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 /*
  36  * When setting this variable to true it enables Two-Dimensional-Paging
  37  * where the hardware walks 2 page tables:
  38  * 1. the guest-virtual to guest-physical
  39  * 2. while doing 1. it walks guest-physical to host-physical
  40  * If the hardware supports that we don't need to do shadow paging.
  41  */
  42 static bool tdp_enabled = false;
  43
  44 #undef MMU_DEBUG
  45
  46 #undef AUDIT
  47
  48 #ifdef AUDIT
  49 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  50 #else
  51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  52 #endif
  53
  54 #ifdef MMU_DEBUG
  55
  56 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  57 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  58
  59 #else
  60
  61 #define pgprintk(x...) do { } while (0)
  62 #define rmap_printk(x...) do { } while (0)
  63
  64 #endif
  65
  66 #if defined(MMU_DEBUG) || defined(AUDIT)
  67 static int dbg = 1;
  68 #endif
  69
  70 #ifndef MMU_DEBUG
  71 #define ASSERT(x) do { } while (0)
  72 #else
  73 #define ASSERT(x)                                                       \
  74         if (!(x)) {                                                     \
  75                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  76                        __FILE__, __LINE__, #x);                         \
  77         }
  78 #endif
  79
  80 #define PT64_PT_BITS 9
  81 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  82 #define PT32_PT_BITS 10
  83 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  84
  85 #define PT_WRITABLE_SHIFT 1
  86
  87 #define PT_PRESENT_MASK (1ULL << 0)
  88 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  89 #define PT_USER_MASK (1ULL << 2)
  90 #define PT_PWT_MASK (1ULL << 3)
  91 #define PT_PCD_MASK (1ULL << 4)
  92 #define PT_ACCESSED_MASK (1ULL << 5)
  93 #define PT_DIRTY_MASK (1ULL << 6)
  94 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  95 #define PT_PAT_MASK (1ULL << 7)
  96 #define PT_GLOBAL_MASK (1ULL << 8)
  97 #define PT64_NX_SHIFT 63
  98 #define PT64_NX_MASK (1ULL << PT64_NX_SHIFT)
  99
 100 #define PT_PAT_SHIFT 7
 101 #define PT_DIR_PAT_SHIFT 12
 102 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
 103
 104 #define PT32_DIR_PSE36_SIZE 4
 105 #define PT32_DIR_PSE36_SHIFT 13
 106 #define PT32_DIR_PSE36_MASK \
 107         (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
 108
 109
 110 #define PT_FIRST_AVAIL_BITS_SHIFT 9
 111 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
 112
 113 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 114
 115 #define PT64_LEVEL_BITS 9
 116
 117 #define PT64_LEVEL_SHIFT(level) \
 118                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
 119
 120 #define PT64_LEVEL_MASK(level) \
 121                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 122
 123 #define PT64_INDEX(address, level)\
 124         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 125
 126
 127 #define PT32_LEVEL_BITS 10
 128
 129 #define PT32_LEVEL_SHIFT(level) \
 130                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 131
 132 #define PT32_LEVEL_MASK(level) \
 133                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 134
 135 #define PT32_INDEX(address, level)\
 136         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 137
 138
 139 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 140 #define PT64_DIR_BASE_ADDR_MASK \
 141         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 142
 143 #define PT32_BASE_ADDR_MASK PAGE_MASK
 144 #define PT32_DIR_BASE_ADDR_MASK \
 145         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 146
 147 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 148                         | PT64_NX_MASK)
 149
 150 #define PFERR_PRESENT_MASK (1U << 0)
 151 #define PFERR_WRITE_MASK (1U << 1)
 152 #define PFERR_USER_MASK (1U << 2)
 153 #define PFERR_FETCH_MASK (1U << 4)
 154
 155 #define PT64_ROOT_LEVEL 4
 156 #define PT32_ROOT_LEVEL 2
 157 #define PT32E_ROOT_LEVEL 3
 158
 159 #define PT_DIRECTORY_LEVEL 2
 160 #define PT_PAGE_TABLE_LEVEL 1
 161
 162 #define RMAP_EXT 4
 163
 164 #define ACC_EXEC_MASK    1
 165 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 166 #define ACC_USER_MASK    PT_USER_MASK
 167 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 168
 169 struct kvm_rmap_desc {
 170         u64 *shadow_ptes[RMAP_EXT];
 171         struct kvm_rmap_desc *more;
 172 };
 173
 174 static struct kmem_cache *pte_chain_cache;
 175 static struct kmem_cache *rmap_desc_cache;
 176 static struct kmem_cache *mmu_page_header_cache;
 177
 178 static u64 __read_mostly shadow_trap_nonpresent_pte;
 179 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 180
 181 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 182 {
 183         shadow_trap_nonpresent_pte = trap_pte;
 184         shadow_notrap_nonpresent_pte = notrap_pte;
 185 }
 186 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 187
 188 static int is_write_protection(struct kvm_vcpu *vcpu)
 189 {
 190         return vcpu->arch.cr0 & X86_CR0_WP;
 191 }
 192
 193 static int is_cpuid_PSE36(void)
 194 {
 195         return 1;
 196 }
 197
 198 static int is_nx(struct kvm_vcpu *vcpu)
 199 {
 200         return vcpu->arch.shadow_efer & EFER_NX;
 201 }
 202
 203 static int is_present_pte(unsigned long pte)
 204 {
 205         return pte & PT_PRESENT_MASK;
 206 }
 207
 208 static int is_shadow_present_pte(u64 pte)
 209 {
 210         return pte != shadow_trap_nonpresent_pte
 211                 && pte != shadow_notrap_nonpresent_pte;
 212 }
 213
 214 static int is_writeble_pte(unsigned long pte)
 215 {
 216         return pte & PT_WRITABLE_MASK;
 217 }
 218
 219 static int is_dirty_pte(unsigned long pte)
 220 {
 221         return pte & PT_DIRTY_MASK;
 222 }
 223
 224 static int is_rmap_pte(u64 pte)
 225 {
 226         return is_shadow_present_pte(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 != 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 & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 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);
 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 {
 686         union kvm_mmu_page_role role;
 687         unsigned index;
 688         unsigned quadrant;
 689         struct hlist_head *bucket;
 690         struct kvm_mmu_page *sp;
 691         struct hlist_node *node;
 692
 693         role.word = 0;
 694         role.glevels = vcpu->arch.mmu.root_level;
 695         role.level = level;
 696         role.metaphysical = metaphysical;
 697         role.access = access;
 698         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
 699                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 700                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 701                 role.quadrant = quadrant;
 702         }
 703         pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
 704                  gfn, role.word);
 705         index = kvm_page_table_hashfn(gfn);
 706         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
 707         hlist_for_each_entry(sp, node, bucket, hash_link)
 708                 if (sp->gfn == gfn && sp->role.word == role.word) {
 709                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
 710                         pgprintk("%s: found\n", __FUNCTION__);
 711                         return sp;
 712                 }
 713         ++vcpu->kvm->stat.mmu_cache_miss;
 714         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
 715         if (!sp)
 716                 return sp;
 717         pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
 718         sp->gfn = gfn;
 719         sp->role = role;
 720         hlist_add_head(&sp->hash_link, bucket);
 721         vcpu->arch.mmu.prefetch_page(vcpu, sp);
 722         if (!metaphysical)
 723                 rmap_write_protect(vcpu->kvm, gfn);
 724         return sp;
 725 }
 726
 727 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 728                                          struct kvm_mmu_page *sp)
 729 {
 730         unsigned i;
 731         u64 *pt;
 732         u64 ent;
 733
 734         pt = sp->spt;
 735
 736         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
 737                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 738                         if (is_shadow_present_pte(pt[i]))
 739                                 rmap_remove(kvm, &pt[i]);
 740                         pt[i] = shadow_trap_nonpresent_pte;
 741                 }
 742                 kvm_flush_remote_tlbs(kvm);
 743                 return;
 744         }
 745
 746         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 747                 ent = pt[i];
 748
 749                 pt[i] = shadow_trap_nonpresent_pte;
 750                 if (!is_shadow_present_pte(ent))
 751                         continue;
 752                 ent &= PT64_BASE_ADDR_MASK;
 753                 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
 754         }
 755         kvm_flush_remote_tlbs(kvm);
 756 }
 757
 758 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
 759 {
 760         mmu_page_remove_parent_pte(sp, parent_pte);
 761 }
 762
 763 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 764 {
 765         int i;
 766
 767         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 768                 if (kvm->vcpus[i])
 769                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
 770 }
 771
 772 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 773 {
 774         u64 *parent_pte;
 775
 776         ++kvm->stat.mmu_shadow_zapped;
 777         while (sp->multimapped || sp->parent_pte) {
 778                 if (!sp->multimapped)
 779                         parent_pte = sp->parent_pte;
 780                 else {
 781                         struct kvm_pte_chain *chain;
 782
 783                         chain = container_of(sp->parent_ptes.first,
 784                                              struct kvm_pte_chain, link);
 785                         parent_pte = chain->parent_ptes[0];
 786                 }
 787                 BUG_ON(!parent_pte);
 788                 kvm_mmu_put_page(sp, parent_pte);
 789                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 790         }
 791         kvm_mmu_page_unlink_children(kvm, sp);
 792         if (!sp->root_count) {
 793                 hlist_del(&sp->hash_link);
 794                 kvm_mmu_free_page(kvm, sp);
 795         } else
 796                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
 797         kvm_mmu_reset_last_pte_updated(kvm);
 798 }
 799
 800 /*
 801  * Changing the number of mmu pages allocated to the vm
 802  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 803  */
 804 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 805 {
 806         /*
 807          * If we set the number of mmu pages to be smaller be than the
 808          * number of actived pages , we must to free some mmu pages before we
 809          * change the value
 810          */
 811
 812         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
 813             kvm_nr_mmu_pages) {
 814                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
 815                                        - kvm->arch.n_free_mmu_pages;
 816
 817                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 818                         struct kvm_mmu_page *page;
 819
 820                         page = container_of(kvm->arch.active_mmu_pages.prev,
 821                                             struct kvm_mmu_page, link);
 822                         kvm_mmu_zap_page(kvm, page);
 823                         n_used_mmu_pages--;
 824                 }
 825                 kvm->arch.n_free_mmu_pages = 0;
 826         }
 827         else
 828                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
 829                                          - kvm->arch.n_alloc_mmu_pages;
 830
 831         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
 832 }
 833
 834 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 835 {
 836         unsigned index;
 837         struct hlist_head *bucket;
 838         struct kvm_mmu_page *sp;
 839         struct hlist_node *node, *n;
 840         int r;
 841
 842         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 843         r = 0;
 844         index = kvm_page_table_hashfn(gfn);
 845         bucket = &kvm->arch.mmu_page_hash[index];
 846         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
 847                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 848                         pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
 849                                  sp->role.word);
 850                         kvm_mmu_zap_page(kvm, sp);
 851                         r = 1;
 852                 }
 853         return r;
 854 }
 855
 856 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
 857 {
 858         struct kvm_mmu_page *sp;
 859
 860         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
 861                 pgprintk("%s: zap %lx %x\n", __FUNCTION__, gfn, sp->role.word);
 862                 kvm_mmu_zap_page(kvm, sp);
 863         }
 864 }
 865
 866 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
 867 {
 868         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
 869         struct kvm_mmu_page *sp = page_header(__pa(pte));
 870
 871         __set_bit(slot, &sp->slot_bitmap);
 872 }
 873
 874 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 875 {
 876         struct page *page;
 877
 878         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
 879
 880         if (gpa == UNMAPPED_GVA)
 881                 return NULL;
 882
 883         down_read(&current->mm->mmap_sem);
 884         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
 885         up_read(&current->mm->mmap_sem);
 886
 887         return page;
 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 = 0;
 897         int was_writeble = is_writeble_pte(*shadow_pte);
 898         hfn_t host_pfn = (*shadow_pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 899
 900         pgprintk("%s: spte %llx access %x write_fault %d"
 901                  " user_fault %d gfn %lx\n",
 902                  __FUNCTION__, *shadow_pte, pt_access,
 903                  write_fault, user_fault, gfn);
 904
 905         if (is_rmap_pte(*shadow_pte)) {
 906                 if (host_pfn != page_to_pfn(page)) {
 907                         pgprintk("hfn old %lx new %lx\n",
 908                                  host_pfn, page_to_pfn(page));
 909                         rmap_remove(vcpu->kvm, shadow_pte);
 910                 }
 911                 else
 912                         was_rmapped = 1;
 913         }
 914
 915         /*
 916          * We don't set the accessed bit, since we sometimes want to see
 917          * whether the guest actually used the pte (in order to detect
 918          * demand paging).
 919          */
 920         spte = PT_PRESENT_MASK | PT_DIRTY_MASK;
 921         if (!dirty)
 922                 pte_access &= ~ACC_WRITE_MASK;
 923         if (!(pte_access & ACC_EXEC_MASK))
 924                 spte |= PT64_NX_MASK;
 925
 926         spte |= PT_PRESENT_MASK;
 927         if (pte_access & ACC_USER_MASK)
 928                 spte |= PT_USER_MASK;
 929
 930         spte |= page_to_phys(page);
 931
 932         if ((pte_access & ACC_WRITE_MASK)
 933             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
 934                 struct kvm_mmu_page *shadow;
 935
 936                 spte |= PT_WRITABLE_MASK;
 937                 if (user_fault) {
 938                         mmu_unshadow(vcpu->kvm, gfn);
 939                         goto unshadowed;
 940                 }
 941
 942                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
 943                 if (shadow) {
 944                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 945                                  __FUNCTION__, gfn);
 946                         pte_access &= ~ACC_WRITE_MASK;
 947                         if (is_writeble_pte(spte)) {
 948                                 spte &= ~PT_WRITABLE_MASK;
 949                                 kvm_x86_ops->tlb_flush(vcpu);
 950                         }
 951                         if (write_fault)
 952                                 *ptwrite = 1;
 953                 }
 954         }
 955
 956 unshadowed:
 957
 958         if (pte_access & ACC_WRITE_MASK)
 959                 mark_page_dirty(vcpu->kvm, gfn);
 960
 961         pgprintk("%s: setting spte %llx\n", __FUNCTION__, spte);
 962         set_shadow_pte(shadow_pte, spte);
 963         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
 964         if (!was_rmapped) {
 965                 rmap_add(vcpu, shadow_pte, gfn);
 966                 if (!is_rmap_pte(*shadow_pte))
 967                         kvm_release_page_clean(page);
 968         } else {
 969                 if (was_writeble)
 970                         kvm_release_page_dirty(page);
 971                 else
 972                         kvm_release_page_clean(page);
 973         }
 974         if (!ptwrite || !*ptwrite)
 975                 vcpu->arch.last_pte_updated = shadow_pte;
 976 }
 977
 978 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 979 {
 980 }
 981
 982 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
 983                            gfn_t gfn, struct page *page, int level)
 984 {
 985         hpa_t table_addr = vcpu->arch.mmu.root_hpa;
 986         int pt_write = 0;
 987
 988         for (; ; level--) {
 989                 u32 index = PT64_INDEX(v, level);
 990                 u64 *table;
 991
 992                 ASSERT(VALID_PAGE(table_addr));
 993                 table = __va(table_addr);
 994
 995                 if (level == 1) {
 996                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
 997                                      0, write, 1, &pt_write, gfn, page);
 998                         return pt_write;
 999                 }
1000
1001                 if (table[index] == shadow_trap_nonpresent_pte) {
1002                         struct kvm_mmu_page *new_table;
1003                         gfn_t pseudo_gfn;
1004
1005                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1006                                 >> PAGE_SHIFT;
1007                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1008                                                      v, level - 1,
1009                                                      1, ACC_ALL, &table[index]);
1010                         if (!new_table) {
1011                                 pgprintk("nonpaging_map: ENOMEM\n");
1012                                 kvm_release_page_clean(page);
1013                                 return -ENOMEM;
1014                         }
1015
1016                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
1017                                 | PT_WRITABLE_MASK | PT_USER_MASK;
1018                 }
1019                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1020         }
1021 }
1022
1023 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1024 {
1025         int r;
1026
1027         struct page *page;
1028
1029         down_read(&vcpu->kvm->slots_lock);
1030
1031         down_read(&current->mm->mmap_sem);
1032         page = gfn_to_page(vcpu->kvm, gfn);
1033         up_read(&current->mm->mmap_sem);
1034
1035         /* mmio */
1036         if (is_error_page(page)) {
1037                 kvm_release_page_clean(page);
1038                 up_read(&vcpu->kvm->slots_lock);
1039                 return 1;
1040         }
1041
1042         spin_lock(&vcpu->kvm->mmu_lock);
1043         kvm_mmu_free_some_pages(vcpu);
1044         r = __direct_map(vcpu, v, write, gfn, page, PT32E_ROOT_LEVEL);
1045         spin_unlock(&vcpu->kvm->mmu_lock);
1046
1047         up_read(&vcpu->kvm->slots_lock);
1048
1049         return r;
1050 }
1051
1052
1053 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1054                                     struct kvm_mmu_page *sp)
1055 {
1056         int i;
1057
1058         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1059                 sp->spt[i] = shadow_trap_nonpresent_pte;
1060 }
1061
1062 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1063 {
1064         int i;
1065         struct kvm_mmu_page *sp;
1066
1067         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1068                 return;
1069         spin_lock(&vcpu->kvm->mmu_lock);
1070 #ifdef CONFIG_X86_64
1071         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1072                 hpa_t root = vcpu->arch.mmu.root_hpa;
1073
1074                 sp = page_header(root);
1075                 --sp->root_count;
1076                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1077                 spin_unlock(&vcpu->kvm->mmu_lock);
1078                 return;
1079         }
1080 #endif
1081         for (i = 0; i < 4; ++i) {
1082                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1083
1084                 if (root) {
1085                         root &= PT64_BASE_ADDR_MASK;
1086                         sp = page_header(root);
1087                         --sp->root_count;
1088                 }
1089                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1090         }
1091         spin_unlock(&vcpu->kvm->mmu_lock);
1092         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1093 }
1094
1095 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1096 {
1097         int i;
1098         gfn_t root_gfn;
1099         struct kvm_mmu_page *sp;
1100
1101         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1102
1103 #ifdef CONFIG_X86_64
1104         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1105                 hpa_t root = vcpu->arch.mmu.root_hpa;
1106
1107                 ASSERT(!VALID_PAGE(root));
1108                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1109                                       PT64_ROOT_LEVEL, 0, ACC_ALL, NULL);
1110                 root = __pa(sp->spt);
1111                 ++sp->root_count;
1112                 vcpu->arch.mmu.root_hpa = root;
1113                 return;
1114         }
1115 #endif
1116         for (i = 0; i < 4; ++i) {
1117                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1118
1119                 ASSERT(!VALID_PAGE(root));
1120                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1121                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1122                                 vcpu->arch.mmu.pae_root[i] = 0;
1123                                 continue;
1124                         }
1125                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1126                 } else if (vcpu->arch.mmu.root_level == 0)
1127                         root_gfn = 0;
1128                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1129                                       PT32_ROOT_LEVEL, !is_paging(vcpu),
1130                                       ACC_ALL, NULL);
1131                 root = __pa(sp->spt);
1132                 ++sp->root_count;
1133                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1134         }
1135         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1136 }
1137
1138 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1139 {
1140         return vaddr;
1141 }
1142
1143 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1144                                 u32 error_code)
1145 {
1146         gfn_t gfn;
1147         int r;
1148
1149         pgprintk("%s: gva %lx error %x\n", __FUNCTION__, gva, error_code);
1150         r = mmu_topup_memory_caches(vcpu);
1151         if (r)
1152                 return r;
1153
1154         ASSERT(vcpu);
1155         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1156
1157         gfn = gva >> PAGE_SHIFT;
1158
1159         return nonpaging_map(vcpu, gva & PAGE_MASK,
1160                              error_code & PFERR_WRITE_MASK, gfn);
1161 }
1162
1163 static void nonpaging_free(struct kvm_vcpu *vcpu)
1164 {
1165         mmu_free_roots(vcpu);
1166 }
1167
1168 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1169 {
1170         struct kvm_mmu *context = &vcpu->arch.mmu;
1171
1172         context->new_cr3 = nonpaging_new_cr3;
1173         context->page_fault = nonpaging_page_fault;
1174         context->gva_to_gpa = nonpaging_gva_to_gpa;
1175         context->free = nonpaging_free;
1176         context->prefetch_page = nonpaging_prefetch_page;
1177         context->root_level = 0;
1178         context->shadow_root_level = PT32E_ROOT_LEVEL;
1179         context->root_hpa = INVALID_PAGE;
1180         return 0;
1181 }
1182
1183 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1184 {
1185         ++vcpu->stat.tlb_flush;
1186         kvm_x86_ops->tlb_flush(vcpu);
1187 }
1188
1189 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1190 {
1191         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->arch.cr3);
1192         mmu_free_roots(vcpu);
1193 }
1194
1195 static void inject_page_fault(struct kvm_vcpu *vcpu,
1196                               u64 addr,
1197                               u32 err_code)
1198 {
1199         kvm_inject_page_fault(vcpu, addr, err_code);
1200 }
1201
1202 static void paging_free(struct kvm_vcpu *vcpu)
1203 {
1204         nonpaging_free(vcpu);
1205 }
1206
1207 #define PTTYPE 64
1208 #include "paging_tmpl.h"
1209 #undef PTTYPE
1210
1211 #define PTTYPE 32
1212 #include "paging_tmpl.h"
1213 #undef PTTYPE
1214
1215 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1216 {
1217         struct kvm_mmu *context = &vcpu->arch.mmu;
1218
1219         ASSERT(is_pae(vcpu));
1220         context->new_cr3 = paging_new_cr3;
1221         context->page_fault = paging64_page_fault;
1222         context->gva_to_gpa = paging64_gva_to_gpa;
1223         context->prefetch_page = paging64_prefetch_page;
1224         context->free = paging_free;
1225         context->root_level = level;
1226         context->shadow_root_level = level;
1227         context->root_hpa = INVALID_PAGE;
1228         return 0;
1229 }
1230
1231 static int paging64_init_context(struct kvm_vcpu *vcpu)
1232 {
1233         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1234 }
1235
1236 static int paging32_init_context(struct kvm_vcpu *vcpu)
1237 {
1238         struct kvm_mmu *context = &vcpu->arch.mmu;
1239
1240         context->new_cr3 = paging_new_cr3;
1241         context->page_fault = paging32_page_fault;
1242         context->gva_to_gpa = paging32_gva_to_gpa;
1243         context->free = paging_free;
1244         context->prefetch_page = paging32_prefetch_page;
1245         context->root_level = PT32_ROOT_LEVEL;
1246         context->shadow_root_level = PT32E_ROOT_LEVEL;
1247         context->root_hpa = INVALID_PAGE;
1248         return 0;
1249 }
1250
1251 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1252 {
1253         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1254 }
1255
1256 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1257 {
1258         ASSERT(vcpu);
1259         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1260
1261         if (!is_paging(vcpu))
1262                 return nonpaging_init_context(vcpu);
1263         else if (is_long_mode(vcpu))
1264                 return paging64_init_context(vcpu);
1265         else if (is_pae(vcpu))
1266                 return paging32E_init_context(vcpu);
1267         else
1268                 return paging32_init_context(vcpu);
1269 }
1270
1271 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1272 {
1273         ASSERT(vcpu);
1274         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1275                 vcpu->arch.mmu.free(vcpu);
1276                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1277         }
1278 }
1279
1280 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1281 {
1282         destroy_kvm_mmu(vcpu);
1283         return init_kvm_mmu(vcpu);
1284 }
1285 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1286
1287 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1288 {
1289         int r;
1290
1291         r = mmu_topup_memory_caches(vcpu);
1292         if (r)
1293                 goto out;
1294         spin_lock(&vcpu->kvm->mmu_lock);
1295         kvm_mmu_free_some_pages(vcpu);
1296         mmu_alloc_roots(vcpu);
1297         spin_unlock(&vcpu->kvm->mmu_lock);
1298         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1299         kvm_mmu_flush_tlb(vcpu);
1300 out:
1301         return r;
1302 }
1303 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1304
1305 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1306 {
1307         mmu_free_roots(vcpu);
1308 }
1309
1310 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1311                                   struct kvm_mmu_page *sp,
1312                                   u64 *spte)
1313 {
1314         u64 pte;
1315         struct kvm_mmu_page *child;
1316
1317         pte = *spte;
1318         if (is_shadow_present_pte(pte)) {
1319                 if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1320                         rmap_remove(vcpu->kvm, spte);
1321                 else {
1322                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1323                         mmu_page_remove_parent_pte(child, spte);
1324                 }
1325         }
1326         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1327 }
1328
1329 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1330                                   struct kvm_mmu_page *sp,
1331                                   u64 *spte,
1332                                   const void *new)
1333 {
1334         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1335                 ++vcpu->kvm->stat.mmu_pde_zapped;
1336                 return;
1337         }
1338
1339         ++vcpu->kvm->stat.mmu_pte_updated;
1340         if (sp->role.glevels == PT32_ROOT_LEVEL)
1341                 paging32_update_pte(vcpu, sp, spte, new);
1342         else
1343                 paging64_update_pte(vcpu, sp, spte, new);
1344 }
1345
1346 static bool need_remote_flush(u64 old, u64 new)
1347 {
1348         if (!is_shadow_present_pte(old))
1349                 return false;
1350         if (!is_shadow_present_pte(new))
1351                 return true;
1352         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1353                 return true;
1354         old ^= PT64_NX_MASK;
1355         new ^= PT64_NX_MASK;
1356         return (old & ~new & PT64_PERM_MASK) != 0;
1357 }
1358
1359 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1360 {
1361         if (need_remote_flush(old, new))
1362                 kvm_flush_remote_tlbs(vcpu->kvm);
1363         else
1364                 kvm_mmu_flush_tlb(vcpu);
1365 }
1366
1367 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1368 {
1369         u64 *spte = vcpu->arch.last_pte_updated;
1370
1371         return !!(spte && (*spte & PT_ACCESSED_MASK));
1372 }
1373
1374 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1375                                           const u8 *new, int bytes)
1376 {
1377         gfn_t gfn;
1378         int r;
1379         u64 gpte = 0;
1380         struct page *page;
1381
1382         if (bytes != 4 && bytes != 8)
1383                 return;
1384
1385         /*
1386          * Assume that the pte write on a page table of the same type
1387          * as the current vcpu paging mode.  This is nearly always true
1388          * (might be false while changing modes).  Note it is verified later
1389          * by update_pte().
1390          */
1391         if (is_pae(vcpu)) {
1392                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1393                 if ((bytes == 4) && (gpa % 4 == 0)) {
1394                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1395                         if (r)
1396                                 return;
1397                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1398                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1399                         memcpy((void *)&gpte, new, 8);
1400                 }
1401         } else {
1402                 if ((bytes == 4) && (gpa % 4 == 0))
1403                         memcpy((void *)&gpte, new, 4);
1404         }
1405         if (!is_present_pte(gpte))
1406                 return;
1407         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1408
1409         down_read(&vcpu->kvm->slots_lock);
1410         page = gfn_to_page(vcpu->kvm, gfn);
1411         up_read(&vcpu->kvm->slots_lock);
1412
1413         if (is_error_page(page)) {
1414                 kvm_release_page_clean(page);
1415                 return;
1416         }
1417         vcpu->arch.update_pte.gfn = gfn;
1418         vcpu->arch.update_pte.page = page;
1419 }
1420
1421 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1422                        const u8 *new, int bytes)
1423 {
1424         gfn_t gfn = gpa >> PAGE_SHIFT;
1425         struct kvm_mmu_page *sp;
1426         struct hlist_node *node, *n;
1427         struct hlist_head *bucket;
1428         unsigned index;
1429         u64 entry, gentry;
1430         u64 *spte;
1431         unsigned offset = offset_in_page(gpa);
1432         unsigned pte_size;
1433         unsigned page_offset;
1434         unsigned misaligned;
1435         unsigned quadrant;
1436         int level;
1437         int flooded = 0;
1438         int npte;
1439         int r;
1440
1441         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1442         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1443         spin_lock(&vcpu->kvm->mmu_lock);
1444         kvm_mmu_free_some_pages(vcpu);
1445         ++vcpu->kvm->stat.mmu_pte_write;
1446         kvm_mmu_audit(vcpu, "pre pte write");
1447         if (gfn == vcpu->arch.last_pt_write_gfn
1448             && !last_updated_pte_accessed(vcpu)) {
1449                 ++vcpu->arch.last_pt_write_count;
1450                 if (vcpu->arch.last_pt_write_count >= 3)
1451                         flooded = 1;
1452         } else {
1453                 vcpu->arch.last_pt_write_gfn = gfn;
1454                 vcpu->arch.last_pt_write_count = 1;
1455                 vcpu->arch.last_pte_updated = NULL;
1456         }
1457         index = kvm_page_table_hashfn(gfn);
1458         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1459         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1460                 if (sp->gfn != gfn || sp->role.metaphysical)
1461                         continue;
1462                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1463                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1464                 misaligned |= bytes < 4;
1465                 if (misaligned || flooded) {
1466                         /*
1467                          * Misaligned accesses are too much trouble to fix
1468                          * up; also, they usually indicate a page is not used
1469                          * as a page table.
1470                          *
1471                          * If we're seeing too many writes to a page,
1472                          * it may no longer be a page table, or we may be
1473                          * forking, in which case it is better to unmap the
1474                          * page.
1475                          */
1476                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1477                                  gpa, bytes, sp->role.word);
1478                         kvm_mmu_zap_page(vcpu->kvm, sp);
1479                         ++vcpu->kvm->stat.mmu_flooded;
1480                         continue;
1481                 }
1482                 page_offset = offset;
1483                 level = sp->role.level;
1484                 npte = 1;
1485                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1486                         page_offset <<= 1;      /* 32->64 */
1487                         /*
1488                          * A 32-bit pde maps 4MB while the shadow pdes map
1489                          * only 2MB.  So we need to double the offset again
1490                          * and zap two pdes instead of one.
1491                          */
1492                         if (level == PT32_ROOT_LEVEL) {
1493                                 page_offset &= ~7; /* kill rounding error */
1494                                 page_offset <<= 1;
1495                                 npte = 2;
1496                         }
1497                         quadrant = page_offset >> PAGE_SHIFT;
1498                         page_offset &= ~PAGE_MASK;
1499                         if (quadrant != sp->role.quadrant)
1500                                 continue;
1501                 }
1502                 spte = &sp->spt[page_offset / sizeof(*spte)];
1503                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1504                         gentry = 0;
1505                         r = kvm_read_guest_atomic(vcpu->kvm,
1506                                                   gpa & ~(u64)(pte_size - 1),
1507                                                   &gentry, pte_size);
1508                         new = (const void *)&gentry;
1509                         if (r < 0)
1510                                 new = NULL;
1511                 }
1512                 while (npte--) {
1513                         entry = *spte;
1514                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1515                         if (new)
1516                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1517                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1518                         ++spte;
1519                 }
1520         }
1521         kvm_mmu_audit(vcpu, "post pte write");
1522         spin_unlock(&vcpu->kvm->mmu_lock);
1523         if (vcpu->arch.update_pte.page) {
1524                 kvm_release_page_clean(vcpu->arch.update_pte.page);
1525                 vcpu->arch.update_pte.page = NULL;
1526         }
1527 }
1528
1529 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1530 {
1531         gpa_t gpa;
1532         int r;
1533
1534         down_read(&vcpu->kvm->slots_lock);
1535         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1536         up_read(&vcpu->kvm->slots_lock);
1537
1538         spin_lock(&vcpu->kvm->mmu_lock);
1539         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1540         spin_unlock(&vcpu->kvm->mmu_lock);
1541         return r;
1542 }
1543
1544 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1545 {
1546         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1547                 struct kvm_mmu_page *sp;
1548
1549                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1550                                   struct kvm_mmu_page, link);
1551                 kvm_mmu_zap_page(vcpu->kvm, sp);
1552                 ++vcpu->kvm->stat.mmu_recycled;
1553         }
1554 }
1555
1556 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1557 {
1558         int r;
1559         enum emulation_result er;
1560
1561         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1562         if (r < 0)
1563                 goto out;
1564
1565         if (!r) {
1566                 r = 1;
1567                 goto out;
1568         }
1569
1570         r = mmu_topup_memory_caches(vcpu);
1571         if (r)
1572                 goto out;
1573
1574         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1575
1576         switch (er) {
1577         case EMULATE_DONE:
1578                 return 1;
1579         case EMULATE_DO_MMIO:
1580                 ++vcpu->stat.mmio_exits;
1581                 return 0;
1582         case EMULATE_FAIL:
1583                 kvm_report_emulation_failure(vcpu, "pagetable");
1584                 return 1;
1585         default:
1586                 BUG();
1587         }
1588 out:
1589         return r;
1590 }
1591 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1592
1593 void kvm_enable_tdp(void)
1594 {
1595         tdp_enabled = true;
1596 }
1597 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
1598
1599 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1600 {
1601         struct kvm_mmu_page *sp;
1602
1603         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1604                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1605                                   struct kvm_mmu_page, link);
1606                 kvm_mmu_zap_page(vcpu->kvm, sp);
1607         }
1608         free_page((unsigned long)vcpu->arch.mmu.pae_root);
1609 }
1610
1611 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1612 {
1613         struct page *page;
1614         int i;
1615
1616         ASSERT(vcpu);
1617
1618         if (vcpu->kvm->arch.n_requested_mmu_pages)
1619                 vcpu->kvm->arch.n_free_mmu_pages =
1620                                         vcpu->kvm->arch.n_requested_mmu_pages;
1621         else
1622                 vcpu->kvm->arch.n_free_mmu_pages =
1623                                         vcpu->kvm->arch.n_alloc_mmu_pages;
1624         /*
1625          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1626          * Therefore we need to allocate shadow page tables in the first
1627          * 4GB of memory, which happens to fit the DMA32 zone.
1628          */
1629         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1630         if (!page)
1631                 goto error_1;
1632         vcpu->arch.mmu.pae_root = page_address(page);
1633         for (i = 0; i < 4; ++i)
1634                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1635
1636         return 0;
1637
1638 error_1:
1639         free_mmu_pages(vcpu);
1640         return -ENOMEM;
1641 }
1642
1643 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1644 {
1645         ASSERT(vcpu);
1646         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1647
1648         return alloc_mmu_pages(vcpu);
1649 }
1650
1651 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1652 {
1653         ASSERT(vcpu);
1654         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1655
1656         return init_kvm_mmu(vcpu);
1657 }
1658
1659 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1660 {
1661         ASSERT(vcpu);
1662
1663         destroy_kvm_mmu(vcpu);
1664         free_mmu_pages(vcpu);
1665         mmu_free_memory_caches(vcpu);
1666 }
1667
1668 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1669 {
1670         struct kvm_mmu_page *sp;
1671
1672         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1673                 int i;
1674                 u64 *pt;
1675
1676                 if (!test_bit(slot, &sp->slot_bitmap))
1677                         continue;
1678
1679                 pt = sp->spt;
1680                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1681                         /* avoid RMW */
1682                         if (pt[i] & PT_WRITABLE_MASK)
1683                                 pt[i] &= ~PT_WRITABLE_MASK;
1684         }
1685 }
1686
1687 void kvm_mmu_zap_all(struct kvm *kvm)
1688 {
1689         struct kvm_mmu_page *sp, *node;
1690
1691         spin_lock(&kvm->mmu_lock);
1692         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1693                 kvm_mmu_zap_page(kvm, sp);
1694         spin_unlock(&kvm->mmu_lock);
1695
1696         kvm_flush_remote_tlbs(kvm);
1697 }
1698
1699 void kvm_mmu_module_exit(void)
1700 {
1701         if (pte_chain_cache)
1702                 kmem_cache_destroy(pte_chain_cache);
1703         if (rmap_desc_cache)
1704                 kmem_cache_destroy(rmap_desc_cache);
1705         if (mmu_page_header_cache)
1706                 kmem_cache_destroy(mmu_page_header_cache);
1707 }
1708
1709 int kvm_mmu_module_init(void)
1710 {
1711         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1712                                             sizeof(struct kvm_pte_chain),
1713                                             0, 0, NULL);
1714         if (!pte_chain_cache)
1715                 goto nomem;
1716         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1717                                             sizeof(struct kvm_rmap_desc),
1718                                             0, 0, NULL);
1719         if (!rmap_desc_cache)
1720                 goto nomem;
1721
1722         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1723                                                   sizeof(struct kvm_mmu_page),
1724                                                   0, 0, NULL);
1725         if (!mmu_page_header_cache)
1726                 goto nomem;
1727
1728         return 0;
1729
1730 nomem:
1731         kvm_mmu_module_exit();
1732         return -ENOMEM;
1733 }
1734
1735 /*
1736  * Caculate mmu pages needed for kvm.
1737  */
1738 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
1739 {
1740         int i;
1741         unsigned int nr_mmu_pages;
1742         unsigned int  nr_pages = 0;
1743
1744         for (i = 0; i < kvm->nmemslots; i++)
1745                 nr_pages += kvm->memslots[i].npages;
1746
1747         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
1748         nr_mmu_pages = max(nr_mmu_pages,
1749                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
1750
1751         return nr_mmu_pages;
1752 }
1753
1754 #ifdef AUDIT
1755
1756 static const char *audit_msg;
1757
1758 static gva_t canonicalize(gva_t gva)
1759 {
1760 #ifdef CONFIG_X86_64
1761         gva = (long long)(gva << 16) >> 16;
1762 #endif
1763         return gva;
1764 }
1765
1766 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1767                                 gva_t va, int level)
1768 {
1769         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1770         int i;
1771         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1772
1773         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1774                 u64 ent = pt[i];
1775
1776                 if (ent == shadow_trap_nonpresent_pte)
1777                         continue;
1778
1779                 va = canonicalize(va);
1780                 if (level > 1) {
1781                         if (ent == shadow_notrap_nonpresent_pte)
1782                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1783                                        " in nonleaf level: levels %d gva %lx"
1784                                        " level %d pte %llx\n", audit_msg,
1785                                        vcpu->arch.mmu.root_level, va, level, ent);
1786
1787                         audit_mappings_page(vcpu, ent, va, level - 1);
1788                 } else {
1789                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
1790                         struct page *page = gpa_to_page(vcpu, gpa);
1791                         hpa_t hpa = page_to_phys(page);
1792
1793                         if (is_shadow_present_pte(ent)
1794                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1795                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1796                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1797                                        audit_msg, vcpu->arch.mmu.root_level,
1798                                        va, gpa, hpa, ent,
1799                                        is_shadow_present_pte(ent));
1800                         else if (ent == shadow_notrap_nonpresent_pte
1801                                  && !is_error_hpa(hpa))
1802                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1803                                        " valid guest gva %lx\n", audit_msg, va);
1804                         kvm_release_page_clean(page);
1805
1806                 }
1807         }
1808 }
1809
1810 static void audit_mappings(struct kvm_vcpu *vcpu)
1811 {
1812         unsigned i;
1813
1814         if (vcpu->arch.mmu.root_level == 4)
1815                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
1816         else
1817                 for (i = 0; i < 4; ++i)
1818                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
1819                                 audit_mappings_page(vcpu,
1820                                                     vcpu->arch.mmu.pae_root[i],
1821                                                     i << 30,
1822                                                     2);
1823 }
1824
1825 static int count_rmaps(struct kvm_vcpu *vcpu)
1826 {
1827         int nmaps = 0;
1828         int i, j, k;
1829
1830         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1831                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1832                 struct kvm_rmap_desc *d;
1833
1834                 for (j = 0; j < m->npages; ++j) {
1835                         unsigned long *rmapp = &m->rmap[j];
1836
1837                         if (!*rmapp)
1838                                 continue;
1839                         if (!(*rmapp & 1)) {
1840                                 ++nmaps;
1841                                 continue;
1842                         }
1843                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1844                         while (d) {
1845                                 for (k = 0; k < RMAP_EXT; ++k)
1846                                         if (d->shadow_ptes[k])
1847                                                 ++nmaps;
1848                                         else
1849                                                 break;
1850                                 d = d->more;
1851                         }
1852                 }
1853         }
1854         return nmaps;
1855 }
1856
1857 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1858 {
1859         int nmaps = 0;
1860         struct kvm_mmu_page *sp;
1861         int i;
1862
1863         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
1864                 u64 *pt = sp->spt;
1865
1866                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
1867                         continue;
1868
1869                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1870                         u64 ent = pt[i];
1871
1872                         if (!(ent & PT_PRESENT_MASK))
1873                                 continue;
1874                         if (!(ent & PT_WRITABLE_MASK))
1875                                 continue;
1876                         ++nmaps;
1877                 }
1878         }
1879         return nmaps;
1880 }
1881
1882 static void audit_rmap(struct kvm_vcpu *vcpu)
1883 {
1884         int n_rmap = count_rmaps(vcpu);
1885         int n_actual = count_writable_mappings(vcpu);
1886
1887         if (n_rmap != n_actual)
1888                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1889                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1890 }
1891
1892 static void audit_write_protection(struct kvm_vcpu *vcpu)
1893 {
1894         struct kvm_mmu_page *sp;
1895         struct kvm_memory_slot *slot;
1896         unsigned long *rmapp;
1897         gfn_t gfn;
1898
1899         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
1900                 if (sp->role.metaphysical)
1901                         continue;
1902
1903                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
1904                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
1905                 rmapp = &slot->rmap[gfn - slot->base_gfn];
1906                 if (*rmapp)
1907                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1908                                " mappings: gfn %lx role %x\n",
1909                                __FUNCTION__, audit_msg, sp->gfn,
1910                                sp->role.word);
1911         }
1912 }
1913
1914 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1915 {
1916         int olddbg = dbg;
1917
1918         dbg = 0;
1919         audit_msg = msg;
1920         audit_rmap(vcpu);
1921         audit_write_protection(vcpu);
1922         audit_mappings(vcpu);
1923         dbg = olddbg;
1924 }
1925
1926 #endif