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 "mmu.h"
  21
  22 #include <linux/kvm_host.h>
  23 #include <linux/types.h>
  24 #include <linux/string.h>
  25 #include <linux/mm.h>
  26 #include <linux/highmem.h>
  27 #include <linux/module.h>
  28 #include <linux/swap.h>
  29 #include <linux/hugetlb.h>
  30 #include <linux/compiler.h>
  31
  32 #include <asm/page.h>
  33 #include <asm/cmpxchg.h>
  34 #include <asm/io.h>
  35 #include <asm/vmx.h>
  36
  37 /*
  38  * When setting this variable to true it enables Two-Dimensional-Paging
  39  * where the hardware walks 2 page tables:
  40  * 1. the guest-virtual to guest-physical
  41  * 2. while doing 1. it walks guest-physical to host-physical
  42  * If the hardware supports that we don't need to do shadow paging.
  43  */
  44 bool tdp_enabled = false;
  45
  46 #undef MMU_DEBUG
  47
  48 #undef AUDIT
  49
  50 #ifdef AUDIT
  51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
  52 #else
  53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
  54 #endif
  55
  56 #ifdef MMU_DEBUG
  57
  58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
  59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
  60
  61 #else
  62
  63 #define pgprintk(x...) do { } while (0)
  64 #define rmap_printk(x...) do { } while (0)
  65
  66 #endif
  67
  68 #if defined(MMU_DEBUG) || defined(AUDIT)
  69 static int dbg = 0;
  70 module_param(dbg, bool, 0644);
  71 #endif
  72
  73 static int oos_shadow = 1;
  74 module_param(oos_shadow, bool, 0644);
  75
  76 #ifndef MMU_DEBUG
  77 #define ASSERT(x) do { } while (0)
  78 #else
  79 #define ASSERT(x)                                                       \
  80         if (!(x)) {                                                     \
  81                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  82                        __FILE__, __LINE__, #x);                         \
  83         }
  84 #endif
  85
  86 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  87 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  88
  89 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  90
  91 #define PT64_LEVEL_BITS 9
  92
  93 #define PT64_LEVEL_SHIFT(level) \
  94                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  95
  96 #define PT64_LEVEL_MASK(level) \
  97                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  98
  99 #define PT64_INDEX(address, level)\
 100         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
 101
 102
 103 #define PT32_LEVEL_BITS 10
 104
 105 #define PT32_LEVEL_SHIFT(level) \
 106                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 107
 108 #define PT32_LEVEL_MASK(level) \
 109                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 110
 111 #define PT32_INDEX(address, level)\
 112         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 113
 114
 115 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 116 #define PT64_DIR_BASE_ADDR_MASK \
 117         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 118
 119 #define PT32_BASE_ADDR_MASK PAGE_MASK
 120 #define PT32_DIR_BASE_ADDR_MASK \
 121         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 122
 123 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 124                         | PT64_NX_MASK)
 125
 126 #define PFERR_PRESENT_MASK (1U << 0)
 127 #define PFERR_WRITE_MASK (1U << 1)
 128 #define PFERR_USER_MASK (1U << 2)
 129 #define PFERR_FETCH_MASK (1U << 4)
 130
 131 #define PT_DIRECTORY_LEVEL 2
 132 #define PT_PAGE_TABLE_LEVEL 1
 133
 134 #define RMAP_EXT 4
 135
 136 #define ACC_EXEC_MASK    1
 137 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 138 #define ACC_USER_MASK    PT_USER_MASK
 139 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 140
 141 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 142
 143 struct kvm_rmap_desc {
 144         u64 *shadow_ptes[RMAP_EXT];
 145         struct kvm_rmap_desc *more;
 146 };
 147
 148 struct kvm_shadow_walk {
 149         int (*entry)(struct kvm_shadow_walk *walk, struct kvm_vcpu *vcpu,
 150                      u64 addr, u64 *spte, int level);
 151 };
 152
 153 struct kvm_unsync_walk {
 154         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 155 };
 156
 157 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 158
 159 static struct kmem_cache *pte_chain_cache;
 160 static struct kmem_cache *rmap_desc_cache;
 161 static struct kmem_cache *mmu_page_header_cache;
 162
 163 static u64 __read_mostly shadow_trap_nonpresent_pte;
 164 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 165 static u64 __read_mostly shadow_base_present_pte;
 166 static u64 __read_mostly shadow_nx_mask;
 167 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 168 static u64 __read_mostly shadow_user_mask;
 169 static u64 __read_mostly shadow_accessed_mask;
 170 static u64 __read_mostly shadow_dirty_mask;
 171 static u64 __read_mostly shadow_mt_mask;
 172
 173 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 174 {
 175         shadow_trap_nonpresent_pte = trap_pte;
 176         shadow_notrap_nonpresent_pte = notrap_pte;
 177 }
 178 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 179
 180 void kvm_mmu_set_base_ptes(u64 base_pte)
 181 {
 182         shadow_base_present_pte = base_pte;
 183 }
 184 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 185
 186 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 187                 u64 dirty_mask, u64 nx_mask, u64 x_mask, u64 mt_mask)
 188 {
 189         shadow_user_mask = user_mask;
 190         shadow_accessed_mask = accessed_mask;
 191         shadow_dirty_mask = dirty_mask;
 192         shadow_nx_mask = nx_mask;
 193         shadow_x_mask = x_mask;
 194         shadow_mt_mask = mt_mask;
 195 }
 196 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 197
 198 static int is_write_protection(struct kvm_vcpu *vcpu)
 199 {
 200         return vcpu->arch.cr0 & X86_CR0_WP;
 201 }
 202
 203 static int is_cpuid_PSE36(void)
 204 {
 205         return 1;
 206 }
 207
 208 static int is_nx(struct kvm_vcpu *vcpu)
 209 {
 210         return vcpu->arch.shadow_efer & EFER_NX;
 211 }
 212
 213 static int is_present_pte(unsigned long pte)
 214 {
 215         return pte & PT_PRESENT_MASK;
 216 }
 217
 218 static int is_shadow_present_pte(u64 pte)
 219 {
 220         return pte != shadow_trap_nonpresent_pte
 221                 && pte != shadow_notrap_nonpresent_pte;
 222 }
 223
 224 static int is_large_pte(u64 pte)
 225 {
 226         return pte & PT_PAGE_SIZE_MASK;
 227 }
 228
 229 static int is_writeble_pte(unsigned long pte)
 230 {
 231         return pte & PT_WRITABLE_MASK;
 232 }
 233
 234 static int is_dirty_pte(unsigned long pte)
 235 {
 236         return pte & shadow_dirty_mask;
 237 }
 238
 239 static int is_rmap_pte(u64 pte)
 240 {
 241         return is_shadow_present_pte(pte);
 242 }
 243
 244 static pfn_t spte_to_pfn(u64 pte)
 245 {
 246         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 247 }
 248
 249 static gfn_t pse36_gfn_delta(u32 gpte)
 250 {
 251         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 252
 253         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 254 }
 255
 256 static void set_shadow_pte(u64 *sptep, u64 spte)
 257 {
 258 #ifdef CONFIG_X86_64
 259         set_64bit((unsigned long *)sptep, spte);
 260 #else
 261         set_64bit((unsigned long long *)sptep, spte);
 262 #endif
 263 }
 264
 265 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 266                                   struct kmem_cache *base_cache, int min)
 267 {
 268         void *obj;
 269
 270         if (cache->nobjs >= min)
 271                 return 0;
 272         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 273                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 274                 if (!obj)
 275                         return -ENOMEM;
 276                 cache->objects[cache->nobjs++] = obj;
 277         }
 278         return 0;
 279 }
 280
 281 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 282 {
 283         while (mc->nobjs)
 284                 kfree(mc->objects[--mc->nobjs]);
 285 }
 286
 287 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 288                                        int min)
 289 {
 290         struct page *page;
 291
 292         if (cache->nobjs >= min)
 293                 return 0;
 294         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 295                 page = alloc_page(GFP_KERNEL);
 296                 if (!page)
 297                         return -ENOMEM;
 298                 set_page_private(page, 0);
 299                 cache->objects[cache->nobjs++] = page_address(page);
 300         }
 301         return 0;
 302 }
 303
 304 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 305 {
 306         while (mc->nobjs)
 307                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 308 }
 309
 310 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 311 {
 312         int r;
 313
 314         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 315                                    pte_chain_cache, 4);
 316         if (r)
 317                 goto out;
 318         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 319                                    rmap_desc_cache, 4);
 320         if (r)
 321                 goto out;
 322         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 323         if (r)
 324                 goto out;
 325         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 326                                    mmu_page_header_cache, 4);
 327 out:
 328         return r;
 329 }
 330
 331 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 332 {
 333         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 334         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 335         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 336         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 337 }
 338
 339 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 340                                     size_t size)
 341 {
 342         void *p;
 343
 344         BUG_ON(!mc->nobjs);
 345         p = mc->objects[--mc->nobjs];
 346         memset(p, 0, size);
 347         return p;
 348 }
 349
 350 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 351 {
 352         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 353                                       sizeof(struct kvm_pte_chain));
 354 }
 355
 356 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 357 {
 358         kfree(pc);
 359 }
 360
 361 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 362 {
 363         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 364                                       sizeof(struct kvm_rmap_desc));
 365 }
 366
 367 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 368 {
 369         kfree(rd);
 370 }
 371
 372 /*
 373  * Return the pointer to the largepage write count for a given
 374  * gfn, handling slots that are not large page aligned.
 375  */
 376 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 377 {
 378         unsigned long idx;
 379
 380         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 381               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 382         return &slot->lpage_info[idx].write_count;
 383 }
 384
 385 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 386 {
 387         int *write_count;
 388
 389         gfn = unalias_gfn(kvm, gfn);
 390         write_count = slot_largepage_idx(gfn,
 391                                          gfn_to_memslot_unaliased(kvm, gfn));
 392         *write_count += 1;
 393 }
 394
 395 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 396 {
 397         int *write_count;
 398
 399         gfn = unalias_gfn(kvm, gfn);
 400         write_count = slot_largepage_idx(gfn,
 401                                          gfn_to_memslot_unaliased(kvm, gfn));
 402         *write_count -= 1;
 403         WARN_ON(*write_count < 0);
 404 }
 405
 406 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 407 {
 408         struct kvm_memory_slot *slot;
 409         int *largepage_idx;
 410
 411         gfn = unalias_gfn(kvm, gfn);
 412         slot = gfn_to_memslot_unaliased(kvm, gfn);
 413         if (slot) {
 414                 largepage_idx = slot_largepage_idx(gfn, slot);
 415                 return *largepage_idx;
 416         }
 417
 418         return 1;
 419 }
 420
 421 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 422 {
 423         struct vm_area_struct *vma;
 424         unsigned long addr;
 425         int ret = 0;
 426
 427         addr = gfn_to_hva(kvm, gfn);
 428         if (kvm_is_error_hva(addr))
 429                 return ret;
 430
 431         down_read(&current->mm->mmap_sem);
 432         vma = find_vma(current->mm, addr);
 433         if (vma && is_vm_hugetlb_page(vma))
 434                 ret = 1;
 435         up_read(&current->mm->mmap_sem);
 436
 437         return ret;
 438 }
 439
 440 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 441 {
 442         struct kvm_memory_slot *slot;
 443
 444         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 445                 return 0;
 446
 447         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 448                 return 0;
 449
 450         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 451         if (slot && slot->dirty_bitmap)
 452                 return 0;
 453
 454         return 1;
 455 }
 456
 457 /*
 458  * Take gfn and return the reverse mapping to it.
 459  * Note: gfn must be unaliased before this function get called
 460  */
 461
 462 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 463 {
 464         struct kvm_memory_slot *slot;
 465         unsigned long idx;
 466
 467         slot = gfn_to_memslot(kvm, gfn);
 468         if (!lpage)
 469                 return &slot->rmap[gfn - slot->base_gfn];
 470
 471         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 472               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 473
 474         return &slot->lpage_info[idx].rmap_pde;
 475 }
 476
 477 /*
 478  * Reverse mapping data structures:
 479  *
 480  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 481  * that points to page_address(page).
 482  *
 483  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 484  * containing more mappings.
 485  */
 486 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 487 {
 488         struct kvm_mmu_page *sp;
 489         struct kvm_rmap_desc *desc;
 490         unsigned long *rmapp;
 491         int i;
 492
 493         if (!is_rmap_pte(*spte))
 494                 return;
 495         gfn = unalias_gfn(vcpu->kvm, gfn);
 496         sp = page_header(__pa(spte));
 497         sp->gfns[spte - sp->spt] = gfn;
 498         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 499         if (!*rmapp) {
 500                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 501                 *rmapp = (unsigned long)spte;
 502         } else if (!(*rmapp & 1)) {
 503                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 504                 desc = mmu_alloc_rmap_desc(vcpu);
 505                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 506                 desc->shadow_ptes[1] = spte;
 507                 *rmapp = (unsigned long)desc | 1;
 508         } else {
 509                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 510                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 511                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 512                         desc = desc->more;
 513                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 514                         desc->more = mmu_alloc_rmap_desc(vcpu);
 515                         desc = desc->more;
 516                 }
 517                 for (i = 0; desc->shadow_ptes[i]; ++i)
 518                         ;
 519                 desc->shadow_ptes[i] = spte;
 520         }
 521 }
 522
 523 static void rmap_desc_remove_entry(unsigned long *rmapp,
 524                                    struct kvm_rmap_desc *desc,
 525                                    int i,
 526                                    struct kvm_rmap_desc *prev_desc)
 527 {
 528         int j;
 529
 530         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 531                 ;
 532         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 533         desc->shadow_ptes[j] = NULL;
 534         if (j != 0)
 535                 return;
 536         if (!prev_desc && !desc->more)
 537                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 538         else
 539                 if (prev_desc)
 540                         prev_desc->more = desc->more;
 541                 else
 542                         *rmapp = (unsigned long)desc->more | 1;
 543         mmu_free_rmap_desc(desc);
 544 }
 545
 546 static void rmap_remove(struct kvm *kvm, u64 *spte)
 547 {
 548         struct kvm_rmap_desc *desc;
 549         struct kvm_rmap_desc *prev_desc;
 550         struct kvm_mmu_page *sp;
 551         pfn_t pfn;
 552         unsigned long *rmapp;
 553         int i;
 554
 555         if (!is_rmap_pte(*spte))
 556                 return;
 557         sp = page_header(__pa(spte));
 558         pfn = spte_to_pfn(*spte);
 559         if (*spte & shadow_accessed_mask)
 560                 kvm_set_pfn_accessed(pfn);
 561         if (is_writeble_pte(*spte))
 562                 kvm_release_pfn_dirty(pfn);
 563         else
 564                 kvm_release_pfn_clean(pfn);
 565         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 566         if (!*rmapp) {
 567                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 568                 BUG();
 569         } else if (!(*rmapp & 1)) {
 570                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 571                 if ((u64 *)*rmapp != spte) {
 572                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 573                                spte, *spte);
 574                         BUG();
 575                 }
 576                 *rmapp = 0;
 577         } else {
 578                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 579                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 580                 prev_desc = NULL;
 581                 while (desc) {
 582                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 583                                 if (desc->shadow_ptes[i] == spte) {
 584                                         rmap_desc_remove_entry(rmapp,
 585                                                                desc, i,
 586                                                                prev_desc);
 587                                         return;
 588                                 }
 589                         prev_desc = desc;
 590                         desc = desc->more;
 591                 }
 592                 BUG();
 593         }
 594 }
 595
 596 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 597 {
 598         struct kvm_rmap_desc *desc;
 599         struct kvm_rmap_desc *prev_desc;
 600         u64 *prev_spte;
 601         int i;
 602
 603         if (!*rmapp)
 604                 return NULL;
 605         else if (!(*rmapp & 1)) {
 606                 if (!spte)
 607                         return (u64 *)*rmapp;
 608                 return NULL;
 609         }
 610         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 611         prev_desc = NULL;
 612         prev_spte = NULL;
 613         while (desc) {
 614                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 615                         if (prev_spte == spte)
 616                                 return desc->shadow_ptes[i];
 617                         prev_spte = desc->shadow_ptes[i];
 618                 }
 619                 desc = desc->more;
 620         }
 621         return NULL;
 622 }
 623
 624 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 625 {
 626         unsigned long *rmapp;
 627         u64 *spte;
 628         int write_protected = 0;
 629
 630         gfn = unalias_gfn(kvm, gfn);
 631         rmapp = gfn_to_rmap(kvm, gfn, 0);
 632
 633         spte = rmap_next(kvm, rmapp, NULL);
 634         while (spte) {
 635                 BUG_ON(!spte);
 636                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 637                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 638                 if (is_writeble_pte(*spte)) {
 639                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 640                         write_protected = 1;
 641                 }
 642                 spte = rmap_next(kvm, rmapp, spte);
 643         }
 644         if (write_protected) {
 645                 pfn_t pfn;
 646
 647                 spte = rmap_next(kvm, rmapp, NULL);
 648                 pfn = spte_to_pfn(*spte);
 649                 kvm_set_pfn_dirty(pfn);
 650         }
 651
 652         /* check for huge page mappings */
 653         rmapp = gfn_to_rmap(kvm, gfn, 1);
 654         spte = rmap_next(kvm, rmapp, NULL);
 655         while (spte) {
 656                 BUG_ON(!spte);
 657                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 658                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 659                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 660                 if (is_writeble_pte(*spte)) {
 661                         rmap_remove(kvm, spte);
 662                         --kvm->stat.lpages;
 663                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 664                         spte = NULL;
 665                         write_protected = 1;
 666                 }
 667                 spte = rmap_next(kvm, rmapp, spte);
 668         }
 669
 670         return write_protected;
 671 }
 672
 673 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 674 {
 675         u64 *spte;
 676         int need_tlb_flush = 0;
 677
 678         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 679                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 680                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 681                 rmap_remove(kvm, spte);
 682                 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 683                 need_tlb_flush = 1;
 684         }
 685         return need_tlb_flush;
 686 }
 687
 688 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 689                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 690 {
 691         int i;
 692         int retval = 0;
 693
 694         /*
 695          * If mmap_sem isn't taken, we can look the memslots with only
 696          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 697          */
 698         for (i = 0; i < kvm->nmemslots; i++) {
 699                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 700                 unsigned long start = memslot->userspace_addr;
 701                 unsigned long end;
 702
 703                 /* mmu_lock protects userspace_addr */
 704                 if (!start)
 705                         continue;
 706
 707                 end = start + (memslot->npages << PAGE_SHIFT);
 708                 if (hva >= start && hva < end) {
 709                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 710                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 711                         retval |= handler(kvm,
 712                                           &memslot->lpage_info[
 713                                                   gfn_offset /
 714                                                   KVM_PAGES_PER_HPAGE].rmap_pde);
 715                 }
 716         }
 717
 718         return retval;
 719 }
 720
 721 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 722 {
 723         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 724 }
 725
 726 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 727 {
 728         u64 *spte;
 729         int young = 0;
 730
 731         /* always return old for EPT */
 732         if (!shadow_accessed_mask)
 733                 return 0;
 734
 735         spte = rmap_next(kvm, rmapp, NULL);
 736         while (spte) {
 737                 int _young;
 738                 u64 _spte = *spte;
 739                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 740                 _young = _spte & PT_ACCESSED_MASK;
 741                 if (_young) {
 742                         young = 1;
 743                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 744                 }
 745                 spte = rmap_next(kvm, rmapp, spte);
 746         }
 747         return young;
 748 }
 749
 750 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 751 {
 752         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 753 }
 754
 755 #ifdef MMU_DEBUG
 756 static int is_empty_shadow_page(u64 *spt)
 757 {
 758         u64 *pos;
 759         u64 *end;
 760
 761         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 762                 if (is_shadow_present_pte(*pos)) {
 763                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 764                                pos, *pos);
 765                         return 0;
 766                 }
 767         return 1;
 768 }
 769 #endif
 770
 771 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 772 {
 773         ASSERT(is_empty_shadow_page(sp->spt));
 774         list_del(&sp->link);
 775         __free_page(virt_to_page(sp->spt));
 776         __free_page(virt_to_page(sp->gfns));
 777         kfree(sp);
 778         ++kvm->arch.n_free_mmu_pages;
 779 }
 780
 781 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 782 {
 783         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 784 }
 785
 786 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 787                                                u64 *parent_pte)
 788 {
 789         struct kvm_mmu_page *sp;
 790
 791         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 792         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 793         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 794         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 795         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 796         INIT_LIST_HEAD(&sp->oos_link);
 797         ASSERT(is_empty_shadow_page(sp->spt));
 798         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 799         sp->multimapped = 0;
 800         sp->global = 1;
 801         sp->parent_pte = parent_pte;
 802         --vcpu->kvm->arch.n_free_mmu_pages;
 803         return sp;
 804 }
 805
 806 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 807                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 808 {
 809         struct kvm_pte_chain *pte_chain;
 810         struct hlist_node *node;
 811         int i;
 812
 813         if (!parent_pte)
 814                 return;
 815         if (!sp->multimapped) {
 816                 u64 *old = sp->parent_pte;
 817
 818                 if (!old) {
 819                         sp->parent_pte = parent_pte;
 820                         return;
 821                 }
 822                 sp->multimapped = 1;
 823                 pte_chain = mmu_alloc_pte_chain(vcpu);
 824                 INIT_HLIST_HEAD(&sp->parent_ptes);
 825                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 826                 pte_chain->parent_ptes[0] = old;
 827         }
 828         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 829                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 830                         continue;
 831                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 832                         if (!pte_chain->parent_ptes[i]) {
 833                                 pte_chain->parent_ptes[i] = parent_pte;
 834                                 return;
 835                         }
 836         }
 837         pte_chain = mmu_alloc_pte_chain(vcpu);
 838         BUG_ON(!pte_chain);
 839         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 840         pte_chain->parent_ptes[0] = parent_pte;
 841 }
 842
 843 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 844                                        u64 *parent_pte)
 845 {
 846         struct kvm_pte_chain *pte_chain;
 847         struct hlist_node *node;
 848         int i;
 849
 850         if (!sp->multimapped) {
 851                 BUG_ON(sp->parent_pte != parent_pte);
 852                 sp->parent_pte = NULL;
 853                 return;
 854         }
 855         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 856                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 857                         if (!pte_chain->parent_ptes[i])
 858                                 break;
 859                         if (pte_chain->parent_ptes[i] != parent_pte)
 860                                 continue;
 861                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 862                                 && pte_chain->parent_ptes[i + 1]) {
 863                                 pte_chain->parent_ptes[i]
 864                                         = pte_chain->parent_ptes[i + 1];
 865                                 ++i;
 866                         }
 867                         pte_chain->parent_ptes[i] = NULL;
 868                         if (i == 0) {
 869                                 hlist_del(&pte_chain->link);
 870                                 mmu_free_pte_chain(pte_chain);
 871                                 if (hlist_empty(&sp->parent_ptes)) {
 872                                         sp->multimapped = 0;
 873                                         sp->parent_pte = NULL;
 874                                 }
 875                         }
 876                         return;
 877                 }
 878         BUG();
 879 }
 880
 881
 882 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 883                             mmu_parent_walk_fn fn)
 884 {
 885         struct kvm_pte_chain *pte_chain;
 886         struct hlist_node *node;
 887         struct kvm_mmu_page *parent_sp;
 888         int i;
 889
 890         if (!sp->multimapped && sp->parent_pte) {
 891                 parent_sp = page_header(__pa(sp->parent_pte));
 892                 fn(vcpu, parent_sp);
 893                 mmu_parent_walk(vcpu, parent_sp, fn);
 894                 return;
 895         }
 896         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 897                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 898                         if (!pte_chain->parent_ptes[i])
 899                                 break;
 900                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 901                         fn(vcpu, parent_sp);
 902                         mmu_parent_walk(vcpu, parent_sp, fn);
 903                 }
 904 }
 905
 906 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
 907 {
 908         unsigned int index;
 909         struct kvm_mmu_page *sp = page_header(__pa(spte));
 910
 911         index = spte - sp->spt;
 912         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
 913                 sp->unsync_children++;
 914         WARN_ON(!sp->unsync_children);
 915 }
 916
 917 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
 918 {
 919         struct kvm_pte_chain *pte_chain;
 920         struct hlist_node *node;
 921         int i;
 922
 923         if (!sp->parent_pte)
 924                 return;
 925
 926         if (!sp->multimapped) {
 927                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
 928                 return;
 929         }
 930
 931         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 932                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 933                         if (!pte_chain->parent_ptes[i])
 934                                 break;
 935                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
 936                 }
 937 }
 938
 939 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
 940 {
 941         kvm_mmu_update_parents_unsync(sp);
 942         return 1;
 943 }
 944
 945 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
 946                                         struct kvm_mmu_page *sp)
 947 {
 948         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
 949         kvm_mmu_update_parents_unsync(sp);
 950 }
 951
 952 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 953                                     struct kvm_mmu_page *sp)
 954 {
 955         int i;
 956
 957         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 958                 sp->spt[i] = shadow_trap_nonpresent_pte;
 959 }
 960
 961 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
 962                                struct kvm_mmu_page *sp)
 963 {
 964         return 1;
 965 }
 966
 967 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
 968 {
 969 }
 970
 971 #define KVM_PAGE_ARRAY_NR 16
 972
 973 struct kvm_mmu_pages {
 974         struct mmu_page_and_offset {
 975                 struct kvm_mmu_page *sp;
 976                 unsigned int idx;
 977         } page[KVM_PAGE_ARRAY_NR];
 978         unsigned int nr;
 979 };
 980
 981 #define for_each_unsync_children(bitmap, idx)           \
 982         for (idx = find_first_bit(bitmap, 512);         \
 983              idx < 512;                                 \
 984              idx = find_next_bit(bitmap, 512, idx+1))
 985
 986 int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
 987                    int idx)
 988 {
 989         int i;
 990
 991         if (sp->unsync)
 992                 for (i=0; i < pvec->nr; i++)
 993                         if (pvec->page[i].sp == sp)
 994                                 return 0;
 995
 996         pvec->page[pvec->nr].sp = sp;
 997         pvec->page[pvec->nr].idx = idx;
 998         pvec->nr++;
 999         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1000 }
1001
1002 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1003                            struct kvm_mmu_pages *pvec)
1004 {
1005         int i, ret, nr_unsync_leaf = 0;
1006
1007         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1008                 u64 ent = sp->spt[i];
1009
1010                 if (is_shadow_present_pte(ent)) {
1011                         struct kvm_mmu_page *child;
1012                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1013
1014                         if (child->unsync_children) {
1015                                 if (mmu_pages_add(pvec, child, i))
1016                                         return -ENOSPC;
1017
1018                                 ret = __mmu_unsync_walk(child, pvec);
1019                                 if (!ret)
1020                                         __clear_bit(i, sp->unsync_child_bitmap);
1021                                 else if (ret > 0)
1022                                         nr_unsync_leaf += ret;
1023                                 else
1024                                         return ret;
1025                         }
1026
1027                         if (child->unsync) {
1028                                 nr_unsync_leaf++;
1029                                 if (mmu_pages_add(pvec, child, i))
1030                                         return -ENOSPC;
1031                         }
1032                 }
1033         }
1034
1035         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1036                 sp->unsync_children = 0;
1037
1038         return nr_unsync_leaf;
1039 }
1040
1041 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1042                            struct kvm_mmu_pages *pvec)
1043 {
1044         if (!sp->unsync_children)
1045                 return 0;
1046
1047         mmu_pages_add(pvec, sp, 0);
1048         return __mmu_unsync_walk(sp, pvec);
1049 }
1050
1051 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1052 {
1053         unsigned index;
1054         struct hlist_head *bucket;
1055         struct kvm_mmu_page *sp;
1056         struct hlist_node *node;
1057
1058         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1059         index = kvm_page_table_hashfn(gfn);
1060         bucket = &kvm->arch.mmu_page_hash[index];
1061         hlist_for_each_entry(sp, node, bucket, hash_link)
1062                 if (sp->gfn == gfn && !sp->role.metaphysical
1063                     && !sp->role.invalid) {
1064                         pgprintk("%s: found role %x\n",
1065                                  __func__, sp->role.word);
1066                         return sp;
1067                 }
1068         return NULL;
1069 }
1070
1071 static void kvm_unlink_unsync_global(struct kvm *kvm, struct kvm_mmu_page *sp)
1072 {
1073         list_del(&sp->oos_link);
1074         --kvm->stat.mmu_unsync_global;
1075 }
1076
1077 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1078 {
1079         WARN_ON(!sp->unsync);
1080         sp->unsync = 0;
1081         if (sp->global)
1082                 kvm_unlink_unsync_global(kvm, sp);
1083         --kvm->stat.mmu_unsync;
1084 }
1085
1086 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1087
1088 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1089 {
1090         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1091                 kvm_mmu_zap_page(vcpu->kvm, sp);
1092                 return 1;
1093         }
1094
1095         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1096                 kvm_flush_remote_tlbs(vcpu->kvm);
1097         kvm_unlink_unsync_page(vcpu->kvm, sp);
1098         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1099                 kvm_mmu_zap_page(vcpu->kvm, sp);
1100                 return 1;
1101         }
1102
1103         kvm_mmu_flush_tlb(vcpu);
1104         return 0;
1105 }
1106
1107 struct mmu_page_path {
1108         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1109         unsigned int idx[PT64_ROOT_LEVEL-1];
1110 };
1111
1112 #define for_each_sp(pvec, sp, parents, i)                       \
1113                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1114                         sp = pvec.page[i].sp;                   \
1115                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1116                         i = mmu_pages_next(&pvec, &parents, i))
1117
1118 int mmu_pages_next(struct kvm_mmu_pages *pvec, struct mmu_page_path *parents,
1119                    int i)
1120 {
1121         int n;
1122
1123         for (n = i+1; n < pvec->nr; n++) {
1124                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1125
1126                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1127                         parents->idx[0] = pvec->page[n].idx;
1128                         return n;
1129                 }
1130
1131                 parents->parent[sp->role.level-2] = sp;
1132                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1133         }
1134
1135         return n;
1136 }
1137
1138 void mmu_pages_clear_parents(struct mmu_page_path *parents)
1139 {
1140         struct kvm_mmu_page *sp;
1141         unsigned int level = 0;
1142
1143         do {
1144                 unsigned int idx = parents->idx[level];
1145
1146                 sp = parents->parent[level];
1147                 if (!sp)
1148                         return;
1149
1150                 --sp->unsync_children;
1151                 WARN_ON((int)sp->unsync_children < 0);
1152                 __clear_bit(idx, sp->unsync_child_bitmap);
1153                 level++;
1154         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1155 }
1156
1157 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1158                                struct mmu_page_path *parents,
1159                                struct kvm_mmu_pages *pvec)
1160 {
1161         parents->parent[parent->role.level-1] = NULL;
1162         pvec->nr = 0;
1163 }
1164
1165 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1166                               struct kvm_mmu_page *parent)
1167 {
1168         int i;
1169         struct kvm_mmu_page *sp;
1170         struct mmu_page_path parents;
1171         struct kvm_mmu_pages pages;
1172
1173         kvm_mmu_pages_init(parent, &parents, &pages);
1174         while (mmu_unsync_walk(parent, &pages)) {
1175                 int protected = 0;
1176
1177                 for_each_sp(pages, sp, parents, i)
1178                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1179
1180                 if (protected)
1181                         kvm_flush_remote_tlbs(vcpu->kvm);
1182
1183                 for_each_sp(pages, sp, parents, i) {
1184                         kvm_sync_page(vcpu, sp);
1185                         mmu_pages_clear_parents(&parents);
1186                 }
1187                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1188                 kvm_mmu_pages_init(parent, &parents, &pages);
1189         }
1190 }
1191
1192 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1193                                              gfn_t gfn,
1194                                              gva_t gaddr,
1195                                              unsigned level,
1196                                              int metaphysical,
1197                                              unsigned access,
1198                                              u64 *parent_pte)
1199 {
1200         union kvm_mmu_page_role role;
1201         unsigned index;
1202         unsigned quadrant;
1203         struct hlist_head *bucket;
1204         struct kvm_mmu_page *sp;
1205         struct hlist_node *node, *tmp;
1206
1207         role.word = 0;
1208         role.glevels = vcpu->arch.mmu.root_level;
1209         role.level = level;
1210         role.metaphysical = metaphysical;
1211         role.access = access;
1212         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1213                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1214                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1215                 role.quadrant = quadrant;
1216         }
1217         pgprintk("%s: looking gfn %lx role %x\n", __func__,
1218                  gfn, role.word);
1219         index = kvm_page_table_hashfn(gfn);
1220         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1221         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1222                 if (sp->gfn == gfn) {
1223                         if (sp->unsync)
1224                                 if (kvm_sync_page(vcpu, sp))
1225                                         continue;
1226
1227                         if (sp->role.word != role.word)
1228                                 continue;
1229
1230                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1231                         if (sp->unsync_children) {
1232                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1233                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1234                         }
1235                         pgprintk("%s: found\n", __func__);
1236                         return sp;
1237                 }
1238         ++vcpu->kvm->stat.mmu_cache_miss;
1239         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1240         if (!sp)
1241                 return sp;
1242         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1243         sp->gfn = gfn;
1244         sp->role = role;
1245         hlist_add_head(&sp->hash_link, bucket);
1246         if (!metaphysical) {
1247                 if (rmap_write_protect(vcpu->kvm, gfn))
1248                         kvm_flush_remote_tlbs(vcpu->kvm);
1249                 account_shadowed(vcpu->kvm, gfn);
1250         }
1251         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1252                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1253         else
1254                 nonpaging_prefetch_page(vcpu, sp);
1255         return sp;
1256 }
1257
1258 static int walk_shadow(struct kvm_shadow_walk *walker,
1259                        struct kvm_vcpu *vcpu, u64 addr)
1260 {
1261         hpa_t shadow_addr;
1262         int level;
1263         int r;
1264         u64 *sptep;
1265         unsigned index;
1266
1267         shadow_addr = vcpu->arch.mmu.root_hpa;
1268         level = vcpu->arch.mmu.shadow_root_level;
1269         if (level == PT32E_ROOT_LEVEL) {
1270                 shadow_addr = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1271                 shadow_addr &= PT64_BASE_ADDR_MASK;
1272                 --level;
1273         }
1274
1275         while (level >= PT_PAGE_TABLE_LEVEL) {
1276                 index = SHADOW_PT_INDEX(addr, level);
1277                 sptep = ((u64 *)__va(shadow_addr)) + index;
1278                 r = walker->entry(walker, vcpu, addr, sptep, level);
1279                 if (r)
1280                         return r;
1281                 shadow_addr = *sptep & PT64_BASE_ADDR_MASK;
1282                 --level;
1283         }
1284         return 0;
1285 }
1286
1287 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1288                                          struct kvm_mmu_page *sp)
1289 {
1290         unsigned i;
1291         u64 *pt;
1292         u64 ent;
1293
1294         pt = sp->spt;
1295
1296         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1297                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1298                         if (is_shadow_present_pte(pt[i]))
1299                                 rmap_remove(kvm, &pt[i]);
1300                         pt[i] = shadow_trap_nonpresent_pte;
1301                 }
1302                 return;
1303         }
1304
1305         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1306                 ent = pt[i];
1307
1308                 if (is_shadow_present_pte(ent)) {
1309                         if (!is_large_pte(ent)) {
1310                                 ent &= PT64_BASE_ADDR_MASK;
1311                                 mmu_page_remove_parent_pte(page_header(ent),
1312                                                            &pt[i]);
1313                         } else {
1314                                 --kvm->stat.lpages;
1315                                 rmap_remove(kvm, &pt[i]);
1316                         }
1317                 }
1318                 pt[i] = shadow_trap_nonpresent_pte;
1319         }
1320 }
1321
1322 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1323 {
1324         mmu_page_remove_parent_pte(sp, parent_pte);
1325 }
1326
1327 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1328 {
1329         int i;
1330
1331         for (i = 0; i < KVM_MAX_VCPUS; ++i)
1332                 if (kvm->vcpus[i])
1333                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
1334 }
1335
1336 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1337 {
1338         u64 *parent_pte;
1339
1340         while (sp->multimapped || sp->parent_pte) {
1341                 if (!sp->multimapped)
1342                         parent_pte = sp->parent_pte;
1343                 else {
1344                         struct kvm_pte_chain *chain;
1345
1346                         chain = container_of(sp->parent_ptes.first,
1347                                              struct kvm_pte_chain, link);
1348                         parent_pte = chain->parent_ptes[0];
1349                 }
1350                 BUG_ON(!parent_pte);
1351                 kvm_mmu_put_page(sp, parent_pte);
1352                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1353         }
1354 }
1355
1356 static int mmu_zap_unsync_children(struct kvm *kvm,
1357                                    struct kvm_mmu_page *parent)
1358 {
1359         int i, zapped = 0;
1360         struct mmu_page_path parents;
1361         struct kvm_mmu_pages pages;
1362
1363         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1364                 return 0;
1365
1366         kvm_mmu_pages_init(parent, &parents, &pages);
1367         while (mmu_unsync_walk(parent, &pages)) {
1368                 struct kvm_mmu_page *sp;
1369
1370                 for_each_sp(pages, sp, parents, i) {
1371                         kvm_mmu_zap_page(kvm, sp);
1372                         mmu_pages_clear_parents(&parents);
1373                 }
1374                 zapped += pages.nr;
1375                 kvm_mmu_pages_init(parent, &parents, &pages);
1376         }
1377
1378         return zapped;
1379 }
1380
1381 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1382 {
1383         int ret;
1384         ++kvm->stat.mmu_shadow_zapped;
1385         ret = mmu_zap_unsync_children(kvm, sp);
1386         kvm_mmu_page_unlink_children(kvm, sp);
1387         kvm_mmu_unlink_parents(kvm, sp);
1388         kvm_flush_remote_tlbs(kvm);
1389         if (!sp->role.invalid && !sp->role.metaphysical)
1390                 unaccount_shadowed(kvm, sp->gfn);
1391         if (sp->unsync)
1392                 kvm_unlink_unsync_page(kvm, sp);
1393         if (!sp->root_count) {
1394                 hlist_del(&sp->hash_link);
1395                 kvm_mmu_free_page(kvm, sp);
1396         } else {
1397                 sp->role.invalid = 1;
1398                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1399                 kvm_reload_remote_mmus(kvm);
1400         }
1401         kvm_mmu_reset_last_pte_updated(kvm);
1402         return ret;
1403 }
1404
1405 /*
1406  * Changing the number of mmu pages allocated to the vm
1407  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1408  */
1409 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1410 {
1411         /*
1412          * If we set the number of mmu pages to be smaller be than the
1413          * number of actived pages , we must to free some mmu pages before we
1414          * change the value
1415          */
1416
1417         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1418             kvm_nr_mmu_pages) {
1419                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1420                                        - kvm->arch.n_free_mmu_pages;
1421
1422                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1423                         struct kvm_mmu_page *page;
1424
1425                         page = container_of(kvm->arch.active_mmu_pages.prev,
1426                                             struct kvm_mmu_page, link);
1427                         kvm_mmu_zap_page(kvm, page);
1428                         n_used_mmu_pages--;
1429                 }
1430                 kvm->arch.n_free_mmu_pages = 0;
1431         }
1432         else
1433                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1434                                          - kvm->arch.n_alloc_mmu_pages;
1435
1436         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1437 }
1438
1439 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1440 {
1441         unsigned index;
1442         struct hlist_head *bucket;
1443         struct kvm_mmu_page *sp;
1444         struct hlist_node *node, *n;
1445         int r;
1446
1447         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1448         r = 0;
1449         index = kvm_page_table_hashfn(gfn);
1450         bucket = &kvm->arch.mmu_page_hash[index];
1451         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1452                 if (sp->gfn == gfn && !sp->role.metaphysical) {
1453                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1454                                  sp->role.word);
1455                         r = 1;
1456                         if (kvm_mmu_zap_page(kvm, sp))
1457                                 n = bucket->first;
1458                 }
1459         return r;
1460 }
1461
1462 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1463 {
1464         struct kvm_mmu_page *sp;
1465
1466         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1467                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1468                 kvm_mmu_zap_page(kvm, sp);
1469         }
1470 }
1471
1472 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1473 {
1474         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1475         struct kvm_mmu_page *sp = page_header(__pa(pte));
1476
1477         __set_bit(slot, sp->slot_bitmap);
1478 }
1479
1480 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1481 {
1482         int i;
1483         u64 *pt = sp->spt;
1484
1485         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1486                 return;
1487
1488         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1489                 if (pt[i] == shadow_notrap_nonpresent_pte)
1490                         set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1491         }
1492 }
1493
1494 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1495 {
1496         struct page *page;
1497
1498         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1499
1500         if (gpa == UNMAPPED_GVA)
1501                 return NULL;
1502
1503         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1504
1505         return page;
1506 }
1507
1508 /*
1509  * The function is based on mtrr_type_lookup() in
1510  * arch/x86/kernel/cpu/mtrr/generic.c
1511  */
1512 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1513                          u64 start, u64 end)
1514 {
1515         int i;
1516         u64 base, mask;
1517         u8 prev_match, curr_match;
1518         int num_var_ranges = KVM_NR_VAR_MTRR;
1519
1520         if (!mtrr_state->enabled)
1521                 return 0xFF;
1522
1523         /* Make end inclusive end, instead of exclusive */
1524         end--;
1525
1526         /* Look in fixed ranges. Just return the type as per start */
1527         if (mtrr_state->have_fixed && (start < 0x100000)) {
1528                 int idx;
1529
1530                 if (start < 0x80000) {
1531                         idx = 0;
1532                         idx += (start >> 16);
1533                         return mtrr_state->fixed_ranges[idx];
1534                 } else if (start < 0xC0000) {
1535                         idx = 1 * 8;
1536                         idx += ((start - 0x80000) >> 14);
1537                         return mtrr_state->fixed_ranges[idx];
1538                 } else if (start < 0x1000000) {
1539                         idx = 3 * 8;
1540                         idx += ((start - 0xC0000) >> 12);
1541                         return mtrr_state->fixed_ranges[idx];
1542                 }
1543         }
1544
1545         /*
1546          * Look in variable ranges
1547          * Look of multiple ranges matching this address and pick type
1548          * as per MTRR precedence
1549          */
1550         if (!(mtrr_state->enabled & 2))
1551                 return mtrr_state->def_type;
1552
1553         prev_match = 0xFF;
1554         for (i = 0; i < num_var_ranges; ++i) {
1555                 unsigned short start_state, end_state;
1556
1557                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1558                         continue;
1559
1560                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1561                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1562                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1563                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1564
1565                 start_state = ((start & mask) == (base & mask));
1566                 end_state = ((end & mask) == (base & mask));
1567                 if (start_state != end_state)
1568                         return 0xFE;
1569
1570                 if ((start & mask) != (base & mask))
1571                         continue;
1572
1573                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1574                 if (prev_match == 0xFF) {
1575                         prev_match = curr_match;
1576                         continue;
1577                 }
1578
1579                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1580                     curr_match == MTRR_TYPE_UNCACHABLE)
1581                         return MTRR_TYPE_UNCACHABLE;
1582
1583                 if ((prev_match == MTRR_TYPE_WRBACK &&
1584                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1585                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1586                      curr_match == MTRR_TYPE_WRBACK)) {
1587                         prev_match = MTRR_TYPE_WRTHROUGH;
1588                         curr_match = MTRR_TYPE_WRTHROUGH;
1589                 }
1590
1591                 if (prev_match != curr_match)
1592                         return MTRR_TYPE_UNCACHABLE;
1593         }
1594
1595         if (prev_match != 0xFF)
1596                 return prev_match;
1597
1598         return mtrr_state->def_type;
1599 }
1600
1601 static u8 get_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1602 {
1603         u8 mtrr;
1604
1605         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1606                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1607         if (mtrr == 0xfe || mtrr == 0xff)
1608                 mtrr = MTRR_TYPE_WRBACK;
1609         return mtrr;
1610 }
1611
1612 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1613 {
1614         unsigned index;
1615         struct hlist_head *bucket;
1616         struct kvm_mmu_page *s;
1617         struct hlist_node *node, *n;
1618
1619         index = kvm_page_table_hashfn(sp->gfn);
1620         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1621         /* don't unsync if pagetable is shadowed with multiple roles */
1622         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1623                 if (s->gfn != sp->gfn || s->role.metaphysical)
1624                         continue;
1625                 if (s->role.word != sp->role.word)
1626                         return 1;
1627         }
1628         ++vcpu->kvm->stat.mmu_unsync;
1629         sp->unsync = 1;
1630
1631         if (sp->global) {
1632                 list_add(&sp->oos_link, &vcpu->kvm->arch.oos_global_pages);
1633                 ++vcpu->kvm->stat.mmu_unsync_global;
1634         } else
1635                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1636
1637         mmu_convert_notrap(sp);
1638         return 0;
1639 }
1640
1641 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1642                                   bool can_unsync)
1643 {
1644         struct kvm_mmu_page *shadow;
1645
1646         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1647         if (shadow) {
1648                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1649                         return 1;
1650                 if (shadow->unsync)
1651                         return 0;
1652                 if (can_unsync && oos_shadow)
1653                         return kvm_unsync_page(vcpu, shadow);
1654                 return 1;
1655         }
1656         return 0;
1657 }
1658
1659 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1660                     unsigned pte_access, int user_fault,
1661                     int write_fault, int dirty, int largepage,
1662                     int global, gfn_t gfn, pfn_t pfn, bool speculative,
1663                     bool can_unsync)
1664 {
1665         u64 spte;
1666         int ret = 0;
1667         u64 mt_mask = shadow_mt_mask;
1668         struct kvm_mmu_page *sp = page_header(__pa(shadow_pte));
1669
1670         if (!global && sp->global) {
1671                 sp->global = 0;
1672                 if (sp->unsync) {
1673                         kvm_unlink_unsync_global(vcpu->kvm, sp);
1674                         kvm_mmu_mark_parents_unsync(vcpu, sp);
1675                 }
1676         }
1677
1678         /*
1679          * We don't set the accessed bit, since we sometimes want to see
1680          * whether the guest actually used the pte (in order to detect
1681          * demand paging).
1682          */
1683         spte = shadow_base_present_pte | shadow_dirty_mask;
1684         if (!speculative)
1685                 spte |= shadow_accessed_mask;
1686         if (!dirty)
1687                 pte_access &= ~ACC_WRITE_MASK;
1688         if (pte_access & ACC_EXEC_MASK)
1689                 spte |= shadow_x_mask;
1690         else
1691                 spte |= shadow_nx_mask;
1692         if (pte_access & ACC_USER_MASK)
1693                 spte |= shadow_user_mask;
1694         if (largepage)
1695                 spte |= PT_PAGE_SIZE_MASK;
1696         if (mt_mask) {
1697                 mt_mask = get_memory_type(vcpu, gfn) <<
1698                           kvm_x86_ops->get_mt_mask_shift();
1699                 spte |= mt_mask;
1700         }
1701
1702         spte |= (u64)pfn << PAGE_SHIFT;
1703
1704         if ((pte_access & ACC_WRITE_MASK)
1705             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1706
1707                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1708                         ret = 1;
1709                         spte = shadow_trap_nonpresent_pte;
1710                         goto set_pte;
1711                 }
1712
1713                 spte |= PT_WRITABLE_MASK;
1714
1715                 /*
1716                  * Optimization: for pte sync, if spte was writable the hash
1717                  * lookup is unnecessary (and expensive). Write protection
1718                  * is responsibility of mmu_get_page / kvm_sync_page.
1719                  * Same reasoning can be applied to dirty page accounting.
1720                  */
1721                 if (!can_unsync && is_writeble_pte(*shadow_pte))
1722                         goto set_pte;
1723
1724                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1725                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1726                                  __func__, gfn);
1727                         ret = 1;
1728                         pte_access &= ~ACC_WRITE_MASK;
1729                         if (is_writeble_pte(spte))
1730                                 spte &= ~PT_WRITABLE_MASK;
1731                 }
1732         }
1733
1734         if (pte_access & ACC_WRITE_MASK)
1735                 mark_page_dirty(vcpu->kvm, gfn);
1736
1737 set_pte:
1738         set_shadow_pte(shadow_pte, spte);
1739         return ret;
1740 }
1741
1742 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1743                          unsigned pt_access, unsigned pte_access,
1744                          int user_fault, int write_fault, int dirty,
1745                          int *ptwrite, int largepage, int global,
1746                          gfn_t gfn, pfn_t pfn, bool speculative)
1747 {
1748         int was_rmapped = 0;
1749         int was_writeble = is_writeble_pte(*shadow_pte);
1750
1751         pgprintk("%s: spte %llx access %x write_fault %d"
1752                  " user_fault %d gfn %lx\n",
1753                  __func__, *shadow_pte, pt_access,
1754                  write_fault, user_fault, gfn);
1755
1756         if (is_rmap_pte(*shadow_pte)) {
1757                 /*
1758                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1759                  * the parent of the now unreachable PTE.
1760                  */
1761                 if (largepage && !is_large_pte(*shadow_pte)) {
1762                         struct kvm_mmu_page *child;
1763                         u64 pte = *shadow_pte;
1764
1765                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1766                         mmu_page_remove_parent_pte(child, shadow_pte);
1767                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1768                         pgprintk("hfn old %lx new %lx\n",
1769                                  spte_to_pfn(*shadow_pte), pfn);
1770                         rmap_remove(vcpu->kvm, shadow_pte);
1771                 } else {
1772                         if (largepage)
1773                                 was_rmapped = is_large_pte(*shadow_pte);
1774                         else
1775                                 was_rmapped = 1;
1776                 }
1777         }
1778         if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1779                       dirty, largepage, global, gfn, pfn, speculative, true)) {
1780                 if (write_fault)
1781                         *ptwrite = 1;
1782                 kvm_x86_ops->tlb_flush(vcpu);
1783         }
1784
1785         pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1786         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1787                  is_large_pte(*shadow_pte)? "2MB" : "4kB",
1788                  is_present_pte(*shadow_pte)?"RW":"R", gfn,
1789                  *shadow_pte, shadow_pte);
1790         if (!was_rmapped && is_large_pte(*shadow_pte))
1791                 ++vcpu->kvm->stat.lpages;
1792
1793         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1794         if (!was_rmapped) {
1795                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1796                 if (!is_rmap_pte(*shadow_pte))
1797                         kvm_release_pfn_clean(pfn);
1798         } else {
1799                 if (was_writeble)
1800                         kvm_release_pfn_dirty(pfn);
1801                 else
1802                         kvm_release_pfn_clean(pfn);
1803         }
1804         if (speculative) {
1805                 vcpu->arch.last_pte_updated = shadow_pte;
1806                 vcpu->arch.last_pte_gfn = gfn;
1807         }
1808 }
1809
1810 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1811 {
1812 }
1813
1814 struct direct_shadow_walk {
1815         struct kvm_shadow_walk walker;
1816         pfn_t pfn;
1817         int write;
1818         int largepage;
1819         int pt_write;
1820 };
1821
1822 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1823                             struct kvm_vcpu *vcpu,
1824                             u64 addr, u64 *sptep, int level)
1825 {
1826         struct direct_shadow_walk *walk =
1827                 container_of(_walk, struct direct_shadow_walk, walker);
1828         struct kvm_mmu_page *sp;
1829         gfn_t pseudo_gfn;
1830         gfn_t gfn = addr >> PAGE_SHIFT;
1831
1832         if (level == PT_PAGE_TABLE_LEVEL
1833             || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1834                 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1835                              0, walk->write, 1, &walk->pt_write,
1836                              walk->largepage, 0, gfn, walk->pfn, false);
1837                 ++vcpu->stat.pf_fixed;
1838                 return 1;
1839         }
1840
1841         if (*sptep == shadow_trap_nonpresent_pte) {
1842                 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1843                 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1844                                       1, ACC_ALL, sptep);
1845                 if (!sp) {
1846                         pgprintk("nonpaging_map: ENOMEM\n");
1847                         kvm_release_pfn_clean(walk->pfn);
1848                         return -ENOMEM;
1849                 }
1850
1851                 set_shadow_pte(sptep,
1852                                __pa(sp->spt)
1853                                | PT_PRESENT_MASK | PT_WRITABLE_MASK
1854                                | shadow_user_mask | shadow_x_mask);
1855         }
1856         return 0;
1857 }
1858
1859 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1860                         int largepage, gfn_t gfn, pfn_t pfn)
1861 {
1862         int r;
1863         struct direct_shadow_walk walker = {
1864                 .walker = { .entry = direct_map_entry, },
1865                 .pfn = pfn,
1866                 .largepage = largepage,
1867                 .write = write,
1868                 .pt_write = 0,
1869         };
1870
1871         r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1872         if (r < 0)
1873                 return r;
1874         return walker.pt_write;
1875 }
1876
1877 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1878 {
1879         int r;
1880         int largepage = 0;
1881         pfn_t pfn;
1882         unsigned long mmu_seq;
1883
1884         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1885                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1886                 largepage = 1;
1887         }
1888
1889         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1890         smp_rmb();
1891         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1892
1893         /* mmio */
1894         if (is_error_pfn(pfn)) {
1895                 kvm_release_pfn_clean(pfn);
1896                 return 1;
1897         }
1898
1899         spin_lock(&vcpu->kvm->mmu_lock);
1900         if (mmu_notifier_retry(vcpu, mmu_seq))
1901                 goto out_unlock;
1902         kvm_mmu_free_some_pages(vcpu);
1903         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1904         spin_unlock(&vcpu->kvm->mmu_lock);
1905
1906
1907         return r;
1908
1909 out_unlock:
1910         spin_unlock(&vcpu->kvm->mmu_lock);
1911         kvm_release_pfn_clean(pfn);
1912         return 0;
1913 }
1914
1915
1916 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1917 {
1918         int i;
1919         struct kvm_mmu_page *sp;
1920
1921         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1922                 return;
1923         spin_lock(&vcpu->kvm->mmu_lock);
1924         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1925                 hpa_t root = vcpu->arch.mmu.root_hpa;
1926
1927                 sp = page_header(root);
1928                 --sp->root_count;
1929                 if (!sp->root_count && sp->role.invalid)
1930                         kvm_mmu_zap_page(vcpu->kvm, sp);
1931                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1932                 spin_unlock(&vcpu->kvm->mmu_lock);
1933                 return;
1934         }
1935         for (i = 0; i < 4; ++i) {
1936                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1937
1938                 if (root) {
1939                         root &= PT64_BASE_ADDR_MASK;
1940                         sp = page_header(root);
1941                         --sp->root_count;
1942                         if (!sp->root_count && sp->role.invalid)
1943                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1944                 }
1945                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1946         }
1947         spin_unlock(&vcpu->kvm->mmu_lock);
1948         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1949 }
1950
1951 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1952 {
1953         int i;
1954         gfn_t root_gfn;
1955         struct kvm_mmu_page *sp;
1956         int metaphysical = 0;
1957
1958         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1959
1960         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1961                 hpa_t root = vcpu->arch.mmu.root_hpa;
1962
1963                 ASSERT(!VALID_PAGE(root));
1964                 if (tdp_enabled)
1965                         metaphysical = 1;
1966                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1967                                       PT64_ROOT_LEVEL, metaphysical,
1968                                       ACC_ALL, NULL);
1969                 root = __pa(sp->spt);
1970                 ++sp->root_count;
1971                 vcpu->arch.mmu.root_hpa = root;
1972                 return;
1973         }
1974         metaphysical = !is_paging(vcpu);
1975         if (tdp_enabled)
1976                 metaphysical = 1;
1977         for (i = 0; i < 4; ++i) {
1978                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1979
1980                 ASSERT(!VALID_PAGE(root));
1981                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1982                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1983                                 vcpu->arch.mmu.pae_root[i] = 0;
1984                                 continue;
1985                         }
1986                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1987                 } else if (vcpu->arch.mmu.root_level == 0)
1988                         root_gfn = 0;
1989                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1990                                       PT32_ROOT_LEVEL, metaphysical,
1991                                       ACC_ALL, NULL);
1992                 root = __pa(sp->spt);
1993                 ++sp->root_count;
1994                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1995         }
1996         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1997 }
1998
1999 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
2000 {
2001         int i;
2002         struct kvm_mmu_page *sp;
2003
2004         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2005                 return;
2006         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2007                 hpa_t root = vcpu->arch.mmu.root_hpa;
2008                 sp = page_header(root);
2009                 mmu_sync_children(vcpu, sp);
2010                 return;
2011         }
2012         for (i = 0; i < 4; ++i) {
2013                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2014
2015                 if (root) {
2016                         root &= PT64_BASE_ADDR_MASK;
2017                         sp = page_header(root);
2018                         mmu_sync_children(vcpu, sp);
2019                 }
2020         }
2021 }
2022
2023 static void mmu_sync_global(struct kvm_vcpu *vcpu)
2024 {
2025         struct kvm *kvm = vcpu->kvm;
2026         struct kvm_mmu_page *sp, *n;
2027
2028         list_for_each_entry_safe(sp, n, &kvm->arch.oos_global_pages, oos_link)
2029                 kvm_sync_page(vcpu, sp);
2030 }
2031
2032 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2033 {
2034         spin_lock(&vcpu->kvm->mmu_lock);
2035         mmu_sync_roots(vcpu);
2036         spin_unlock(&vcpu->kvm->mmu_lock);
2037 }
2038
2039 void kvm_mmu_sync_global(struct kvm_vcpu *vcpu)
2040 {
2041         spin_lock(&vcpu->kvm->mmu_lock);
2042         mmu_sync_global(vcpu);
2043         spin_unlock(&vcpu->kvm->mmu_lock);
2044 }
2045
2046 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2047 {
2048         return vaddr;
2049 }
2050
2051 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2052                                 u32 error_code)
2053 {
2054         gfn_t gfn;
2055         int r;
2056
2057         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2058         r = mmu_topup_memory_caches(vcpu);
2059         if (r)
2060                 return r;
2061
2062         ASSERT(vcpu);
2063         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2064
2065         gfn = gva >> PAGE_SHIFT;
2066
2067         return nonpaging_map(vcpu, gva & PAGE_MASK,
2068                              error_code & PFERR_WRITE_MASK, gfn);
2069 }
2070
2071 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2072                                 u32 error_code)
2073 {
2074         pfn_t pfn;
2075         int r;
2076         int largepage = 0;
2077         gfn_t gfn = gpa >> PAGE_SHIFT;
2078         unsigned long mmu_seq;
2079
2080         ASSERT(vcpu);
2081         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2082
2083         r = mmu_topup_memory_caches(vcpu);
2084         if (r)
2085                 return r;
2086
2087         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
2088                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2089                 largepage = 1;
2090         }
2091         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2092         smp_rmb();
2093         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2094         if (is_error_pfn(pfn)) {
2095                 kvm_release_pfn_clean(pfn);
2096                 return 1;
2097         }
2098         spin_lock(&vcpu->kvm->mmu_lock);
2099         if (mmu_notifier_retry(vcpu, mmu_seq))
2100                 goto out_unlock;
2101         kvm_mmu_free_some_pages(vcpu);
2102         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2103                          largepage, gfn, pfn);
2104         spin_unlock(&vcpu->kvm->mmu_lock);
2105
2106         return r;
2107
2108 out_unlock:
2109         spin_unlock(&vcpu->kvm->mmu_lock);
2110         kvm_release_pfn_clean(pfn);
2111         return 0;
2112 }
2113
2114 static void nonpaging_free(struct kvm_vcpu *vcpu)
2115 {
2116         mmu_free_roots(vcpu);
2117 }
2118
2119 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2120 {
2121         struct kvm_mmu *context = &vcpu->arch.mmu;
2122
2123         context->new_cr3 = nonpaging_new_cr3;
2124         context->page_fault = nonpaging_page_fault;
2125         context->gva_to_gpa = nonpaging_gva_to_gpa;
2126         context->free = nonpaging_free;
2127         context->prefetch_page = nonpaging_prefetch_page;
2128         context->sync_page = nonpaging_sync_page;
2129         context->invlpg = nonpaging_invlpg;
2130         context->root_level = 0;
2131         context->shadow_root_level = PT32E_ROOT_LEVEL;
2132         context->root_hpa = INVALID_PAGE;
2133         return 0;
2134 }
2135
2136 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2137 {
2138         ++vcpu->stat.tlb_flush;
2139         kvm_x86_ops->tlb_flush(vcpu);
2140 }
2141
2142 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2143 {
2144         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2145         mmu_free_roots(vcpu);
2146 }
2147
2148 static void inject_page_fault(struct kvm_vcpu *vcpu,
2149                               u64 addr,
2150                               u32 err_code)
2151 {
2152         kvm_inject_page_fault(vcpu, addr, err_code);
2153 }
2154
2155 static void paging_free(struct kvm_vcpu *vcpu)
2156 {
2157         nonpaging_free(vcpu);
2158 }
2159
2160 #define PTTYPE 64
2161 #include "paging_tmpl.h"
2162 #undef PTTYPE
2163
2164 #define PTTYPE 32
2165 #include "paging_tmpl.h"
2166 #undef PTTYPE
2167
2168 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2169 {
2170         struct kvm_mmu *context = &vcpu->arch.mmu;
2171
2172         ASSERT(is_pae(vcpu));
2173         context->new_cr3 = paging_new_cr3;
2174         context->page_fault = paging64_page_fault;
2175         context->gva_to_gpa = paging64_gva_to_gpa;
2176         context->prefetch_page = paging64_prefetch_page;
2177         context->sync_page = paging64_sync_page;
2178         context->invlpg = paging64_invlpg;
2179         context->free = paging_free;
2180         context->root_level = level;
2181         context->shadow_root_level = level;
2182         context->root_hpa = INVALID_PAGE;
2183         return 0;
2184 }
2185
2186 static int paging64_init_context(struct kvm_vcpu *vcpu)
2187 {
2188         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2189 }
2190
2191 static int paging32_init_context(struct kvm_vcpu *vcpu)
2192 {
2193         struct kvm_mmu *context = &vcpu->arch.mmu;
2194
2195         context->new_cr3 = paging_new_cr3;
2196         context->page_fault = paging32_page_fault;
2197         context->gva_to_gpa = paging32_gva_to_gpa;
2198         context->free = paging_free;
2199         context->prefetch_page = paging32_prefetch_page;
2200         context->sync_page = paging32_sync_page;
2201         context->invlpg = paging32_invlpg;
2202         context->root_level = PT32_ROOT_LEVEL;
2203         context->shadow_root_level = PT32E_ROOT_LEVEL;
2204         context->root_hpa = INVALID_PAGE;
2205         return 0;
2206 }
2207
2208 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2209 {
2210         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2211 }
2212
2213 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2214 {
2215         struct kvm_mmu *context = &vcpu->arch.mmu;
2216
2217         context->new_cr3 = nonpaging_new_cr3;
2218         context->page_fault = tdp_page_fault;
2219         context->free = nonpaging_free;
2220         context->prefetch_page = nonpaging_prefetch_page;
2221         context->sync_page = nonpaging_sync_page;
2222         context->invlpg = nonpaging_invlpg;
2223         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2224         context->root_hpa = INVALID_PAGE;
2225
2226         if (!is_paging(vcpu)) {
2227                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2228                 context->root_level = 0;
2229         } else if (is_long_mode(vcpu)) {
2230                 context->gva_to_gpa = paging64_gva_to_gpa;
2231                 context->root_level = PT64_ROOT_LEVEL;
2232         } else if (is_pae(vcpu)) {
2233                 context->gva_to_gpa = paging64_gva_to_gpa;
2234                 context->root_level = PT32E_ROOT_LEVEL;
2235         } else {
2236                 context->gva_to_gpa = paging32_gva_to_gpa;
2237                 context->root_level = PT32_ROOT_LEVEL;
2238         }
2239
2240         return 0;
2241 }
2242
2243 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2244 {
2245         ASSERT(vcpu);
2246         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2247
2248         if (!is_paging(vcpu))
2249                 return nonpaging_init_context(vcpu);
2250         else if (is_long_mode(vcpu))
2251                 return paging64_init_context(vcpu);
2252         else if (is_pae(vcpu))
2253                 return paging32E_init_context(vcpu);
2254         else
2255                 return paging32_init_context(vcpu);
2256 }
2257
2258 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2259 {
2260         vcpu->arch.update_pte.pfn = bad_pfn;
2261
2262         if (tdp_enabled)
2263                 return init_kvm_tdp_mmu(vcpu);
2264         else
2265                 return init_kvm_softmmu(vcpu);
2266 }
2267
2268 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2269 {
2270         ASSERT(vcpu);
2271         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2272                 vcpu->arch.mmu.free(vcpu);
2273                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2274         }
2275 }
2276
2277 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2278 {
2279         destroy_kvm_mmu(vcpu);
2280         return init_kvm_mmu(vcpu);
2281 }
2282 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2283
2284 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2285 {
2286         int r;
2287
2288         r = mmu_topup_memory_caches(vcpu);
2289         if (r)
2290                 goto out;
2291         spin_lock(&vcpu->kvm->mmu_lock);
2292         kvm_mmu_free_some_pages(vcpu);
2293         mmu_alloc_roots(vcpu);
2294         mmu_sync_roots(vcpu);
2295         spin_unlock(&vcpu->kvm->mmu_lock);
2296         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2297         kvm_mmu_flush_tlb(vcpu);
2298 out:
2299         return r;
2300 }
2301 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2302
2303 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2304 {
2305         mmu_free_roots(vcpu);
2306 }
2307
2308 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2309                                   struct kvm_mmu_page *sp,
2310                                   u64 *spte)
2311 {
2312         u64 pte;
2313         struct kvm_mmu_page *child;
2314
2315         pte = *spte;
2316         if (is_shadow_present_pte(pte)) {
2317                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2318                     is_large_pte(pte))
2319                         rmap_remove(vcpu->kvm, spte);
2320                 else {
2321                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2322                         mmu_page_remove_parent_pte(child, spte);
2323                 }
2324         }
2325         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2326         if (is_large_pte(pte))
2327                 --vcpu->kvm->stat.lpages;
2328 }
2329
2330 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2331                                   struct kvm_mmu_page *sp,
2332                                   u64 *spte,
2333                                   const void *new)
2334 {
2335         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2336                 if (!vcpu->arch.update_pte.largepage ||
2337                     sp->role.glevels == PT32_ROOT_LEVEL) {
2338                         ++vcpu->kvm->stat.mmu_pde_zapped;
2339                         return;
2340                 }
2341         }
2342
2343         ++vcpu->kvm->stat.mmu_pte_updated;
2344         if (sp->role.glevels == PT32_ROOT_LEVEL)
2345                 paging32_update_pte(vcpu, sp, spte, new);
2346         else
2347                 paging64_update_pte(vcpu, sp, spte, new);
2348 }
2349
2350 static bool need_remote_flush(u64 old, u64 new)
2351 {
2352         if (!is_shadow_present_pte(old))
2353                 return false;
2354         if (!is_shadow_present_pte(new))
2355                 return true;
2356         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2357                 return true;
2358         old ^= PT64_NX_MASK;
2359         new ^= PT64_NX_MASK;
2360         return (old & ~new & PT64_PERM_MASK) != 0;
2361 }
2362
2363 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2364 {
2365         if (need_remote_flush(old, new))
2366                 kvm_flush_remote_tlbs(vcpu->kvm);
2367         else
2368                 kvm_mmu_flush_tlb(vcpu);
2369 }
2370
2371 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2372 {
2373         u64 *spte = vcpu->arch.last_pte_updated;
2374
2375         return !!(spte && (*spte & shadow_accessed_mask));
2376 }
2377
2378 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2379                                           const u8 *new, int bytes)
2380 {
2381         gfn_t gfn;
2382         int r;
2383         u64 gpte = 0;
2384         pfn_t pfn;
2385
2386         vcpu->arch.update_pte.largepage = 0;
2387
2388         if (bytes != 4 && bytes != 8)
2389                 return;
2390
2391         /*
2392          * Assume that the pte write on a page table of the same type
2393          * as the current vcpu paging mode.  This is nearly always true
2394          * (might be false while changing modes).  Note it is verified later
2395          * by update_pte().
2396          */
2397         if (is_pae(vcpu)) {
2398                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2399                 if ((bytes == 4) && (gpa % 4 == 0)) {
2400                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2401                         if (r)
2402                                 return;
2403                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2404                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2405                         memcpy((void *)&gpte, new, 8);
2406                 }
2407         } else {
2408                 if ((bytes == 4) && (gpa % 4 == 0))
2409                         memcpy((void *)&gpte, new, 4);
2410         }
2411         if (!is_present_pte(gpte))
2412                 return;
2413         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2414
2415         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2416                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2417                 vcpu->arch.update_pte.largepage = 1;
2418         }
2419         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2420         smp_rmb();
2421         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2422
2423         if (is_error_pfn(pfn)) {
2424                 kvm_release_pfn_clean(pfn);
2425                 return;
2426         }
2427         vcpu->arch.update_pte.gfn = gfn;
2428         vcpu->arch.update_pte.pfn = pfn;
2429 }
2430
2431 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2432 {
2433         u64 *spte = vcpu->arch.last_pte_updated;
2434
2435         if (spte
2436             && vcpu->arch.last_pte_gfn == gfn
2437             && shadow_accessed_mask
2438             && !(*spte & shadow_accessed_mask)
2439             && is_shadow_present_pte(*spte))
2440                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2441 }
2442
2443 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2444                        const u8 *new, int bytes)
2445 {
2446         gfn_t gfn = gpa >> PAGE_SHIFT;
2447         struct kvm_mmu_page *sp;
2448         struct hlist_node *node, *n;
2449         struct hlist_head *bucket;
2450         unsigned index;
2451         u64 entry, gentry;
2452         u64 *spte;
2453         unsigned offset = offset_in_page(gpa);
2454         unsigned pte_size;
2455         unsigned page_offset;
2456         unsigned misaligned;
2457         unsigned quadrant;
2458         int level;
2459         int flooded = 0;
2460         int npte;
2461         int r;
2462
2463         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2464         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2465         spin_lock(&vcpu->kvm->mmu_lock);
2466         kvm_mmu_access_page(vcpu, gfn);
2467         kvm_mmu_free_some_pages(vcpu);
2468         ++vcpu->kvm->stat.mmu_pte_write;
2469         kvm_mmu_audit(vcpu, "pre pte write");
2470         if (gfn == vcpu->arch.last_pt_write_gfn
2471             && !last_updated_pte_accessed(vcpu)) {
2472                 ++vcpu->arch.last_pt_write_count;
2473                 if (vcpu->arch.last_pt_write_count >= 3)
2474                         flooded = 1;
2475         } else {
2476                 vcpu->arch.last_pt_write_gfn = gfn;
2477                 vcpu->arch.last_pt_write_count = 1;
2478                 vcpu->arch.last_pte_updated = NULL;
2479         }
2480         index = kvm_page_table_hashfn(gfn);
2481         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2482         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2483                 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
2484                         continue;
2485                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2486                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2487                 misaligned |= bytes < 4;
2488                 if (misaligned || flooded) {
2489                         /*
2490                          * Misaligned accesses are too much trouble to fix
2491                          * up; also, they usually indicate a page is not used
2492                          * as a page table.
2493                          *
2494                          * If we're seeing too many writes to a page,
2495                          * it may no longer be a page table, or we may be
2496                          * forking, in which case it is better to unmap the
2497                          * page.
2498                          */
2499                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2500                                  gpa, bytes, sp->role.word);
2501                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2502                                 n = bucket->first;
2503                         ++vcpu->kvm->stat.mmu_flooded;
2504                         continue;
2505                 }
2506                 page_offset = offset;
2507                 level = sp->role.level;
2508                 npte = 1;
2509                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2510                         page_offset <<= 1;      /* 32->64 */
2511                         /*
2512                          * A 32-bit pde maps 4MB while the shadow pdes map
2513                          * only 2MB.  So we need to double the offset again
2514                          * and zap two pdes instead of one.
2515                          */
2516                         if (level == PT32_ROOT_LEVEL) {
2517                                 page_offset &= ~7; /* kill rounding error */
2518                                 page_offset <<= 1;
2519                                 npte = 2;
2520                         }
2521                         quadrant = page_offset >> PAGE_SHIFT;
2522                         page_offset &= ~PAGE_MASK;
2523                         if (quadrant != sp->role.quadrant)
2524                                 continue;
2525                 }
2526                 spte = &sp->spt[page_offset / sizeof(*spte)];
2527                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2528                         gentry = 0;
2529                         r = kvm_read_guest_atomic(vcpu->kvm,
2530                                                   gpa & ~(u64)(pte_size - 1),
2531                                                   &gentry, pte_size);
2532                         new = (const void *)&gentry;
2533                         if (r < 0)
2534                                 new = NULL;
2535                 }
2536                 while (npte--) {
2537                         entry = *spte;
2538                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2539                         if (new)
2540                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2541                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2542                         ++spte;
2543                 }
2544         }
2545         kvm_mmu_audit(vcpu, "post pte write");
2546         spin_unlock(&vcpu->kvm->mmu_lock);
2547         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2548                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2549                 vcpu->arch.update_pte.pfn = bad_pfn;
2550         }
2551 }
2552
2553 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2554 {
2555         gpa_t gpa;
2556         int r;
2557
2558         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2559
2560         spin_lock(&vcpu->kvm->mmu_lock);
2561         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2562         spin_unlock(&vcpu->kvm->mmu_lock);
2563         return r;
2564 }
2565 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2566
2567 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2568 {
2569         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2570                 struct kvm_mmu_page *sp;
2571
2572                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2573                                   struct kvm_mmu_page, link);
2574                 kvm_mmu_zap_page(vcpu->kvm, sp);
2575                 ++vcpu->kvm->stat.mmu_recycled;
2576         }
2577 }
2578
2579 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2580 {
2581         int r;
2582         enum emulation_result er;
2583
2584         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2585         if (r < 0)
2586                 goto out;
2587
2588         if (!r) {
2589                 r = 1;
2590                 goto out;
2591         }
2592
2593         r = mmu_topup_memory_caches(vcpu);
2594         if (r)
2595                 goto out;
2596
2597         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2598
2599         switch (er) {
2600         case EMULATE_DONE:
2601                 return 1;
2602         case EMULATE_DO_MMIO:
2603                 ++vcpu->stat.mmio_exits;
2604                 return 0;
2605         case EMULATE_FAIL:
2606                 kvm_report_emulation_failure(vcpu, "pagetable");
2607                 return 1;
2608         default:
2609                 BUG();
2610         }
2611 out:
2612         return r;
2613 }
2614 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2615
2616 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2617 {
2618         spin_lock(&vcpu->kvm->mmu_lock);
2619         vcpu->arch.mmu.invlpg(vcpu, gva);
2620         spin_unlock(&vcpu->kvm->mmu_lock);
2621         kvm_mmu_flush_tlb(vcpu);
2622         ++vcpu->stat.invlpg;
2623 }
2624 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2625
2626 void kvm_enable_tdp(void)
2627 {
2628         tdp_enabled = true;
2629 }
2630 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2631
2632 void kvm_disable_tdp(void)
2633 {
2634         tdp_enabled = false;
2635 }
2636 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2637
2638 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2639 {
2640         struct kvm_mmu_page *sp;
2641
2642         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2643                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2644                                   struct kvm_mmu_page, link);
2645                 kvm_mmu_zap_page(vcpu->kvm, sp);
2646                 cond_resched();
2647         }
2648         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2649 }
2650
2651 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2652 {
2653         struct page *page;
2654         int i;
2655
2656         ASSERT(vcpu);
2657
2658         if (vcpu->kvm->arch.n_requested_mmu_pages)
2659                 vcpu->kvm->arch.n_free_mmu_pages =
2660                                         vcpu->kvm->arch.n_requested_mmu_pages;
2661         else
2662                 vcpu->kvm->arch.n_free_mmu_pages =
2663                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2664         /*
2665          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2666          * Therefore we need to allocate shadow page tables in the first
2667          * 4GB of memory, which happens to fit the DMA32 zone.
2668          */
2669         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2670         if (!page)
2671                 goto error_1;
2672         vcpu->arch.mmu.pae_root = page_address(page);
2673         for (i = 0; i < 4; ++i)
2674                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2675
2676         return 0;
2677
2678 error_1:
2679         free_mmu_pages(vcpu);
2680         return -ENOMEM;
2681 }
2682
2683 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2684 {
2685         ASSERT(vcpu);
2686         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2687
2688         return alloc_mmu_pages(vcpu);
2689 }
2690
2691 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2692 {
2693         ASSERT(vcpu);
2694         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2695
2696         return init_kvm_mmu(vcpu);
2697 }
2698
2699 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2700 {
2701         ASSERT(vcpu);
2702
2703         destroy_kvm_mmu(vcpu);
2704         free_mmu_pages(vcpu);
2705         mmu_free_memory_caches(vcpu);
2706 }
2707
2708 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2709 {
2710         struct kvm_mmu_page *sp;
2711
2712         spin_lock(&kvm->mmu_lock);
2713         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2714                 int i;
2715                 u64 *pt;
2716
2717                 if (!test_bit(slot, sp->slot_bitmap))
2718                         continue;
2719
2720                 pt = sp->spt;
2721                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2722                         /* avoid RMW */
2723                         if (pt[i] & PT_WRITABLE_MASK)
2724                                 pt[i] &= ~PT_WRITABLE_MASK;
2725         }
2726         kvm_flush_remote_tlbs(kvm);
2727         spin_unlock(&kvm->mmu_lock);
2728 }
2729
2730 void kvm_mmu_zap_all(struct kvm *kvm)
2731 {
2732         struct kvm_mmu_page *sp, *node;
2733
2734         spin_lock(&kvm->mmu_lock);
2735         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2736                 if (kvm_mmu_zap_page(kvm, sp))
2737                         node = container_of(kvm->arch.active_mmu_pages.next,
2738                                             struct kvm_mmu_page, link);
2739         spin_unlock(&kvm->mmu_lock);
2740
2741         kvm_flush_remote_tlbs(kvm);
2742 }
2743
2744 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2745 {
2746         struct kvm_mmu_page *page;
2747
2748         page = container_of(kvm->arch.active_mmu_pages.prev,
2749                             struct kvm_mmu_page, link);
2750         kvm_mmu_zap_page(kvm, page);
2751 }
2752
2753 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2754 {
2755         struct kvm *kvm;
2756         struct kvm *kvm_freed = NULL;
2757         int cache_count = 0;
2758
2759         spin_lock(&kvm_lock);
2760
2761         list_for_each_entry(kvm, &vm_list, vm_list) {
2762                 int npages;
2763
2764                 if (!down_read_trylock(&kvm->slots_lock))
2765                         continue;
2766                 spin_lock(&kvm->mmu_lock);
2767                 npages = kvm->arch.n_alloc_mmu_pages -
2768                          kvm->arch.n_free_mmu_pages;
2769                 cache_count += npages;
2770                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2771                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2772                         cache_count--;
2773                         kvm_freed = kvm;
2774                 }
2775                 nr_to_scan--;
2776
2777                 spin_unlock(&kvm->mmu_lock);
2778                 up_read(&kvm->slots_lock);
2779         }
2780         if (kvm_freed)
2781                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2782
2783         spin_unlock(&kvm_lock);
2784
2785         return cache_count;
2786 }
2787
2788 static struct shrinker mmu_shrinker = {
2789         .shrink = mmu_shrink,
2790         .seeks = DEFAULT_SEEKS * 10,
2791 };
2792
2793 static void mmu_destroy_caches(void)
2794 {
2795         if (pte_chain_cache)
2796                 kmem_cache_destroy(pte_chain_cache);
2797         if (rmap_desc_cache)
2798                 kmem_cache_destroy(rmap_desc_cache);
2799         if (mmu_page_header_cache)
2800                 kmem_cache_destroy(mmu_page_header_cache);
2801 }
2802
2803 void kvm_mmu_module_exit(void)
2804 {
2805         mmu_destroy_caches();
2806         unregister_shrinker(&mmu_shrinker);
2807 }
2808
2809 int kvm_mmu_module_init(void)
2810 {
2811         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2812                                             sizeof(struct kvm_pte_chain),
2813                                             0, 0, NULL);
2814         if (!pte_chain_cache)
2815                 goto nomem;
2816         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2817                                             sizeof(struct kvm_rmap_desc),
2818                                             0, 0, NULL);
2819         if (!rmap_desc_cache)
2820                 goto nomem;
2821
2822         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2823                                                   sizeof(struct kvm_mmu_page),
2824                                                   0, 0, NULL);
2825         if (!mmu_page_header_cache)
2826                 goto nomem;
2827
2828         register_shrinker(&mmu_shrinker);
2829
2830         return 0;
2831
2832 nomem:
2833         mmu_destroy_caches();
2834         return -ENOMEM;
2835 }
2836
2837 /*
2838  * Caculate mmu pages needed for kvm.
2839  */
2840 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2841 {
2842         int i;
2843         unsigned int nr_mmu_pages;
2844         unsigned int  nr_pages = 0;
2845
2846         for (i = 0; i < kvm->nmemslots; i++)
2847                 nr_pages += kvm->memslots[i].npages;
2848
2849         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2850         nr_mmu_pages = max(nr_mmu_pages,
2851                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2852
2853         return nr_mmu_pages;
2854 }
2855
2856 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2857                                 unsigned len)
2858 {
2859         if (len > buffer->len)
2860                 return NULL;
2861         return buffer->ptr;
2862 }
2863
2864 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2865                                 unsigned len)
2866 {
2867         void *ret;
2868
2869         ret = pv_mmu_peek_buffer(buffer, len);
2870         if (!ret)
2871                 return ret;
2872         buffer->ptr += len;
2873         buffer->len -= len;
2874         buffer->processed += len;
2875         return ret;
2876 }
2877
2878 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2879                              gpa_t addr, gpa_t value)
2880 {
2881         int bytes = 8;
2882         int r;
2883
2884         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2885                 bytes = 4;
2886
2887         r = mmu_topup_memory_caches(vcpu);
2888         if (r)
2889                 return r;
2890
2891         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2892                 return -EFAULT;
2893
2894         return 1;
2895 }
2896
2897 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2898 {
2899         kvm_x86_ops->tlb_flush(vcpu);
2900         set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
2901         return 1;
2902 }
2903
2904 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2905 {
2906         spin_lock(&vcpu->kvm->mmu_lock);
2907         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2908         spin_unlock(&vcpu->kvm->mmu_lock);
2909         return 1;
2910 }
2911
2912 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2913                              struct kvm_pv_mmu_op_buffer *buffer)
2914 {
2915         struct kvm_mmu_op_header *header;
2916
2917         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2918         if (!header)
2919                 return 0;
2920         switch (header->op) {
2921         case KVM_MMU_OP_WRITE_PTE: {
2922                 struct kvm_mmu_op_write_pte *wpte;
2923
2924                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2925                 if (!wpte)
2926                         return 0;
2927                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2928                                         wpte->pte_val);
2929         }
2930         case KVM_MMU_OP_FLUSH_TLB: {
2931                 struct kvm_mmu_op_flush_tlb *ftlb;
2932
2933                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2934                 if (!ftlb)
2935                         return 0;
2936                 return kvm_pv_mmu_flush_tlb(vcpu);
2937         }
2938         case KVM_MMU_OP_RELEASE_PT: {
2939                 struct kvm_mmu_op_release_pt *rpt;
2940
2941                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2942                 if (!rpt)
2943                         return 0;
2944                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2945         }
2946         default: return 0;
2947         }
2948 }
2949
2950 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2951                   gpa_t addr, unsigned long *ret)
2952 {
2953         int r;
2954         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2955
2956         buffer->ptr = buffer->buf;
2957         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
2958         buffer->processed = 0;
2959
2960         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
2961         if (r)
2962                 goto out;
2963
2964         while (buffer->len) {
2965                 r = kvm_pv_mmu_op_one(vcpu, buffer);
2966                 if (r < 0)
2967                         goto out;
2968                 if (r == 0)
2969                         break;
2970         }
2971
2972         r = 1;
2973 out:
2974         *ret = buffer->processed;
2975         return r;
2976 }
2977
2978 #ifdef AUDIT
2979
2980 static const char *audit_msg;
2981
2982 static gva_t canonicalize(gva_t gva)
2983 {
2984 #ifdef CONFIG_X86_64
2985         gva = (long long)(gva << 16) >> 16;
2986 #endif
2987         return gva;
2988 }
2989
2990 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2991                                 gva_t va, int level)
2992 {
2993         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2994         int i;
2995         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2996
2997         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2998                 u64 ent = pt[i];
2999
3000                 if (ent == shadow_trap_nonpresent_pte)
3001                         continue;
3002
3003                 va = canonicalize(va);
3004                 if (level > 1) {
3005                         if (ent == shadow_notrap_nonpresent_pte)
3006                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
3007                                        " in nonleaf level: levels %d gva %lx"
3008                                        " level %d pte %llx\n", audit_msg,
3009                                        vcpu->arch.mmu.root_level, va, level, ent);
3010
3011                         audit_mappings_page(vcpu, ent, va, level - 1);
3012                 } else {
3013                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3014                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
3015
3016                         if (is_shadow_present_pte(ent)
3017                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3018                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3019                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3020                                        audit_msg, vcpu->arch.mmu.root_level,
3021                                        va, gpa, hpa, ent,
3022                                        is_shadow_present_pte(ent));
3023                         else if (ent == shadow_notrap_nonpresent_pte
3024                                  && !is_error_hpa(hpa))
3025                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3026                                        " valid guest gva %lx\n", audit_msg, va);
3027                         kvm_release_pfn_clean(pfn);
3028
3029                 }
3030         }
3031 }
3032
3033 static void audit_mappings(struct kvm_vcpu *vcpu)
3034 {
3035         unsigned i;
3036
3037         if (vcpu->arch.mmu.root_level == 4)
3038                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3039         else
3040                 for (i = 0; i < 4; ++i)
3041                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3042                                 audit_mappings_page(vcpu,
3043                                                     vcpu->arch.mmu.pae_root[i],
3044                                                     i << 30,
3045                                                     2);
3046 }
3047
3048 static int count_rmaps(struct kvm_vcpu *vcpu)
3049 {
3050         int nmaps = 0;
3051         int i, j, k;
3052
3053         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3054                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3055                 struct kvm_rmap_desc *d;
3056
3057                 for (j = 0; j < m->npages; ++j) {
3058                         unsigned long *rmapp = &m->rmap[j];
3059
3060                         if (!*rmapp)
3061                                 continue;
3062                         if (!(*rmapp & 1)) {
3063                                 ++nmaps;
3064                                 continue;
3065                         }
3066                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3067                         while (d) {
3068                                 for (k = 0; k < RMAP_EXT; ++k)
3069                                         if (d->shadow_ptes[k])
3070                                                 ++nmaps;
3071                                         else
3072                                                 break;
3073                                 d = d->more;
3074                         }
3075                 }
3076         }
3077         return nmaps;
3078 }
3079
3080 static int count_writable_mappings(struct kvm_vcpu *vcpu)
3081 {
3082         int nmaps = 0;
3083         struct kvm_mmu_page *sp;
3084         int i;
3085
3086         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3087                 u64 *pt = sp->spt;
3088
3089                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3090                         continue;
3091
3092                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3093                         u64 ent = pt[i];
3094
3095                         if (!(ent & PT_PRESENT_MASK))
3096                                 continue;
3097                         if (!(ent & PT_WRITABLE_MASK))
3098                                 continue;
3099                         ++nmaps;
3100                 }
3101         }
3102         return nmaps;
3103 }
3104
3105 static void audit_rmap(struct kvm_vcpu *vcpu)
3106 {
3107         int n_rmap = count_rmaps(vcpu);
3108         int n_actual = count_writable_mappings(vcpu);
3109
3110         if (n_rmap != n_actual)
3111                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
3112                        __func__, audit_msg, n_rmap, n_actual);
3113 }
3114
3115 static void audit_write_protection(struct kvm_vcpu *vcpu)
3116 {
3117         struct kvm_mmu_page *sp;
3118         struct kvm_memory_slot *slot;
3119         unsigned long *rmapp;
3120         gfn_t gfn;
3121
3122         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3123                 if (sp->role.metaphysical)
3124                         continue;
3125
3126                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3127                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3128                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3129                 if (*rmapp)
3130                         printk(KERN_ERR "%s: (%s) shadow page has writable"
3131                                " mappings: gfn %lx role %x\n",
3132                                __func__, audit_msg, sp->gfn,
3133                                sp->role.word);
3134         }
3135 }
3136
3137 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3138 {
3139         int olddbg = dbg;
3140
3141         dbg = 0;
3142         audit_msg = msg;
3143         audit_rmap(vcpu);
3144         audit_write_protection(vcpu);
3145         audit_mappings(vcpu);
3146         dbg = olddbg;
3147 }
3148
3149 #endif