err.no Git - linux-2.6/blob - arch/x86/kvm/mmu.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * MMU support
   8  *
   9  * Copyright (C) 2006 Qumranet, Inc.
  10  *
  11  * Authors:
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *   Avi Kivity   <avi@qumranet.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2.  See
  16  * the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "vmx.h"
  21 #include "mmu.h"
  22
  23 #include <linux/kvm_host.h>
  24 #include <linux/types.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/highmem.h>
  28 #include <linux/module.h>
  29 #include <linux/swap.h>
  30 #include <linux/hugetlb.h>
  31 #include <linux/compiler.h>
  32
  33 #include <asm/page.h>
  34 #include <asm/cmpxchg.h>
  35 #include <asm/io.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 = 1;
  70 #endif
  71
  72 #ifndef MMU_DEBUG
  73 #define ASSERT(x) do { } while (0)
  74 #else
  75 #define ASSERT(x)                                                       \
  76         if (!(x)) {                                                     \
  77                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  78                        __FILE__, __LINE__, #x);                         \
  79         }
  80 #endif
  81
  82 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  83 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  84
  85 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  86
  87 #define PT64_LEVEL_BITS 9
  88
  89 #define PT64_LEVEL_SHIFT(level) \
  90                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  91
  92 #define PT64_LEVEL_MASK(level) \
  93                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  94
  95 #define PT64_INDEX(address, level)\
  96         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
  97
  98
  99 #define PT32_LEVEL_BITS 10
 100
 101 #define PT32_LEVEL_SHIFT(level) \
 102                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 103
 104 #define PT32_LEVEL_MASK(level) \
 105                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 106
 107 #define PT32_INDEX(address, level)\
 108         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 109
 110
 111 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 112 #define PT64_DIR_BASE_ADDR_MASK \
 113         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 114
 115 #define PT32_BASE_ADDR_MASK PAGE_MASK
 116 #define PT32_DIR_BASE_ADDR_MASK \
 117         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 118
 119 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 120                         | PT64_NX_MASK)
 121
 122 #define PFERR_PRESENT_MASK (1U << 0)
 123 #define PFERR_WRITE_MASK (1U << 1)
 124 #define PFERR_USER_MASK (1U << 2)
 125 #define PFERR_FETCH_MASK (1U << 4)
 126
 127 #define PT_DIRECTORY_LEVEL 2
 128 #define PT_PAGE_TABLE_LEVEL 1
 129
 130 #define RMAP_EXT 4
 131
 132 #define ACC_EXEC_MASK    1
 133 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 134 #define ACC_USER_MASK    PT_USER_MASK
 135 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 136
 137 struct kvm_pv_mmu_op_buffer {
 138         void *ptr;
 139         unsigned len;
 140         unsigned processed;
 141         char buf[512] __aligned(sizeof(long));
 142 };
 143
 144 struct kvm_rmap_desc {
 145         u64 *shadow_ptes[RMAP_EXT];
 146         struct kvm_rmap_desc *more;
 147 };
 148
 149 static struct kmem_cache *pte_chain_cache;
 150 static struct kmem_cache *rmap_desc_cache;
 151 static struct kmem_cache *mmu_page_header_cache;
 152
 153 static u64 __read_mostly shadow_trap_nonpresent_pte;
 154 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 155 static u64 __read_mostly shadow_base_present_pte;
 156 static u64 __read_mostly shadow_nx_mask;
 157 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 158 static u64 __read_mostly shadow_user_mask;
 159 static u64 __read_mostly shadow_accessed_mask;
 160 static u64 __read_mostly shadow_dirty_mask;
 161
 162 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 163 {
 164         shadow_trap_nonpresent_pte = trap_pte;
 165         shadow_notrap_nonpresent_pte = notrap_pte;
 166 }
 167 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 168
 169 void kvm_mmu_set_base_ptes(u64 base_pte)
 170 {
 171         shadow_base_present_pte = base_pte;
 172 }
 173 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 174
 175 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 176                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 177 {
 178         shadow_user_mask = user_mask;
 179         shadow_accessed_mask = accessed_mask;
 180         shadow_dirty_mask = dirty_mask;
 181         shadow_nx_mask = nx_mask;
 182         shadow_x_mask = x_mask;
 183 }
 184 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 185
 186 static int is_write_protection(struct kvm_vcpu *vcpu)
 187 {
 188         return vcpu->arch.cr0 & X86_CR0_WP;
 189 }
 190
 191 static int is_cpuid_PSE36(void)
 192 {
 193         return 1;
 194 }
 195
 196 static int is_nx(struct kvm_vcpu *vcpu)
 197 {
 198         return vcpu->arch.shadow_efer & EFER_NX;
 199 }
 200
 201 static int is_present_pte(unsigned long pte)
 202 {
 203         return pte & PT_PRESENT_MASK;
 204 }
 205
 206 static int is_shadow_present_pte(u64 pte)
 207 {
 208         return pte != shadow_trap_nonpresent_pte
 209                 && pte != shadow_notrap_nonpresent_pte;
 210 }
 211
 212 static int is_large_pte(u64 pte)
 213 {
 214         return pte & PT_PAGE_SIZE_MASK;
 215 }
 216
 217 static int is_writeble_pte(unsigned long pte)
 218 {
 219         return pte & PT_WRITABLE_MASK;
 220 }
 221
 222 static int is_dirty_pte(unsigned long pte)
 223 {
 224         return pte & shadow_dirty_mask;
 225 }
 226
 227 static int is_rmap_pte(u64 pte)
 228 {
 229         return is_shadow_present_pte(pte);
 230 }
 231
 232 static pfn_t spte_to_pfn(u64 pte)
 233 {
 234         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 235 }
 236
 237 static gfn_t pse36_gfn_delta(u32 gpte)
 238 {
 239         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 240
 241         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 242 }
 243
 244 static void set_shadow_pte(u64 *sptep, u64 spte)
 245 {
 246 #ifdef CONFIG_X86_64
 247         set_64bit((unsigned long *)sptep, spte);
 248 #else
 249         set_64bit((unsigned long long *)sptep, spte);
 250 #endif
 251 }
 252
 253 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 254                                   struct kmem_cache *base_cache, int min)
 255 {
 256         void *obj;
 257
 258         if (cache->nobjs >= min)
 259                 return 0;
 260         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 261                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 262                 if (!obj)
 263                         return -ENOMEM;
 264                 cache->objects[cache->nobjs++] = obj;
 265         }
 266         return 0;
 267 }
 268
 269 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 270 {
 271         while (mc->nobjs)
 272                 kfree(mc->objects[--mc->nobjs]);
 273 }
 274
 275 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 276                                        int min)
 277 {
 278         struct page *page;
 279
 280         if (cache->nobjs >= min)
 281                 return 0;
 282         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 283                 page = alloc_page(GFP_KERNEL);
 284                 if (!page)
 285                         return -ENOMEM;
 286                 set_page_private(page, 0);
 287                 cache->objects[cache->nobjs++] = page_address(page);
 288         }
 289         return 0;
 290 }
 291
 292 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 293 {
 294         while (mc->nobjs)
 295                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 296 }
 297
 298 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 299 {
 300         int r;
 301
 302         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 303                                    pte_chain_cache, 4);
 304         if (r)
 305                 goto out;
 306         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 307                                    rmap_desc_cache, 1);
 308         if (r)
 309                 goto out;
 310         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 311         if (r)
 312                 goto out;
 313         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 314                                    mmu_page_header_cache, 4);
 315 out:
 316         return r;
 317 }
 318
 319 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 320 {
 321         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 322         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 323         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 324         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 325 }
 326
 327 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 328                                     size_t size)
 329 {
 330         void *p;
 331
 332         BUG_ON(!mc->nobjs);
 333         p = mc->objects[--mc->nobjs];
 334         memset(p, 0, size);
 335         return p;
 336 }
 337
 338 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 339 {
 340         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 341                                       sizeof(struct kvm_pte_chain));
 342 }
 343
 344 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 345 {
 346         kfree(pc);
 347 }
 348
 349 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 350 {
 351         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 352                                       sizeof(struct kvm_rmap_desc));
 353 }
 354
 355 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 356 {
 357         kfree(rd);
 358 }
 359
 360 /*
 361  * Return the pointer to the largepage write count for a given
 362  * gfn, handling slots that are not large page aligned.
 363  */
 364 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 365 {
 366         unsigned long idx;
 367
 368         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 369               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 370         return &slot->lpage_info[idx].write_count;
 371 }
 372
 373 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 374 {
 375         int *write_count;
 376
 377         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 378         *write_count += 1;
 379 }
 380
 381 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 382 {
 383         int *write_count;
 384
 385         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 386         *write_count -= 1;
 387         WARN_ON(*write_count < 0);
 388 }
 389
 390 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 391 {
 392         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
 393         int *largepage_idx;
 394
 395         if (slot) {
 396                 largepage_idx = slot_largepage_idx(gfn, slot);
 397                 return *largepage_idx;
 398         }
 399
 400         return 1;
 401 }
 402
 403 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 404 {
 405         struct vm_area_struct *vma;
 406         unsigned long addr;
 407
 408         addr = gfn_to_hva(kvm, gfn);
 409         if (kvm_is_error_hva(addr))
 410                 return 0;
 411
 412         vma = find_vma(current->mm, addr);
 413         if (vma && is_vm_hugetlb_page(vma))
 414                 return 1;
 415
 416         return 0;
 417 }
 418
 419 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 420 {
 421         struct kvm_memory_slot *slot;
 422
 423         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 424                 return 0;
 425
 426         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 427                 return 0;
 428
 429         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 430         if (slot && slot->dirty_bitmap)
 431                 return 0;
 432
 433         return 1;
 434 }
 435
 436 /*
 437  * Take gfn and return the reverse mapping to it.
 438  * Note: gfn must be unaliased before this function get called
 439  */
 440
 441 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 442 {
 443         struct kvm_memory_slot *slot;
 444         unsigned long idx;
 445
 446         slot = gfn_to_memslot(kvm, gfn);
 447         if (!lpage)
 448                 return &slot->rmap[gfn - slot->base_gfn];
 449
 450         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 451               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 452
 453         return &slot->lpage_info[idx].rmap_pde;
 454 }
 455
 456 /*
 457  * Reverse mapping data structures:
 458  *
 459  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 460  * that points to page_address(page).
 461  *
 462  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 463  * containing more mappings.
 464  */
 465 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 466 {
 467         struct kvm_mmu_page *sp;
 468         struct kvm_rmap_desc *desc;
 469         unsigned long *rmapp;
 470         int i;
 471
 472         if (!is_rmap_pte(*spte))
 473                 return;
 474         gfn = unalias_gfn(vcpu->kvm, gfn);
 475         sp = page_header(__pa(spte));
 476         sp->gfns[spte - sp->spt] = gfn;
 477         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 478         if (!*rmapp) {
 479                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 480                 *rmapp = (unsigned long)spte;
 481         } else if (!(*rmapp & 1)) {
 482                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 483                 desc = mmu_alloc_rmap_desc(vcpu);
 484                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 485                 desc->shadow_ptes[1] = spte;
 486                 *rmapp = (unsigned long)desc | 1;
 487         } else {
 488                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 489                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 490                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 491                         desc = desc->more;
 492                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 493                         desc->more = mmu_alloc_rmap_desc(vcpu);
 494                         desc = desc->more;
 495                 }
 496                 for (i = 0; desc->shadow_ptes[i]; ++i)
 497                         ;
 498                 desc->shadow_ptes[i] = spte;
 499         }
 500 }
 501
 502 static void rmap_desc_remove_entry(unsigned long *rmapp,
 503                                    struct kvm_rmap_desc *desc,
 504                                    int i,
 505                                    struct kvm_rmap_desc *prev_desc)
 506 {
 507         int j;
 508
 509         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 510                 ;
 511         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 512         desc->shadow_ptes[j] = NULL;
 513         if (j != 0)
 514                 return;
 515         if (!prev_desc && !desc->more)
 516                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 517         else
 518                 if (prev_desc)
 519                         prev_desc->more = desc->more;
 520                 else
 521                         *rmapp = (unsigned long)desc->more | 1;
 522         mmu_free_rmap_desc(desc);
 523 }
 524
 525 static void rmap_remove(struct kvm *kvm, u64 *spte)
 526 {
 527         struct kvm_rmap_desc *desc;
 528         struct kvm_rmap_desc *prev_desc;
 529         struct kvm_mmu_page *sp;
 530         pfn_t pfn;
 531         unsigned long *rmapp;
 532         int i;
 533
 534         if (!is_rmap_pte(*spte))
 535                 return;
 536         sp = page_header(__pa(spte));
 537         pfn = spte_to_pfn(*spte);
 538         if (*spte & shadow_accessed_mask)
 539                 kvm_set_pfn_accessed(pfn);
 540         if (is_writeble_pte(*spte))
 541                 kvm_release_pfn_dirty(pfn);
 542         else
 543                 kvm_release_pfn_clean(pfn);
 544         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 545         if (!*rmapp) {
 546                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 547                 BUG();
 548         } else if (!(*rmapp & 1)) {
 549                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 550                 if ((u64 *)*rmapp != spte) {
 551                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 552                                spte, *spte);
 553                         BUG();
 554                 }
 555                 *rmapp = 0;
 556         } else {
 557                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 558                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 559                 prev_desc = NULL;
 560                 while (desc) {
 561                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 562                                 if (desc->shadow_ptes[i] == spte) {
 563                                         rmap_desc_remove_entry(rmapp,
 564                                                                desc, i,
 565                                                                prev_desc);
 566                                         return;
 567                                 }
 568                         prev_desc = desc;
 569                         desc = desc->more;
 570                 }
 571                 BUG();
 572         }
 573 }
 574
 575 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 576 {
 577         struct kvm_rmap_desc *desc;
 578         struct kvm_rmap_desc *prev_desc;
 579         u64 *prev_spte;
 580         int i;
 581
 582         if (!*rmapp)
 583                 return NULL;
 584         else if (!(*rmapp & 1)) {
 585                 if (!spte)
 586                         return (u64 *)*rmapp;
 587                 return NULL;
 588         }
 589         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 590         prev_desc = NULL;
 591         prev_spte = NULL;
 592         while (desc) {
 593                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 594                         if (prev_spte == spte)
 595                                 return desc->shadow_ptes[i];
 596                         prev_spte = desc->shadow_ptes[i];
 597                 }
 598                 desc = desc->more;
 599         }
 600         return NULL;
 601 }
 602
 603 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 604 {
 605         unsigned long *rmapp;
 606         u64 *spte;
 607         int write_protected = 0;
 608
 609         gfn = unalias_gfn(kvm, gfn);
 610         rmapp = gfn_to_rmap(kvm, gfn, 0);
 611
 612         spte = rmap_next(kvm, rmapp, NULL);
 613         while (spte) {
 614                 BUG_ON(!spte);
 615                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 616                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 617                 if (is_writeble_pte(*spte)) {
 618                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 619                         write_protected = 1;
 620                 }
 621                 spte = rmap_next(kvm, rmapp, spte);
 622         }
 623         if (write_protected) {
 624                 pfn_t pfn;
 625
 626                 spte = rmap_next(kvm, rmapp, NULL);
 627                 pfn = spte_to_pfn(*spte);
 628                 kvm_set_pfn_dirty(pfn);
 629         }
 630
 631         /* check for huge page mappings */
 632         rmapp = gfn_to_rmap(kvm, gfn, 1);
 633         spte = rmap_next(kvm, rmapp, NULL);
 634         while (spte) {
 635                 BUG_ON(!spte);
 636                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 637                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 638                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 639                 if (is_writeble_pte(*spte)) {
 640                         rmap_remove(kvm, spte);
 641                         --kvm->stat.lpages;
 642                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 643                         spte = NULL;
 644                         write_protected = 1;
 645                 }
 646                 spte = rmap_next(kvm, rmapp, spte);
 647         }
 648
 649         if (write_protected)
 650                 kvm_flush_remote_tlbs(kvm);
 651
 652         account_shadowed(kvm, gfn);
 653 }
 654
 655 #ifdef MMU_DEBUG
 656 static int is_empty_shadow_page(u64 *spt)
 657 {
 658         u64 *pos;
 659         u64 *end;
 660
 661         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 662                 if (is_shadow_present_pte(*pos)) {
 663                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 664                                pos, *pos);
 665                         return 0;
 666                 }
 667         return 1;
 668 }
 669 #endif
 670
 671 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 672 {
 673         ASSERT(is_empty_shadow_page(sp->spt));
 674         list_del(&sp->link);
 675         __free_page(virt_to_page(sp->spt));
 676         __free_page(virt_to_page(sp->gfns));
 677         kfree(sp);
 678         ++kvm->arch.n_free_mmu_pages;
 679 }
 680
 681 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 682 {
 683         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 684 }
 685
 686 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 687                                                u64 *parent_pte)
 688 {
 689         struct kvm_mmu_page *sp;
 690
 691         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 692         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 693         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 694         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 695         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 696         ASSERT(is_empty_shadow_page(sp->spt));
 697         sp->slot_bitmap = 0;
 698         sp->multimapped = 0;
 699         sp->parent_pte = parent_pte;
 700         --vcpu->kvm->arch.n_free_mmu_pages;
 701         return sp;
 702 }
 703
 704 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 705                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 706 {
 707         struct kvm_pte_chain *pte_chain;
 708         struct hlist_node *node;
 709         int i;
 710
 711         if (!parent_pte)
 712                 return;
 713         if (!sp->multimapped) {
 714                 u64 *old = sp->parent_pte;
 715
 716                 if (!old) {
 717                         sp->parent_pte = parent_pte;
 718                         return;
 719                 }
 720                 sp->multimapped = 1;
 721                 pte_chain = mmu_alloc_pte_chain(vcpu);
 722                 INIT_HLIST_HEAD(&sp->parent_ptes);
 723                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 724                 pte_chain->parent_ptes[0] = old;
 725         }
 726         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 727                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 728                         continue;
 729                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 730                         if (!pte_chain->parent_ptes[i]) {
 731                                 pte_chain->parent_ptes[i] = parent_pte;
 732                                 return;
 733                         }
 734         }
 735         pte_chain = mmu_alloc_pte_chain(vcpu);
 736         BUG_ON(!pte_chain);
 737         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 738         pte_chain->parent_ptes[0] = parent_pte;
 739 }
 740
 741 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 742                                        u64 *parent_pte)
 743 {
 744         struct kvm_pte_chain *pte_chain;
 745         struct hlist_node *node;
 746         int i;
 747
 748         if (!sp->multimapped) {
 749                 BUG_ON(sp->parent_pte != parent_pte);
 750                 sp->parent_pte = NULL;
 751                 return;
 752         }
 753         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 754                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 755                         if (!pte_chain->parent_ptes[i])
 756                                 break;
 757                         if (pte_chain->parent_ptes[i] != parent_pte)
 758                                 continue;
 759                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 760                                 && pte_chain->parent_ptes[i + 1]) {
 761                                 pte_chain->parent_ptes[i]
 762                                         = pte_chain->parent_ptes[i + 1];
 763                                 ++i;
 764                         }
 765                         pte_chain->parent_ptes[i] = NULL;
 766                         if (i == 0) {
 767                                 hlist_del(&pte_chain->link);
 768                                 mmu_free_pte_chain(pte_chain);
 769                                 if (hlist_empty(&sp->parent_ptes)) {
 770                                         sp->multimapped = 0;
 771                                         sp->parent_pte = NULL;
 772                                 }
 773                         }
 774                         return;
 775                 }
 776         BUG();
 777 }
 778
 779 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 780                                     struct kvm_mmu_page *sp)
 781 {
 782         int i;
 783
 784         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 785                 sp->spt[i] = shadow_trap_nonpresent_pte;
 786 }
 787
 788 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
 789 {
 790         unsigned index;
 791         struct hlist_head *bucket;
 792         struct kvm_mmu_page *sp;
 793         struct hlist_node *node;
 794
 795         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 796         index = kvm_page_table_hashfn(gfn);
 797         bucket = &kvm->arch.mmu_page_hash[index];
 798         hlist_for_each_entry(sp, node, bucket, hash_link)
 799                 if (sp->gfn == gfn && !sp->role.metaphysical
 800                     && !sp->role.invalid) {
 801                         pgprintk("%s: found role %x\n",
 802                                  __func__, sp->role.word);
 803                         return sp;
 804                 }
 805         return NULL;
 806 }
 807
 808 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 809                                              gfn_t gfn,
 810                                              gva_t gaddr,
 811                                              unsigned level,
 812                                              int metaphysical,
 813                                              unsigned access,
 814                                              u64 *parent_pte)
 815 {
 816         union kvm_mmu_page_role role;
 817         unsigned index;
 818         unsigned quadrant;
 819         struct hlist_head *bucket;
 820         struct kvm_mmu_page *sp;
 821         struct hlist_node *node;
 822
 823         role.word = 0;
 824         role.glevels = vcpu->arch.mmu.root_level;
 825         role.level = level;
 826         role.metaphysical = metaphysical;
 827         role.access = access;
 828         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
 829                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 830                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 831                 role.quadrant = quadrant;
 832         }
 833         pgprintk("%s: looking gfn %lx role %x\n", __func__,
 834                  gfn, role.word);
 835         index = kvm_page_table_hashfn(gfn);
 836         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
 837         hlist_for_each_entry(sp, node, bucket, hash_link)
 838                 if (sp->gfn == gfn && sp->role.word == role.word) {
 839                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
 840                         pgprintk("%s: found\n", __func__);
 841                         return sp;
 842                 }
 843         ++vcpu->kvm->stat.mmu_cache_miss;
 844         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
 845         if (!sp)
 846                 return sp;
 847         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
 848         sp->gfn = gfn;
 849         sp->role = role;
 850         hlist_add_head(&sp->hash_link, bucket);
 851         if (!metaphysical)
 852                 rmap_write_protect(vcpu->kvm, gfn);
 853         vcpu->arch.mmu.prefetch_page(vcpu, sp);
 854         return sp;
 855 }
 856
 857 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 858                                          struct kvm_mmu_page *sp)
 859 {
 860         unsigned i;
 861         u64 *pt;
 862         u64 ent;
 863
 864         pt = sp->spt;
 865
 866         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
 867                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 868                         if (is_shadow_present_pte(pt[i]))
 869                                 rmap_remove(kvm, &pt[i]);
 870                         pt[i] = shadow_trap_nonpresent_pte;
 871                 }
 872                 kvm_flush_remote_tlbs(kvm);
 873                 return;
 874         }
 875
 876         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 877                 ent = pt[i];
 878
 879                 if (is_shadow_present_pte(ent)) {
 880                         if (!is_large_pte(ent)) {
 881                                 ent &= PT64_BASE_ADDR_MASK;
 882                                 mmu_page_remove_parent_pte(page_header(ent),
 883                                                            &pt[i]);
 884                         } else {
 885                                 --kvm->stat.lpages;
 886                                 rmap_remove(kvm, &pt[i]);
 887                         }
 888                 }
 889                 pt[i] = shadow_trap_nonpresent_pte;
 890         }
 891         kvm_flush_remote_tlbs(kvm);
 892 }
 893
 894 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
 895 {
 896         mmu_page_remove_parent_pte(sp, parent_pte);
 897 }
 898
 899 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 900 {
 901         int i;
 902
 903         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 904                 if (kvm->vcpus[i])
 905                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
 906 }
 907
 908 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 909 {
 910         u64 *parent_pte;
 911
 912         ++kvm->stat.mmu_shadow_zapped;
 913         while (sp->multimapped || sp->parent_pte) {
 914                 if (!sp->multimapped)
 915                         parent_pte = sp->parent_pte;
 916                 else {
 917                         struct kvm_pte_chain *chain;
 918
 919                         chain = container_of(sp->parent_ptes.first,
 920                                              struct kvm_pte_chain, link);
 921                         parent_pte = chain->parent_ptes[0];
 922                 }
 923                 BUG_ON(!parent_pte);
 924                 kvm_mmu_put_page(sp, parent_pte);
 925                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 926         }
 927         kvm_mmu_page_unlink_children(kvm, sp);
 928         if (!sp->root_count) {
 929                 if (!sp->role.metaphysical)
 930                         unaccount_shadowed(kvm, sp->gfn);
 931                 hlist_del(&sp->hash_link);
 932                 kvm_mmu_free_page(kvm, sp);
 933         } else {
 934                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
 935                 sp->role.invalid = 1;
 936                 kvm_reload_remote_mmus(kvm);
 937         }
 938         kvm_mmu_reset_last_pte_updated(kvm);
 939 }
 940
 941 /*
 942  * Changing the number of mmu pages allocated to the vm
 943  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 944  */
 945 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 946 {
 947         /*
 948          * If we set the number of mmu pages to be smaller be than the
 949          * number of actived pages , we must to free some mmu pages before we
 950          * change the value
 951          */
 952
 953         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
 954             kvm_nr_mmu_pages) {
 955                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
 956                                        - kvm->arch.n_free_mmu_pages;
 957
 958                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 959                         struct kvm_mmu_page *page;
 960
 961                         page = container_of(kvm->arch.active_mmu_pages.prev,
 962                                             struct kvm_mmu_page, link);
 963                         kvm_mmu_zap_page(kvm, page);
 964                         n_used_mmu_pages--;
 965                 }
 966                 kvm->arch.n_free_mmu_pages = 0;
 967         }
 968         else
 969                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
 970                                          - kvm->arch.n_alloc_mmu_pages;
 971
 972         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
 973 }
 974
 975 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 976 {
 977         unsigned index;
 978         struct hlist_head *bucket;
 979         struct kvm_mmu_page *sp;
 980         struct hlist_node *node, *n;
 981         int r;
 982
 983         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 984         r = 0;
 985         index = kvm_page_table_hashfn(gfn);
 986         bucket = &kvm->arch.mmu_page_hash[index];
 987         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
 988                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 989                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
 990                                  sp->role.word);
 991                         kvm_mmu_zap_page(kvm, sp);
 992                         r = 1;
 993                 }
 994         return r;
 995 }
 996
 997 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
 998 {
 999         struct kvm_mmu_page *sp;
1000
1001         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1002                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1003                 kvm_mmu_zap_page(kvm, sp);
1004         }
1005 }
1006
1007 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1008 {
1009         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1010         struct kvm_mmu_page *sp = page_header(__pa(pte));
1011
1012         __set_bit(slot, &sp->slot_bitmap);
1013 }
1014
1015 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1016 {
1017         struct page *page;
1018
1019         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1020
1021         if (gpa == UNMAPPED_GVA)
1022                 return NULL;
1023
1024         down_read(&current->mm->mmap_sem);
1025         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1026         up_read(&current->mm->mmap_sem);
1027
1028         return page;
1029 }
1030
1031 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1032                          unsigned pt_access, unsigned pte_access,
1033                          int user_fault, int write_fault, int dirty,
1034                          int *ptwrite, int largepage, gfn_t gfn,
1035                          pfn_t pfn, bool speculative)
1036 {
1037         u64 spte;
1038         int was_rmapped = 0;
1039         int was_writeble = is_writeble_pte(*shadow_pte);
1040
1041         pgprintk("%s: spte %llx access %x write_fault %d"
1042                  " user_fault %d gfn %lx\n",
1043                  __func__, *shadow_pte, pt_access,
1044                  write_fault, user_fault, gfn);
1045
1046         if (is_rmap_pte(*shadow_pte)) {
1047                 /*
1048                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1049                  * the parent of the now unreachable PTE.
1050                  */
1051                 if (largepage && !is_large_pte(*shadow_pte)) {
1052                         struct kvm_mmu_page *child;
1053                         u64 pte = *shadow_pte;
1054
1055                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1056                         mmu_page_remove_parent_pte(child, shadow_pte);
1057                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1058                         pgprintk("hfn old %lx new %lx\n",
1059                                  spte_to_pfn(*shadow_pte), pfn);
1060                         rmap_remove(vcpu->kvm, shadow_pte);
1061                 } else {
1062                         if (largepage)
1063                                 was_rmapped = is_large_pte(*shadow_pte);
1064                         else
1065                                 was_rmapped = 1;
1066                 }
1067         }
1068
1069         /*
1070          * We don't set the accessed bit, since we sometimes want to see
1071          * whether the guest actually used the pte (in order to detect
1072          * demand paging).
1073          */
1074         spte = shadow_base_present_pte | shadow_dirty_mask;
1075         if (!speculative)
1076                 pte_access |= PT_ACCESSED_MASK;
1077         if (!dirty)
1078                 pte_access &= ~ACC_WRITE_MASK;
1079         if (pte_access & ACC_EXEC_MASK)
1080                 spte |= shadow_x_mask;
1081         else
1082                 spte |= shadow_nx_mask;
1083         if (pte_access & ACC_USER_MASK)
1084                 spte |= shadow_user_mask;
1085         if (largepage)
1086                 spte |= PT_PAGE_SIZE_MASK;
1087
1088         spte |= (u64)pfn << PAGE_SHIFT;
1089
1090         if ((pte_access & ACC_WRITE_MASK)
1091             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1092                 struct kvm_mmu_page *shadow;
1093
1094                 spte |= PT_WRITABLE_MASK;
1095
1096                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1097                 if (shadow ||
1098                    (largepage && has_wrprotected_page(vcpu->kvm, gfn))) {
1099                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1100                                  __func__, gfn);
1101                         pte_access &= ~ACC_WRITE_MASK;
1102                         if (is_writeble_pte(spte)) {
1103                                 spte &= ~PT_WRITABLE_MASK;
1104                                 kvm_x86_ops->tlb_flush(vcpu);
1105                         }
1106                         if (write_fault)
1107                                 *ptwrite = 1;
1108                 }
1109         }
1110
1111         if (pte_access & ACC_WRITE_MASK)
1112                 mark_page_dirty(vcpu->kvm, gfn);
1113
1114         pgprintk("%s: setting spte %llx\n", __func__, spte);
1115         pgprintk("instantiating %s PTE (%s) at %d (%llx) addr %llx\n",
1116                  (spte&PT_PAGE_SIZE_MASK)? "2MB" : "4kB",
1117                  (spte&PT_WRITABLE_MASK)?"RW":"R", gfn, spte, shadow_pte);
1118         set_shadow_pte(shadow_pte, spte);
1119         if (!was_rmapped && (spte & PT_PAGE_SIZE_MASK)
1120             && (spte & PT_PRESENT_MASK))
1121                 ++vcpu->kvm->stat.lpages;
1122
1123         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1124         if (!was_rmapped) {
1125                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1126                 if (!is_rmap_pte(*shadow_pte))
1127                         kvm_release_pfn_clean(pfn);
1128         } else {
1129                 if (was_writeble)
1130                         kvm_release_pfn_dirty(pfn);
1131                 else
1132                         kvm_release_pfn_clean(pfn);
1133         }
1134         if (speculative) {
1135                 vcpu->arch.last_pte_updated = shadow_pte;
1136                 vcpu->arch.last_pte_gfn = gfn;
1137         }
1138 }
1139
1140 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1141 {
1142 }
1143
1144 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1145                            int largepage, gfn_t gfn, pfn_t pfn,
1146                            int level)
1147 {
1148         hpa_t table_addr = vcpu->arch.mmu.root_hpa;
1149         int pt_write = 0;
1150
1151         for (; ; level--) {
1152                 u32 index = PT64_INDEX(v, level);
1153                 u64 *table;
1154
1155                 ASSERT(VALID_PAGE(table_addr));
1156                 table = __va(table_addr);
1157
1158                 if (level == 1) {
1159                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1160                                      0, write, 1, &pt_write, 0, gfn, pfn, false);
1161                         return pt_write;
1162                 }
1163
1164                 if (largepage && level == 2) {
1165                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1166                                      0, write, 1, &pt_write, 1, gfn, pfn, false);
1167                         return pt_write;
1168                 }
1169
1170                 if (table[index] == shadow_trap_nonpresent_pte) {
1171                         struct kvm_mmu_page *new_table;
1172                         gfn_t pseudo_gfn;
1173
1174                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1175                                 >> PAGE_SHIFT;
1176                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1177                                                      v, level - 1,
1178                                                      1, ACC_ALL, &table[index]);
1179                         if (!new_table) {
1180                                 pgprintk("nonpaging_map: ENOMEM\n");
1181                                 kvm_release_pfn_clean(pfn);
1182                                 return -ENOMEM;
1183                         }
1184
1185                         table[index] = __pa(new_table->spt)
1186                                 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1187                                 | shadow_user_mask | shadow_x_mask;
1188                 }
1189                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1190         }
1191 }
1192
1193 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1194 {
1195         int r;
1196         int largepage = 0;
1197         pfn_t pfn;
1198
1199         down_read(&current->mm->mmap_sem);
1200         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1201                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1202                 largepage = 1;
1203         }
1204
1205         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1206         up_read(&current->mm->mmap_sem);
1207
1208         /* mmio */
1209         if (is_error_pfn(pfn)) {
1210                 kvm_release_pfn_clean(pfn);
1211                 return 1;
1212         }
1213
1214         spin_lock(&vcpu->kvm->mmu_lock);
1215         kvm_mmu_free_some_pages(vcpu);
1216         r = __direct_map(vcpu, v, write, largepage, gfn, pfn,
1217                          PT32E_ROOT_LEVEL);
1218         spin_unlock(&vcpu->kvm->mmu_lock);
1219
1220
1221         return r;
1222 }
1223
1224
1225 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1226 {
1227         int i;
1228         struct kvm_mmu_page *sp;
1229
1230         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1231                 return;
1232         spin_lock(&vcpu->kvm->mmu_lock);
1233         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1234                 hpa_t root = vcpu->arch.mmu.root_hpa;
1235
1236                 sp = page_header(root);
1237                 --sp->root_count;
1238                 if (!sp->root_count && sp->role.invalid)
1239                         kvm_mmu_zap_page(vcpu->kvm, sp);
1240                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1241                 spin_unlock(&vcpu->kvm->mmu_lock);
1242                 return;
1243         }
1244         for (i = 0; i < 4; ++i) {
1245                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1246
1247                 if (root) {
1248                         root &= PT64_BASE_ADDR_MASK;
1249                         sp = page_header(root);
1250                         --sp->root_count;
1251                         if (!sp->root_count && sp->role.invalid)
1252                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1253                 }
1254                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1255         }
1256         spin_unlock(&vcpu->kvm->mmu_lock);
1257         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1258 }
1259
1260 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1261 {
1262         int i;
1263         gfn_t root_gfn;
1264         struct kvm_mmu_page *sp;
1265         int metaphysical = 0;
1266
1267         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1268
1269         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1270                 hpa_t root = vcpu->arch.mmu.root_hpa;
1271
1272                 ASSERT(!VALID_PAGE(root));
1273                 if (tdp_enabled)
1274                         metaphysical = 1;
1275                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1276                                       PT64_ROOT_LEVEL, metaphysical,
1277                                       ACC_ALL, NULL);
1278                 root = __pa(sp->spt);
1279                 ++sp->root_count;
1280                 vcpu->arch.mmu.root_hpa = root;
1281                 return;
1282         }
1283         metaphysical = !is_paging(vcpu);
1284         if (tdp_enabled)
1285                 metaphysical = 1;
1286         for (i = 0; i < 4; ++i) {
1287                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1288
1289                 ASSERT(!VALID_PAGE(root));
1290                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1291                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1292                                 vcpu->arch.mmu.pae_root[i] = 0;
1293                                 continue;
1294                         }
1295                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1296                 } else if (vcpu->arch.mmu.root_level == 0)
1297                         root_gfn = 0;
1298                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1299                                       PT32_ROOT_LEVEL, metaphysical,
1300                                       ACC_ALL, NULL);
1301                 root = __pa(sp->spt);
1302                 ++sp->root_count;
1303                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1304         }
1305         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1306 }
1307
1308 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1309 {
1310         return vaddr;
1311 }
1312
1313 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1314                                 u32 error_code)
1315 {
1316         gfn_t gfn;
1317         int r;
1318
1319         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1320         r = mmu_topup_memory_caches(vcpu);
1321         if (r)
1322                 return r;
1323
1324         ASSERT(vcpu);
1325         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1326
1327         gfn = gva >> PAGE_SHIFT;
1328
1329         return nonpaging_map(vcpu, gva & PAGE_MASK,
1330                              error_code & PFERR_WRITE_MASK, gfn);
1331 }
1332
1333 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1334                                 u32 error_code)
1335 {
1336         pfn_t pfn;
1337         int r;
1338         int largepage = 0;
1339         gfn_t gfn = gpa >> PAGE_SHIFT;
1340
1341         ASSERT(vcpu);
1342         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1343
1344         r = mmu_topup_memory_caches(vcpu);
1345         if (r)
1346                 return r;
1347
1348         down_read(&current->mm->mmap_sem);
1349         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1350                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1351                 largepage = 1;
1352         }
1353         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1354         up_read(&current->mm->mmap_sem);
1355         if (is_error_pfn(pfn)) {
1356                 kvm_release_pfn_clean(pfn);
1357                 return 1;
1358         }
1359         spin_lock(&vcpu->kvm->mmu_lock);
1360         kvm_mmu_free_some_pages(vcpu);
1361         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1362                          largepage, gfn, pfn, kvm_x86_ops->get_tdp_level());
1363         spin_unlock(&vcpu->kvm->mmu_lock);
1364
1365         return r;
1366 }
1367
1368 static void nonpaging_free(struct kvm_vcpu *vcpu)
1369 {
1370         mmu_free_roots(vcpu);
1371 }
1372
1373 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1374 {
1375         struct kvm_mmu *context = &vcpu->arch.mmu;
1376
1377         context->new_cr3 = nonpaging_new_cr3;
1378         context->page_fault = nonpaging_page_fault;
1379         context->gva_to_gpa = nonpaging_gva_to_gpa;
1380         context->free = nonpaging_free;
1381         context->prefetch_page = nonpaging_prefetch_page;
1382         context->root_level = 0;
1383         context->shadow_root_level = PT32E_ROOT_LEVEL;
1384         context->root_hpa = INVALID_PAGE;
1385         return 0;
1386 }
1387
1388 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1389 {
1390         ++vcpu->stat.tlb_flush;
1391         kvm_x86_ops->tlb_flush(vcpu);
1392 }
1393
1394 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1395 {
1396         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1397         mmu_free_roots(vcpu);
1398 }
1399
1400 static void inject_page_fault(struct kvm_vcpu *vcpu,
1401                               u64 addr,
1402                               u32 err_code)
1403 {
1404         kvm_inject_page_fault(vcpu, addr, err_code);
1405 }
1406
1407 static void paging_free(struct kvm_vcpu *vcpu)
1408 {
1409         nonpaging_free(vcpu);
1410 }
1411
1412 #define PTTYPE 64
1413 #include "paging_tmpl.h"
1414 #undef PTTYPE
1415
1416 #define PTTYPE 32
1417 #include "paging_tmpl.h"
1418 #undef PTTYPE
1419
1420 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1421 {
1422         struct kvm_mmu *context = &vcpu->arch.mmu;
1423
1424         ASSERT(is_pae(vcpu));
1425         context->new_cr3 = paging_new_cr3;
1426         context->page_fault = paging64_page_fault;
1427         context->gva_to_gpa = paging64_gva_to_gpa;
1428         context->prefetch_page = paging64_prefetch_page;
1429         context->free = paging_free;
1430         context->root_level = level;
1431         context->shadow_root_level = level;
1432         context->root_hpa = INVALID_PAGE;
1433         return 0;
1434 }
1435
1436 static int paging64_init_context(struct kvm_vcpu *vcpu)
1437 {
1438         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1439 }
1440
1441 static int paging32_init_context(struct kvm_vcpu *vcpu)
1442 {
1443         struct kvm_mmu *context = &vcpu->arch.mmu;
1444
1445         context->new_cr3 = paging_new_cr3;
1446         context->page_fault = paging32_page_fault;
1447         context->gva_to_gpa = paging32_gva_to_gpa;
1448         context->free = paging_free;
1449         context->prefetch_page = paging32_prefetch_page;
1450         context->root_level = PT32_ROOT_LEVEL;
1451         context->shadow_root_level = PT32E_ROOT_LEVEL;
1452         context->root_hpa = INVALID_PAGE;
1453         return 0;
1454 }
1455
1456 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1457 {
1458         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1459 }
1460
1461 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1462 {
1463         struct kvm_mmu *context = &vcpu->arch.mmu;
1464
1465         context->new_cr3 = nonpaging_new_cr3;
1466         context->page_fault = tdp_page_fault;
1467         context->free = nonpaging_free;
1468         context->prefetch_page = nonpaging_prefetch_page;
1469         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
1470         context->root_hpa = INVALID_PAGE;
1471
1472         if (!is_paging(vcpu)) {
1473                 context->gva_to_gpa = nonpaging_gva_to_gpa;
1474                 context->root_level = 0;
1475         } else if (is_long_mode(vcpu)) {
1476                 context->gva_to_gpa = paging64_gva_to_gpa;
1477                 context->root_level = PT64_ROOT_LEVEL;
1478         } else if (is_pae(vcpu)) {
1479                 context->gva_to_gpa = paging64_gva_to_gpa;
1480                 context->root_level = PT32E_ROOT_LEVEL;
1481         } else {
1482                 context->gva_to_gpa = paging32_gva_to_gpa;
1483                 context->root_level = PT32_ROOT_LEVEL;
1484         }
1485
1486         return 0;
1487 }
1488
1489 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1490 {
1491         ASSERT(vcpu);
1492         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1493
1494         if (!is_paging(vcpu))
1495                 return nonpaging_init_context(vcpu);
1496         else if (is_long_mode(vcpu))
1497                 return paging64_init_context(vcpu);
1498         else if (is_pae(vcpu))
1499                 return paging32E_init_context(vcpu);
1500         else
1501                 return paging32_init_context(vcpu);
1502 }
1503
1504 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1505 {
1506         vcpu->arch.update_pte.pfn = bad_pfn;
1507
1508         if (tdp_enabled)
1509                 return init_kvm_tdp_mmu(vcpu);
1510         else
1511                 return init_kvm_softmmu(vcpu);
1512 }
1513
1514 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1515 {
1516         ASSERT(vcpu);
1517         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1518                 vcpu->arch.mmu.free(vcpu);
1519                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1520         }
1521 }
1522
1523 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1524 {
1525         destroy_kvm_mmu(vcpu);
1526         return init_kvm_mmu(vcpu);
1527 }
1528 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1529
1530 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1531 {
1532         int r;
1533
1534         r = mmu_topup_memory_caches(vcpu);
1535         if (r)
1536                 goto out;
1537         spin_lock(&vcpu->kvm->mmu_lock);
1538         kvm_mmu_free_some_pages(vcpu);
1539         mmu_alloc_roots(vcpu);
1540         spin_unlock(&vcpu->kvm->mmu_lock);
1541         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1542         kvm_mmu_flush_tlb(vcpu);
1543 out:
1544         return r;
1545 }
1546 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1547
1548 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1549 {
1550         mmu_free_roots(vcpu);
1551 }
1552
1553 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1554                                   struct kvm_mmu_page *sp,
1555                                   u64 *spte)
1556 {
1557         u64 pte;
1558         struct kvm_mmu_page *child;
1559
1560         pte = *spte;
1561         if (is_shadow_present_pte(pte)) {
1562                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
1563                     is_large_pte(pte))
1564                         rmap_remove(vcpu->kvm, spte);
1565                 else {
1566                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1567                         mmu_page_remove_parent_pte(child, spte);
1568                 }
1569         }
1570         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1571         if (is_large_pte(pte))
1572                 --vcpu->kvm->stat.lpages;
1573 }
1574
1575 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1576                                   struct kvm_mmu_page *sp,
1577                                   u64 *spte,
1578                                   const void *new)
1579 {
1580         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1581                 if (!vcpu->arch.update_pte.largepage ||
1582                     sp->role.glevels == PT32_ROOT_LEVEL) {
1583                         ++vcpu->kvm->stat.mmu_pde_zapped;
1584                         return;
1585                 }
1586         }
1587
1588         ++vcpu->kvm->stat.mmu_pte_updated;
1589         if (sp->role.glevels == PT32_ROOT_LEVEL)
1590                 paging32_update_pte(vcpu, sp, spte, new);
1591         else
1592                 paging64_update_pte(vcpu, sp, spte, new);
1593 }
1594
1595 static bool need_remote_flush(u64 old, u64 new)
1596 {
1597         if (!is_shadow_present_pte(old))
1598                 return false;
1599         if (!is_shadow_present_pte(new))
1600                 return true;
1601         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1602                 return true;
1603         old ^= PT64_NX_MASK;
1604         new ^= PT64_NX_MASK;
1605         return (old & ~new & PT64_PERM_MASK) != 0;
1606 }
1607
1608 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1609 {
1610         if (need_remote_flush(old, new))
1611                 kvm_flush_remote_tlbs(vcpu->kvm);
1612         else
1613                 kvm_mmu_flush_tlb(vcpu);
1614 }
1615
1616 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1617 {
1618         u64 *spte = vcpu->arch.last_pte_updated;
1619
1620         return !!(spte && (*spte & shadow_accessed_mask));
1621 }
1622
1623 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1624                                           const u8 *new, int bytes)
1625 {
1626         gfn_t gfn;
1627         int r;
1628         u64 gpte = 0;
1629         pfn_t pfn;
1630
1631         vcpu->arch.update_pte.largepage = 0;
1632
1633         if (bytes != 4 && bytes != 8)
1634                 return;
1635
1636         /*
1637          * Assume that the pte write on a page table of the same type
1638          * as the current vcpu paging mode.  This is nearly always true
1639          * (might be false while changing modes).  Note it is verified later
1640          * by update_pte().
1641          */
1642         if (is_pae(vcpu)) {
1643                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1644                 if ((bytes == 4) && (gpa % 4 == 0)) {
1645                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1646                         if (r)
1647                                 return;
1648                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1649                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1650                         memcpy((void *)&gpte, new, 8);
1651                 }
1652         } else {
1653                 if ((bytes == 4) && (gpa % 4 == 0))
1654                         memcpy((void *)&gpte, new, 4);
1655         }
1656         if (!is_present_pte(gpte))
1657                 return;
1658         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1659
1660         down_read(&current->mm->mmap_sem);
1661         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
1662                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1663                 vcpu->arch.update_pte.largepage = 1;
1664         }
1665         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1666         up_read(&current->mm->mmap_sem);
1667
1668         if (is_error_pfn(pfn)) {
1669                 kvm_release_pfn_clean(pfn);
1670                 return;
1671         }
1672         vcpu->arch.update_pte.gfn = gfn;
1673         vcpu->arch.update_pte.pfn = pfn;
1674 }
1675
1676 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
1677 {
1678         u64 *spte = vcpu->arch.last_pte_updated;
1679
1680         if (spte
1681             && vcpu->arch.last_pte_gfn == gfn
1682             && shadow_accessed_mask
1683             && !(*spte & shadow_accessed_mask)
1684             && is_shadow_present_pte(*spte))
1685                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
1686 }
1687
1688 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1689                        const u8 *new, int bytes)
1690 {
1691         gfn_t gfn = gpa >> PAGE_SHIFT;
1692         struct kvm_mmu_page *sp;
1693         struct hlist_node *node, *n;
1694         struct hlist_head *bucket;
1695         unsigned index;
1696         u64 entry, gentry;
1697         u64 *spte;
1698         unsigned offset = offset_in_page(gpa);
1699         unsigned pte_size;
1700         unsigned page_offset;
1701         unsigned misaligned;
1702         unsigned quadrant;
1703         int level;
1704         int flooded = 0;
1705         int npte;
1706         int r;
1707
1708         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
1709         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1710         spin_lock(&vcpu->kvm->mmu_lock);
1711         kvm_mmu_access_page(vcpu, gfn);
1712         kvm_mmu_free_some_pages(vcpu);
1713         ++vcpu->kvm->stat.mmu_pte_write;
1714         kvm_mmu_audit(vcpu, "pre pte write");
1715         if (gfn == vcpu->arch.last_pt_write_gfn
1716             && !last_updated_pte_accessed(vcpu)) {
1717                 ++vcpu->arch.last_pt_write_count;
1718                 if (vcpu->arch.last_pt_write_count >= 3)
1719                         flooded = 1;
1720         } else {
1721                 vcpu->arch.last_pt_write_gfn = gfn;
1722                 vcpu->arch.last_pt_write_count = 1;
1723                 vcpu->arch.last_pte_updated = NULL;
1724         }
1725         index = kvm_page_table_hashfn(gfn);
1726         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1727         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1728                 if (sp->gfn != gfn || sp->role.metaphysical)
1729                         continue;
1730                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1731                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1732                 misaligned |= bytes < 4;
1733                 if (misaligned || flooded) {
1734                         /*
1735                          * Misaligned accesses are too much trouble to fix
1736                          * up; also, they usually indicate a page is not used
1737                          * as a page table.
1738                          *
1739                          * If we're seeing too many writes to a page,
1740                          * it may no longer be a page table, or we may be
1741                          * forking, in which case it is better to unmap the
1742                          * page.
1743                          */
1744                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1745                                  gpa, bytes, sp->role.word);
1746                         kvm_mmu_zap_page(vcpu->kvm, sp);
1747                         ++vcpu->kvm->stat.mmu_flooded;
1748                         continue;
1749                 }
1750                 page_offset = offset;
1751                 level = sp->role.level;
1752                 npte = 1;
1753                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1754                         page_offset <<= 1;      /* 32->64 */
1755                         /*
1756                          * A 32-bit pde maps 4MB while the shadow pdes map
1757                          * only 2MB.  So we need to double the offset again
1758                          * and zap two pdes instead of one.
1759                          */
1760                         if (level == PT32_ROOT_LEVEL) {
1761                                 page_offset &= ~7; /* kill rounding error */
1762                                 page_offset <<= 1;
1763                                 npte = 2;
1764                         }
1765                         quadrant = page_offset >> PAGE_SHIFT;
1766                         page_offset &= ~PAGE_MASK;
1767                         if (quadrant != sp->role.quadrant)
1768                                 continue;
1769                 }
1770                 spte = &sp->spt[page_offset / sizeof(*spte)];
1771                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1772                         gentry = 0;
1773                         r = kvm_read_guest_atomic(vcpu->kvm,
1774                                                   gpa & ~(u64)(pte_size - 1),
1775                                                   &gentry, pte_size);
1776                         new = (const void *)&gentry;
1777                         if (r < 0)
1778                                 new = NULL;
1779                 }
1780                 while (npte--) {
1781                         entry = *spte;
1782                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1783                         if (new)
1784                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1785                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1786                         ++spte;
1787                 }
1788         }
1789         kvm_mmu_audit(vcpu, "post pte write");
1790         spin_unlock(&vcpu->kvm->mmu_lock);
1791         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
1792                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
1793                 vcpu->arch.update_pte.pfn = bad_pfn;
1794         }
1795 }
1796
1797 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1798 {
1799         gpa_t gpa;
1800         int r;
1801
1802         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1803
1804         spin_lock(&vcpu->kvm->mmu_lock);
1805         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1806         spin_unlock(&vcpu->kvm->mmu_lock);
1807         return r;
1808 }
1809
1810 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1811 {
1812         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1813                 struct kvm_mmu_page *sp;
1814
1815                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1816                                   struct kvm_mmu_page, link);
1817                 kvm_mmu_zap_page(vcpu->kvm, sp);
1818                 ++vcpu->kvm->stat.mmu_recycled;
1819         }
1820 }
1821
1822 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1823 {
1824         int r;
1825         enum emulation_result er;
1826
1827         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1828         if (r < 0)
1829                 goto out;
1830
1831         if (!r) {
1832                 r = 1;
1833                 goto out;
1834         }
1835
1836         r = mmu_topup_memory_caches(vcpu);
1837         if (r)
1838                 goto out;
1839
1840         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1841
1842         switch (er) {
1843         case EMULATE_DONE:
1844                 return 1;
1845         case EMULATE_DO_MMIO:
1846                 ++vcpu->stat.mmio_exits;
1847                 return 0;
1848         case EMULATE_FAIL:
1849                 kvm_report_emulation_failure(vcpu, "pagetable");
1850                 return 1;
1851         default:
1852                 BUG();
1853         }
1854 out:
1855         return r;
1856 }
1857 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1858
1859 void kvm_enable_tdp(void)
1860 {
1861         tdp_enabled = true;
1862 }
1863 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
1864
1865 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1866 {
1867         struct kvm_mmu_page *sp;
1868
1869         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1870                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1871                                   struct kvm_mmu_page, link);
1872                 kvm_mmu_zap_page(vcpu->kvm, sp);
1873                 cond_resched();
1874         }
1875         free_page((unsigned long)vcpu->arch.mmu.pae_root);
1876 }
1877
1878 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1879 {
1880         struct page *page;
1881         int i;
1882
1883         ASSERT(vcpu);
1884
1885         if (vcpu->kvm->arch.n_requested_mmu_pages)
1886                 vcpu->kvm->arch.n_free_mmu_pages =
1887                                         vcpu->kvm->arch.n_requested_mmu_pages;
1888         else
1889                 vcpu->kvm->arch.n_free_mmu_pages =
1890                                         vcpu->kvm->arch.n_alloc_mmu_pages;
1891         /*
1892          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1893          * Therefore we need to allocate shadow page tables in the first
1894          * 4GB of memory, which happens to fit the DMA32 zone.
1895          */
1896         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1897         if (!page)
1898                 goto error_1;
1899         vcpu->arch.mmu.pae_root = page_address(page);
1900         for (i = 0; i < 4; ++i)
1901                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1902
1903         return 0;
1904
1905 error_1:
1906         free_mmu_pages(vcpu);
1907         return -ENOMEM;
1908 }
1909
1910 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1911 {
1912         ASSERT(vcpu);
1913         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1914
1915         return alloc_mmu_pages(vcpu);
1916 }
1917
1918 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1919 {
1920         ASSERT(vcpu);
1921         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1922
1923         return init_kvm_mmu(vcpu);
1924 }
1925
1926 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1927 {
1928         ASSERT(vcpu);
1929
1930         destroy_kvm_mmu(vcpu);
1931         free_mmu_pages(vcpu);
1932         mmu_free_memory_caches(vcpu);
1933 }
1934
1935 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1936 {
1937         struct kvm_mmu_page *sp;
1938
1939         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1940                 int i;
1941                 u64 *pt;
1942
1943                 if (!test_bit(slot, &sp->slot_bitmap))
1944                         continue;
1945
1946                 pt = sp->spt;
1947                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1948                         /* avoid RMW */
1949                         if (pt[i] & PT_WRITABLE_MASK)
1950                                 pt[i] &= ~PT_WRITABLE_MASK;
1951         }
1952 }
1953
1954 void kvm_mmu_zap_all(struct kvm *kvm)
1955 {
1956         struct kvm_mmu_page *sp, *node;
1957
1958         spin_lock(&kvm->mmu_lock);
1959         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1960                 kvm_mmu_zap_page(kvm, sp);
1961         spin_unlock(&kvm->mmu_lock);
1962
1963         kvm_flush_remote_tlbs(kvm);
1964 }
1965
1966 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
1967 {
1968         struct kvm_mmu_page *page;
1969
1970         page = container_of(kvm->arch.active_mmu_pages.prev,
1971                             struct kvm_mmu_page, link);
1972         kvm_mmu_zap_page(kvm, page);
1973 }
1974
1975 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
1976 {
1977         struct kvm *kvm;
1978         struct kvm *kvm_freed = NULL;
1979         int cache_count = 0;
1980
1981         spin_lock(&kvm_lock);
1982
1983         list_for_each_entry(kvm, &vm_list, vm_list) {
1984                 int npages;
1985
1986                 spin_lock(&kvm->mmu_lock);
1987                 npages = kvm->arch.n_alloc_mmu_pages -
1988                          kvm->arch.n_free_mmu_pages;
1989                 cache_count += npages;
1990                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
1991                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
1992                         cache_count--;
1993                         kvm_freed = kvm;
1994                 }
1995                 nr_to_scan--;
1996
1997                 spin_unlock(&kvm->mmu_lock);
1998         }
1999         if (kvm_freed)
2000                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2001
2002         spin_unlock(&kvm_lock);
2003
2004         return cache_count;
2005 }
2006
2007 static struct shrinker mmu_shrinker = {
2008         .shrink = mmu_shrink,
2009         .seeks = DEFAULT_SEEKS * 10,
2010 };
2011
2012 static void mmu_destroy_caches(void)
2013 {
2014         if (pte_chain_cache)
2015                 kmem_cache_destroy(pte_chain_cache);
2016         if (rmap_desc_cache)
2017                 kmem_cache_destroy(rmap_desc_cache);
2018         if (mmu_page_header_cache)
2019                 kmem_cache_destroy(mmu_page_header_cache);
2020 }
2021
2022 void kvm_mmu_module_exit(void)
2023 {
2024         mmu_destroy_caches();
2025         unregister_shrinker(&mmu_shrinker);
2026 }
2027
2028 int kvm_mmu_module_init(void)
2029 {
2030         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2031                                             sizeof(struct kvm_pte_chain),
2032                                             0, 0, NULL);
2033         if (!pte_chain_cache)
2034                 goto nomem;
2035         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2036                                             sizeof(struct kvm_rmap_desc),
2037                                             0, 0, NULL);
2038         if (!rmap_desc_cache)
2039                 goto nomem;
2040
2041         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2042                                                   sizeof(struct kvm_mmu_page),
2043                                                   0, 0, NULL);
2044         if (!mmu_page_header_cache)
2045                 goto nomem;
2046
2047         register_shrinker(&mmu_shrinker);
2048
2049         return 0;
2050
2051 nomem:
2052         mmu_destroy_caches();
2053         return -ENOMEM;
2054 }
2055
2056 /*
2057  * Caculate mmu pages needed for kvm.
2058  */
2059 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2060 {
2061         int i;
2062         unsigned int nr_mmu_pages;
2063         unsigned int  nr_pages = 0;
2064
2065         for (i = 0; i < kvm->nmemslots; i++)
2066                 nr_pages += kvm->memslots[i].npages;
2067
2068         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2069         nr_mmu_pages = max(nr_mmu_pages,
2070                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2071
2072         return nr_mmu_pages;
2073 }
2074
2075 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2076                                 unsigned len)
2077 {
2078         if (len > buffer->len)
2079                 return NULL;
2080         return buffer->ptr;
2081 }
2082
2083 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2084                                 unsigned len)
2085 {
2086         void *ret;
2087
2088         ret = pv_mmu_peek_buffer(buffer, len);
2089         if (!ret)
2090                 return ret;
2091         buffer->ptr += len;
2092         buffer->len -= len;
2093         buffer->processed += len;
2094         return ret;
2095 }
2096
2097 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2098                              gpa_t addr, gpa_t value)
2099 {
2100         int bytes = 8;
2101         int r;
2102
2103         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2104                 bytes = 4;
2105
2106         r = mmu_topup_memory_caches(vcpu);
2107         if (r)
2108                 return r;
2109
2110         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2111                 return -EFAULT;
2112
2113         return 1;
2114 }
2115
2116 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2117 {
2118         kvm_x86_ops->tlb_flush(vcpu);
2119         return 1;
2120 }
2121
2122 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2123 {
2124         spin_lock(&vcpu->kvm->mmu_lock);
2125         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2126         spin_unlock(&vcpu->kvm->mmu_lock);
2127         return 1;
2128 }
2129
2130 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2131                              struct kvm_pv_mmu_op_buffer *buffer)
2132 {
2133         struct kvm_mmu_op_header *header;
2134
2135         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2136         if (!header)
2137                 return 0;
2138         switch (header->op) {
2139         case KVM_MMU_OP_WRITE_PTE: {
2140                 struct kvm_mmu_op_write_pte *wpte;
2141
2142                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2143                 if (!wpte)
2144                         return 0;
2145                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2146                                         wpte->pte_val);
2147         }
2148         case KVM_MMU_OP_FLUSH_TLB: {
2149                 struct kvm_mmu_op_flush_tlb *ftlb;
2150
2151                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2152                 if (!ftlb)
2153                         return 0;
2154                 return kvm_pv_mmu_flush_tlb(vcpu);
2155         }
2156         case KVM_MMU_OP_RELEASE_PT: {
2157                 struct kvm_mmu_op_release_pt *rpt;
2158
2159                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2160                 if (!rpt)
2161                         return 0;
2162                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2163         }
2164         default: return 0;
2165         }
2166 }
2167
2168 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2169                   gpa_t addr, unsigned long *ret)
2170 {
2171         int r;
2172         struct kvm_pv_mmu_op_buffer buffer;
2173
2174         buffer.ptr = buffer.buf;
2175         buffer.len = min_t(unsigned long, bytes, sizeof buffer.buf);
2176         buffer.processed = 0;
2177
2178         r = kvm_read_guest(vcpu->kvm, addr, buffer.buf, buffer.len);
2179         if (r)
2180                 goto out;
2181
2182         while (buffer.len) {
2183                 r = kvm_pv_mmu_op_one(vcpu, &buffer);
2184                 if (r < 0)
2185                         goto out;
2186                 if (r == 0)
2187                         break;
2188         }
2189
2190         r = 1;
2191 out:
2192         *ret = buffer.processed;
2193         return r;
2194 }
2195
2196 #ifdef AUDIT
2197
2198 static const char *audit_msg;
2199
2200 static gva_t canonicalize(gva_t gva)
2201 {
2202 #ifdef CONFIG_X86_64
2203         gva = (long long)(gva << 16) >> 16;
2204 #endif
2205         return gva;
2206 }
2207
2208 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2209                                 gva_t va, int level)
2210 {
2211         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2212         int i;
2213         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2214
2215         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2216                 u64 ent = pt[i];
2217
2218                 if (ent == shadow_trap_nonpresent_pte)
2219                         continue;
2220
2221                 va = canonicalize(va);
2222                 if (level > 1) {
2223                         if (ent == shadow_notrap_nonpresent_pte)
2224                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2225                                        " in nonleaf level: levels %d gva %lx"
2226                                        " level %d pte %llx\n", audit_msg,
2227                                        vcpu->arch.mmu.root_level, va, level, ent);
2228
2229                         audit_mappings_page(vcpu, ent, va, level - 1);
2230                 } else {
2231                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2232                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2233
2234                         if (is_shadow_present_pte(ent)
2235                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2236                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2237                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2238                                        audit_msg, vcpu->arch.mmu.root_level,
2239                                        va, gpa, hpa, ent,
2240                                        is_shadow_present_pte(ent));
2241                         else if (ent == shadow_notrap_nonpresent_pte
2242                                  && !is_error_hpa(hpa))
2243                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2244                                        " valid guest gva %lx\n", audit_msg, va);
2245                         kvm_release_pfn_clean(pfn);
2246
2247                 }
2248         }
2249 }
2250
2251 static void audit_mappings(struct kvm_vcpu *vcpu)
2252 {
2253         unsigned i;
2254
2255         if (vcpu->arch.mmu.root_level == 4)
2256                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2257         else
2258                 for (i = 0; i < 4; ++i)
2259                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2260                                 audit_mappings_page(vcpu,
2261                                                     vcpu->arch.mmu.pae_root[i],
2262                                                     i << 30,
2263                                                     2);
2264 }
2265
2266 static int count_rmaps(struct kvm_vcpu *vcpu)
2267 {
2268         int nmaps = 0;
2269         int i, j, k;
2270
2271         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2272                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2273                 struct kvm_rmap_desc *d;
2274
2275                 for (j = 0; j < m->npages; ++j) {
2276                         unsigned long *rmapp = &m->rmap[j];
2277
2278                         if (!*rmapp)
2279                                 continue;
2280                         if (!(*rmapp & 1)) {
2281                                 ++nmaps;
2282                                 continue;
2283                         }
2284                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2285                         while (d) {
2286                                 for (k = 0; k < RMAP_EXT; ++k)
2287                                         if (d->shadow_ptes[k])
2288                                                 ++nmaps;
2289                                         else
2290                                                 break;
2291                                 d = d->more;
2292                         }
2293                 }
2294         }
2295         return nmaps;
2296 }
2297
2298 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2299 {
2300         int nmaps = 0;
2301         struct kvm_mmu_page *sp;
2302         int i;
2303
2304         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2305                 u64 *pt = sp->spt;
2306
2307                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2308                         continue;
2309
2310                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2311                         u64 ent = pt[i];
2312
2313                         if (!(ent & PT_PRESENT_MASK))
2314                                 continue;
2315                         if (!(ent & PT_WRITABLE_MASK))
2316                                 continue;
2317                         ++nmaps;
2318                 }
2319         }
2320         return nmaps;
2321 }
2322
2323 static void audit_rmap(struct kvm_vcpu *vcpu)
2324 {
2325         int n_rmap = count_rmaps(vcpu);
2326         int n_actual = count_writable_mappings(vcpu);
2327
2328         if (n_rmap != n_actual)
2329                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2330                        __func__, audit_msg, n_rmap, n_actual);
2331 }
2332
2333 static void audit_write_protection(struct kvm_vcpu *vcpu)
2334 {
2335         struct kvm_mmu_page *sp;
2336         struct kvm_memory_slot *slot;
2337         unsigned long *rmapp;
2338         gfn_t gfn;
2339
2340         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2341                 if (sp->role.metaphysical)
2342                         continue;
2343
2344                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2345                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2346                 rmapp = &slot->rmap[gfn - slot->base_gfn];
2347                 if (*rmapp)
2348                         printk(KERN_ERR "%s: (%s) shadow page has writable"
2349                                " mappings: gfn %lx role %x\n",
2350                                __func__, audit_msg, sp->gfn,
2351                                sp->role.word);
2352         }
2353 }
2354
2355 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2356 {
2357         int olddbg = dbg;
2358
2359         dbg = 0;
2360         audit_msg = msg;
2361         audit_rmap(vcpu);
2362         audit_write_protection(vcpu);
2363         audit_mappings(vcpu);
2364         dbg = olddbg;
2365 }
2366
2367 #endif