SAFE public projects git trees. - safe/jmp/linux-2.6/blob - arch/x86/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "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_RSVD_MASK (1U << 3)
 130 #define PFERR_FETCH_MASK (1U << 4)
 131
 132 #define PT_DIRECTORY_LEVEL 2
 133 #define PT_PAGE_TABLE_LEVEL 1
 134
 135 #define RMAP_EXT 4
 136
 137 #define ACC_EXEC_MASK    1
 138 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 139 #define ACC_USER_MASK    PT_USER_MASK
 140 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 141
 142 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 143
 144 struct kvm_rmap_desc {
 145         u64 *shadow_ptes[RMAP_EXT];
 146         struct kvm_rmap_desc *more;
 147 };
 148
 149 struct kvm_shadow_walk_iterator {
 150         u64 addr;
 151         hpa_t shadow_addr;
 152         int level;
 153         u64 *sptep;
 154         unsigned index;
 155 };
 156
 157 #define for_each_shadow_entry(_vcpu, _addr, _walker)    \
 158         for (shadow_walk_init(&(_walker), _vcpu, _addr);        \
 159              shadow_walk_okay(&(_walker));                      \
 160              shadow_walk_next(&(_walker)))
 161
 162
 163 struct kvm_unsync_walk {
 164         int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
 165 };
 166
 167 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
 168
 169 static struct kmem_cache *pte_chain_cache;
 170 static struct kmem_cache *rmap_desc_cache;
 171 static struct kmem_cache *mmu_page_header_cache;
 172
 173 static u64 __read_mostly shadow_trap_nonpresent_pte;
 174 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 175 static u64 __read_mostly shadow_base_present_pte;
 176 static u64 __read_mostly shadow_nx_mask;
 177 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 178 static u64 __read_mostly shadow_user_mask;
 179 static u64 __read_mostly shadow_accessed_mask;
 180 static u64 __read_mostly shadow_dirty_mask;
 181 static u64 __read_mostly shadow_mt_mask;
 182
 183 static inline u64 rsvd_bits(int s, int e)
 184 {
 185         return ((1ULL << (e - s + 1)) - 1) << s;
 186 }
 187
 188 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 189 {
 190         shadow_trap_nonpresent_pte = trap_pte;
 191         shadow_notrap_nonpresent_pte = notrap_pte;
 192 }
 193 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 194
 195 void kvm_mmu_set_base_ptes(u64 base_pte)
 196 {
 197         shadow_base_present_pte = base_pte;
 198 }
 199 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 200
 201 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 202                 u64 dirty_mask, u64 nx_mask, u64 x_mask, u64 mt_mask)
 203 {
 204         shadow_user_mask = user_mask;
 205         shadow_accessed_mask = accessed_mask;
 206         shadow_dirty_mask = dirty_mask;
 207         shadow_nx_mask = nx_mask;
 208         shadow_x_mask = x_mask;
 209         shadow_mt_mask = mt_mask;
 210 }
 211 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 212
 213 static int is_write_protection(struct kvm_vcpu *vcpu)
 214 {
 215         return vcpu->arch.cr0 & X86_CR0_WP;
 216 }
 217
 218 static int is_cpuid_PSE36(void)
 219 {
 220         return 1;
 221 }
 222
 223 static int is_nx(struct kvm_vcpu *vcpu)
 224 {
 225         return vcpu->arch.shadow_efer & EFER_NX;
 226 }
 227
 228 static int is_shadow_present_pte(u64 pte)
 229 {
 230         return pte != shadow_trap_nonpresent_pte
 231                 && pte != shadow_notrap_nonpresent_pte;
 232 }
 233
 234 static int is_large_pte(u64 pte)
 235 {
 236         return pte & PT_PAGE_SIZE_MASK;
 237 }
 238
 239 static int is_writeble_pte(unsigned long pte)
 240 {
 241         return pte & PT_WRITABLE_MASK;
 242 }
 243
 244 static int is_dirty_pte(unsigned long pte)
 245 {
 246         return pte & shadow_dirty_mask;
 247 }
 248
 249 static int is_rmap_pte(u64 pte)
 250 {
 251         return is_shadow_present_pte(pte);
 252 }
 253
 254 static pfn_t spte_to_pfn(u64 pte)
 255 {
 256         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 257 }
 258
 259 static gfn_t pse36_gfn_delta(u32 gpte)
 260 {
 261         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 262
 263         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 264 }
 265
 266 static void set_shadow_pte(u64 *sptep, u64 spte)
 267 {
 268 #ifdef CONFIG_X86_64
 269         set_64bit((unsigned long *)sptep, spte);
 270 #else
 271         set_64bit((unsigned long long *)sptep, spte);
 272 #endif
 273 }
 274
 275 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 276                                   struct kmem_cache *base_cache, int min)
 277 {
 278         void *obj;
 279
 280         if (cache->nobjs >= min)
 281                 return 0;
 282         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 283                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 284                 if (!obj)
 285                         return -ENOMEM;
 286                 cache->objects[cache->nobjs++] = obj;
 287         }
 288         return 0;
 289 }
 290
 291 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 292 {
 293         while (mc->nobjs)
 294                 kfree(mc->objects[--mc->nobjs]);
 295 }
 296
 297 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 298                                        int min)
 299 {
 300         struct page *page;
 301
 302         if (cache->nobjs >= min)
 303                 return 0;
 304         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 305                 page = alloc_page(GFP_KERNEL);
 306                 if (!page)
 307                         return -ENOMEM;
 308                 set_page_private(page, 0);
 309                 cache->objects[cache->nobjs++] = page_address(page);
 310         }
 311         return 0;
 312 }
 313
 314 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 315 {
 316         while (mc->nobjs)
 317                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 318 }
 319
 320 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 321 {
 322         int r;
 323
 324         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 325                                    pte_chain_cache, 4);
 326         if (r)
 327                 goto out;
 328         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 329                                    rmap_desc_cache, 4);
 330         if (r)
 331                 goto out;
 332         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 333         if (r)
 334                 goto out;
 335         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 336                                    mmu_page_header_cache, 4);
 337 out:
 338         return r;
 339 }
 340
 341 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 342 {
 343         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 344         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 345         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 346         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 347 }
 348
 349 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 350                                     size_t size)
 351 {
 352         void *p;
 353
 354         BUG_ON(!mc->nobjs);
 355         p = mc->objects[--mc->nobjs];
 356         return p;
 357 }
 358
 359 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 360 {
 361         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 362                                       sizeof(struct kvm_pte_chain));
 363 }
 364
 365 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 366 {
 367         kfree(pc);
 368 }
 369
 370 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 371 {
 372         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 373                                       sizeof(struct kvm_rmap_desc));
 374 }
 375
 376 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 377 {
 378         kfree(rd);
 379 }
 380
 381 /*
 382  * Return the pointer to the largepage write count for a given
 383  * gfn, handling slots that are not large page aligned.
 384  */
 385 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 386 {
 387         unsigned long idx;
 388
 389         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 390               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 391         return &slot->lpage_info[idx].write_count;
 392 }
 393
 394 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 395 {
 396         int *write_count;
 397
 398         gfn = unalias_gfn(kvm, gfn);
 399         write_count = slot_largepage_idx(gfn,
 400                                          gfn_to_memslot_unaliased(kvm, gfn));
 401         *write_count += 1;
 402 }
 403
 404 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 405 {
 406         int *write_count;
 407
 408         gfn = unalias_gfn(kvm, gfn);
 409         write_count = slot_largepage_idx(gfn,
 410                                          gfn_to_memslot_unaliased(kvm, gfn));
 411         *write_count -= 1;
 412         WARN_ON(*write_count < 0);
 413 }
 414
 415 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 416 {
 417         struct kvm_memory_slot *slot;
 418         int *largepage_idx;
 419
 420         gfn = unalias_gfn(kvm, gfn);
 421         slot = gfn_to_memslot_unaliased(kvm, gfn);
 422         if (slot) {
 423                 largepage_idx = slot_largepage_idx(gfn, slot);
 424                 return *largepage_idx;
 425         }
 426
 427         return 1;
 428 }
 429
 430 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 431 {
 432         struct vm_area_struct *vma;
 433         unsigned long addr;
 434         int ret = 0;
 435
 436         addr = gfn_to_hva(kvm, gfn);
 437         if (kvm_is_error_hva(addr))
 438                 return ret;
 439
 440         down_read(&current->mm->mmap_sem);
 441         vma = find_vma(current->mm, addr);
 442         if (vma && is_vm_hugetlb_page(vma))
 443                 ret = 1;
 444         up_read(&current->mm->mmap_sem);
 445
 446         return ret;
 447 }
 448
 449 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 450 {
 451         struct kvm_memory_slot *slot;
 452
 453         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 454                 return 0;
 455
 456         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 457                 return 0;
 458
 459         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 460         if (slot && slot->dirty_bitmap)
 461                 return 0;
 462
 463         return 1;
 464 }
 465
 466 /*
 467  * Take gfn and return the reverse mapping to it.
 468  * Note: gfn must be unaliased before this function get called
 469  */
 470
 471 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 472 {
 473         struct kvm_memory_slot *slot;
 474         unsigned long idx;
 475
 476         slot = gfn_to_memslot(kvm, gfn);
 477         if (!lpage)
 478                 return &slot->rmap[gfn - slot->base_gfn];
 479
 480         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 481               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 482
 483         return &slot->lpage_info[idx].rmap_pde;
 484 }
 485
 486 /*
 487  * Reverse mapping data structures:
 488  *
 489  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 490  * that points to page_address(page).
 491  *
 492  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 493  * containing more mappings.
 494  */
 495 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 496 {
 497         struct kvm_mmu_page *sp;
 498         struct kvm_rmap_desc *desc;
 499         unsigned long *rmapp;
 500         int i;
 501
 502         if (!is_rmap_pte(*spte))
 503                 return;
 504         gfn = unalias_gfn(vcpu->kvm, gfn);
 505         sp = page_header(__pa(spte));
 506         sp->gfns[spte - sp->spt] = gfn;
 507         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 508         if (!*rmapp) {
 509                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 510                 *rmapp = (unsigned long)spte;
 511         } else if (!(*rmapp & 1)) {
 512                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 513                 desc = mmu_alloc_rmap_desc(vcpu);
 514                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 515                 desc->shadow_ptes[1] = spte;
 516                 *rmapp = (unsigned long)desc | 1;
 517         } else {
 518                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 519                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 520                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 521                         desc = desc->more;
 522                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 523                         desc->more = mmu_alloc_rmap_desc(vcpu);
 524                         desc = desc->more;
 525                 }
 526                 for (i = 0; desc->shadow_ptes[i]; ++i)
 527                         ;
 528                 desc->shadow_ptes[i] = spte;
 529         }
 530 }
 531
 532 static void rmap_desc_remove_entry(unsigned long *rmapp,
 533                                    struct kvm_rmap_desc *desc,
 534                                    int i,
 535                                    struct kvm_rmap_desc *prev_desc)
 536 {
 537         int j;
 538
 539         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 540                 ;
 541         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 542         desc->shadow_ptes[j] = NULL;
 543         if (j != 0)
 544                 return;
 545         if (!prev_desc && !desc->more)
 546                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 547         else
 548                 if (prev_desc)
 549                         prev_desc->more = desc->more;
 550                 else
 551                         *rmapp = (unsigned long)desc->more | 1;
 552         mmu_free_rmap_desc(desc);
 553 }
 554
 555 static void rmap_remove(struct kvm *kvm, u64 *spte)
 556 {
 557         struct kvm_rmap_desc *desc;
 558         struct kvm_rmap_desc *prev_desc;
 559         struct kvm_mmu_page *sp;
 560         pfn_t pfn;
 561         unsigned long *rmapp;
 562         int i;
 563
 564         if (!is_rmap_pte(*spte))
 565                 return;
 566         sp = page_header(__pa(spte));
 567         pfn = spte_to_pfn(*spte);
 568         if (*spte & shadow_accessed_mask)
 569                 kvm_set_pfn_accessed(pfn);
 570         if (is_writeble_pte(*spte))
 571                 kvm_release_pfn_dirty(pfn);
 572         else
 573                 kvm_release_pfn_clean(pfn);
 574         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 575         if (!*rmapp) {
 576                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 577                 BUG();
 578         } else if (!(*rmapp & 1)) {
 579                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 580                 if ((u64 *)*rmapp != spte) {
 581                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 582                                spte, *spte);
 583                         BUG();
 584                 }
 585                 *rmapp = 0;
 586         } else {
 587                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 588                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 589                 prev_desc = NULL;
 590                 while (desc) {
 591                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 592                                 if (desc->shadow_ptes[i] == spte) {
 593                                         rmap_desc_remove_entry(rmapp,
 594                                                                desc, i,
 595                                                                prev_desc);
 596                                         return;
 597                                 }
 598                         prev_desc = desc;
 599                         desc = desc->more;
 600                 }
 601                 BUG();
 602         }
 603 }
 604
 605 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 606 {
 607         struct kvm_rmap_desc *desc;
 608         struct kvm_rmap_desc *prev_desc;
 609         u64 *prev_spte;
 610         int i;
 611
 612         if (!*rmapp)
 613                 return NULL;
 614         else if (!(*rmapp & 1)) {
 615                 if (!spte)
 616                         return (u64 *)*rmapp;
 617                 return NULL;
 618         }
 619         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 620         prev_desc = NULL;
 621         prev_spte = NULL;
 622         while (desc) {
 623                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 624                         if (prev_spte == spte)
 625                                 return desc->shadow_ptes[i];
 626                         prev_spte = desc->shadow_ptes[i];
 627                 }
 628                 desc = desc->more;
 629         }
 630         return NULL;
 631 }
 632
 633 static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 634 {
 635         unsigned long *rmapp;
 636         u64 *spte;
 637         int write_protected = 0;
 638
 639         gfn = unalias_gfn(kvm, gfn);
 640         rmapp = gfn_to_rmap(kvm, gfn, 0);
 641
 642         spte = rmap_next(kvm, rmapp, NULL);
 643         while (spte) {
 644                 BUG_ON(!spte);
 645                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 646                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 647                 if (is_writeble_pte(*spte)) {
 648                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 649                         write_protected = 1;
 650                 }
 651                 spte = rmap_next(kvm, rmapp, spte);
 652         }
 653         if (write_protected) {
 654                 pfn_t pfn;
 655
 656                 spte = rmap_next(kvm, rmapp, NULL);
 657                 pfn = spte_to_pfn(*spte);
 658                 kvm_set_pfn_dirty(pfn);
 659         }
 660
 661         /* check for huge page mappings */
 662         rmapp = gfn_to_rmap(kvm, gfn, 1);
 663         spte = rmap_next(kvm, rmapp, NULL);
 664         while (spte) {
 665                 BUG_ON(!spte);
 666                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 667                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 668                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 669                 if (is_writeble_pte(*spte)) {
 670                         rmap_remove(kvm, spte);
 671                         --kvm->stat.lpages;
 672                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 673                         spte = NULL;
 674                         write_protected = 1;
 675                 }
 676                 spte = rmap_next(kvm, rmapp, spte);
 677         }
 678
 679         return write_protected;
 680 }
 681
 682 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
 683 {
 684         u64 *spte;
 685         int need_tlb_flush = 0;
 686
 687         while ((spte = rmap_next(kvm, rmapp, NULL))) {
 688                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 689                 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
 690                 rmap_remove(kvm, spte);
 691                 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 692                 need_tlb_flush = 1;
 693         }
 694         return need_tlb_flush;
 695 }
 696
 697 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 698                           int (*handler)(struct kvm *kvm, unsigned long *rmapp))
 699 {
 700         int i;
 701         int retval = 0;
 702
 703         /*
 704          * If mmap_sem isn't taken, we can look the memslots with only
 705          * the mmu_lock by skipping over the slots with userspace_addr == 0.
 706          */
 707         for (i = 0; i < kvm->nmemslots; i++) {
 708                 struct kvm_memory_slot *memslot = &kvm->memslots[i];
 709                 unsigned long start = memslot->userspace_addr;
 710                 unsigned long end;
 711
 712                 /* mmu_lock protects userspace_addr */
 713                 if (!start)
 714                         continue;
 715
 716                 end = start + (memslot->npages << PAGE_SHIFT);
 717                 if (hva >= start && hva < end) {
 718                         gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
 719                         retval |= handler(kvm, &memslot->rmap[gfn_offset]);
 720                         retval |= handler(kvm,
 721                                           &memslot->lpage_info[
 722                                                   gfn_offset /
 723                                                   KVM_PAGES_PER_HPAGE].rmap_pde);
 724                 }
 725         }
 726
 727         return retval;
 728 }
 729
 730 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 731 {
 732         return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 733 }
 734
 735 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
 736 {
 737         u64 *spte;
 738         int young = 0;
 739
 740         /* always return old for EPT */
 741         if (!shadow_accessed_mask)
 742                 return 0;
 743
 744         spte = rmap_next(kvm, rmapp, NULL);
 745         while (spte) {
 746                 int _young;
 747                 u64 _spte = *spte;
 748                 BUG_ON(!(_spte & PT_PRESENT_MASK));
 749                 _young = _spte & PT_ACCESSED_MASK;
 750                 if (_young) {
 751                         young = 1;
 752                         clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
 753                 }
 754                 spte = rmap_next(kvm, rmapp, spte);
 755         }
 756         return young;
 757 }
 758
 759 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 760 {
 761         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 762 }
 763
 764 #ifdef MMU_DEBUG
 765 static int is_empty_shadow_page(u64 *spt)
 766 {
 767         u64 *pos;
 768         u64 *end;
 769
 770         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 771                 if (is_shadow_present_pte(*pos)) {
 772                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 773                                pos, *pos);
 774                         return 0;
 775                 }
 776         return 1;
 777 }
 778 #endif
 779
 780 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 781 {
 782         ASSERT(is_empty_shadow_page(sp->spt));
 783         list_del(&sp->link);
 784         __free_page(virt_to_page(sp->spt));
 785         __free_page(virt_to_page(sp->gfns));
 786         kfree(sp);
 787         ++kvm->arch.n_free_mmu_pages;
 788 }
 789
 790 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 791 {
 792         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 793 }
 794
 795 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 796                                                u64 *parent_pte)
 797 {
 798         struct kvm_mmu_page *sp;
 799
 800         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 801         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 802         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 803         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 804         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 805         INIT_LIST_HEAD(&sp->oos_link);
 806         bitmap_zero(sp->slot_bitmap, KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS);
 807         sp->multimapped = 0;
 808         sp->parent_pte = parent_pte;
 809         --vcpu->kvm->arch.n_free_mmu_pages;
 810         return sp;
 811 }
 812
 813 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 814                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 815 {
 816         struct kvm_pte_chain *pte_chain;
 817         struct hlist_node *node;
 818         int i;
 819
 820         if (!parent_pte)
 821                 return;
 822         if (!sp->multimapped) {
 823                 u64 *old = sp->parent_pte;
 824
 825                 if (!old) {
 826                         sp->parent_pte = parent_pte;
 827                         return;
 828                 }
 829                 sp->multimapped = 1;
 830                 pte_chain = mmu_alloc_pte_chain(vcpu);
 831                 INIT_HLIST_HEAD(&sp->parent_ptes);
 832                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 833                 pte_chain->parent_ptes[0] = old;
 834         }
 835         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 836                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 837                         continue;
 838                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 839                         if (!pte_chain->parent_ptes[i]) {
 840                                 pte_chain->parent_ptes[i] = parent_pte;
 841                                 return;
 842                         }
 843         }
 844         pte_chain = mmu_alloc_pte_chain(vcpu);
 845         BUG_ON(!pte_chain);
 846         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 847         pte_chain->parent_ptes[0] = parent_pte;
 848 }
 849
 850 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 851                                        u64 *parent_pte)
 852 {
 853         struct kvm_pte_chain *pte_chain;
 854         struct hlist_node *node;
 855         int i;
 856
 857         if (!sp->multimapped) {
 858                 BUG_ON(sp->parent_pte != parent_pte);
 859                 sp->parent_pte = NULL;
 860                 return;
 861         }
 862         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 863                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 864                         if (!pte_chain->parent_ptes[i])
 865                                 break;
 866                         if (pte_chain->parent_ptes[i] != parent_pte)
 867                                 continue;
 868                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 869                                 && pte_chain->parent_ptes[i + 1]) {
 870                                 pte_chain->parent_ptes[i]
 871                                         = pte_chain->parent_ptes[i + 1];
 872                                 ++i;
 873                         }
 874                         pte_chain->parent_ptes[i] = NULL;
 875                         if (i == 0) {
 876                                 hlist_del(&pte_chain->link);
 877                                 mmu_free_pte_chain(pte_chain);
 878                                 if (hlist_empty(&sp->parent_ptes)) {
 879                                         sp->multimapped = 0;
 880                                         sp->parent_pte = NULL;
 881                                 }
 882                         }
 883                         return;
 884                 }
 885         BUG();
 886 }
 887
 888
 889 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 890                             mmu_parent_walk_fn fn)
 891 {
 892         struct kvm_pte_chain *pte_chain;
 893         struct hlist_node *node;
 894         struct kvm_mmu_page *parent_sp;
 895         int i;
 896
 897         if (!sp->multimapped && sp->parent_pte) {
 898                 parent_sp = page_header(__pa(sp->parent_pte));
 899                 fn(vcpu, parent_sp);
 900                 mmu_parent_walk(vcpu, parent_sp, fn);
 901                 return;
 902         }
 903         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 904                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 905                         if (!pte_chain->parent_ptes[i])
 906                                 break;
 907                         parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
 908                         fn(vcpu, parent_sp);
 909                         mmu_parent_walk(vcpu, parent_sp, fn);
 910                 }
 911 }
 912
 913 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
 914 {
 915         unsigned int index;
 916         struct kvm_mmu_page *sp = page_header(__pa(spte));
 917
 918         index = spte - sp->spt;
 919         if (!__test_and_set_bit(index, sp->unsync_child_bitmap))
 920                 sp->unsync_children++;
 921         WARN_ON(!sp->unsync_children);
 922 }
 923
 924 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
 925 {
 926         struct kvm_pte_chain *pte_chain;
 927         struct hlist_node *node;
 928         int i;
 929
 930         if (!sp->parent_pte)
 931                 return;
 932
 933         if (!sp->multimapped) {
 934                 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
 935                 return;
 936         }
 937
 938         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 939                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 940                         if (!pte_chain->parent_ptes[i])
 941                                 break;
 942                         kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
 943                 }
 944 }
 945
 946 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
 947 {
 948         kvm_mmu_update_parents_unsync(sp);
 949         return 1;
 950 }
 951
 952 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
 953                                         struct kvm_mmu_page *sp)
 954 {
 955         mmu_parent_walk(vcpu, sp, unsync_walk_fn);
 956         kvm_mmu_update_parents_unsync(sp);
 957 }
 958
 959 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 960                                     struct kvm_mmu_page *sp)
 961 {
 962         int i;
 963
 964         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 965                 sp->spt[i] = shadow_trap_nonpresent_pte;
 966 }
 967
 968 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
 969                                struct kvm_mmu_page *sp)
 970 {
 971         return 1;
 972 }
 973
 974 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
 975 {
 976 }
 977
 978 #define KVM_PAGE_ARRAY_NR 16
 979
 980 struct kvm_mmu_pages {
 981         struct mmu_page_and_offset {
 982                 struct kvm_mmu_page *sp;
 983                 unsigned int idx;
 984         } page[KVM_PAGE_ARRAY_NR];
 985         unsigned int nr;
 986 };
 987
 988 #define for_each_unsync_children(bitmap, idx)           \
 989         for (idx = find_first_bit(bitmap, 512);         \
 990              idx < 512;                                 \
 991              idx = find_next_bit(bitmap, 512, idx+1))
 992
 993 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
 994                          int idx)
 995 {
 996         int i;
 997
 998         if (sp->unsync)
 999                 for (i=0; i < pvec->nr; i++)
1000                         if (pvec->page[i].sp == sp)
1001                                 return 0;
1002
1003         pvec->page[pvec->nr].sp = sp;
1004         pvec->page[pvec->nr].idx = idx;
1005         pvec->nr++;
1006         return (pvec->nr == KVM_PAGE_ARRAY_NR);
1007 }
1008
1009 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1010                            struct kvm_mmu_pages *pvec)
1011 {
1012         int i, ret, nr_unsync_leaf = 0;
1013
1014         for_each_unsync_children(sp->unsync_child_bitmap, i) {
1015                 u64 ent = sp->spt[i];
1016
1017                 if (is_shadow_present_pte(ent) && !is_large_pte(ent)) {
1018                         struct kvm_mmu_page *child;
1019                         child = page_header(ent & PT64_BASE_ADDR_MASK);
1020
1021                         if (child->unsync_children) {
1022                                 if (mmu_pages_add(pvec, child, i))
1023                                         return -ENOSPC;
1024
1025                                 ret = __mmu_unsync_walk(child, pvec);
1026                                 if (!ret)
1027                                         __clear_bit(i, sp->unsync_child_bitmap);
1028                                 else if (ret > 0)
1029                                         nr_unsync_leaf += ret;
1030                                 else
1031                                         return ret;
1032                         }
1033
1034                         if (child->unsync) {
1035                                 nr_unsync_leaf++;
1036                                 if (mmu_pages_add(pvec, child, i))
1037                                         return -ENOSPC;
1038                         }
1039                 }
1040         }
1041
1042         if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
1043                 sp->unsync_children = 0;
1044
1045         return nr_unsync_leaf;
1046 }
1047
1048 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1049                            struct kvm_mmu_pages *pvec)
1050 {
1051         if (!sp->unsync_children)
1052                 return 0;
1053
1054         mmu_pages_add(pvec, sp, 0);
1055         return __mmu_unsync_walk(sp, pvec);
1056 }
1057
1058 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1059 {
1060         unsigned index;
1061         struct hlist_head *bucket;
1062         struct kvm_mmu_page *sp;
1063         struct hlist_node *node;
1064
1065         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1066         index = kvm_page_table_hashfn(gfn);
1067         bucket = &kvm->arch.mmu_page_hash[index];
1068         hlist_for_each_entry(sp, node, bucket, hash_link)
1069                 if (sp->gfn == gfn && !sp->role.direct
1070                     && !sp->role.invalid) {
1071                         pgprintk("%s: found role %x\n",
1072                                  __func__, sp->role.word);
1073                         return sp;
1074                 }
1075         return NULL;
1076 }
1077
1078 static void kvm_unlink_unsync_global(struct kvm *kvm, struct kvm_mmu_page *sp)
1079 {
1080         list_del(&sp->oos_link);
1081         --kvm->stat.mmu_unsync_global;
1082 }
1083
1084 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1085 {
1086         WARN_ON(!sp->unsync);
1087         sp->unsync = 0;
1088         if (sp->global)
1089                 kvm_unlink_unsync_global(kvm, sp);
1090         --kvm->stat.mmu_unsync;
1091 }
1092
1093 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1094
1095 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1096 {
1097         if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1098                 kvm_mmu_zap_page(vcpu->kvm, sp);
1099                 return 1;
1100         }
1101
1102         if (rmap_write_protect(vcpu->kvm, sp->gfn))
1103                 kvm_flush_remote_tlbs(vcpu->kvm);
1104         kvm_unlink_unsync_page(vcpu->kvm, sp);
1105         if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1106                 kvm_mmu_zap_page(vcpu->kvm, sp);
1107                 return 1;
1108         }
1109
1110         kvm_mmu_flush_tlb(vcpu);
1111         return 0;
1112 }
1113
1114 struct mmu_page_path {
1115         struct kvm_mmu_page *parent[PT64_ROOT_LEVEL-1];
1116         unsigned int idx[PT64_ROOT_LEVEL-1];
1117 };
1118
1119 #define for_each_sp(pvec, sp, parents, i)                       \
1120                 for (i = mmu_pages_next(&pvec, &parents, -1),   \
1121                         sp = pvec.page[i].sp;                   \
1122                         i < pvec.nr && ({ sp = pvec.page[i].sp; 1;});   \
1123                         i = mmu_pages_next(&pvec, &parents, i))
1124
1125 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1126                           struct mmu_page_path *parents,
1127                           int i)
1128 {
1129         int n;
1130
1131         for (n = i+1; n < pvec->nr; n++) {
1132                 struct kvm_mmu_page *sp = pvec->page[n].sp;
1133
1134                 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1135                         parents->idx[0] = pvec->page[n].idx;
1136                         return n;
1137                 }
1138
1139                 parents->parent[sp->role.level-2] = sp;
1140                 parents->idx[sp->role.level-1] = pvec->page[n].idx;
1141         }
1142
1143         return n;
1144 }
1145
1146 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1147 {
1148         struct kvm_mmu_page *sp;
1149         unsigned int level = 0;
1150
1151         do {
1152                 unsigned int idx = parents->idx[level];
1153
1154                 sp = parents->parent[level];
1155                 if (!sp)
1156                         return;
1157
1158                 --sp->unsync_children;
1159                 WARN_ON((int)sp->unsync_children < 0);
1160                 __clear_bit(idx, sp->unsync_child_bitmap);
1161                 level++;
1162         } while (level < PT64_ROOT_LEVEL-1 && !sp->unsync_children);
1163 }
1164
1165 static void kvm_mmu_pages_init(struct kvm_mmu_page *parent,
1166                                struct mmu_page_path *parents,
1167                                struct kvm_mmu_pages *pvec)
1168 {
1169         parents->parent[parent->role.level-1] = NULL;
1170         pvec->nr = 0;
1171 }
1172
1173 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1174                               struct kvm_mmu_page *parent)
1175 {
1176         int i;
1177         struct kvm_mmu_page *sp;
1178         struct mmu_page_path parents;
1179         struct kvm_mmu_pages pages;
1180
1181         kvm_mmu_pages_init(parent, &parents, &pages);
1182         while (mmu_unsync_walk(parent, &pages)) {
1183                 int protected = 0;
1184
1185                 for_each_sp(pages, sp, parents, i)
1186                         protected |= rmap_write_protect(vcpu->kvm, sp->gfn);
1187
1188                 if (protected)
1189                         kvm_flush_remote_tlbs(vcpu->kvm);
1190
1191                 for_each_sp(pages, sp, parents, i) {
1192                         kvm_sync_page(vcpu, sp);
1193                         mmu_pages_clear_parents(&parents);
1194                 }
1195                 cond_resched_lock(&vcpu->kvm->mmu_lock);
1196                 kvm_mmu_pages_init(parent, &parents, &pages);
1197         }
1198 }
1199
1200 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1201                                              gfn_t gfn,
1202                                              gva_t gaddr,
1203                                              unsigned level,
1204                                              int direct,
1205                                              unsigned access,
1206                                              u64 *parent_pte)
1207 {
1208         union kvm_mmu_page_role role;
1209         unsigned index;
1210         unsigned quadrant;
1211         struct hlist_head *bucket;
1212         struct kvm_mmu_page *sp;
1213         struct hlist_node *node, *tmp;
1214
1215         role = vcpu->arch.mmu.base_role;
1216         role.level = level;
1217         role.direct = direct;
1218         role.access = access;
1219         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1220                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1221                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1222                 role.quadrant = quadrant;
1223         }
1224         pgprintk("%s: looking gfn %lx role %x\n", __func__,
1225                  gfn, role.word);
1226         index = kvm_page_table_hashfn(gfn);
1227         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1228         hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1229                 if (sp->gfn == gfn) {
1230                         if (sp->unsync)
1231                                 if (kvm_sync_page(vcpu, sp))
1232                                         continue;
1233
1234                         if (sp->role.word != role.word)
1235                                 continue;
1236
1237                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1238                         if (sp->unsync_children) {
1239                                 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1240                                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1241                         }
1242                         pgprintk("%s: found\n", __func__);
1243                         return sp;
1244                 }
1245         ++vcpu->kvm->stat.mmu_cache_miss;
1246         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1247         if (!sp)
1248                 return sp;
1249         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1250         sp->gfn = gfn;
1251         sp->role = role;
1252         sp->global = 0;
1253         hlist_add_head(&sp->hash_link, bucket);
1254         if (!direct) {
1255                 if (rmap_write_protect(vcpu->kvm, gfn))
1256                         kvm_flush_remote_tlbs(vcpu->kvm);
1257                 account_shadowed(vcpu->kvm, gfn);
1258         }
1259         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1260                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1261         else
1262                 nonpaging_prefetch_page(vcpu, sp);
1263         return sp;
1264 }
1265
1266 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
1267                              struct kvm_vcpu *vcpu, u64 addr)
1268 {
1269         iterator->addr = addr;
1270         iterator->shadow_addr = vcpu->arch.mmu.root_hpa;
1271         iterator->level = vcpu->arch.mmu.shadow_root_level;
1272         if (iterator->level == PT32E_ROOT_LEVEL) {
1273                 iterator->shadow_addr
1274                         = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1275                 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
1276                 --iterator->level;
1277                 if (!iterator->shadow_addr)
1278                         iterator->level = 0;
1279         }
1280 }
1281
1282 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
1283 {
1284         if (iterator->level < PT_PAGE_TABLE_LEVEL)
1285                 return false;
1286         iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
1287         iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
1288         return true;
1289 }
1290
1291 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
1292 {
1293         iterator->shadow_addr = *iterator->sptep & PT64_BASE_ADDR_MASK;
1294         --iterator->level;
1295 }
1296
1297 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1298                                          struct kvm_mmu_page *sp)
1299 {
1300         unsigned i;
1301         u64 *pt;
1302         u64 ent;
1303
1304         pt = sp->spt;
1305
1306         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1307                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1308                         if (is_shadow_present_pte(pt[i]))
1309                                 rmap_remove(kvm, &pt[i]);
1310                         pt[i] = shadow_trap_nonpresent_pte;
1311                 }
1312                 return;
1313         }
1314
1315         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1316                 ent = pt[i];
1317
1318                 if (is_shadow_present_pte(ent)) {
1319                         if (!is_large_pte(ent)) {
1320                                 ent &= PT64_BASE_ADDR_MASK;
1321                                 mmu_page_remove_parent_pte(page_header(ent),
1322                                                            &pt[i]);
1323                         } else {
1324                                 --kvm->stat.lpages;
1325                                 rmap_remove(kvm, &pt[i]);
1326                         }
1327                 }
1328                 pt[i] = shadow_trap_nonpresent_pte;
1329         }
1330 }
1331
1332 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1333 {
1334         mmu_page_remove_parent_pte(sp, parent_pte);
1335 }
1336
1337 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1338 {
1339         int i;
1340
1341         for (i = 0; i < KVM_MAX_VCPUS; ++i)
1342                 if (kvm->vcpus[i])
1343                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
1344 }
1345
1346 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1347 {
1348         u64 *parent_pte;
1349
1350         while (sp->multimapped || sp->parent_pte) {
1351                 if (!sp->multimapped)
1352                         parent_pte = sp->parent_pte;
1353                 else {
1354                         struct kvm_pte_chain *chain;
1355
1356                         chain = container_of(sp->parent_ptes.first,
1357                                              struct kvm_pte_chain, link);
1358                         parent_pte = chain->parent_ptes[0];
1359                 }
1360                 BUG_ON(!parent_pte);
1361                 kvm_mmu_put_page(sp, parent_pte);
1362                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1363         }
1364 }
1365
1366 static int mmu_zap_unsync_children(struct kvm *kvm,
1367                                    struct kvm_mmu_page *parent)
1368 {
1369         int i, zapped = 0;
1370         struct mmu_page_path parents;
1371         struct kvm_mmu_pages pages;
1372
1373         if (parent->role.level == PT_PAGE_TABLE_LEVEL)
1374                 return 0;
1375
1376         kvm_mmu_pages_init(parent, &parents, &pages);
1377         while (mmu_unsync_walk(parent, &pages)) {
1378                 struct kvm_mmu_page *sp;
1379
1380                 for_each_sp(pages, sp, parents, i) {
1381                         kvm_mmu_zap_page(kvm, sp);
1382                         mmu_pages_clear_parents(&parents);
1383                 }
1384                 zapped += pages.nr;
1385                 kvm_mmu_pages_init(parent, &parents, &pages);
1386         }
1387
1388         return zapped;
1389 }
1390
1391 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1392 {
1393         int ret;
1394         ++kvm->stat.mmu_shadow_zapped;
1395         ret = mmu_zap_unsync_children(kvm, sp);
1396         kvm_mmu_page_unlink_children(kvm, sp);
1397         kvm_mmu_unlink_parents(kvm, sp);
1398         kvm_flush_remote_tlbs(kvm);
1399         if (!sp->role.invalid && !sp->role.direct)
1400                 unaccount_shadowed(kvm, sp->gfn);
1401         if (sp->unsync)
1402                 kvm_unlink_unsync_page(kvm, sp);
1403         if (!sp->root_count) {
1404                 hlist_del(&sp->hash_link);
1405                 kvm_mmu_free_page(kvm, sp);
1406         } else {
1407                 sp->role.invalid = 1;
1408                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1409                 kvm_reload_remote_mmus(kvm);
1410         }
1411         kvm_mmu_reset_last_pte_updated(kvm);
1412         return ret;
1413 }
1414
1415 /*
1416  * Changing the number of mmu pages allocated to the vm
1417  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1418  */
1419 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1420 {
1421         /*
1422          * If we set the number of mmu pages to be smaller be than the
1423          * number of actived pages , we must to free some mmu pages before we
1424          * change the value
1425          */
1426
1427         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1428             kvm_nr_mmu_pages) {
1429                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1430                                        - kvm->arch.n_free_mmu_pages;
1431
1432                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1433                         struct kvm_mmu_page *page;
1434
1435                         page = container_of(kvm->arch.active_mmu_pages.prev,
1436                                             struct kvm_mmu_page, link);
1437                         kvm_mmu_zap_page(kvm, page);
1438                         n_used_mmu_pages--;
1439                 }
1440                 kvm->arch.n_free_mmu_pages = 0;
1441         }
1442         else
1443                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1444                                          - kvm->arch.n_alloc_mmu_pages;
1445
1446         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1447 }
1448
1449 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1450 {
1451         unsigned index;
1452         struct hlist_head *bucket;
1453         struct kvm_mmu_page *sp;
1454         struct hlist_node *node, *n;
1455         int r;
1456
1457         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1458         r = 0;
1459         index = kvm_page_table_hashfn(gfn);
1460         bucket = &kvm->arch.mmu_page_hash[index];
1461         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1462                 if (sp->gfn == gfn && !sp->role.direct) {
1463                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1464                                  sp->role.word);
1465                         r = 1;
1466                         if (kvm_mmu_zap_page(kvm, sp))
1467                                 n = bucket->first;
1468                 }
1469         return r;
1470 }
1471
1472 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1473 {
1474         unsigned index;
1475         struct hlist_head *bucket;
1476         struct kvm_mmu_page *sp;
1477         struct hlist_node *node, *nn;
1478
1479         index = kvm_page_table_hashfn(gfn);
1480         bucket = &kvm->arch.mmu_page_hash[index];
1481         hlist_for_each_entry_safe(sp, node, nn, bucket, hash_link) {
1482                 if (sp->gfn == gfn && !sp->role.direct
1483                     && !sp->role.invalid) {
1484                         pgprintk("%s: zap %lx %x\n",
1485                                  __func__, gfn, sp->role.word);
1486                         kvm_mmu_zap_page(kvm, sp);
1487                 }
1488         }
1489 }
1490
1491 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1492 {
1493         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1494         struct kvm_mmu_page *sp = page_header(__pa(pte));
1495
1496         __set_bit(slot, sp->slot_bitmap);
1497 }
1498
1499 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1500 {
1501         int i;
1502         u64 *pt = sp->spt;
1503
1504         if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1505                 return;
1506
1507         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1508                 if (pt[i] == shadow_notrap_nonpresent_pte)
1509                         set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1510         }
1511 }
1512
1513 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1514 {
1515         struct page *page;
1516
1517         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1518
1519         if (gpa == UNMAPPED_GVA)
1520                 return NULL;
1521
1522         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1523
1524         return page;
1525 }
1526
1527 /*
1528  * The function is based on mtrr_type_lookup() in
1529  * arch/x86/kernel/cpu/mtrr/generic.c
1530  */
1531 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1532                          u64 start, u64 end)
1533 {
1534         int i;
1535         u64 base, mask;
1536         u8 prev_match, curr_match;
1537         int num_var_ranges = KVM_NR_VAR_MTRR;
1538
1539         if (!mtrr_state->enabled)
1540                 return 0xFF;
1541
1542         /* Make end inclusive end, instead of exclusive */
1543         end--;
1544
1545         /* Look in fixed ranges. Just return the type as per start */
1546         if (mtrr_state->have_fixed && (start < 0x100000)) {
1547                 int idx;
1548
1549                 if (start < 0x80000) {
1550                         idx = 0;
1551                         idx += (start >> 16);
1552                         return mtrr_state->fixed_ranges[idx];
1553                 } else if (start < 0xC0000) {
1554                         idx = 1 * 8;
1555                         idx += ((start - 0x80000) >> 14);
1556                         return mtrr_state->fixed_ranges[idx];
1557                 } else if (start < 0x1000000) {
1558                         idx = 3 * 8;
1559                         idx += ((start - 0xC0000) >> 12);
1560                         return mtrr_state->fixed_ranges[idx];
1561                 }
1562         }
1563
1564         /*
1565          * Look in variable ranges
1566          * Look of multiple ranges matching this address and pick type
1567          * as per MTRR precedence
1568          */
1569         if (!(mtrr_state->enabled & 2))
1570                 return mtrr_state->def_type;
1571
1572         prev_match = 0xFF;
1573         for (i = 0; i < num_var_ranges; ++i) {
1574                 unsigned short start_state, end_state;
1575
1576                 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1577                         continue;
1578
1579                 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1580                        (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1581                 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1582                        (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1583
1584                 start_state = ((start & mask) == (base & mask));
1585                 end_state = ((end & mask) == (base & mask));
1586                 if (start_state != end_state)
1587                         return 0xFE;
1588
1589                 if ((start & mask) != (base & mask))
1590                         continue;
1591
1592                 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1593                 if (prev_match == 0xFF) {
1594                         prev_match = curr_match;
1595                         continue;
1596                 }
1597
1598                 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1599                     curr_match == MTRR_TYPE_UNCACHABLE)
1600                         return MTRR_TYPE_UNCACHABLE;
1601
1602                 if ((prev_match == MTRR_TYPE_WRBACK &&
1603                      curr_match == MTRR_TYPE_WRTHROUGH) ||
1604                     (prev_match == MTRR_TYPE_WRTHROUGH &&
1605                      curr_match == MTRR_TYPE_WRBACK)) {
1606                         prev_match = MTRR_TYPE_WRTHROUGH;
1607                         curr_match = MTRR_TYPE_WRTHROUGH;
1608                 }
1609
1610                 if (prev_match != curr_match)
1611                         return MTRR_TYPE_UNCACHABLE;
1612         }
1613
1614         if (prev_match != 0xFF)
1615                 return prev_match;
1616
1617         return mtrr_state->def_type;
1618 }
1619
1620 static u8 get_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1621 {
1622         u8 mtrr;
1623
1624         mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1625                              (gfn << PAGE_SHIFT) + PAGE_SIZE);
1626         if (mtrr == 0xfe || mtrr == 0xff)
1627                 mtrr = MTRR_TYPE_WRBACK;
1628         return mtrr;
1629 }
1630
1631 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1632 {
1633         unsigned index;
1634         struct hlist_head *bucket;
1635         struct kvm_mmu_page *s;
1636         struct hlist_node *node, *n;
1637
1638         index = kvm_page_table_hashfn(sp->gfn);
1639         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1640         /* don't unsync if pagetable is shadowed with multiple roles */
1641         hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1642                 if (s->gfn != sp->gfn || s->role.direct)
1643                         continue;
1644                 if (s->role.word != sp->role.word)
1645                         return 1;
1646         }
1647         ++vcpu->kvm->stat.mmu_unsync;
1648         sp->unsync = 1;
1649
1650         if (sp->global) {
1651                 list_add(&sp->oos_link, &vcpu->kvm->arch.oos_global_pages);
1652                 ++vcpu->kvm->stat.mmu_unsync_global;
1653         } else
1654                 kvm_mmu_mark_parents_unsync(vcpu, sp);
1655
1656         mmu_convert_notrap(sp);
1657         return 0;
1658 }
1659
1660 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1661                                   bool can_unsync)
1662 {
1663         struct kvm_mmu_page *shadow;
1664
1665         shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1666         if (shadow) {
1667                 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1668                         return 1;
1669                 if (shadow->unsync)
1670                         return 0;
1671                 if (can_unsync && oos_shadow)
1672                         return kvm_unsync_page(vcpu, shadow);
1673                 return 1;
1674         }
1675         return 0;
1676 }
1677
1678 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1679                     unsigned pte_access, int user_fault,
1680                     int write_fault, int dirty, int largepage,
1681                     int global, gfn_t gfn, pfn_t pfn, bool speculative,
1682                     bool can_unsync)
1683 {
1684         u64 spte;
1685         int ret = 0;
1686         u64 mt_mask = shadow_mt_mask;
1687         struct kvm_mmu_page *sp = page_header(__pa(shadow_pte));
1688
1689         if (!global && sp->global) {
1690                 sp->global = 0;
1691                 if (sp->unsync) {
1692                         kvm_unlink_unsync_global(vcpu->kvm, sp);
1693                         kvm_mmu_mark_parents_unsync(vcpu, sp);
1694                 }
1695         }
1696
1697         /*
1698          * We don't set the accessed bit, since we sometimes want to see
1699          * whether the guest actually used the pte (in order to detect
1700          * demand paging).
1701          */
1702         spte = shadow_base_present_pte | shadow_dirty_mask;
1703         if (!speculative)
1704                 spte |= shadow_accessed_mask;
1705         if (!dirty)
1706                 pte_access &= ~ACC_WRITE_MASK;
1707         if (pte_access & ACC_EXEC_MASK)
1708                 spte |= shadow_x_mask;
1709         else
1710                 spte |= shadow_nx_mask;
1711         if (pte_access & ACC_USER_MASK)
1712                 spte |= shadow_user_mask;
1713         if (largepage)
1714                 spte |= PT_PAGE_SIZE_MASK;
1715         if (mt_mask) {
1716                 if (!kvm_is_mmio_pfn(pfn)) {
1717                         mt_mask = get_memory_type(vcpu, gfn) <<
1718                                 kvm_x86_ops->get_mt_mask_shift();
1719                         mt_mask |= VMX_EPT_IGMT_BIT;
1720                 } else
1721                         mt_mask = MTRR_TYPE_UNCACHABLE <<
1722                                 kvm_x86_ops->get_mt_mask_shift();
1723                 spte |= mt_mask;
1724         }
1725
1726         spte |= (u64)pfn << PAGE_SHIFT;
1727
1728         if ((pte_access & ACC_WRITE_MASK)
1729             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1730
1731                 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1732                         ret = 1;
1733                         spte = shadow_trap_nonpresent_pte;
1734                         goto set_pte;
1735                 }
1736
1737                 spte |= PT_WRITABLE_MASK;
1738
1739                 /*
1740                  * Optimization: for pte sync, if spte was writable the hash
1741                  * lookup is unnecessary (and expensive). Write protection
1742                  * is responsibility of mmu_get_page / kvm_sync_page.
1743                  * Same reasoning can be applied to dirty page accounting.
1744                  */
1745                 if (!can_unsync && is_writeble_pte(*shadow_pte))
1746                         goto set_pte;
1747
1748                 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1749                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1750                                  __func__, gfn);
1751                         ret = 1;
1752                         pte_access &= ~ACC_WRITE_MASK;
1753                         if (is_writeble_pte(spte))
1754                                 spte &= ~PT_WRITABLE_MASK;
1755                 }
1756         }
1757
1758         if (pte_access & ACC_WRITE_MASK)
1759                 mark_page_dirty(vcpu->kvm, gfn);
1760
1761 set_pte:
1762         set_shadow_pte(shadow_pte, spte);
1763         return ret;
1764 }
1765
1766 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1767                          unsigned pt_access, unsigned pte_access,
1768                          int user_fault, int write_fault, int dirty,
1769                          int *ptwrite, int largepage, int global,
1770                          gfn_t gfn, pfn_t pfn, bool speculative)
1771 {
1772         int was_rmapped = 0;
1773         int was_writeble = is_writeble_pte(*shadow_pte);
1774
1775         pgprintk("%s: spte %llx access %x write_fault %d"
1776                  " user_fault %d gfn %lx\n",
1777                  __func__, *shadow_pte, pt_access,
1778                  write_fault, user_fault, gfn);
1779
1780         if (is_rmap_pte(*shadow_pte)) {
1781                 /*
1782                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1783                  * the parent of the now unreachable PTE.
1784                  */
1785                 if (largepage && !is_large_pte(*shadow_pte)) {
1786                         struct kvm_mmu_page *child;
1787                         u64 pte = *shadow_pte;
1788
1789                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1790                         mmu_page_remove_parent_pte(child, shadow_pte);
1791                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1792                         pgprintk("hfn old %lx new %lx\n",
1793                                  spte_to_pfn(*shadow_pte), pfn);
1794                         rmap_remove(vcpu->kvm, shadow_pte);
1795                 } else
1796                         was_rmapped = 1;
1797         }
1798         if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1799                       dirty, largepage, global, gfn, pfn, speculative, true)) {
1800                 if (write_fault)
1801                         *ptwrite = 1;
1802                 kvm_x86_ops->tlb_flush(vcpu);
1803         }
1804
1805         pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1806         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1807                  is_large_pte(*shadow_pte)? "2MB" : "4kB",
1808                  is_present_pte(*shadow_pte)?"RW":"R", gfn,
1809                  *shadow_pte, shadow_pte);
1810         if (!was_rmapped && is_large_pte(*shadow_pte))
1811                 ++vcpu->kvm->stat.lpages;
1812
1813         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1814         if (!was_rmapped) {
1815                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1816                 if (!is_rmap_pte(*shadow_pte))
1817                         kvm_release_pfn_clean(pfn);
1818         } else {
1819                 if (was_writeble)
1820                         kvm_release_pfn_dirty(pfn);
1821                 else
1822                         kvm_release_pfn_clean(pfn);
1823         }
1824         if (speculative) {
1825                 vcpu->arch.last_pte_updated = shadow_pte;
1826                 vcpu->arch.last_pte_gfn = gfn;
1827         }
1828 }
1829
1830 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1831 {
1832 }
1833
1834 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1835                         int largepage, gfn_t gfn, pfn_t pfn)
1836 {
1837         struct kvm_shadow_walk_iterator iterator;
1838         struct kvm_mmu_page *sp;
1839         int pt_write = 0;
1840         gfn_t pseudo_gfn;
1841
1842         for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) {
1843                 if (iterator.level == PT_PAGE_TABLE_LEVEL
1844                     || (largepage && iterator.level == PT_DIRECTORY_LEVEL)) {
1845                         mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, ACC_ALL,
1846                                      0, write, 1, &pt_write,
1847                                      largepage, 0, gfn, pfn, false);
1848                         ++vcpu->stat.pf_fixed;
1849                         break;
1850                 }
1851
1852                 if (*iterator.sptep == shadow_trap_nonpresent_pte) {
1853                         pseudo_gfn = (iterator.addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1854                         sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr,
1855                                               iterator.level - 1,
1856                                               1, ACC_ALL, iterator.sptep);
1857                         if (!sp) {
1858                                 pgprintk("nonpaging_map: ENOMEM\n");
1859                                 kvm_release_pfn_clean(pfn);
1860                                 return -ENOMEM;
1861                         }
1862
1863                         set_shadow_pte(iterator.sptep,
1864                                        __pa(sp->spt)
1865                                        | PT_PRESENT_MASK | PT_WRITABLE_MASK
1866                                        | shadow_user_mask | shadow_x_mask);
1867                 }
1868         }
1869         return pt_write;
1870 }
1871
1872 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1873 {
1874         int r;
1875         int largepage = 0;
1876         pfn_t pfn;
1877         unsigned long mmu_seq;
1878
1879         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1880                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1881                 largepage = 1;
1882         }
1883
1884         mmu_seq = vcpu->kvm->mmu_notifier_seq;
1885         smp_rmb();
1886         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1887
1888         /* mmio */
1889         if (is_error_pfn(pfn)) {
1890                 kvm_release_pfn_clean(pfn);
1891                 return 1;
1892         }
1893
1894         spin_lock(&vcpu->kvm->mmu_lock);
1895         if (mmu_notifier_retry(vcpu, mmu_seq))
1896                 goto out_unlock;
1897         kvm_mmu_free_some_pages(vcpu);
1898         r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1899         spin_unlock(&vcpu->kvm->mmu_lock);
1900
1901
1902         return r;
1903
1904 out_unlock:
1905         spin_unlock(&vcpu->kvm->mmu_lock);
1906         kvm_release_pfn_clean(pfn);
1907         return 0;
1908 }
1909
1910
1911 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1912 {
1913         int i;
1914         struct kvm_mmu_page *sp;
1915
1916         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1917                 return;
1918         spin_lock(&vcpu->kvm->mmu_lock);
1919         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1920                 hpa_t root = vcpu->arch.mmu.root_hpa;
1921
1922                 sp = page_header(root);
1923                 --sp->root_count;
1924                 if (!sp->root_count && sp->role.invalid)
1925                         kvm_mmu_zap_page(vcpu->kvm, sp);
1926                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1927                 spin_unlock(&vcpu->kvm->mmu_lock);
1928                 return;
1929         }
1930         for (i = 0; i < 4; ++i) {
1931                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1932
1933                 if (root) {
1934                         root &= PT64_BASE_ADDR_MASK;
1935                         sp = page_header(root);
1936                         --sp->root_count;
1937                         if (!sp->root_count && sp->role.invalid)
1938                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1939                 }
1940                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1941         }
1942         spin_unlock(&vcpu->kvm->mmu_lock);
1943         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1944 }
1945
1946 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1947 {
1948         int i;
1949         gfn_t root_gfn;
1950         struct kvm_mmu_page *sp;
1951         int direct = 0;
1952
1953         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1954
1955         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1956                 hpa_t root = vcpu->arch.mmu.root_hpa;
1957
1958                 ASSERT(!VALID_PAGE(root));
1959                 if (tdp_enabled)
1960                         direct = 1;
1961                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1962                                       PT64_ROOT_LEVEL, direct,
1963                                       ACC_ALL, NULL);
1964                 root = __pa(sp->spt);
1965                 ++sp->root_count;
1966                 vcpu->arch.mmu.root_hpa = root;
1967                 return;
1968         }
1969         direct = !is_paging(vcpu);
1970         if (tdp_enabled)
1971                 direct = 1;
1972         for (i = 0; i < 4; ++i) {
1973                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1974
1975                 ASSERT(!VALID_PAGE(root));
1976                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1977                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1978                                 vcpu->arch.mmu.pae_root[i] = 0;
1979                                 continue;
1980                         }
1981                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1982                 } else if (vcpu->arch.mmu.root_level == 0)
1983                         root_gfn = 0;
1984                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1985                                       PT32_ROOT_LEVEL, direct,
1986                                       ACC_ALL, NULL);
1987                 root = __pa(sp->spt);
1988                 ++sp->root_count;
1989                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1990         }
1991         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1992 }
1993
1994 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
1995 {
1996         int i;
1997         struct kvm_mmu_page *sp;
1998
1999         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
2000                 return;
2001         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
2002                 hpa_t root = vcpu->arch.mmu.root_hpa;
2003                 sp = page_header(root);
2004                 mmu_sync_children(vcpu, sp);
2005                 return;
2006         }
2007         for (i = 0; i < 4; ++i) {
2008                 hpa_t root = vcpu->arch.mmu.pae_root[i];
2009
2010                 if (root) {
2011                         root &= PT64_BASE_ADDR_MASK;
2012                         sp = page_header(root);
2013                         mmu_sync_children(vcpu, sp);
2014                 }
2015         }
2016 }
2017
2018 static void mmu_sync_global(struct kvm_vcpu *vcpu)
2019 {
2020         struct kvm *kvm = vcpu->kvm;
2021         struct kvm_mmu_page *sp, *n;
2022
2023         list_for_each_entry_safe(sp, n, &kvm->arch.oos_global_pages, oos_link)
2024                 kvm_sync_page(vcpu, sp);
2025 }
2026
2027 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
2028 {
2029         spin_lock(&vcpu->kvm->mmu_lock);
2030         mmu_sync_roots(vcpu);
2031         spin_unlock(&vcpu->kvm->mmu_lock);
2032 }
2033
2034 void kvm_mmu_sync_global(struct kvm_vcpu *vcpu)
2035 {
2036         spin_lock(&vcpu->kvm->mmu_lock);
2037         mmu_sync_global(vcpu);
2038         spin_unlock(&vcpu->kvm->mmu_lock);
2039 }
2040
2041 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
2042 {
2043         return vaddr;
2044 }
2045
2046 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
2047                                 u32 error_code)
2048 {
2049         gfn_t gfn;
2050         int r;
2051
2052         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
2053         r = mmu_topup_memory_caches(vcpu);
2054         if (r)
2055                 return r;
2056
2057         ASSERT(vcpu);
2058         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2059
2060         gfn = gva >> PAGE_SHIFT;
2061
2062         return nonpaging_map(vcpu, gva & PAGE_MASK,
2063                              error_code & PFERR_WRITE_MASK, gfn);
2064 }
2065
2066 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
2067                                 u32 error_code)
2068 {
2069         pfn_t pfn;
2070         int r;
2071         int largepage = 0;
2072         gfn_t gfn = gpa >> PAGE_SHIFT;
2073         unsigned long mmu_seq;
2074
2075         ASSERT(vcpu);
2076         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
2077
2078         r = mmu_topup_memory_caches(vcpu);
2079         if (r)
2080                 return r;
2081
2082         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
2083                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2084                 largepage = 1;
2085         }
2086         mmu_seq = vcpu->kvm->mmu_notifier_seq;
2087         smp_rmb();
2088         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2089         if (is_error_pfn(pfn)) {
2090                 kvm_release_pfn_clean(pfn);
2091                 return 1;
2092         }
2093         spin_lock(&vcpu->kvm->mmu_lock);
2094         if (mmu_notifier_retry(vcpu, mmu_seq))
2095                 goto out_unlock;
2096         kvm_mmu_free_some_pages(vcpu);
2097         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
2098                          largepage, gfn, pfn);
2099         spin_unlock(&vcpu->kvm->mmu_lock);
2100
2101         return r;
2102
2103 out_unlock:
2104         spin_unlock(&vcpu->kvm->mmu_lock);
2105         kvm_release_pfn_clean(pfn);
2106         return 0;
2107 }
2108
2109 static void nonpaging_free(struct kvm_vcpu *vcpu)
2110 {
2111         mmu_free_roots(vcpu);
2112 }
2113
2114 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
2115 {
2116         struct kvm_mmu *context = &vcpu->arch.mmu;
2117
2118         context->new_cr3 = nonpaging_new_cr3;
2119         context->page_fault = nonpaging_page_fault;
2120         context->gva_to_gpa = nonpaging_gva_to_gpa;
2121         context->free = nonpaging_free;
2122         context->prefetch_page = nonpaging_prefetch_page;
2123         context->sync_page = nonpaging_sync_page;
2124         context->invlpg = nonpaging_invlpg;
2125         context->root_level = 0;
2126         context->shadow_root_level = PT32E_ROOT_LEVEL;
2127         context->root_hpa = INVALID_PAGE;
2128         return 0;
2129 }
2130
2131 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2132 {
2133         ++vcpu->stat.tlb_flush;
2134         kvm_x86_ops->tlb_flush(vcpu);
2135 }
2136
2137 static void paging_new_cr3(struct kvm_vcpu *vcpu)
2138 {
2139         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
2140         mmu_free_roots(vcpu);
2141 }
2142
2143 static void inject_page_fault(struct kvm_vcpu *vcpu,
2144                               u64 addr,
2145                               u32 err_code)
2146 {
2147         kvm_inject_page_fault(vcpu, addr, err_code);
2148 }
2149
2150 static void paging_free(struct kvm_vcpu *vcpu)
2151 {
2152         nonpaging_free(vcpu);
2153 }
2154
2155 static bool is_rsvd_bits_set(struct kvm_vcpu *vcpu, u64 gpte, int level)
2156 {
2157         int bit7;
2158
2159         bit7 = (gpte >> 7) & 1;
2160         return (gpte & vcpu->arch.mmu.rsvd_bits_mask[bit7][level-1]) != 0;
2161 }
2162
2163 #define PTTYPE 64
2164 #include "paging_tmpl.h"
2165 #undef PTTYPE
2166
2167 #define PTTYPE 32
2168 #include "paging_tmpl.h"
2169 #undef PTTYPE
2170
2171 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu, int level)
2172 {
2173         struct kvm_mmu *context = &vcpu->arch.mmu;
2174         int maxphyaddr = cpuid_maxphyaddr(vcpu);
2175         u64 exb_bit_rsvd = 0;
2176
2177         if (!is_nx(vcpu))
2178                 exb_bit_rsvd = rsvd_bits(63, 63);
2179         switch (level) {
2180         case PT32_ROOT_LEVEL:
2181                 /* no rsvd bits for 2 level 4K page table entries */
2182                 context->rsvd_bits_mask[0][1] = 0;
2183                 context->rsvd_bits_mask[0][0] = 0;
2184                 if (is_cpuid_PSE36())
2185                         /* 36bits PSE 4MB page */
2186                         context->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
2187                 else
2188                         /* 32 bits PSE 4MB page */
2189                         context->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
2190                 context->rsvd_bits_mask[1][0] = ~0ull;
2191                 break;
2192         case PT32E_ROOT_LEVEL:
2193                 context->rsvd_bits_mask[0][2] =
2194                         rsvd_bits(maxphyaddr, 63) |
2195                         rsvd_bits(7, 8) | rsvd_bits(1, 2);      /* PDPTE */
2196                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2197                         rsvd_bits(maxphyaddr, 62);      /* PDE */
2198                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2199                         rsvd_bits(maxphyaddr, 62);      /* PTE */
2200                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2201                         rsvd_bits(maxphyaddr, 62) |
2202                         rsvd_bits(13, 20);              /* large page */
2203                 context->rsvd_bits_mask[1][0] = ~0ull;
2204                 break;
2205         case PT64_ROOT_LEVEL:
2206                 context->rsvd_bits_mask[0][3] = exb_bit_rsvd |
2207                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2208                 context->rsvd_bits_mask[0][2] = exb_bit_rsvd |
2209                         rsvd_bits(maxphyaddr, 51) | rsvd_bits(7, 8);
2210                 context->rsvd_bits_mask[0][1] = exb_bit_rsvd |
2211                         rsvd_bits(maxphyaddr, 51);
2212                 context->rsvd_bits_mask[0][0] = exb_bit_rsvd |
2213                         rsvd_bits(maxphyaddr, 51);
2214                 context->rsvd_bits_mask[1][3] = context->rsvd_bits_mask[0][3];
2215                 context->rsvd_bits_mask[1][2] = context->rsvd_bits_mask[0][2];
2216                 context->rsvd_bits_mask[1][1] = exb_bit_rsvd |
2217                         rsvd_bits(maxphyaddr, 51) |
2218                         rsvd_bits(13, 20);              /* large page */
2219                 context->rsvd_bits_mask[1][0] = ~0ull;
2220                 break;
2221         }
2222 }
2223
2224 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2225 {
2226         struct kvm_mmu *context = &vcpu->arch.mmu;
2227
2228         ASSERT(is_pae(vcpu));
2229         context->new_cr3 = paging_new_cr3;
2230         context->page_fault = paging64_page_fault;
2231         context->gva_to_gpa = paging64_gva_to_gpa;
2232         context->prefetch_page = paging64_prefetch_page;
2233         context->sync_page = paging64_sync_page;
2234         context->invlpg = paging64_invlpg;
2235         context->free = paging_free;
2236         context->root_level = level;
2237         context->shadow_root_level = level;
2238         context->root_hpa = INVALID_PAGE;
2239         return 0;
2240 }
2241
2242 static int paging64_init_context(struct kvm_vcpu *vcpu)
2243 {
2244         reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2245         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2246 }
2247
2248 static int paging32_init_context(struct kvm_vcpu *vcpu)
2249 {
2250         struct kvm_mmu *context = &vcpu->arch.mmu;
2251
2252         reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2253         context->new_cr3 = paging_new_cr3;
2254         context->page_fault = paging32_page_fault;
2255         context->gva_to_gpa = paging32_gva_to_gpa;
2256         context->free = paging_free;
2257         context->prefetch_page = paging32_prefetch_page;
2258         context->sync_page = paging32_sync_page;
2259         context->invlpg = paging32_invlpg;
2260         context->root_level = PT32_ROOT_LEVEL;
2261         context->shadow_root_level = PT32E_ROOT_LEVEL;
2262         context->root_hpa = INVALID_PAGE;
2263         return 0;
2264 }
2265
2266 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2267 {
2268         reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2269         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2270 }
2271
2272 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2273 {
2274         struct kvm_mmu *context = &vcpu->arch.mmu;
2275
2276         context->new_cr3 = nonpaging_new_cr3;
2277         context->page_fault = tdp_page_fault;
2278         context->free = nonpaging_free;
2279         context->prefetch_page = nonpaging_prefetch_page;
2280         context->sync_page = nonpaging_sync_page;
2281         context->invlpg = nonpaging_invlpg;
2282         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2283         context->root_hpa = INVALID_PAGE;
2284
2285         if (!is_paging(vcpu)) {
2286                 context->gva_to_gpa = nonpaging_gva_to_gpa;
2287                 context->root_level = 0;
2288         } else if (is_long_mode(vcpu)) {
2289                 reset_rsvds_bits_mask(vcpu, PT64_ROOT_LEVEL);
2290                 context->gva_to_gpa = paging64_gva_to_gpa;
2291                 context->root_level = PT64_ROOT_LEVEL;
2292         } else if (is_pae(vcpu)) {
2293                 reset_rsvds_bits_mask(vcpu, PT32E_ROOT_LEVEL);
2294                 context->gva_to_gpa = paging64_gva_to_gpa;
2295                 context->root_level = PT32E_ROOT_LEVEL;
2296         } else {
2297                 reset_rsvds_bits_mask(vcpu, PT32_ROOT_LEVEL);
2298                 context->gva_to_gpa = paging32_gva_to_gpa;
2299                 context->root_level = PT32_ROOT_LEVEL;
2300         }
2301
2302         return 0;
2303 }
2304
2305 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2306 {
2307         int r;
2308
2309         ASSERT(vcpu);
2310         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2311
2312         if (!is_paging(vcpu))
2313                 r = nonpaging_init_context(vcpu);
2314         else if (is_long_mode(vcpu))
2315                 r = paging64_init_context(vcpu);
2316         else if (is_pae(vcpu))
2317                 r = paging32E_init_context(vcpu);
2318         else
2319                 r = paging32_init_context(vcpu);
2320
2321         vcpu->arch.mmu.base_role.glevels = vcpu->arch.mmu.root_level;
2322
2323         return r;
2324 }
2325
2326 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2327 {
2328         vcpu->arch.update_pte.pfn = bad_pfn;
2329
2330         if (tdp_enabled)
2331                 return init_kvm_tdp_mmu(vcpu);
2332         else
2333                 return init_kvm_softmmu(vcpu);
2334 }
2335
2336 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2337 {
2338         ASSERT(vcpu);
2339         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2340                 vcpu->arch.mmu.free(vcpu);
2341                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2342         }
2343 }
2344
2345 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2346 {
2347         destroy_kvm_mmu(vcpu);
2348         return init_kvm_mmu(vcpu);
2349 }
2350 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2351
2352 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2353 {
2354         int r;
2355
2356         r = mmu_topup_memory_caches(vcpu);
2357         if (r)
2358                 goto out;
2359         spin_lock(&vcpu->kvm->mmu_lock);
2360         kvm_mmu_free_some_pages(vcpu);
2361         mmu_alloc_roots(vcpu);
2362         mmu_sync_roots(vcpu);
2363         spin_unlock(&vcpu->kvm->mmu_lock);
2364         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2365         kvm_mmu_flush_tlb(vcpu);
2366 out:
2367         return r;
2368 }
2369 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2370
2371 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2372 {
2373         mmu_free_roots(vcpu);
2374 }
2375
2376 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2377                                   struct kvm_mmu_page *sp,
2378                                   u64 *spte)
2379 {
2380         u64 pte;
2381         struct kvm_mmu_page *child;
2382
2383         pte = *spte;
2384         if (is_shadow_present_pte(pte)) {
2385                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2386                     is_large_pte(pte))
2387                         rmap_remove(vcpu->kvm, spte);
2388                 else {
2389                         child = page_header(pte & PT64_BASE_ADDR_MASK);
2390                         mmu_page_remove_parent_pte(child, spte);
2391                 }
2392         }
2393         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2394         if (is_large_pte(pte))
2395                 --vcpu->kvm->stat.lpages;
2396 }
2397
2398 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2399                                   struct kvm_mmu_page *sp,
2400                                   u64 *spte,
2401                                   const void *new)
2402 {
2403         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2404                 if (!vcpu->arch.update_pte.largepage ||
2405                     sp->role.glevels == PT32_ROOT_LEVEL) {
2406                         ++vcpu->kvm->stat.mmu_pde_zapped;
2407                         return;
2408                 }
2409         }
2410
2411         ++vcpu->kvm->stat.mmu_pte_updated;
2412         if (sp->role.glevels == PT32_ROOT_LEVEL)
2413                 paging32_update_pte(vcpu, sp, spte, new);
2414         else
2415                 paging64_update_pte(vcpu, sp, spte, new);
2416 }
2417
2418 static bool need_remote_flush(u64 old, u64 new)
2419 {
2420         if (!is_shadow_present_pte(old))
2421                 return false;
2422         if (!is_shadow_present_pte(new))
2423                 return true;
2424         if ((old ^ new) & PT64_BASE_ADDR_MASK)
2425                 return true;
2426         old ^= PT64_NX_MASK;
2427         new ^= PT64_NX_MASK;
2428         return (old & ~new & PT64_PERM_MASK) != 0;
2429 }
2430
2431 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2432 {
2433         if (need_remote_flush(old, new))
2434                 kvm_flush_remote_tlbs(vcpu->kvm);
2435         else
2436                 kvm_mmu_flush_tlb(vcpu);
2437 }
2438
2439 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2440 {
2441         u64 *spte = vcpu->arch.last_pte_updated;
2442
2443         return !!(spte && (*spte & shadow_accessed_mask));
2444 }
2445
2446 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2447                                           const u8 *new, int bytes)
2448 {
2449         gfn_t gfn;
2450         int r;
2451         u64 gpte = 0;
2452         pfn_t pfn;
2453
2454         vcpu->arch.update_pte.largepage = 0;
2455
2456         if (bytes != 4 && bytes != 8)
2457                 return;
2458
2459         /*
2460          * Assume that the pte write on a page table of the same type
2461          * as the current vcpu paging mode.  This is nearly always true
2462          * (might be false while changing modes).  Note it is verified later
2463          * by update_pte().
2464          */
2465         if (is_pae(vcpu)) {
2466                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2467                 if ((bytes == 4) && (gpa % 4 == 0)) {
2468                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2469                         if (r)
2470                                 return;
2471                         memcpy((void *)&gpte + (gpa % 8), new, 4);
2472                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2473                         memcpy((void *)&gpte, new, 8);
2474                 }
2475         } else {
2476                 if ((bytes == 4) && (gpa % 4 == 0))
2477                         memcpy((void *)&gpte, new, 4);
2478         }
2479         if (!is_present_pte(gpte))
2480                 return;
2481         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2482
2483         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2484                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2485                 vcpu->arch.update_pte.largepage = 1;
2486         }
2487         vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2488         smp_rmb();
2489         pfn = gfn_to_pfn(vcpu->kvm, gfn);
2490
2491         if (is_error_pfn(pfn)) {
2492                 kvm_release_pfn_clean(pfn);
2493                 return;
2494         }
2495         vcpu->arch.update_pte.gfn = gfn;
2496         vcpu->arch.update_pte.pfn = pfn;
2497 }
2498
2499 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2500 {
2501         u64 *spte = vcpu->arch.last_pte_updated;
2502
2503         if (spte
2504             && vcpu->arch.last_pte_gfn == gfn
2505             && shadow_accessed_mask
2506             && !(*spte & shadow_accessed_mask)
2507             && is_shadow_present_pte(*spte))
2508                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2509 }
2510
2511 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2512                        const u8 *new, int bytes,
2513                        bool guest_initiated)
2514 {
2515         gfn_t gfn = gpa >> PAGE_SHIFT;
2516         struct kvm_mmu_page *sp;
2517         struct hlist_node *node, *n;
2518         struct hlist_head *bucket;
2519         unsigned index;
2520         u64 entry, gentry;
2521         u64 *spte;
2522         unsigned offset = offset_in_page(gpa);
2523         unsigned pte_size;
2524         unsigned page_offset;
2525         unsigned misaligned;
2526         unsigned quadrant;
2527         int level;
2528         int flooded = 0;
2529         int npte;
2530         int r;
2531
2532         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2533         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2534         spin_lock(&vcpu->kvm->mmu_lock);
2535         kvm_mmu_access_page(vcpu, gfn);
2536         kvm_mmu_free_some_pages(vcpu);
2537         ++vcpu->kvm->stat.mmu_pte_write;
2538         kvm_mmu_audit(vcpu, "pre pte write");
2539         if (guest_initiated) {
2540                 if (gfn == vcpu->arch.last_pt_write_gfn
2541                     && !last_updated_pte_accessed(vcpu)) {
2542                         ++vcpu->arch.last_pt_write_count;
2543                         if (vcpu->arch.last_pt_write_count >= 3)
2544                                 flooded = 1;
2545                 } else {
2546                         vcpu->arch.last_pt_write_gfn = gfn;
2547                         vcpu->arch.last_pt_write_count = 1;
2548                         vcpu->arch.last_pte_updated = NULL;
2549                 }
2550         }
2551         index = kvm_page_table_hashfn(gfn);
2552         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2553         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2554                 if (sp->gfn != gfn || sp->role.direct || sp->role.invalid)
2555                         continue;
2556                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2557                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2558                 misaligned |= bytes < 4;
2559                 if (misaligned || flooded) {
2560                         /*
2561                          * Misaligned accesses are too much trouble to fix
2562                          * up; also, they usually indicate a page is not used
2563                          * as a page table.
2564                          *
2565                          * If we're seeing too many writes to a page,
2566                          * it may no longer be a page table, or we may be
2567                          * forking, in which case it is better to unmap the
2568                          * page.
2569                          */
2570                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2571                                  gpa, bytes, sp->role.word);
2572                         if (kvm_mmu_zap_page(vcpu->kvm, sp))
2573                                 n = bucket->first;
2574                         ++vcpu->kvm->stat.mmu_flooded;
2575                         continue;
2576                 }
2577                 page_offset = offset;
2578                 level = sp->role.level;
2579                 npte = 1;
2580                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2581                         page_offset <<= 1;      /* 32->64 */
2582                         /*
2583                          * A 32-bit pde maps 4MB while the shadow pdes map
2584                          * only 2MB.  So we need to double the offset again
2585                          * and zap two pdes instead of one.
2586                          */
2587                         if (level == PT32_ROOT_LEVEL) {
2588                                 page_offset &= ~7; /* kill rounding error */
2589                                 page_offset <<= 1;
2590                                 npte = 2;
2591                         }
2592                         quadrant = page_offset >> PAGE_SHIFT;
2593                         page_offset &= ~PAGE_MASK;
2594                         if (quadrant != sp->role.quadrant)
2595                                 continue;
2596                 }
2597                 spte = &sp->spt[page_offset / sizeof(*spte)];
2598                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2599                         gentry = 0;
2600                         r = kvm_read_guest_atomic(vcpu->kvm,
2601                                                   gpa & ~(u64)(pte_size - 1),
2602                                                   &gentry, pte_size);
2603                         new = (const void *)&gentry;
2604                         if (r < 0)
2605                                 new = NULL;
2606                 }
2607                 while (npte--) {
2608                         entry = *spte;
2609                         mmu_pte_write_zap_pte(vcpu, sp, spte);
2610                         if (new)
2611                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2612                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2613                         ++spte;
2614                 }
2615         }
2616         kvm_mmu_audit(vcpu, "post pte write");
2617         spin_unlock(&vcpu->kvm->mmu_lock);
2618         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2619                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2620                 vcpu->arch.update_pte.pfn = bad_pfn;
2621         }
2622 }
2623
2624 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2625 {
2626         gpa_t gpa;
2627         int r;
2628
2629         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2630
2631         spin_lock(&vcpu->kvm->mmu_lock);
2632         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2633         spin_unlock(&vcpu->kvm->mmu_lock);
2634         return r;
2635 }
2636 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2637
2638 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2639 {
2640         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2641                 struct kvm_mmu_page *sp;
2642
2643                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2644                                   struct kvm_mmu_page, link);
2645                 kvm_mmu_zap_page(vcpu->kvm, sp);
2646                 ++vcpu->kvm->stat.mmu_recycled;
2647         }
2648 }
2649
2650 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2651 {
2652         int r;
2653         enum emulation_result er;
2654
2655         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2656         if (r < 0)
2657                 goto out;
2658
2659         if (!r) {
2660                 r = 1;
2661                 goto out;
2662         }
2663
2664         r = mmu_topup_memory_caches(vcpu);
2665         if (r)
2666                 goto out;
2667
2668         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2669
2670         switch (er) {
2671         case EMULATE_DONE:
2672                 return 1;
2673         case EMULATE_DO_MMIO:
2674                 ++vcpu->stat.mmio_exits;
2675                 return 0;
2676         case EMULATE_FAIL:
2677                 kvm_report_emulation_failure(vcpu, "pagetable");
2678                 return 1;
2679         default:
2680                 BUG();
2681         }
2682 out:
2683         return r;
2684 }
2685 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2686
2687 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2688 {
2689         vcpu->arch.mmu.invlpg(vcpu, gva);
2690         kvm_mmu_flush_tlb(vcpu);
2691         ++vcpu->stat.invlpg;
2692 }
2693 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2694
2695 void kvm_enable_tdp(void)
2696 {
2697         tdp_enabled = true;
2698 }
2699 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2700
2701 void kvm_disable_tdp(void)
2702 {
2703         tdp_enabled = false;
2704 }
2705 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2706
2707 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2708 {
2709         free_page((unsigned long)vcpu->arch.mmu.pae_root);
2710 }
2711
2712 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2713 {
2714         struct page *page;
2715         int i;
2716
2717         ASSERT(vcpu);
2718
2719         if (vcpu->kvm->arch.n_requested_mmu_pages)
2720                 vcpu->kvm->arch.n_free_mmu_pages =
2721                                         vcpu->kvm->arch.n_requested_mmu_pages;
2722         else
2723                 vcpu->kvm->arch.n_free_mmu_pages =
2724                                         vcpu->kvm->arch.n_alloc_mmu_pages;
2725         /*
2726          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2727          * Therefore we need to allocate shadow page tables in the first
2728          * 4GB of memory, which happens to fit the DMA32 zone.
2729          */
2730         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2731         if (!page)
2732                 goto error_1;
2733         vcpu->arch.mmu.pae_root = page_address(page);
2734         for (i = 0; i < 4; ++i)
2735                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2736
2737         return 0;
2738
2739 error_1:
2740         free_mmu_pages(vcpu);
2741         return -ENOMEM;
2742 }
2743
2744 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2745 {
2746         ASSERT(vcpu);
2747         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2748
2749         return alloc_mmu_pages(vcpu);
2750 }
2751
2752 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2753 {
2754         ASSERT(vcpu);
2755         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2756
2757         return init_kvm_mmu(vcpu);
2758 }
2759
2760 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2761 {
2762         ASSERT(vcpu);
2763
2764         destroy_kvm_mmu(vcpu);
2765         free_mmu_pages(vcpu);
2766         mmu_free_memory_caches(vcpu);
2767 }
2768
2769 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2770 {
2771         struct kvm_mmu_page *sp;
2772
2773         spin_lock(&kvm->mmu_lock);
2774         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2775                 int i;
2776                 u64 *pt;
2777
2778                 if (!test_bit(slot, sp->slot_bitmap))
2779                         continue;
2780
2781                 pt = sp->spt;
2782                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2783                         /* avoid RMW */
2784                         if (pt[i] & PT_WRITABLE_MASK)
2785                                 pt[i] &= ~PT_WRITABLE_MASK;
2786         }
2787         kvm_flush_remote_tlbs(kvm);
2788         spin_unlock(&kvm->mmu_lock);
2789 }
2790
2791 void kvm_mmu_zap_all(struct kvm *kvm)
2792 {
2793         struct kvm_mmu_page *sp, *node;
2794
2795         spin_lock(&kvm->mmu_lock);
2796         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2797                 if (kvm_mmu_zap_page(kvm, sp))
2798                         node = container_of(kvm->arch.active_mmu_pages.next,
2799                                             struct kvm_mmu_page, link);
2800         spin_unlock(&kvm->mmu_lock);
2801
2802         kvm_flush_remote_tlbs(kvm);
2803 }
2804
2805 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2806 {
2807         struct kvm_mmu_page *page;
2808
2809         page = container_of(kvm->arch.active_mmu_pages.prev,
2810                             struct kvm_mmu_page, link);
2811         kvm_mmu_zap_page(kvm, page);
2812 }
2813
2814 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2815 {
2816         struct kvm *kvm;
2817         struct kvm *kvm_freed = NULL;
2818         int cache_count = 0;
2819
2820         spin_lock(&kvm_lock);
2821
2822         list_for_each_entry(kvm, &vm_list, vm_list) {
2823                 int npages;
2824
2825                 if (!down_read_trylock(&kvm->slots_lock))
2826                         continue;
2827                 spin_lock(&kvm->mmu_lock);
2828                 npages = kvm->arch.n_alloc_mmu_pages -
2829                          kvm->arch.n_free_mmu_pages;
2830                 cache_count += npages;
2831                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2832                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
2833                         cache_count--;
2834                         kvm_freed = kvm;
2835                 }
2836                 nr_to_scan--;
2837
2838                 spin_unlock(&kvm->mmu_lock);
2839                 up_read(&kvm->slots_lock);
2840         }
2841         if (kvm_freed)
2842                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2843
2844         spin_unlock(&kvm_lock);
2845
2846         return cache_count;
2847 }
2848
2849 static struct shrinker mmu_shrinker = {
2850         .shrink = mmu_shrink,
2851         .seeks = DEFAULT_SEEKS * 10,
2852 };
2853
2854 static void mmu_destroy_caches(void)
2855 {
2856         if (pte_chain_cache)
2857                 kmem_cache_destroy(pte_chain_cache);
2858         if (rmap_desc_cache)
2859                 kmem_cache_destroy(rmap_desc_cache);
2860         if (mmu_page_header_cache)
2861                 kmem_cache_destroy(mmu_page_header_cache);
2862 }
2863
2864 void kvm_mmu_module_exit(void)
2865 {
2866         mmu_destroy_caches();
2867         unregister_shrinker(&mmu_shrinker);
2868 }
2869
2870 int kvm_mmu_module_init(void)
2871 {
2872         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2873                                             sizeof(struct kvm_pte_chain),
2874                                             0, 0, NULL);
2875         if (!pte_chain_cache)
2876                 goto nomem;
2877         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2878                                             sizeof(struct kvm_rmap_desc),
2879                                             0, 0, NULL);
2880         if (!rmap_desc_cache)
2881                 goto nomem;
2882
2883         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2884                                                   sizeof(struct kvm_mmu_page),
2885                                                   0, 0, NULL);
2886         if (!mmu_page_header_cache)
2887                 goto nomem;
2888
2889         register_shrinker(&mmu_shrinker);
2890
2891         return 0;
2892
2893 nomem:
2894         mmu_destroy_caches();
2895         return -ENOMEM;
2896 }
2897
2898 /*
2899  * Caculate mmu pages needed for kvm.
2900  */
2901 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2902 {
2903         int i;
2904         unsigned int nr_mmu_pages;
2905         unsigned int  nr_pages = 0;
2906
2907         for (i = 0; i < kvm->nmemslots; i++)
2908                 nr_pages += kvm->memslots[i].npages;
2909
2910         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2911         nr_mmu_pages = max(nr_mmu_pages,
2912                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2913
2914         return nr_mmu_pages;
2915 }
2916
2917 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2918                                 unsigned len)
2919 {
2920         if (len > buffer->len)
2921                 return NULL;
2922         return buffer->ptr;
2923 }
2924
2925 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2926                                 unsigned len)
2927 {
2928         void *ret;
2929
2930         ret = pv_mmu_peek_buffer(buffer, len);
2931         if (!ret)
2932                 return ret;
2933         buffer->ptr += len;
2934         buffer->len -= len;
2935         buffer->processed += len;
2936         return ret;
2937 }
2938
2939 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2940                              gpa_t addr, gpa_t value)
2941 {
2942         int bytes = 8;
2943         int r;
2944
2945         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2946                 bytes = 4;
2947
2948         r = mmu_topup_memory_caches(vcpu);
2949         if (r)
2950                 return r;
2951
2952         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2953                 return -EFAULT;
2954
2955         return 1;
2956 }
2957
2958 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2959 {
2960         kvm_set_cr3(vcpu, vcpu->arch.cr3);
2961         return 1;
2962 }
2963
2964 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2965 {
2966         spin_lock(&vcpu->kvm->mmu_lock);
2967         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2968         spin_unlock(&vcpu->kvm->mmu_lock);
2969         return 1;
2970 }
2971
2972 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2973                              struct kvm_pv_mmu_op_buffer *buffer)
2974 {
2975         struct kvm_mmu_op_header *header;
2976
2977         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2978         if (!header)
2979                 return 0;
2980         switch (header->op) {
2981         case KVM_MMU_OP_WRITE_PTE: {
2982                 struct kvm_mmu_op_write_pte *wpte;
2983
2984                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2985                 if (!wpte)
2986                         return 0;
2987                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2988                                         wpte->pte_val);
2989         }
2990         case KVM_MMU_OP_FLUSH_TLB: {
2991                 struct kvm_mmu_op_flush_tlb *ftlb;
2992
2993                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2994                 if (!ftlb)
2995                         return 0;
2996                 return kvm_pv_mmu_flush_tlb(vcpu);
2997         }
2998         case KVM_MMU_OP_RELEASE_PT: {
2999                 struct kvm_mmu_op_release_pt *rpt;
3000
3001                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
3002                 if (!rpt)
3003                         return 0;
3004                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
3005         }
3006         default: return 0;
3007         }
3008 }
3009
3010 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
3011                   gpa_t addr, unsigned long *ret)
3012 {
3013         int r;
3014         struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
3015
3016         buffer->ptr = buffer->buf;
3017         buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
3018         buffer->processed = 0;
3019
3020         r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
3021         if (r)
3022                 goto out;
3023
3024         while (buffer->len) {
3025                 r = kvm_pv_mmu_op_one(vcpu, buffer);
3026                 if (r < 0)
3027                         goto out;
3028                 if (r == 0)
3029                         break;
3030         }
3031
3032         r = 1;
3033 out:
3034         *ret = buffer->processed;
3035         return r;
3036 }
3037
3038 #ifdef AUDIT
3039
3040 static const char *audit_msg;
3041
3042 static gva_t canonicalize(gva_t gva)
3043 {
3044 #ifdef CONFIG_X86_64
3045         gva = (long long)(gva << 16) >> 16;
3046 #endif
3047         return gva;
3048 }
3049
3050 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
3051                                 gva_t va, int level)
3052 {
3053         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
3054         int i;
3055         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
3056
3057         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
3058                 u64 ent = pt[i];
3059
3060                 if (ent == shadow_trap_nonpresent_pte)
3061                         continue;
3062
3063                 va = canonicalize(va);
3064                 if (level > 1) {
3065                         if (ent == shadow_notrap_nonpresent_pte)
3066                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
3067                                        " in nonleaf level: levels %d gva %lx"
3068                                        " level %d pte %llx\n", audit_msg,
3069                                        vcpu->arch.mmu.root_level, va, level, ent);
3070
3071                         audit_mappings_page(vcpu, ent, va, level - 1);
3072                 } else {
3073                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
3074                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
3075
3076                         if (is_shadow_present_pte(ent)
3077                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
3078                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
3079                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3080                                        audit_msg, vcpu->arch.mmu.root_level,
3081                                        va, gpa, hpa, ent,
3082                                        is_shadow_present_pte(ent));
3083                         else if (ent == shadow_notrap_nonpresent_pte
3084                                  && !is_error_hpa(hpa))
3085                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
3086                                        " valid guest gva %lx\n", audit_msg, va);
3087                         kvm_release_pfn_clean(pfn);
3088
3089                 }
3090         }
3091 }
3092
3093 static void audit_mappings(struct kvm_vcpu *vcpu)
3094 {
3095         unsigned i;
3096
3097         if (vcpu->arch.mmu.root_level == 4)
3098                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
3099         else
3100                 for (i = 0; i < 4; ++i)
3101                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
3102                                 audit_mappings_page(vcpu,
3103                                                     vcpu->arch.mmu.pae_root[i],
3104                                                     i << 30,
3105                                                     2);
3106 }
3107
3108 static int count_rmaps(struct kvm_vcpu *vcpu)
3109 {
3110         int nmaps = 0;
3111         int i, j, k;
3112
3113         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
3114                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
3115                 struct kvm_rmap_desc *d;
3116
3117                 for (j = 0; j < m->npages; ++j) {
3118                         unsigned long *rmapp = &m->rmap[j];
3119
3120                         if (!*rmapp)
3121                                 continue;
3122                         if (!(*rmapp & 1)) {
3123                                 ++nmaps;
3124                                 continue;
3125                         }
3126                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
3127                         while (d) {
3128                                 for (k = 0; k < RMAP_EXT; ++k)
3129                                         if (d->shadow_ptes[k])
3130                                                 ++nmaps;
3131                                         else
3132                                                 break;
3133                                 d = d->more;
3134                         }
3135                 }
3136         }
3137         return nmaps;
3138 }
3139
3140 static int count_writable_mappings(struct kvm_vcpu *vcpu)
3141 {
3142         int nmaps = 0;
3143         struct kvm_mmu_page *sp;
3144         int i;
3145
3146         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3147                 u64 *pt = sp->spt;
3148
3149                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
3150                         continue;
3151
3152                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
3153                         u64 ent = pt[i];
3154
3155                         if (!(ent & PT_PRESENT_MASK))
3156                                 continue;
3157                         if (!(ent & PT_WRITABLE_MASK))
3158                                 continue;
3159                         ++nmaps;
3160                 }
3161         }
3162         return nmaps;
3163 }
3164
3165 static void audit_rmap(struct kvm_vcpu *vcpu)
3166 {
3167         int n_rmap = count_rmaps(vcpu);
3168         int n_actual = count_writable_mappings(vcpu);
3169
3170         if (n_rmap != n_actual)
3171                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
3172                        __func__, audit_msg, n_rmap, n_actual);
3173 }
3174
3175 static void audit_write_protection(struct kvm_vcpu *vcpu)
3176 {
3177         struct kvm_mmu_page *sp;
3178         struct kvm_memory_slot *slot;
3179         unsigned long *rmapp;
3180         gfn_t gfn;
3181
3182         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
3183                 if (sp->role.direct)
3184                         continue;
3185
3186                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
3187                 slot = gfn_to_memslot_unaliased(vcpu->kvm, sp->gfn);
3188                 rmapp = &slot->rmap[gfn - slot->base_gfn];
3189                 if (*rmapp)
3190                         printk(KERN_ERR "%s: (%s) shadow page has writable"
3191                                " mappings: gfn %lx role %x\n",
3192                                __func__, audit_msg, sp->gfn,
3193                                sp->role.word);
3194         }
3195 }
3196
3197 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
3198 {
3199         int olddbg = dbg;
3200
3201         dbg = 0;
3202         audit_msg = msg;
3203         audit_rmap(vcpu);
3204         audit_write_protection(vcpu);
3205         audit_mappings(vcpu);
3206         dbg = olddbg;
3207 }
3208
3209 #endif