arch/x86/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "vmx.h"
  21 #include "mmu.h"
  22
  23 #include <linux/kvm_host.h>
  24 #include <linux/types.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/highmem.h>
  28 #include <linux/module.h>
  29 #include <linux/swap.h>
  30
  31 #include <asm/page.h>
  32 #include <asm/cmpxchg.h>
  33 #include <asm/io.h>
  34
  35 #undef MMU_DEBUG
  36
  37 #undef AUDIT
  38
  39 #ifdef AUDIT
  40 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  41 #else
  42 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  43 #endif
  44
  45 #ifdef MMU_DEBUG
  46
  47 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  48 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  49
  50 #else
  51
  52 #define pgprintk(x...) do { } while (0)
  53 #define rmap_printk(x...) do { } while (0)
  54
  55 #endif
  56
  57 #if defined(MMU_DEBUG) || defined(AUDIT)
  58 static int dbg = 1;
  59 #endif
  60
  61 #ifndef MMU_DEBUG
  62 #define ASSERT(x) do { } while (0)
  63 #else
  64 #define ASSERT(x)                                                       \
  65         if (!(x)) {                                                     \
  66                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  67                        __FILE__, __LINE__, #x);                         \
  68         }
  69 #endif
  70
  71 #define PT64_PT_BITS 9
  72 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
  73 #define PT32_PT_BITS 10
  74 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
  75
  76 #define PT_WRITABLE_SHIFT 1
  77
  78 #define PT_PRESENT_MASK (1ULL << 0)
  79 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
  80 #define PT_USER_MASK (1ULL << 2)
  81 #define PT_PWT_MASK (1ULL << 3)
  82 #define PT_PCD_MASK (1ULL << 4)
  83 #define PT_ACCESSED_MASK (1ULL << 5)
  84 #define PT_DIRTY_MASK (1ULL << 6)
  85 #define PT_PAGE_SIZE_MASK (1ULL << 7)
  86 #define PT_PAT_MASK (1ULL << 7)
  87 #define PT_GLOBAL_MASK (1ULL << 8)
  88 #define PT64_NX_SHIFT 63
  89 #define PT64_NX_MASK (1ULL << PT64_NX_SHIFT)
  90
  91 #define PT_PAT_SHIFT 7
  92 #define PT_DIR_PAT_SHIFT 12
  93 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
  94
  95 #define PT32_DIR_PSE36_SIZE 4
  96 #define PT32_DIR_PSE36_SHIFT 13
  97 #define PT32_DIR_PSE36_MASK \
  98         (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
  99
 100
 101 #define PT_FIRST_AVAIL_BITS_SHIFT 9
 102 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
 103
 104 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
 105
 106 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
 107
 108 #define PT64_LEVEL_BITS 9
 109
 110 #define PT64_LEVEL_SHIFT(level) \
 111                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
 112
 113 #define PT64_LEVEL_MASK(level) \
 114                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
 115
 116 #define PT64_INDEX(address, level)\
 117         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 118
 119
 120 #define PT32_LEVEL_BITS 10
 121
 122 #define PT32_LEVEL_SHIFT(level) \
 123                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 124
 125 #define PT32_LEVEL_MASK(level) \
 126                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 127
 128 #define PT32_INDEX(address, level)\
 129         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 130
 131
 132 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 133 #define PT64_DIR_BASE_ADDR_MASK \
 134         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 135
 136 #define PT32_BASE_ADDR_MASK PAGE_MASK
 137 #define PT32_DIR_BASE_ADDR_MASK \
 138         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 139
 140 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 141                         | PT64_NX_MASK)
 142
 143 #define PFERR_PRESENT_MASK (1U << 0)
 144 #define PFERR_WRITE_MASK (1U << 1)
 145 #define PFERR_USER_MASK (1U << 2)
 146 #define PFERR_FETCH_MASK (1U << 4)
 147
 148 #define PT64_ROOT_LEVEL 4
 149 #define PT32_ROOT_LEVEL 2
 150 #define PT32E_ROOT_LEVEL 3
 151
 152 #define PT_DIRECTORY_LEVEL 2
 153 #define PT_PAGE_TABLE_LEVEL 1
 154
 155 #define RMAP_EXT 4
 156
 157 #define ACC_EXEC_MASK    1
 158 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 159 #define ACC_USER_MASK    PT_USER_MASK
 160 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 161
 162 struct kvm_rmap_desc {
 163         u64 *shadow_ptes[RMAP_EXT];
 164         struct kvm_rmap_desc *more;
 165 };
 166
 167 static struct kmem_cache *pte_chain_cache;
 168 static struct kmem_cache *rmap_desc_cache;
 169 static struct kmem_cache *mmu_page_header_cache;
 170
 171 static u64 __read_mostly shadow_trap_nonpresent_pte;
 172 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 173
 174 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 175 {
 176         shadow_trap_nonpresent_pte = trap_pte;
 177         shadow_notrap_nonpresent_pte = notrap_pte;
 178 }
 179 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 180
 181 static int is_write_protection(struct kvm_vcpu *vcpu)
 182 {
 183         return vcpu->arch.cr0 & X86_CR0_WP;
 184 }
 185
 186 static int is_cpuid_PSE36(void)
 187 {
 188         return 1;
 189 }
 190
 191 static int is_nx(struct kvm_vcpu *vcpu)
 192 {
 193         return vcpu->arch.shadow_efer & EFER_NX;
 194 }
 195
 196 static int is_present_pte(unsigned long pte)
 197 {
 198         return pte & PT_PRESENT_MASK;
 199 }
 200
 201 static int is_shadow_present_pte(u64 pte)
 202 {
 203         pte &= ~PT_SHADOW_IO_MARK;
 204         return pte != shadow_trap_nonpresent_pte
 205                 && pte != shadow_notrap_nonpresent_pte;
 206 }
 207
 208 static int is_writeble_pte(unsigned long pte)
 209 {
 210         return pte & PT_WRITABLE_MASK;
 211 }
 212
 213 static int is_dirty_pte(unsigned long pte)
 214 {
 215         return pte & PT_DIRTY_MASK;
 216 }
 217
 218 static int is_io_pte(unsigned long pte)
 219 {
 220         return pte & PT_SHADOW_IO_MARK;
 221 }
 222
 223 static int is_rmap_pte(u64 pte)
 224 {
 225         return pte != shadow_trap_nonpresent_pte
 226                 && pte != shadow_notrap_nonpresent_pte;
 227 }
 228
 229 static gfn_t pse36_gfn_delta(u32 gpte)
 230 {
 231         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 232
 233         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 234 }
 235
 236 static void set_shadow_pte(u64 *sptep, u64 spte)
 237 {
 238 #ifdef CONFIG_X86_64
 239         set_64bit((unsigned long *)sptep, spte);
 240 #else
 241         set_64bit((unsigned long long *)sptep, spte);
 242 #endif
 243 }
 244
 245 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 246                                   struct kmem_cache *base_cache, int min)
 247 {
 248         void *obj;
 249
 250         if (cache->nobjs >= min)
 251                 return 0;
 252         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 253                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 254                 if (!obj)
 255                         return -ENOMEM;
 256                 cache->objects[cache->nobjs++] = obj;
 257         }
 258         return 0;
 259 }
 260
 261 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 262 {
 263         while (mc->nobjs)
 264                 kfree(mc->objects[--mc->nobjs]);
 265 }
 266
 267 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 268                                        int min)
 269 {
 270         struct page *page;
 271
 272         if (cache->nobjs >= min)
 273                 return 0;
 274         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 275                 page = alloc_page(GFP_KERNEL);
 276                 if (!page)
 277                         return -ENOMEM;
 278                 set_page_private(page, 0);
 279                 cache->objects[cache->nobjs++] = page_address(page);
 280         }
 281         return 0;
 282 }
 283
 284 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 285 {
 286         while (mc->nobjs)
 287                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 288 }
 289
 290 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 291 {
 292         int r;
 293
 294         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 295                                    pte_chain_cache, 4);
 296         if (r)
 297                 goto out;
 298         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 299                                    rmap_desc_cache, 1);
 300         if (r)
 301                 goto out;
 302         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 303         if (r)
 304                 goto out;
 305         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 306                                    mmu_page_header_cache, 4);
 307 out:
 308         return r;
 309 }
 310
 311 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 312 {
 313         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 314         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 315         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 316         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 317 }
 318
 319 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 320                                     size_t size)
 321 {
 322         void *p;
 323
 324         BUG_ON(!mc->nobjs);
 325         p = mc->objects[--mc->nobjs];
 326         memset(p, 0, size);
 327         return p;
 328 }
 329
 330 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 331 {
 332         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 333                                       sizeof(struct kvm_pte_chain));
 334 }
 335
 336 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 337 {
 338         kfree(pc);
 339 }
 340
 341 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 342 {
 343         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 344                                       sizeof(struct kvm_rmap_desc));
 345 }
 346
 347 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 348 {
 349         kfree(rd);
 350 }
 351
 352 /*
 353  * Take gfn and return the reverse mapping to it.
 354  * Note: gfn must be unaliased before this function get called
 355  */
 356
 357 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
 358 {
 359         struct kvm_memory_slot *slot;
 360
 361         slot = gfn_to_memslot(kvm, gfn);
 362         return &slot->rmap[gfn - slot->base_gfn];
 363 }
 364
 365 /*
 366  * Reverse mapping data structures:
 367  *
 368  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 369  * that points to page_address(page).
 370  *
 371  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 372  * containing more mappings.
 373  */
 374 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
 375 {
 376         struct kvm_mmu_page *sp;
 377         struct kvm_rmap_desc *desc;
 378         unsigned long *rmapp;
 379         int i;
 380
 381         if (!is_rmap_pte(*spte))
 382                 return;
 383         gfn = unalias_gfn(vcpu->kvm, gfn);
 384         sp = page_header(__pa(spte));
 385         sp->gfns[spte - sp->spt] = gfn;
 386         rmapp = gfn_to_rmap(vcpu->kvm, gfn);
 387         if (!*rmapp) {
 388                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 389                 *rmapp = (unsigned long)spte;
 390         } else if (!(*rmapp & 1)) {
 391                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 392                 desc = mmu_alloc_rmap_desc(vcpu);
 393                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 394                 desc->shadow_ptes[1] = spte;
 395                 *rmapp = (unsigned long)desc | 1;
 396         } else {
 397                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 398                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 399                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 400                         desc = desc->more;
 401                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 402                         desc->more = mmu_alloc_rmap_desc(vcpu);
 403                         desc = desc->more;
 404                 }
 405                 for (i = 0; desc->shadow_ptes[i]; ++i)
 406                         ;
 407                 desc->shadow_ptes[i] = spte;
 408         }
 409 }
 410
 411 static void rmap_desc_remove_entry(unsigned long *rmapp,
 412                                    struct kvm_rmap_desc *desc,
 413                                    int i,
 414                                    struct kvm_rmap_desc *prev_desc)
 415 {
 416         int j;
 417
 418         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 419                 ;
 420         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 421         desc->shadow_ptes[j] = NULL;
 422         if (j != 0)
 423                 return;
 424         if (!prev_desc && !desc->more)
 425                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 426         else
 427                 if (prev_desc)
 428                         prev_desc->more = desc->more;
 429                 else
 430                         *rmapp = (unsigned long)desc->more | 1;
 431         mmu_free_rmap_desc(desc);
 432 }
 433
 434 static void rmap_remove(struct kvm *kvm, u64 *spte)
 435 {
 436         struct kvm_rmap_desc *desc;
 437         struct kvm_rmap_desc *prev_desc;
 438         struct kvm_mmu_page *sp;
 439         struct page *page;
 440         unsigned long *rmapp;
 441         int i;
 442
 443         if (!is_rmap_pte(*spte))
 444                 return;
 445         sp = page_header(__pa(spte));
 446         page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
 447         mark_page_accessed(page);
 448         if (is_writeble_pte(*spte))
 449                 kvm_release_page_dirty(page);
 450         else
 451                 kvm_release_page_clean(page);
 452         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt]);
 453         if (!*rmapp) {
 454                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 455                 BUG();
 456         } else if (!(*rmapp & 1)) {
 457                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 458                 if ((u64 *)*rmapp != spte) {
 459                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 460                                spte, *spte);
 461                         BUG();
 462                 }
 463                 *rmapp = 0;
 464         } else {
 465                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 466                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 467                 prev_desc = NULL;
 468                 while (desc) {
 469                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 470                                 if (desc->shadow_ptes[i] == spte) {
 471                                         rmap_desc_remove_entry(rmapp,
 472                                                                desc, i,
 473                                                                prev_desc);
 474                                         return;
 475                                 }
 476                         prev_desc = desc;
 477                         desc = desc->more;
 478                 }
 479                 BUG();
 480         }
 481 }
 482
 483 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 484 {
 485         struct kvm_rmap_desc *desc;
 486         struct kvm_rmap_desc *prev_desc;
 487         u64 *prev_spte;
 488         int i;
 489
 490         if (!*rmapp)
 491                 return NULL;
 492         else if (!(*rmapp & 1)) {
 493                 if (!spte)
 494                         return (u64 *)*rmapp;
 495                 return NULL;
 496         }
 497         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 498         prev_desc = NULL;
 499         prev_spte = NULL;
 500         while (desc) {
 501                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 502                         if (prev_spte == spte)
 503                                 return desc->shadow_ptes[i];
 504                         prev_spte = desc->shadow_ptes[i];
 505                 }
 506                 desc = desc->more;
 507         }
 508         return NULL;
 509 }
 510
 511 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 512 {
 513         unsigned long *rmapp;
 514         u64 *spte;
 515         int write_protected = 0;
 516
 517         gfn = unalias_gfn(kvm, gfn);
 518         rmapp = gfn_to_rmap(kvm, gfn);
 519
 520         spte = rmap_next(kvm, rmapp, NULL);
 521         while (spte) {
 522                 BUG_ON(!spte);
 523                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 524                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 525                 if (is_writeble_pte(*spte)) {
 526                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 527                         write_protected = 1;
 528                 }
 529                 spte = rmap_next(kvm, rmapp, spte);
 530         }
 531         if (write_protected)
 532                 kvm_flush_remote_tlbs(kvm);
 533 }
 534
 535 #ifdef MMU_DEBUG
 536 static int is_empty_shadow_page(u64 *spt)
 537 {
 538         u64 *pos;
 539         u64 *end;
 540
 541         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 542                 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
 543                         printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
 544                                pos, *pos);
 545                         return 0;
 546                 }
 547         return 1;
 548 }
 549 #endif
 550
 551 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 552 {
 553         ASSERT(is_empty_shadow_page(sp->spt));
 554         list_del(&sp->link);
 555         __free_page(virt_to_page(sp->spt));
 556         __free_page(virt_to_page(sp->gfns));
 557         kfree(sp);
 558         ++kvm->arch.n_free_mmu_pages;
 559 }
 560
 561 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 562 {
 563         return gfn;
 564 }
 565
 566 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 567                                                u64 *parent_pte)
 568 {
 569         struct kvm_mmu_page *sp;
 570
 571         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 572         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 573         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 574         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 575         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 576         ASSERT(is_empty_shadow_page(sp->spt));
 577         sp->slot_bitmap = 0;
 578         sp->multimapped = 0;
 579         sp->parent_pte = parent_pte;
 580         --vcpu->kvm->arch.n_free_mmu_pages;
 581         return sp;
 582 }
 583
 584 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 585                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 586 {
 587         struct kvm_pte_chain *pte_chain;
 588         struct hlist_node *node;
 589         int i;
 590
 591         if (!parent_pte)
 592                 return;
 593         if (!sp->multimapped) {
 594                 u64 *old = sp->parent_pte;
 595
 596                 if (!old) {
 597                         sp->parent_pte = parent_pte;
 598                         return;
 599                 }
 600                 sp->multimapped = 1;
 601                 pte_chain = mmu_alloc_pte_chain(vcpu);
 602                 INIT_HLIST_HEAD(&sp->parent_ptes);
 603                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 604                 pte_chain->parent_ptes[0] = old;
 605         }
 606         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 607                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 608                         continue;
 609                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 610                         if (!pte_chain->parent_ptes[i]) {
 611                                 pte_chain->parent_ptes[i] = parent_pte;
 612                                 return;
 613                         }
 614         }
 615         pte_chain = mmu_alloc_pte_chain(vcpu);
 616         BUG_ON(!pte_chain);
 617         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 618         pte_chain->parent_ptes[0] = parent_pte;
 619 }
 620
 621 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 622                                        u64 *parent_pte)
 623 {
 624         struct kvm_pte_chain *pte_chain;
 625         struct hlist_node *node;
 626         int i;
 627
 628         if (!sp->multimapped) {
 629                 BUG_ON(sp->parent_pte != parent_pte);
 630                 sp->parent_pte = NULL;
 631                 return;
 632         }
 633         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 634                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 635                         if (!pte_chain->parent_ptes[i])
 636                                 break;
 637                         if (pte_chain->parent_ptes[i] != parent_pte)
 638                                 continue;
 639                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 640                                 && pte_chain->parent_ptes[i + 1]) {
 641                                 pte_chain->parent_ptes[i]
 642                                         = pte_chain->parent_ptes[i + 1];
 643                                 ++i;
 644                         }
 645                         pte_chain->parent_ptes[i] = NULL;
 646                         if (i == 0) {
 647                                 hlist_del(&pte_chain->link);
 648                                 mmu_free_pte_chain(pte_chain);
 649                                 if (hlist_empty(&sp->parent_ptes)) {
 650                                         sp->multimapped = 0;
 651                                         sp->parent_pte = NULL;
 652                                 }
 653                         }
 654                         return;
 655                 }
 656         BUG();
 657 }
 658
 659 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
 660 {
 661         unsigned index;
 662         struct hlist_head *bucket;
 663         struct kvm_mmu_page *sp;
 664         struct hlist_node *node;
 665
 666         pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
 667         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
 668         bucket = &kvm->arch.mmu_page_hash[index];
 669         hlist_for_each_entry(sp, node, bucket, hash_link)
 670                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 671                         pgprintk("%s: found role %x\n",
 672                                  __FUNCTION__, sp->role.word);
 673                         return sp;
 674                 }
 675         return NULL;
 676 }
 677
 678 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 679                                              gfn_t gfn,
 680                                              gva_t gaddr,
 681                                              unsigned level,
 682                                              int metaphysical,
 683                                              unsigned access,
 684                                              u64 *parent_pte)
 685 {
 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) % KVM_NUM_MMU_PAGES;
 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) % KVM_NUM_MMU_PAGES;
 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 = is_rmap_pte(*shadow_pte);
 897         int was_writeble = is_writeble_pte(*shadow_pte);
 898
 899         pgprintk("%s: spte %llx access %x write_fault %d"
 900                  " user_fault %d gfn %lx\n",
 901                  __FUNCTION__, *shadow_pte, pt_access,
 902                  write_fault, user_fault, gfn);
 903
 904         /*
 905          * We don't set the accessed bit, since we sometimes want to see
 906          * whether the guest actually used the pte (in order to detect
 907          * demand paging).
 908          */
 909         spte = PT_PRESENT_MASK | PT_DIRTY_MASK;
 910         if (!dirty)
 911                 pte_access &= ~ACC_WRITE_MASK;
 912         if (!(pte_access & ACC_EXEC_MASK))
 913                 spte |= PT64_NX_MASK;
 914
 915         spte |= PT_PRESENT_MASK;
 916         if (pte_access & ACC_USER_MASK)
 917                 spte |= PT_USER_MASK;
 918
 919         if (is_error_page(page)) {
 920                 set_shadow_pte(shadow_pte,
 921                                shadow_trap_nonpresent_pte | PT_SHADOW_IO_MARK);
 922                 kvm_release_page_clean(page);
 923                 return;
 924         }
 925
 926         spte |= page_to_phys(page);
 927
 928         if ((pte_access & ACC_WRITE_MASK)
 929             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
 930                 struct kvm_mmu_page *shadow;
 931
 932                 spte |= PT_WRITABLE_MASK;
 933                 if (user_fault) {
 934                         mmu_unshadow(vcpu->kvm, gfn);
 935                         goto unshadowed;
 936                 }
 937
 938                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
 939                 if (shadow) {
 940                         pgprintk("%s: found shadow page for %lx, marking ro\n",
 941                                  __FUNCTION__, gfn);
 942                         pte_access &= ~ACC_WRITE_MASK;
 943                         if (is_writeble_pte(spte)) {
 944                                 spte &= ~PT_WRITABLE_MASK;
 945                                 kvm_x86_ops->tlb_flush(vcpu);
 946                         }
 947                         if (write_fault)
 948                                 *ptwrite = 1;
 949                 }
 950         }
 951
 952 unshadowed:
 953
 954         if (pte_access & ACC_WRITE_MASK)
 955                 mark_page_dirty(vcpu->kvm, gfn);
 956
 957         pgprintk("%s: setting spte %llx\n", __FUNCTION__, spte);
 958         set_shadow_pte(shadow_pte, spte);
 959         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
 960         if (!was_rmapped) {
 961                 rmap_add(vcpu, shadow_pte, gfn);
 962                 if (!is_rmap_pte(*shadow_pte))
 963                         kvm_release_page_clean(page);
 964         } else {
 965                 if (was_writeble)
 966                         kvm_release_page_dirty(page);
 967                 else
 968                         kvm_release_page_clean(page);
 969         }
 970         if (!ptwrite || !*ptwrite)
 971                 vcpu->arch.last_pte_updated = shadow_pte;
 972 }
 973
 974 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 975 {
 976 }
 977
 978 static int __nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write,
 979                            gfn_t gfn, struct page *page)
 980 {
 981         int level = PT32E_ROOT_LEVEL;
 982         hpa_t table_addr = vcpu->arch.mmu.root_hpa;
 983         int pt_write = 0;
 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                         if (!new_table) {
1008                                 pgprintk("nonpaging_map: ENOMEM\n");
1009                                 kvm_release_page_clean(page);
1010                                 return -ENOMEM;
1011                         }
1012
1013                         table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
1014                                 | PT_WRITABLE_MASK | PT_USER_MASK;
1015                 }
1016                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1017         }
1018 }
1019
1020 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1021 {
1022         int r;
1023
1024         struct page *page;
1025
1026         down_read(&vcpu->kvm->slots_lock);
1027
1028         down_read(&current->mm->mmap_sem);
1029         page = gfn_to_page(vcpu->kvm, gfn);
1030         up_read(&current->mm->mmap_sem);
1031
1032         spin_lock(&vcpu->kvm->mmu_lock);
1033         kvm_mmu_free_some_pages(vcpu);
1034         r = __nonpaging_map(vcpu, v, write, gfn, page);
1035         spin_unlock(&vcpu->kvm->mmu_lock);
1036
1037         up_read(&vcpu->kvm->slots_lock);
1038
1039         return r;
1040 }
1041
1042
1043 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1044                                     struct kvm_mmu_page *sp)
1045 {
1046         int i;
1047
1048         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1049                 sp->spt[i] = shadow_trap_nonpresent_pte;
1050 }
1051
1052 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1053 {
1054         int i;
1055         struct kvm_mmu_page *sp;
1056
1057         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1058                 return;
1059         spin_lock(&vcpu->kvm->mmu_lock);
1060 #ifdef CONFIG_X86_64
1061         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1062                 hpa_t root = vcpu->arch.mmu.root_hpa;
1063
1064                 sp = page_header(root);
1065                 --sp->root_count;
1066                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1067                 spin_unlock(&vcpu->kvm->mmu_lock);
1068                 return;
1069         }
1070 #endif
1071         for (i = 0; i < 4; ++i) {
1072                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1073
1074                 if (root) {
1075                         root &= PT64_BASE_ADDR_MASK;
1076                         sp = page_header(root);
1077                         --sp->root_count;
1078                 }
1079                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1080         }
1081         spin_unlock(&vcpu->kvm->mmu_lock);
1082         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1083 }
1084
1085 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1086 {
1087         int i;
1088         gfn_t root_gfn;
1089         struct kvm_mmu_page *sp;
1090
1091         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1092
1093 #ifdef CONFIG_X86_64
1094         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1095                 hpa_t root = vcpu->arch.mmu.root_hpa;
1096
1097                 ASSERT(!VALID_PAGE(root));
1098                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1099                                       PT64_ROOT_LEVEL, 0, ACC_ALL, NULL);
1100                 root = __pa(sp->spt);
1101                 ++sp->root_count;
1102                 vcpu->arch.mmu.root_hpa = root;
1103                 return;
1104         }
1105 #endif
1106         for (i = 0; i < 4; ++i) {
1107                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1108
1109                 ASSERT(!VALID_PAGE(root));
1110                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1111                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1112                                 vcpu->arch.mmu.pae_root[i] = 0;
1113                                 continue;
1114                         }
1115                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1116                 } else if (vcpu->arch.mmu.root_level == 0)
1117                         root_gfn = 0;
1118                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1119                                       PT32_ROOT_LEVEL, !is_paging(vcpu),
1120                                       ACC_ALL, NULL);
1121                 root = __pa(sp->spt);
1122                 ++sp->root_count;
1123                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1124         }
1125         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1126 }
1127
1128 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1129 {
1130         return vaddr;
1131 }
1132
1133 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1134                                 u32 error_code)
1135 {
1136         gfn_t gfn;
1137         int r;
1138
1139         pgprintk("%s: gva %lx error %x\n", __FUNCTION__, gva, error_code);
1140         r = mmu_topup_memory_caches(vcpu);
1141         if (r)
1142                 return r;
1143
1144         ASSERT(vcpu);
1145         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1146
1147         gfn = gva >> PAGE_SHIFT;
1148
1149         return nonpaging_map(vcpu, gva & PAGE_MASK,
1150                              error_code & PFERR_WRITE_MASK, gfn);
1151 }
1152
1153 static void nonpaging_free(struct kvm_vcpu *vcpu)
1154 {
1155         mmu_free_roots(vcpu);
1156 }
1157
1158 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1159 {
1160         struct kvm_mmu *context = &vcpu->arch.mmu;
1161
1162         context->new_cr3 = nonpaging_new_cr3;
1163         context->page_fault = nonpaging_page_fault;
1164         context->gva_to_gpa = nonpaging_gva_to_gpa;
1165         context->free = nonpaging_free;
1166         context->prefetch_page = nonpaging_prefetch_page;
1167         context->root_level = 0;
1168         context->shadow_root_level = PT32E_ROOT_LEVEL;
1169         context->root_hpa = INVALID_PAGE;
1170         return 0;
1171 }
1172
1173 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1174 {
1175         ++vcpu->stat.tlb_flush;
1176         kvm_x86_ops->tlb_flush(vcpu);
1177 }
1178
1179 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1180 {
1181         pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->arch.cr3);
1182         mmu_free_roots(vcpu);
1183 }
1184
1185 static void inject_page_fault(struct kvm_vcpu *vcpu,
1186                               u64 addr,
1187                               u32 err_code)
1188 {
1189         kvm_inject_page_fault(vcpu, addr, err_code);
1190 }
1191
1192 static void paging_free(struct kvm_vcpu *vcpu)
1193 {
1194         nonpaging_free(vcpu);
1195 }
1196
1197 #define PTTYPE 64
1198 #include "paging_tmpl.h"
1199 #undef PTTYPE
1200
1201 #define PTTYPE 32
1202 #include "paging_tmpl.h"
1203 #undef PTTYPE
1204
1205 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1206 {
1207         struct kvm_mmu *context = &vcpu->arch.mmu;
1208
1209         ASSERT(is_pae(vcpu));
1210         context->new_cr3 = paging_new_cr3;
1211         context->page_fault = paging64_page_fault;
1212         context->gva_to_gpa = paging64_gva_to_gpa;
1213         context->prefetch_page = paging64_prefetch_page;
1214         context->free = paging_free;
1215         context->root_level = level;
1216         context->shadow_root_level = level;
1217         context->root_hpa = INVALID_PAGE;
1218         return 0;
1219 }
1220
1221 static int paging64_init_context(struct kvm_vcpu *vcpu)
1222 {
1223         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1224 }
1225
1226 static int paging32_init_context(struct kvm_vcpu *vcpu)
1227 {
1228         struct kvm_mmu *context = &vcpu->arch.mmu;
1229
1230         context->new_cr3 = paging_new_cr3;
1231         context->page_fault = paging32_page_fault;
1232         context->gva_to_gpa = paging32_gva_to_gpa;
1233         context->free = paging_free;
1234         context->prefetch_page = paging32_prefetch_page;
1235         context->root_level = PT32_ROOT_LEVEL;
1236         context->shadow_root_level = PT32E_ROOT_LEVEL;
1237         context->root_hpa = INVALID_PAGE;
1238         return 0;
1239 }
1240
1241 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1242 {
1243         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1244 }
1245
1246 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1247 {
1248         ASSERT(vcpu);
1249         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1250
1251         if (!is_paging(vcpu))
1252                 return nonpaging_init_context(vcpu);
1253         else if (is_long_mode(vcpu))
1254                 return paging64_init_context(vcpu);
1255         else if (is_pae(vcpu))
1256                 return paging32E_init_context(vcpu);
1257         else
1258                 return paging32_init_context(vcpu);
1259 }
1260
1261 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1262 {
1263         ASSERT(vcpu);
1264         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1265                 vcpu->arch.mmu.free(vcpu);
1266                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1267         }
1268 }
1269
1270 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1271 {
1272         destroy_kvm_mmu(vcpu);
1273         return init_kvm_mmu(vcpu);
1274 }
1275 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1276
1277 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1278 {
1279         int r;
1280
1281         r = mmu_topup_memory_caches(vcpu);
1282         if (r)
1283                 goto out;
1284         spin_lock(&vcpu->kvm->mmu_lock);
1285         kvm_mmu_free_some_pages(vcpu);
1286         mmu_alloc_roots(vcpu);
1287         spin_unlock(&vcpu->kvm->mmu_lock);
1288         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1289         kvm_mmu_flush_tlb(vcpu);
1290 out:
1291         return r;
1292 }
1293 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1294
1295 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1296 {
1297         mmu_free_roots(vcpu);
1298 }
1299
1300 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1301                                   struct kvm_mmu_page *sp,
1302                                   u64 *spte)
1303 {
1304         u64 pte;
1305         struct kvm_mmu_page *child;
1306
1307         pte = *spte;
1308         if (is_shadow_present_pte(pte)) {
1309                 if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1310                         rmap_remove(vcpu->kvm, spte);
1311                 else {
1312                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1313                         mmu_page_remove_parent_pte(child, spte);
1314                 }
1315         }
1316         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1317 }
1318
1319 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1320                                   struct kvm_mmu_page *sp,
1321                                   u64 *spte,
1322                                   const void *new, int bytes,
1323                                   int offset_in_pte)
1324 {
1325         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1326                 ++vcpu->kvm->stat.mmu_pde_zapped;
1327                 return;
1328         }
1329
1330         ++vcpu->kvm->stat.mmu_pte_updated;
1331         if (sp->role.glevels == PT32_ROOT_LEVEL)
1332                 paging32_update_pte(vcpu, sp, spte, new, bytes, offset_in_pte);
1333         else
1334                 paging64_update_pte(vcpu, sp, spte, new, bytes, offset_in_pte);
1335 }
1336
1337 static bool need_remote_flush(u64 old, u64 new)
1338 {
1339         if (!is_shadow_present_pte(old))
1340                 return false;
1341         if (!is_shadow_present_pte(new))
1342                 return true;
1343         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1344                 return true;
1345         old ^= PT64_NX_MASK;
1346         new ^= PT64_NX_MASK;
1347         return (old & ~new & PT64_PERM_MASK) != 0;
1348 }
1349
1350 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1351 {
1352         if (need_remote_flush(old, new))
1353                 kvm_flush_remote_tlbs(vcpu->kvm);
1354         else
1355                 kvm_mmu_flush_tlb(vcpu);
1356 }
1357
1358 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1359 {
1360         u64 *spte = vcpu->arch.last_pte_updated;
1361
1362         return !!(spte && (*spte & PT_ACCESSED_MASK));
1363 }
1364
1365 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1366                                           const u8 *new, int bytes)
1367 {
1368         gfn_t gfn;
1369         int r;
1370         u64 gpte = 0;
1371         struct page *page;
1372
1373         if (bytes != 4 && bytes != 8)
1374                 return;
1375
1376         /*
1377          * Assume that the pte write on a page table of the same type
1378          * as the current vcpu paging mode.  This is nearly always true
1379          * (might be false while changing modes).  Note it is verified later
1380          * by update_pte().
1381          */
1382         if (is_pae(vcpu)) {
1383                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1384                 if ((bytes == 4) && (gpa % 4 == 0)) {
1385                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1386                         if (r)
1387                                 return;
1388                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1389                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1390                         memcpy((void *)&gpte, new, 8);
1391                 }
1392         } else {
1393                 if ((bytes == 4) && (gpa % 4 == 0))
1394                         memcpy((void *)&gpte, new, 4);
1395         }
1396         if (!is_present_pte(gpte))
1397                 return;
1398         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1399
1400         down_read(&current->mm->mmap_sem);
1401         page = gfn_to_page(vcpu->kvm, gfn);
1402         up_read(&current->mm->mmap_sem);
1403
1404         vcpu->arch.update_pte.gfn = gfn;
1405         vcpu->arch.update_pte.page = gfn_to_page(vcpu->kvm, gfn);
1406 }
1407
1408 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1409                        const u8 *new, int bytes)
1410 {
1411         gfn_t gfn = gpa >> PAGE_SHIFT;
1412         struct kvm_mmu_page *sp;
1413         struct hlist_node *node, *n;
1414         struct hlist_head *bucket;
1415         unsigned index;
1416         u64 entry;
1417         u64 *spte;
1418         unsigned offset = offset_in_page(gpa);
1419         unsigned pte_size;
1420         unsigned page_offset;
1421         unsigned misaligned;
1422         unsigned quadrant;
1423         int level;
1424         int flooded = 0;
1425         int npte;
1426
1427         pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1428         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1429         spin_lock(&vcpu->kvm->mmu_lock);
1430         kvm_mmu_free_some_pages(vcpu);
1431         ++vcpu->kvm->stat.mmu_pte_write;
1432         kvm_mmu_audit(vcpu, "pre pte write");
1433         if (gfn == vcpu->arch.last_pt_write_gfn
1434             && !last_updated_pte_accessed(vcpu)) {
1435                 ++vcpu->arch.last_pt_write_count;
1436                 if (vcpu->arch.last_pt_write_count >= 3)
1437                         flooded = 1;
1438         } else {
1439                 vcpu->arch.last_pt_write_gfn = gfn;
1440                 vcpu->arch.last_pt_write_count = 1;
1441                 vcpu->arch.last_pte_updated = NULL;
1442         }
1443         index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1444         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1445         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1446                 if (sp->gfn != gfn || sp->role.metaphysical)
1447                         continue;
1448                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1449                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1450                 misaligned |= bytes < 4;
1451                 if (misaligned || flooded) {
1452                         /*
1453                          * Misaligned accesses are too much trouble to fix
1454                          * up; also, they usually indicate a page is not used
1455                          * as a page table.
1456                          *
1457                          * If we're seeing too many writes to a page,
1458                          * it may no longer be a page table, or we may be
1459                          * forking, in which case it is better to unmap the
1460                          * page.
1461                          */
1462                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1463                                  gpa, bytes, sp->role.word);
1464                         kvm_mmu_zap_page(vcpu->kvm, sp);
1465                         ++vcpu->kvm->stat.mmu_flooded;
1466                         continue;
1467                 }
1468                 page_offset = offset;
1469                 level = sp->role.level;
1470                 npte = 1;
1471                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1472                         page_offset <<= 1;      /* 32->64 */
1473                         /*
1474                          * A 32-bit pde maps 4MB while the shadow pdes map
1475                          * only 2MB.  So we need to double the offset again
1476                          * and zap two pdes instead of one.
1477                          */
1478                         if (level == PT32_ROOT_LEVEL) {
1479                                 page_offset &= ~7; /* kill rounding error */
1480                                 page_offset <<= 1;
1481                                 npte = 2;
1482                         }
1483                         quadrant = page_offset >> PAGE_SHIFT;
1484                         page_offset &= ~PAGE_MASK;
1485                         if (quadrant != sp->role.quadrant)
1486                                 continue;
1487                 }
1488                 spte = &sp->spt[page_offset / sizeof(*spte)];
1489                 while (npte--) {
1490                         entry = *spte;
1491                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1492                         mmu_pte_write_new_pte(vcpu, sp, spte, new, bytes,
1493                                               page_offset & (pte_size - 1));
1494                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1495                         ++spte;
1496                 }
1497         }
1498         kvm_mmu_audit(vcpu, "post pte write");
1499         spin_unlock(&vcpu->kvm->mmu_lock);
1500         if (vcpu->arch.update_pte.page) {
1501                 kvm_release_page_clean(vcpu->arch.update_pte.page);
1502                 vcpu->arch.update_pte.page = NULL;
1503         }
1504 }
1505
1506 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1507 {
1508         gpa_t gpa;
1509         int r;
1510
1511         down_read(&vcpu->kvm->slots_lock);
1512         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1513         up_read(&vcpu->kvm->slots_lock);
1514
1515         spin_lock(&vcpu->kvm->mmu_lock);
1516         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1517         spin_unlock(&vcpu->kvm->mmu_lock);
1518         return r;
1519 }
1520
1521 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1522 {
1523         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1524                 struct kvm_mmu_page *sp;
1525
1526                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1527                                   struct kvm_mmu_page, link);
1528                 kvm_mmu_zap_page(vcpu->kvm, sp);
1529                 ++vcpu->kvm->stat.mmu_recycled;
1530         }
1531 }
1532
1533 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1534 {
1535         int r;
1536         enum emulation_result er;
1537
1538         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1539         if (r < 0)
1540                 goto out;
1541
1542         if (!r) {
1543                 r = 1;
1544                 goto out;
1545         }
1546
1547         r = mmu_topup_memory_caches(vcpu);
1548         if (r)
1549                 goto out;
1550
1551         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1552
1553         switch (er) {
1554         case EMULATE_DONE:
1555                 return 1;
1556         case EMULATE_DO_MMIO:
1557                 ++vcpu->stat.mmio_exits;
1558                 return 0;
1559         case EMULATE_FAIL:
1560                 kvm_report_emulation_failure(vcpu, "pagetable");
1561                 return 1;
1562         default:
1563                 BUG();
1564         }
1565 out:
1566         return r;
1567 }
1568 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1569
1570 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1571 {
1572         struct kvm_mmu_page *sp;
1573
1574         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1575                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1576                                   struct kvm_mmu_page, link);
1577                 kvm_mmu_zap_page(vcpu->kvm, sp);
1578         }
1579         free_page((unsigned long)vcpu->arch.mmu.pae_root);
1580 }
1581
1582 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1583 {
1584         struct page *page;
1585         int i;
1586
1587         ASSERT(vcpu);
1588
1589         if (vcpu->kvm->arch.n_requested_mmu_pages)
1590                 vcpu->kvm->arch.n_free_mmu_pages =
1591                                         vcpu->kvm->arch.n_requested_mmu_pages;
1592         else
1593                 vcpu->kvm->arch.n_free_mmu_pages =
1594                                         vcpu->kvm->arch.n_alloc_mmu_pages;
1595         /*
1596          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1597          * Therefore we need to allocate shadow page tables in the first
1598          * 4GB of memory, which happens to fit the DMA32 zone.
1599          */
1600         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1601         if (!page)
1602                 goto error_1;
1603         vcpu->arch.mmu.pae_root = page_address(page);
1604         for (i = 0; i < 4; ++i)
1605                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1606
1607         return 0;
1608
1609 error_1:
1610         free_mmu_pages(vcpu);
1611         return -ENOMEM;
1612 }
1613
1614 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1615 {
1616         ASSERT(vcpu);
1617         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1618
1619         return alloc_mmu_pages(vcpu);
1620 }
1621
1622 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1623 {
1624         ASSERT(vcpu);
1625         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1626
1627         return init_kvm_mmu(vcpu);
1628 }
1629
1630 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1631 {
1632         ASSERT(vcpu);
1633
1634         destroy_kvm_mmu(vcpu);
1635         free_mmu_pages(vcpu);
1636         mmu_free_memory_caches(vcpu);
1637 }
1638
1639 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1640 {
1641         struct kvm_mmu_page *sp;
1642
1643         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1644                 int i;
1645                 u64 *pt;
1646
1647                 if (!test_bit(slot, &sp->slot_bitmap))
1648                         continue;
1649
1650                 pt = sp->spt;
1651                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1652                         /* avoid RMW */
1653                         if (pt[i] & PT_WRITABLE_MASK)
1654                                 pt[i] &= ~PT_WRITABLE_MASK;
1655         }
1656 }
1657
1658 void kvm_mmu_zap_all(struct kvm *kvm)
1659 {
1660         struct kvm_mmu_page *sp, *node;
1661
1662         spin_lock(&kvm->mmu_lock);
1663         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1664                 kvm_mmu_zap_page(kvm, sp);
1665         spin_unlock(&kvm->mmu_lock);
1666
1667         kvm_flush_remote_tlbs(kvm);
1668 }
1669
1670 void kvm_mmu_module_exit(void)
1671 {
1672         if (pte_chain_cache)
1673                 kmem_cache_destroy(pte_chain_cache);
1674         if (rmap_desc_cache)
1675                 kmem_cache_destroy(rmap_desc_cache);
1676         if (mmu_page_header_cache)
1677                 kmem_cache_destroy(mmu_page_header_cache);
1678 }
1679
1680 int kvm_mmu_module_init(void)
1681 {
1682         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1683                                             sizeof(struct kvm_pte_chain),
1684                                             0, 0, NULL);
1685         if (!pte_chain_cache)
1686                 goto nomem;
1687         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1688                                             sizeof(struct kvm_rmap_desc),
1689                                             0, 0, NULL);
1690         if (!rmap_desc_cache)
1691                 goto nomem;
1692
1693         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1694                                                   sizeof(struct kvm_mmu_page),
1695                                                   0, 0, NULL);
1696         if (!mmu_page_header_cache)
1697                 goto nomem;
1698
1699         return 0;
1700
1701 nomem:
1702         kvm_mmu_module_exit();
1703         return -ENOMEM;
1704 }
1705
1706 /*
1707  * Caculate mmu pages needed for kvm.
1708  */
1709 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
1710 {
1711         int i;
1712         unsigned int nr_mmu_pages;
1713         unsigned int  nr_pages = 0;
1714
1715         for (i = 0; i < kvm->nmemslots; i++)
1716                 nr_pages += kvm->memslots[i].npages;
1717
1718         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
1719         nr_mmu_pages = max(nr_mmu_pages,
1720                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
1721
1722         return nr_mmu_pages;
1723 }
1724
1725 #ifdef AUDIT
1726
1727 static const char *audit_msg;
1728
1729 static gva_t canonicalize(gva_t gva)
1730 {
1731 #ifdef CONFIG_X86_64
1732         gva = (long long)(gva << 16) >> 16;
1733 #endif
1734         return gva;
1735 }
1736
1737 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1738                                 gva_t va, int level)
1739 {
1740         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1741         int i;
1742         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1743
1744         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1745                 u64 ent = pt[i];
1746
1747                 if (ent == shadow_trap_nonpresent_pte)
1748                         continue;
1749
1750                 va = canonicalize(va);
1751                 if (level > 1) {
1752                         if (ent == shadow_notrap_nonpresent_pte)
1753                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
1754                                        " in nonleaf level: levels %d gva %lx"
1755                                        " level %d pte %llx\n", audit_msg,
1756                                        vcpu->arch.mmu.root_level, va, level, ent);
1757
1758                         audit_mappings_page(vcpu, ent, va, level - 1);
1759                 } else {
1760                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
1761                         struct page *page = gpa_to_page(vcpu, gpa);
1762                         hpa_t hpa = page_to_phys(page);
1763
1764                         if (is_shadow_present_pte(ent)
1765                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
1766                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
1767                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1768                                        audit_msg, vcpu->arch.mmu.root_level,
1769                                        va, gpa, hpa, ent,
1770                                        is_shadow_present_pte(ent));
1771                         else if (ent == shadow_notrap_nonpresent_pte
1772                                  && !is_error_hpa(hpa))
1773                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
1774                                        " valid guest gva %lx\n", audit_msg, va);
1775                         kvm_release_page_clean(page);
1776
1777                 }
1778         }
1779 }
1780
1781 static void audit_mappings(struct kvm_vcpu *vcpu)
1782 {
1783         unsigned i;
1784
1785         if (vcpu->arch.mmu.root_level == 4)
1786                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
1787         else
1788                 for (i = 0; i < 4; ++i)
1789                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
1790                                 audit_mappings_page(vcpu,
1791                                                     vcpu->arch.mmu.pae_root[i],
1792                                                     i << 30,
1793                                                     2);
1794 }
1795
1796 static int count_rmaps(struct kvm_vcpu *vcpu)
1797 {
1798         int nmaps = 0;
1799         int i, j, k;
1800
1801         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1802                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1803                 struct kvm_rmap_desc *d;
1804
1805                 for (j = 0; j < m->npages; ++j) {
1806                         unsigned long *rmapp = &m->rmap[j];
1807
1808                         if (!*rmapp)
1809                                 continue;
1810                         if (!(*rmapp & 1)) {
1811                                 ++nmaps;
1812                                 continue;
1813                         }
1814                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1815                         while (d) {
1816                                 for (k = 0; k < RMAP_EXT; ++k)
1817                                         if (d->shadow_ptes[k])
1818                                                 ++nmaps;
1819                                         else
1820                                                 break;
1821                                 d = d->more;
1822                         }
1823                 }
1824         }
1825         return nmaps;
1826 }
1827
1828 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1829 {
1830         int nmaps = 0;
1831         struct kvm_mmu_page *sp;
1832         int i;
1833
1834         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
1835                 u64 *pt = sp->spt;
1836
1837                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
1838                         continue;
1839
1840                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1841                         u64 ent = pt[i];
1842
1843                         if (!(ent & PT_PRESENT_MASK))
1844                                 continue;
1845                         if (!(ent & PT_WRITABLE_MASK))
1846                                 continue;
1847                         ++nmaps;
1848                 }
1849         }
1850         return nmaps;
1851 }
1852
1853 static void audit_rmap(struct kvm_vcpu *vcpu)
1854 {
1855         int n_rmap = count_rmaps(vcpu);
1856         int n_actual = count_writable_mappings(vcpu);
1857
1858         if (n_rmap != n_actual)
1859                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1860                        __FUNCTION__, audit_msg, n_rmap, n_actual);
1861 }
1862
1863 static void audit_write_protection(struct kvm_vcpu *vcpu)
1864 {
1865         struct kvm_mmu_page *sp;
1866         struct kvm_memory_slot *slot;
1867         unsigned long *rmapp;
1868         gfn_t gfn;
1869
1870         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
1871                 if (sp->role.metaphysical)
1872                         continue;
1873
1874                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
1875                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
1876                 rmapp = &slot->rmap[gfn - slot->base_gfn];
1877                 if (*rmapp)
1878                         printk(KERN_ERR "%s: (%s) shadow page has writable"
1879                                " mappings: gfn %lx role %x\n",
1880                                __FUNCTION__, audit_msg, sp->gfn,
1881                                sp->role.word);
1882         }
1883 }
1884
1885 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1886 {
1887         int olddbg = dbg;
1888
1889         dbg = 0;
1890         audit_msg = msg;
1891         audit_rmap(vcpu);
1892         audit_write_protection(vcpu);
1893         audit_mappings(vcpu);
1894         dbg = olddbg;
1895 }
1896
1897 #endif