) {
1117                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1118
1119                 ASSERT(!VALID_PAGE(root));
1120                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1121                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1122                                 vcpu->arch.mmu.pae_root[i] = 0;
1123                                 continue;
1124                         }
1125                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1126                 } else if (vcpu->arch.mmu.root_level == 0)
1127                         root_gfn = 0;
1128                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1129                                       PT32_ROOT_LEVEL, !is_paging(vcpu),
1130                                       ACC_ALL, NULL);
1131                 root = __pa(sp->spt);
1132                 ++sp->root_count;
1133                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1134         }
1135         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1136 }
1137
1138 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1139 {
1140         return vaddr;
1141 }
1142
1143 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1144                                 u32 error_code)
1145 {
1146         gfn_t gfn;
1147         int r;
1148
1149         pgprintk("%s: gva %lx error %x\n", __FUNCTION__, gva, error_code);
1150         r = mmu_topup_memory_caches(vcpu);
1151         if (r)
1152                 return r;
1153
1154         ASSERT(vcpu);
1155         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1156
1157         gfn = gva >> PAGE_SHIFT;
1158
1159         return nonpaging_map(vcpu, gva & PAGE_MASK,
1160                              error_code & PFERR_WRITE_MASK, gfn);
1161 }
1162
1163 static void nonpaging_free(struct kvm_vcpu *vcpu)
1164 {
1165         mmu_free_roots(vcpu);
1166 }
1167
1168 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1169 {
1170         struct kvm_mmu *context = &vcpu->arch.mmu;
1171
1172         context->new_cr3 = nonpaging_new_cr3;
1173         context->page_fault = nonpaging_page_fault;
1174         context->gva_to_gpa = nonpaging_gva_to_gpa;
1175         context->free = nonpaging_free;
1176         context->prefetch_page = nonpaging_prefetch_page;
1177         context->root_level = 0;
1178         context->shadow_root_level = PT32E_ROOT_LEVEL;
1179         context->root_hpa = INVALID_PAGE;
1180         return 0;
1181 }
1182
1183 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1184 {
1185         ++vcpu->stat.tlb_flush;
1186         kvm_x86_ops->tlb_flush(vcpu);
1187 }
1188
1189 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1190 {
1191         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->arch.cr3);
1192         mmu_free_roots(vcpu);
1193 }
1194
1195 static void inject_page_fault(struct kvm_vcpu *vcpu,
1196                               u64 addr,
1197                               u32 err_code)
1198 {
1199         kvm_inject_page_fault(vcpu, addr, err_code);
1200 }
1201
1202 static void paging_free(struct kvm_vcpu *vcpu)
1203 {
1204         nonpaging_free(vcpu);
1205 }
1206
1207 #define PTTYPE 64
1208 #include "paging_tmpl.h"
1209 #undef PTTYPE
1210
1211 #define PTTYPE 32
1212 #include "paging_tmpl.h"
1213 #undef PTTYPE
1214
1215 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1216 {
1217         struct kvm_mmu *context = &vcpu->arch.mmu;
1218
1219         ASSERT(is_pae(vcpu));
1220         context->new_cr3 = paging_new_cr3;
1221         context->page_fault = paging64_page_fault;
1222         context->gva_to_gpa = paging64_gva_to_gpa;
1223         context->prefetch_page = paging64_prefetch_page;
1224         context->free = paging_free;
1225         context->root_level = level;
1226         context->shadow_root_level = level;
1227         context->root_hpa = INVALID_PAGE;
1228         return 0;
1229 }
1230
1231 static int paging64_init_context(struct kvm_vcpu *vcpu)
1232 {
1233         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1234 }
1235
1236 static int paging32_init_context(struct kvm_vcpu *vcpu)
1237 {
1238         struct kvm_mmu *context = &vcpu->arch.mmu;
1239
1240         context->new_cr3 = paging_new_cr3;
1241         context->page_fault = paging32_page_fault;
1242         context->gva_to_gpa = paging32_gva_to_gpa;
1243         context->free = paging_free;
1244         context->prefetch_page = paging32_prefetch_page;
1245         context->root_level = PT32_ROOT_LEVEL;
1246         context->shadow_root_level = PT32E_ROOT_LEVEL;
1247         context->root_hpa = INVALID_PAGE;
1248         return 0;
1249 }
1250
1251 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1252 {
1253         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1254 }
1255
1256 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1257 {
1258         ASSERT(vcpu);
1259         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1260
1261         if (!is_paging(vcpu))
1262                 return nonpaging_init_context(vcpu);
1263         else if (is_long_mode(vcpu))
1264                 return paging64_init_context(vcpu);
1265         else if (is_pae(vcpu))
1266                 return paging32E_init_context(vcpu);
1267         else
1268                 return paging32_init_context(vcpu);
1269 }
1270
1271 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1272 {
1273         ASSERT(vcpu);
1274         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1275                 vcpu->arch.mmu.free(vcpu);
1276                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1277         }
1278 }
1279
1280 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1281 {
1282         destroy_kvm_mmu(vcpu);
1283         return init_kvm_mmu(vcpu);
1284 }
1285 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1286
1287 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1288 {
1289         int r;
1290
1291         r = mmu_topup_memory_caches(vcpu);
1292         if (r)
1293                 goto out;
1294         spin_lock(&vcpu->kvm->mmu_lock);
1295         kvm_mmu_free_some_pages(vcpu);
1296         mmu_alloc_roots(vcpu);
1297         spin_unlock(&vcpu->kvm->mmu_lock);
1298         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1299         kvm_mmu_flush_tlb(vcpu);
1300 out:
1301         return r;
1302 }
1303 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1304
1305 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1306 {
1307         mmu_free_roots(vcpu);
1308 }
1309
1310 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1311                                   struct kvm_mmu_page *sp,
1312                                   u64 *spte)
1313 {
1314         u64 pte;
1315         struct kvm_mmu_page *child;
1316
1317         pte = *spte;
1318         if (is_shadow_present_pte(pte)) {
1319                 if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1320                         rmap_remove(vcpu->kvm, spte);
1321                 else {
1322                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1323                         mmu_page_remove_parent_pte(child, spte);
1324                 }
1325         }
1326         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1327 }
1328
1329 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1330                                   struct kvm_mmu_page *sp,
1331                                   u64 *spte,
1332                                   const void *new)
1333 {
1334         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1335                 ++vcpu->kvm->stat.mmu_pde_zapped;
1336                 return;
1337         }
1338
1339         ++vcpu->kvm->stat.mmu_pte_updated;
1340         if (sp->role.glevels == PT32_ROOT_LEVEL)
1341                 paging32_update_pte(vcpu, sp, spte, new);
1342         else
1343                 paging64_update_pte(vcpu, sp, spte, new);
1344 }
1345
1346 static bool need_remote_flush(u64 old, u64 new)
1347 {
1348         if (!is_shadow_present_pte(old))
1349                 return false;
1350         if (!is_shadow_present_pte(new))
1351                 return true;
1352         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1353                 return true;
1354         old ^= PT64_NX_MASK;
1355         new ^= PT64_NX_MASK;
1356         return (old & ~new & PT64_PERM_MASK) != 0;
1357 }
1358
1359 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1360 {
1361         if (need_remote_flush(old, new))
1362                 kvm_flush_remote_tlbs(vcpu->kvm);
1363         else
1364                 kvm_mmu_flush_tlb(vcpu);
1365 }
1366
1367 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1368 {
1369         u64 *spte = vcpu->arch.last_pte_updated;
1370
1371         return !!(spte && (*spte & PT_ACCESSED_MASK));
1372 }
1373
1374 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1375                                           const u8 *new, int bytes)
1376 {
1377         gfn_t gfn;
1378         int r;
1379         u64 gpte = 0;
1380         struct page *page;
1381
1382         if (bytes != 4 && bytes != 8)
1383                 return;
1384
1385         /*
1386          * Assume that the pte write on a page table of the same type
1387          * as the current vcpu paging mode.  This is nearly always true
1388          * (might be false while changing modes).  Note it is verified later
1389          * by update_pte().
1390          */
1391         if (is_pae(vcpu)) {
1392                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1393                 if ((bytes == 4) && (gpa % 4 == 0)) {
1394                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1395                         if (r)
1396                                 return;
1397                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1398                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1399                         memcpy((void *)&gpte, new, 8);
1400                 }
1401         } else {
1402                 if ((bytes == 4) && (gpa % 4 == 0))
1403                         memcpy((void *)&gpte, new, 4);
1404         }
1405         if (!is_present_pte(gpte))
1406                 return;
1407         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1408
1409         down_read(&current->mm->mmap_sem);
1410         page = gfn_to_page(vcpu->kvm, gfn);
1411         up_read(&current->mm->mmap_sem);
1412
1413         vcpu->arch.update_pte.gfn = gfn;
1414         vcpu->arch.update_pte.page = page;
1415 }
1416
1417 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1418                        const u8 *new, int bytes)
1419 {
1420         gfn_t gfn = gpa >> PAGE_SHIFT;
1421         struct kvm_mmu_page *sp;
1422         struct hlist_node *node, *n;
1423         struct hlist_head *bucket;
1424         unsigned index;
1425         u64 entry, gentry;
1426         u64 *spte;
1427         unsigned offset = offset_in_page(gpa);
1428         unsigned pte_size;
1429         unsigned page_offset;
1430         unsigned misaligned;
1431         unsigned quadrant;
1432         int level;
1433         int flooded = 0;
1434         int npte;
1435         int r;
1436
1437         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1438         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1439         spin_lock(&vcpu->kvm->mmu_lock);
1440         kvm_mmu_free_some_pages(vcpu);
1441         ++vcpu->kvm->stat.mmu_pte_write;
1442         kvm_mmu_audit(vcpu, "pre pte write");
1443         if (gfn == vcpu->arch.last_pt_write_gfn
1444             && !last_updated_pte_accessed(vcpu)) {
1445                 ++vcpu->arch.last_pt_write_count;
1446                 if (vcpu->arch.last_pt_write_count >= 3)
1447                         flooded = 1;
1448         } else {
1449                 vcpu->arch.last_pt_write_gfn = gfn;
1450                 vcpu->arch.last_pt_write_count = 1;
1451                 vcpu->arch.last_pte_updated = NULL;
1452         }
1453         index = kvm_page_table_hashfn(gfn);
1454         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1455         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1456                 if (sp->gfn != gfn || sp->role.metaphysical)
1457                         continue;
1458                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1459                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1460                 misaligned |= bytes < 4;
1461                 if (misaligned || flooded) {
1462                         /*
1463                          * Misaligned accesses are too much trouble to fix
1464                          * up; also, they usually indicate a page is not used
1465                          * as a page table.
1466                          *
1467                          * If we're seeing too many writes to a page,
1468                          * it may no longer be a page table, or we may be
1469                          * forking, in which case it is better to unmap the
1470                          * page.
1471                          */
1472                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1473                                  gpa, bytes, sp->role.word);
1474                         kvm_mmu_zap_page(vcpu->kvm, sp);
1475                         ++vcpu->kvm->stat.mmu_flooded;
1476                         continue;
1477                 }
1478                 page_offset = offset;
1479                 level = sp->role.level;
1480                 npte = 1;
1481                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1482                         page_offset <<= 1;      /* 32->64 */
1483                         /*
1484                          * A 32-bit pde maps 4MB while the shadow pdes map
1485                          * only 2MB.  So we need to double the offset again
1486                          * and zap two pdes instead of one.
1487                          */
1488                         if (level == PT32_ROOT_LEVEL) {
1489                                 page_offset &= ~7; /* kill rounding error */
1490                                 page_offset <<= 1;
1491                                 npte = 2;
1492                         }
1493                         quadrant = page_offset >> PAGE_SHIFT;
1494                         page_offset &= ~PAGE_MASK;
1495                         if (quadrant != sp->role.quadrant)
1496                                 continue;
1497                 }
1498                 spte = &sp->spt[page_offset / sizeof(*spte)];
1499                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1500                         gentry = 0;
1501                         r = kvm_read_guest_atomic(vcpu->kvm,
1502                                                   gpa & ~(u64)(pte_size - 1),
1503                                                   &gentry, pte_size);
1504                         new = (const void *)&gentry;
1505                         if (r < 0)
1506                                 new = NULL;
1507                 }
1508                 while (npte--) {
1509                         entry = *spte;
1510                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1511                         if (new)
1512                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1513                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1514                         ++spte;
1515                 }
1516         }
1517         kvm_mmu_audit(vcpu, "post pte write");
1518         spin_unlock(&vcpu->kvm->mmu_lock);
1519         if (vcpu->arch.update_pte.page) {
1520                 kvm_release_page_clean(vcpu->arch.update_pte.page);
1521                 vcpu->arch.update_pte.page = NULL;
1522         }
1523 }
1524
1525 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1526 {
1527         gpa_t gpa;
1528         int r;
1529
1530         down_read(&vcpu->kvm->slots_lock);
1531         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1532         up_read(&vcpu->kvm->slots_lock);
1533
1534         spin_lock(&vcpu->kvm->mmu_lock);
1535         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1536         spin_unlock(&vcpu->kvm->mmu_lock);
1537         return r;
1538 }
1539
1540 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1541 {
1542         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1543                 struct kvm_mmu_page *sp;
1544
1545                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1546                                   struct kvm_mmu_page, link);
1547                 kvm_mmu_zap_page(vcpu->kvm, sp);
1548                 ++vcpu->kvm->stat.mmu_recycled;
1549         }
1550 }
1551
1552 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1553 {
1554         int r;
1555         enum emulation_result er;
1556
1557         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1558         if (r < 0)
1559                 goto out;
1560
1561         if (!r) {
1562                 r = 1;
1563                 goto out;
1564         }
1565
1566         r = mmu_topup_memory_caches(vcpu);
1567         if (r)
1568                 goto out;
1569
1570         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1571
1572         switch (er) {
1573         case EMULATE_DONE:
1574                 return 1;
1575         case EMULATE_DO_MMIO:
1576                 ++vcpu->stat.mmio_exits;
1577                 return 0;
1578         case EMULATE_FAIL:
1579                 kvm_report_emulation_failure(vcpu, "pagetable");
1580                 return 1;
1581         default:
1582                 BUG();
1583         }
1584 out:
1585         return r;
1586 }
1587 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1588
1589 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1590 {
1591         struct kvm_mmu_page *sp;
1592
1593         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1594                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1595                                   struct kvm_mmu_page, link);
1596                 kvm_mmu_zap_page(vcpu->kvm, sp);
1597         }
1598         free_page((unsigned long)vcpu->arch.mmu.pae_root);
1599 }
1600
1601 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1602 {
1603         struct page *page;
1604         int i;
1605
1606         ASSERT(vcpu);
1607
1608         if (vcpu->kvm->arch.n_requested_mmu_pages)
1609                 vcpu->kvm->arch.n_free_mmu_pages =
1610                                         vcpu->kvm->arch.n_requested_mmu_pages;
1611         else
1612                 vcpu->kvm->arch.n_free_mmu_pages =
1613                                         vcpu->kvm->arch.n_alloc_mmu_pages;
1614         /*
1615          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1616          * Therefore we need to allocate shadow page tables in the first
1617          * 4GB of memory, which happens to fit the DMA32 zone.
1618          */
1619         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1620         if (!page)
1621                 goto error_1;
1622         vcpu->arch.mmu.pae_root = page_address(page);
1623         for (i = 0; i < 4; ++i)
1624                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1625
1626         return 0;
1627
1628 error_1:
1629         free_mmu_pages(vcpu);
1630         return -ENOMEM;
1631 }
1632
1633 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1634 {
1635         ASSERT(vcpu);
1636         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1637
1638         return alloc_mmu_pages(vcpu);
1639 }
1640
1641 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1642 {
1643         ASSERT(vcpu);
1644         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1645
1646         return init_kvm_mmu(vcpu);
1647 }
1648
1649 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1650 {
1651         ASSERT(vcpu);
1652
1653         destroy_kvm_mmu(vcpu);
1654         free_mmu_pages(vcpu);
1655         mmu_free_memory_caches(vcpu);
1656 }
1657
1658 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1659 {
1660         struct kvm_mmu_page *sp;
1661
1662         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1663                 int i;
1664                 u64 *pt;
1665
1666                 if (!test_bit(slot, &sp->slot_bitmap))
1667                         continue;
1668
1669                 pt = sp->spt;
1670                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1671                         /* avoid RMW */
1672                         if (pt[i] & PT_WRITABLE_MASK)
1673                                 pt[i] &= ~PT_WRITABLE_MASK;
1674         }
1675 }
1676
1677 void kvm_mmu_zap_all(struct kvm *kvm)
1678 {
1679         struct kvm_mmu_page *sp, *node;
1680
1681         spin_lock(&kvm->mmu_lock);
1682         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1683                 kvm_mmu_zap_page(kvm, sp);
1684         spin_unlock(&kvm->mmu_lock);
1685
1686         kvm_flush_remote_tlbs(kvm);
1687 }
1688
1689 void kvm_mmu_module_exit(void)
1690 {
1691         if (pte_chain_cache)
1692                 kmem_cache_destroy(pte_chain_cache);
1693         if (rmap_desc_cache)
1694                 kmem_cache_destroy(rmap_desc_cache);
1695         if (mmu_page_header_cache)
1696                 kmem_cache_destroy(mmu_page_header_cache);
1697 }
1698
1699 int kvm_mmu_module_init(void)
1700 {
1701         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1702                                             sizeof(struct kvm_pte_chain),
1703                                             0, 0, NULL);
1704         if (!pte_chain_cache)
1705                 goto nomem;
1706         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1707                                             sizeof(struct kvm_rmap_desc),
1708                                             0, 0, NULL);
1709         if (!rmap_desc_cache)
1710                 goto nomem;
1711
1712         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1713                                                   sizeof(struct kvm_mmu_page),
1714                                                   0, 0, NULL);
1715         if (!mmu_page_header_cache)
1716                 goto nomem;
1717
1718         return 0;
1719
1720 nomem:
1721         kvm_mmu_module_exit();
1722         return -ENOMEM;
1723 }
1724
1725 /*
1726  * Caculate mmu pages needed for kvm.
1727  */
1728 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
1729 {
1730         int i;
1731         unsigned int nr_mmu_pages;
1732         unsigned int  nr_pages = 0;
1733
1734         for (i = 0; i < kvm->nmemslots; i++)
1735                 nr_pages += kvm->memslots[i].npages;
1736
1737         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
1738         nr_mmu_pages = max(nr_mmu_pages,
1739                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
1740
1741         return nr_mmu_pages;
1742 }
1743
1744 #ifdef AUDIT
1745
1746 static const char *audit_msg;
1747
1748 static gva_t canonicalize(gva_t gva)
1749 {
1750 #ifdef CONFIG_X86_64
1751         gva = (long long)(gva << 16) >> 16;
1752 #endif
1753         return gva;
1754 }
1755
1756 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1757                                 gva_t va, int level)
1758 {
1759         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1760         int i;
1761         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1762
1763         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1764                 u64 ent = pt[i];
1765
1766                 if (ent == shadow_trap_nonpresent_pte)
1767                         continue;
1768
1769                 va = canonicalize(va);
1770                 if (level > 1) {
1771                         if (ent == shadow_notrap_nonpresent_pte)
1772                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1773                                        " in nonleaf level: levels %d gva %lx"
1774                                        " level %d pte %llx\n", audit_msg,
1775                                        vcpu->arch.mmu.root_level, va, level, ent);
1776
1777                         audit_mappings_page(vcpu, ent, va, level - 1);
1778                 } else {
1779                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
1780                         struct page *page = gpa_to_page(vcpu, gpa);
1781                         hpa_t hpa = page_to_phys(page);
1782
1783                         if (is_shadow_present_pte(ent)
1784                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1785                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1786                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1787                                        audit_msg, vcpu->arch.mmu.root_level,
1788                                        va, gpa, hpa, ent,
1789                                        is_shadow_present_pte(ent));
1790                         else if (ent == shadow_notrap_nonpresent_pte
1791                                  && !is_error_hpa(hpa))
1792                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1793                                        " valid guest gva %lx\n", audit_msg, va);
1794                         kvm_release_page_clean(page);
1795
1796                 }
1797         }
1798 }
1799
1800 static void audit_mappings(struct kvm_vcpu *vcpu)
1801 {
1802         unsigned i;
1803
1804         if (vcpu->arch.mmu.root_level == 4)
1805                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
1806         else
1807                 for (i = 0; i < 4; ++i)
1808                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
1809                                 audit_mappings_page(vcpu,
1810                                                     vcpu->arch.mmu.pae_root[i],
1811                                                     i << 30,
1812                                                     2);
1813 }
1814
1815 static int count_rmaps(struct kvm_vcpu *vcpu)
1816 {
1817         int nmaps = 0;
1818         int i, j, k;
1819
1820         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1821                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1822                 struct kvm_rmap_desc *d;
1823
1824                 for (j = 0; j < m->npages; ++j) {
1825                         unsigned long *rmapp = &m->rmap[j];
1826
1827                         if (!*rmapp)
1828                                 continue;
1829                         if (!(*rmapp & 1)) {
1830                                 ++nmaps;
1831                                 continue;
1832                         }
1833                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1834                         while (d) {
1835                                 for (k = 0; k < RMAP_EXT; ++k)
1836                                         if (d->shadow_ptes[k])
1837                                                 ++nmaps;
1838                                         else
1839                                                 break;
1840                                 d = d->more;
1841                         }
1842                 }
1843         }
1844         return nmaps;
1845 }
1846
1847 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1848 {
1849         int nmaps = 0;
1850         struct kvm_mmu_page *sp;
1851         int i;
1852
1853         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
1854                 u64 *pt = sp->spt;
1855
1856                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
1857                         continue;
1858
1859                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1860                         u64 ent = pt[i];
1861
1862                         if (!(ent & PT_PRESENT_MASK))
1863                                 continue;
1864                         if (!(ent & PT_WRITABLE_MASK))
1865                                 continue;
1866                         ++nmaps;
1867                 }
1868         }
1869         return nmaps;
1870 }
1871
1872 static void audit_rmap(struct kvm_vcpu *vcpu)
1873 {
1874         int n_rmap = count_rmaps(vcpu);
1875         int n_actual = count_writable_mappings(vcpu);
1876
1877         if (n_rmap != n_actual)
1878                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1879                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1880 }
1881
1882 static void audit_write_protection(struct kvm_vcpu *vcpu)
1883 {
1884         struct kvm_mmu_page *sp;
1885         struct kvm_memory_slot *slot;
1886         unsigned long *rmapp;
1887         gfn_t gfn;
1888
1889         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
1890                 if (sp->role.metaphysical)
1891                         continue;
1892
1893                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
1894                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
1895                 rmapp = &slot->rmap[gfn - slot->base_gfn];
1896                 if (*rmapp)
1897                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1898                                " mappings: gfn %lx role %x\n",
1899                                __FUNCTION__, audit_msg, sp->gfn,
1900                                sp->role.word);
1901         }
1902 }
1903
1904 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1905 {
1906         int olddbg = dbg;
1907
1908         dbg = 0;
1909         audit_msg = msg;
1910         audit_rmap(vcpu);
1911         audit_write_protection(vcpu);
1912         audit_mappings(vcpu);
1913         dbg = olddbg;
1914 }
1915
1916 #endif