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