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