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 "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 = 0;
  70 module_param(dbg, bool, 0644);
  71 #endif
  72
  73 #ifndef MMU_DEBUG
  74 #define ASSERT(x) do { } while (0)
  75 #else
  76 #define ASSERT(x)                                                       \
  77         if (!(x)) {                                                     \
  78                 printk(KERN_WARNING "assertion failed %s:%d: %s\n",     \
  79                        __FILE__, __LINE__, #x);                         \
  80         }
  81 #endif
  82
  83 #define PT_FIRST_AVAIL_BITS_SHIFT 9
  84 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
  85
  86 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
  87
  88 #define PT64_LEVEL_BITS 9
  89
  90 #define PT64_LEVEL_SHIFT(level) \
  91                 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
  92
  93 #define PT64_LEVEL_MASK(level) \
  94                 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
  95
  96 #define PT64_INDEX(address, level)\
  97         (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
  98
  99
 100 #define PT32_LEVEL_BITS 10
 101
 102 #define PT32_LEVEL_SHIFT(level) \
 103                 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
 104
 105 #define PT32_LEVEL_MASK(level) \
 106                 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
 107
 108 #define PT32_INDEX(address, level)\
 109         (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
 110
 111
 112 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
 113 #define PT64_DIR_BASE_ADDR_MASK \
 114         (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
 115
 116 #define PT32_BASE_ADDR_MASK PAGE_MASK
 117 #define PT32_DIR_BASE_ADDR_MASK \
 118         (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
 119
 120 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
 121                         | PT64_NX_MASK)
 122
 123 #define PFERR_PRESENT_MASK (1U << 0)
 124 #define PFERR_WRITE_MASK (1U << 1)
 125 #define PFERR_USER_MASK (1U << 2)
 126 #define PFERR_FETCH_MASK (1U << 4)
 127
 128 #define PT_DIRECTORY_LEVEL 2
 129 #define PT_PAGE_TABLE_LEVEL 1
 130
 131 #define RMAP_EXT 4
 132
 133 #define ACC_EXEC_MASK    1
 134 #define ACC_WRITE_MASK   PT_WRITABLE_MASK
 135 #define ACC_USER_MASK    PT_USER_MASK
 136 #define ACC_ALL          (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
 137
 138 struct kvm_pv_mmu_op_buffer {
 139         void *ptr;
 140         unsigned len;
 141         unsigned processed;
 142         char buf[512] __aligned(sizeof(long));
 143 };
 144
 145 struct kvm_rmap_desc {
 146         u64 *shadow_ptes[RMAP_EXT];
 147         struct kvm_rmap_desc *more;
 148 };
 149
 150 static struct kmem_cache *pte_chain_cache;
 151 static struct kmem_cache *rmap_desc_cache;
 152 static struct kmem_cache *mmu_page_header_cache;
 153
 154 static u64 __read_mostly shadow_trap_nonpresent_pte;
 155 static u64 __read_mostly shadow_notrap_nonpresent_pte;
 156 static u64 __read_mostly shadow_base_present_pte;
 157 static u64 __read_mostly shadow_nx_mask;
 158 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
 159 static u64 __read_mostly shadow_user_mask;
 160 static u64 __read_mostly shadow_accessed_mask;
 161 static u64 __read_mostly shadow_dirty_mask;
 162
 163 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
 164 {
 165         shadow_trap_nonpresent_pte = trap_pte;
 166         shadow_notrap_nonpresent_pte = notrap_pte;
 167 }
 168 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
 169
 170 void kvm_mmu_set_base_ptes(u64 base_pte)
 171 {
 172         shadow_base_present_pte = base_pte;
 173 }
 174 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
 175
 176 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
 177                 u64 dirty_mask, u64 nx_mask, u64 x_mask)
 178 {
 179         shadow_user_mask = user_mask;
 180         shadow_accessed_mask = accessed_mask;
 181         shadow_dirty_mask = dirty_mask;
 182         shadow_nx_mask = nx_mask;
 183         shadow_x_mask = x_mask;
 184 }
 185 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
 186
 187 static int is_write_protection(struct kvm_vcpu *vcpu)
 188 {
 189         return vcpu->arch.cr0 & X86_CR0_WP;
 190 }
 191
 192 static int is_cpuid_PSE36(void)
 193 {
 194         return 1;
 195 }
 196
 197 static int is_nx(struct kvm_vcpu *vcpu)
 198 {
 199         return vcpu->arch.shadow_efer & EFER_NX;
 200 }
 201
 202 static int is_present_pte(unsigned long pte)
 203 {
 204         return pte & PT_PRESENT_MASK;
 205 }
 206
 207 static int is_shadow_present_pte(u64 pte)
 208 {
 209         return pte != shadow_trap_nonpresent_pte
 210                 && pte != shadow_notrap_nonpresent_pte;
 211 }
 212
 213 static int is_large_pte(u64 pte)
 214 {
 215         return pte & PT_PAGE_SIZE_MASK;
 216 }
 217
 218 static int is_writeble_pte(unsigned long pte)
 219 {
 220         return pte & PT_WRITABLE_MASK;
 221 }
 222
 223 static int is_dirty_pte(unsigned long pte)
 224 {
 225         return pte & shadow_dirty_mask;
 226 }
 227
 228 static int is_rmap_pte(u64 pte)
 229 {
 230         return is_shadow_present_pte(pte);
 231 }
 232
 233 static pfn_t spte_to_pfn(u64 pte)
 234 {
 235         return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 236 }
 237
 238 static gfn_t pse36_gfn_delta(u32 gpte)
 239 {
 240         int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
 241
 242         return (gpte & PT32_DIR_PSE36_MASK) << shift;
 243 }
 244
 245 static void set_shadow_pte(u64 *sptep, u64 spte)
 246 {
 247 #ifdef CONFIG_X86_64
 248         set_64bit((unsigned long *)sptep, spte);
 249 #else
 250         set_64bit((unsigned long long *)sptep, spte);
 251 #endif
 252 }
 253
 254 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
 255                                   struct kmem_cache *base_cache, int min)
 256 {
 257         void *obj;
 258
 259         if (cache->nobjs >= min)
 260                 return 0;
 261         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 262                 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
 263                 if (!obj)
 264                         return -ENOMEM;
 265                 cache->objects[cache->nobjs++] = obj;
 266         }
 267         return 0;
 268 }
 269
 270 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
 271 {
 272         while (mc->nobjs)
 273                 kfree(mc->objects[--mc->nobjs]);
 274 }
 275
 276 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
 277                                        int min)
 278 {
 279         struct page *page;
 280
 281         if (cache->nobjs >= min)
 282                 return 0;
 283         while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
 284                 page = alloc_page(GFP_KERNEL);
 285                 if (!page)
 286                         return -ENOMEM;
 287                 set_page_private(page, 0);
 288                 cache->objects[cache->nobjs++] = page_address(page);
 289         }
 290         return 0;
 291 }
 292
 293 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
 294 {
 295         while (mc->nobjs)
 296                 free_page((unsigned long)mc->objects[--mc->nobjs]);
 297 }
 298
 299 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
 300 {
 301         int r;
 302
 303         r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
 304                                    pte_chain_cache, 4);
 305         if (r)
 306                 goto out;
 307         r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
 308                                    rmap_desc_cache, 1);
 309         if (r)
 310                 goto out;
 311         r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
 312         if (r)
 313                 goto out;
 314         r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
 315                                    mmu_page_header_cache, 4);
 316 out:
 317         return r;
 318 }
 319
 320 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
 321 {
 322         mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
 323         mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
 324         mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
 325         mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
 326 }
 327
 328 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
 329                                     size_t size)
 330 {
 331         void *p;
 332
 333         BUG_ON(!mc->nobjs);
 334         p = mc->objects[--mc->nobjs];
 335         memset(p, 0, size);
 336         return p;
 337 }
 338
 339 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
 340 {
 341         return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
 342                                       sizeof(struct kvm_pte_chain));
 343 }
 344
 345 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
 346 {
 347         kfree(pc);
 348 }
 349
 350 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
 351 {
 352         return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
 353                                       sizeof(struct kvm_rmap_desc));
 354 }
 355
 356 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
 357 {
 358         kfree(rd);
 359 }
 360
 361 /*
 362  * Return the pointer to the largepage write count for a given
 363  * gfn, handling slots that are not large page aligned.
 364  */
 365 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
 366 {
 367         unsigned long idx;
 368
 369         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 370               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 371         return &slot->lpage_info[idx].write_count;
 372 }
 373
 374 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
 375 {
 376         int *write_count;
 377
 378         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 379         *write_count += 1;
 380 }
 381
 382 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
 383 {
 384         int *write_count;
 385
 386         write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
 387         *write_count -= 1;
 388         WARN_ON(*write_count < 0);
 389 }
 390
 391 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
 392 {
 393         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
 394         int *largepage_idx;
 395
 396         if (slot) {
 397                 largepage_idx = slot_largepage_idx(gfn, slot);
 398                 return *largepage_idx;
 399         }
 400
 401         return 1;
 402 }
 403
 404 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
 405 {
 406         struct vm_area_struct *vma;
 407         unsigned long addr;
 408
 409         addr = gfn_to_hva(kvm, gfn);
 410         if (kvm_is_error_hva(addr))
 411                 return 0;
 412
 413         vma = find_vma(current->mm, addr);
 414         if (vma && is_vm_hugetlb_page(vma))
 415                 return 1;
 416
 417         return 0;
 418 }
 419
 420 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
 421 {
 422         struct kvm_memory_slot *slot;
 423
 424         if (has_wrprotected_page(vcpu->kvm, large_gfn))
 425                 return 0;
 426
 427         if (!host_largepage_backed(vcpu->kvm, large_gfn))
 428                 return 0;
 429
 430         slot = gfn_to_memslot(vcpu->kvm, large_gfn);
 431         if (slot && slot->dirty_bitmap)
 432                 return 0;
 433
 434         return 1;
 435 }
 436
 437 /*
 438  * Take gfn and return the reverse mapping to it.
 439  * Note: gfn must be unaliased before this function get called
 440  */
 441
 442 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 443 {
 444         struct kvm_memory_slot *slot;
 445         unsigned long idx;
 446
 447         slot = gfn_to_memslot(kvm, gfn);
 448         if (!lpage)
 449                 return &slot->rmap[gfn - slot->base_gfn];
 450
 451         idx = (gfn / KVM_PAGES_PER_HPAGE) -
 452               (slot->base_gfn / KVM_PAGES_PER_HPAGE);
 453
 454         return &slot->lpage_info[idx].rmap_pde;
 455 }
 456
 457 /*
 458  * Reverse mapping data structures:
 459  *
 460  * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
 461  * that points to page_address(page).
 462  *
 463  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
 464  * containing more mappings.
 465  */
 466 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 467 {
 468         struct kvm_mmu_page *sp;
 469         struct kvm_rmap_desc *desc;
 470         unsigned long *rmapp;
 471         int i;
 472
 473         if (!is_rmap_pte(*spte))
 474                 return;
 475         gfn = unalias_gfn(vcpu->kvm, gfn);
 476         sp = page_header(__pa(spte));
 477         sp->gfns[spte - sp->spt] = gfn;
 478         rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
 479         if (!*rmapp) {
 480                 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
 481                 *rmapp = (unsigned long)spte;
 482         } else if (!(*rmapp & 1)) {
 483                 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
 484                 desc = mmu_alloc_rmap_desc(vcpu);
 485                 desc->shadow_ptes[0] = (u64 *)*rmapp;
 486                 desc->shadow_ptes[1] = spte;
 487                 *rmapp = (unsigned long)desc | 1;
 488         } else {
 489                 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
 490                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 491                 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
 492                         desc = desc->more;
 493                 if (desc->shadow_ptes[RMAP_EXT-1]) {
 494                         desc->more = mmu_alloc_rmap_desc(vcpu);
 495                         desc = desc->more;
 496                 }
 497                 for (i = 0; desc->shadow_ptes[i]; ++i)
 498                         ;
 499                 desc->shadow_ptes[i] = spte;
 500         }
 501 }
 502
 503 static void rmap_desc_remove_entry(unsigned long *rmapp,
 504                                    struct kvm_rmap_desc *desc,
 505                                    int i,
 506                                    struct kvm_rmap_desc *prev_desc)
 507 {
 508         int j;
 509
 510         for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
 511                 ;
 512         desc->shadow_ptes[i] = desc->shadow_ptes[j];
 513         desc->shadow_ptes[j] = NULL;
 514         if (j != 0)
 515                 return;
 516         if (!prev_desc && !desc->more)
 517                 *rmapp = (unsigned long)desc->shadow_ptes[0];
 518         else
 519                 if (prev_desc)
 520                         prev_desc->more = desc->more;
 521                 else
 522                         *rmapp = (unsigned long)desc->more | 1;
 523         mmu_free_rmap_desc(desc);
 524 }
 525
 526 static void rmap_remove(struct kvm *kvm, u64 *spte)
 527 {
 528         struct kvm_rmap_desc *desc;
 529         struct kvm_rmap_desc *prev_desc;
 530         struct kvm_mmu_page *sp;
 531         pfn_t pfn;
 532         unsigned long *rmapp;
 533         int i;
 534
 535         if (!is_rmap_pte(*spte))
 536                 return;
 537         sp = page_header(__pa(spte));
 538         pfn = spte_to_pfn(*spte);
 539         if (*spte & shadow_accessed_mask)
 540                 kvm_set_pfn_accessed(pfn);
 541         if (is_writeble_pte(*spte))
 542                 kvm_release_pfn_dirty(pfn);
 543         else
 544                 kvm_release_pfn_clean(pfn);
 545         rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
 546         if (!*rmapp) {
 547                 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
 548                 BUG();
 549         } else if (!(*rmapp & 1)) {
 550                 rmap_printk("rmap_remove:  %p %llx 1->0\n", spte, *spte);
 551                 if ((u64 *)*rmapp != spte) {
 552                         printk(KERN_ERR "rmap_remove:  %p %llx 1->BUG\n",
 553                                spte, *spte);
 554                         BUG();
 555                 }
 556                 *rmapp = 0;
 557         } else {
 558                 rmap_printk("rmap_remove:  %p %llx many->many\n", spte, *spte);
 559                 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 560                 prev_desc = NULL;
 561                 while (desc) {
 562                         for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
 563                                 if (desc->shadow_ptes[i] == spte) {
 564                                         rmap_desc_remove_entry(rmapp,
 565                                                                desc, i,
 566                                                                prev_desc);
 567                                         return;
 568                                 }
 569                         prev_desc = desc;
 570                         desc = desc->more;
 571                 }
 572                 BUG();
 573         }
 574 }
 575
 576 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
 577 {
 578         struct kvm_rmap_desc *desc;
 579         struct kvm_rmap_desc *prev_desc;
 580         u64 *prev_spte;
 581         int i;
 582
 583         if (!*rmapp)
 584                 return NULL;
 585         else if (!(*rmapp & 1)) {
 586                 if (!spte)
 587                         return (u64 *)*rmapp;
 588                 return NULL;
 589         }
 590         desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
 591         prev_desc = NULL;
 592         prev_spte = NULL;
 593         while (desc) {
 594                 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
 595                         if (prev_spte == spte)
 596                                 return desc->shadow_ptes[i];
 597                         prev_spte = desc->shadow_ptes[i];
 598                 }
 599                 desc = desc->more;
 600         }
 601         return NULL;
 602 }
 603
 604 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
 605 {
 606         unsigned long *rmapp;
 607         u64 *spte;
 608         int write_protected = 0;
 609
 610         gfn = unalias_gfn(kvm, gfn);
 611         rmapp = gfn_to_rmap(kvm, gfn, 0);
 612
 613         spte = rmap_next(kvm, rmapp, NULL);
 614         while (spte) {
 615                 BUG_ON(!spte);
 616                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 617                 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
 618                 if (is_writeble_pte(*spte)) {
 619                         set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
 620                         write_protected = 1;
 621                 }
 622                 spte = rmap_next(kvm, rmapp, spte);
 623         }
 624         if (write_protected) {
 625                 pfn_t pfn;
 626
 627                 spte = rmap_next(kvm, rmapp, NULL);
 628                 pfn = spte_to_pfn(*spte);
 629                 kvm_set_pfn_dirty(pfn);
 630         }
 631
 632         /* check for huge page mappings */
 633         rmapp = gfn_to_rmap(kvm, gfn, 1);
 634         spte = rmap_next(kvm, rmapp, NULL);
 635         while (spte) {
 636                 BUG_ON(!spte);
 637                 BUG_ON(!(*spte & PT_PRESENT_MASK));
 638                 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
 639                 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
 640                 if (is_writeble_pte(*spte)) {
 641                         rmap_remove(kvm, spte);
 642                         --kvm->stat.lpages;
 643                         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
 644                         spte = NULL;
 645                         write_protected = 1;
 646                 }
 647                 spte = rmap_next(kvm, rmapp, spte);
 648         }
 649
 650         if (write_protected)
 651                 kvm_flush_remote_tlbs(kvm);
 652
 653         account_shadowed(kvm, gfn);
 654 }
 655
 656 #ifdef MMU_DEBUG
 657 static int is_empty_shadow_page(u64 *spt)
 658 {
 659         u64 *pos;
 660         u64 *end;
 661
 662         for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
 663                 if (is_shadow_present_pte(*pos)) {
 664                         printk(KERN_ERR "%s: %p %llx\n", __func__,
 665                                pos, *pos);
 666                         return 0;
 667                 }
 668         return 1;
 669 }
 670 #endif
 671
 672 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 673 {
 674         ASSERT(is_empty_shadow_page(sp->spt));
 675         list_del(&sp->link);
 676         __free_page(virt_to_page(sp->spt));
 677         __free_page(virt_to_page(sp->gfns));
 678         kfree(sp);
 679         ++kvm->arch.n_free_mmu_pages;
 680 }
 681
 682 static unsigned kvm_page_table_hashfn(gfn_t gfn)
 683 {
 684         return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
 685 }
 686
 687 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
 688                                                u64 *parent_pte)
 689 {
 690         struct kvm_mmu_page *sp;
 691
 692         sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
 693         sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 694         sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
 695         set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
 696         list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
 697         ASSERT(is_empty_shadow_page(sp->spt));
 698         sp->slot_bitmap = 0;
 699         sp->multimapped = 0;
 700         sp->parent_pte = parent_pte;
 701         --vcpu->kvm->arch.n_free_mmu_pages;
 702         return sp;
 703 }
 704
 705 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
 706                                     struct kvm_mmu_page *sp, u64 *parent_pte)
 707 {
 708         struct kvm_pte_chain *pte_chain;
 709         struct hlist_node *node;
 710         int i;
 711
 712         if (!parent_pte)
 713                 return;
 714         if (!sp->multimapped) {
 715                 u64 *old = sp->parent_pte;
 716
 717                 if (!old) {
 718                         sp->parent_pte = parent_pte;
 719                         return;
 720                 }
 721                 sp->multimapped = 1;
 722                 pte_chain = mmu_alloc_pte_chain(vcpu);
 723                 INIT_HLIST_HEAD(&sp->parent_ptes);
 724                 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 725                 pte_chain->parent_ptes[0] = old;
 726         }
 727         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
 728                 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
 729                         continue;
 730                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
 731                         if (!pte_chain->parent_ptes[i]) {
 732                                 pte_chain->parent_ptes[i] = parent_pte;
 733                                 return;
 734                         }
 735         }
 736         pte_chain = mmu_alloc_pte_chain(vcpu);
 737         BUG_ON(!pte_chain);
 738         hlist_add_head(&pte_chain->link, &sp->parent_ptes);
 739         pte_chain->parent_ptes[0] = parent_pte;
 740 }
 741
 742 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
 743                                        u64 *parent_pte)
 744 {
 745         struct kvm_pte_chain *pte_chain;
 746         struct hlist_node *node;
 747         int i;
 748
 749         if (!sp->multimapped) {
 750                 BUG_ON(sp->parent_pte != parent_pte);
 751                 sp->parent_pte = NULL;
 752                 return;
 753         }
 754         hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
 755                 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
 756                         if (!pte_chain->parent_ptes[i])
 757                                 break;
 758                         if (pte_chain->parent_ptes[i] != parent_pte)
 759                                 continue;
 760                         while (i + 1 < NR_PTE_CHAIN_ENTRIES
 761                                 && pte_chain->parent_ptes[i + 1]) {
 762                                 pte_chain->parent_ptes[i]
 763                                         = pte_chain->parent_ptes[i + 1];
 764                                 ++i;
 765                         }
 766                         pte_chain->parent_ptes[i] = NULL;
 767                         if (i == 0) {
 768                                 hlist_del(&pte_chain->link);
 769                                 mmu_free_pte_chain(pte_chain);
 770                                 if (hlist_empty(&sp->parent_ptes)) {
 771                                         sp->multimapped = 0;
 772                                         sp->parent_pte = NULL;
 773                                 }
 774                         }
 775                         return;
 776                 }
 777         BUG();
 778 }
 779
 780 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
 781                                     struct kvm_mmu_page *sp)
 782 {
 783         int i;
 784
 785         for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
 786                 sp->spt[i] = shadow_trap_nonpresent_pte;
 787 }
 788
 789 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
 790 {
 791         unsigned index;
 792         struct hlist_head *bucket;
 793         struct kvm_mmu_page *sp;
 794         struct hlist_node *node;
 795
 796         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 797         index = kvm_page_table_hashfn(gfn);
 798         bucket = &kvm->arch.mmu_page_hash[index];
 799         hlist_for_each_entry(sp, node, bucket, hash_link)
 800                 if (sp->gfn == gfn && !sp->role.metaphysical
 801                     && !sp->role.invalid) {
 802                         pgprintk("%s: found role %x\n",
 803                                  __func__, sp->role.word);
 804                         return sp;
 805                 }
 806         return NULL;
 807 }
 808
 809 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
 810                                              gfn_t gfn,
 811                                              gva_t gaddr,
 812                                              unsigned level,
 813                                              int metaphysical,
 814                                              unsigned access,
 815                                              u64 *parent_pte)
 816 {
 817         union kvm_mmu_page_role role;
 818         unsigned index;
 819         unsigned quadrant;
 820         struct hlist_head *bucket;
 821         struct kvm_mmu_page *sp;
 822         struct hlist_node *node;
 823
 824         role.word = 0;
 825         role.glevels = vcpu->arch.mmu.root_level;
 826         role.level = level;
 827         role.metaphysical = metaphysical;
 828         role.access = access;
 829         if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
 830                 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
 831                 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
 832                 role.quadrant = quadrant;
 833         }
 834         pgprintk("%s: looking gfn %lx role %x\n", __func__,
 835                  gfn, role.word);
 836         index = kvm_page_table_hashfn(gfn);
 837         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
 838         hlist_for_each_entry(sp, node, bucket, hash_link)
 839                 if (sp->gfn == gfn && sp->role.word == role.word) {
 840                         mmu_page_add_parent_pte(vcpu, sp, parent_pte);
 841                         pgprintk("%s: found\n", __func__);
 842                         return sp;
 843                 }
 844         ++vcpu->kvm->stat.mmu_cache_miss;
 845         sp = kvm_mmu_alloc_page(vcpu, parent_pte);
 846         if (!sp)
 847                 return sp;
 848         pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
 849         sp->gfn = gfn;
 850         sp->role = role;
 851         hlist_add_head(&sp->hash_link, bucket);
 852         if (!metaphysical)
 853                 rmap_write_protect(vcpu->kvm, gfn);
 854         if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
 855                 vcpu->arch.mmu.prefetch_page(vcpu, sp);
 856         else
 857                 nonpaging_prefetch_page(vcpu, sp);
 858         return sp;
 859 }
 860
 861 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
 862                                          struct kvm_mmu_page *sp)
 863 {
 864         unsigned i;
 865         u64 *pt;
 866         u64 ent;
 867
 868         pt = sp->spt;
 869
 870         if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
 871                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 872                         if (is_shadow_present_pte(pt[i]))
 873                                 rmap_remove(kvm, &pt[i]);
 874                         pt[i] = shadow_trap_nonpresent_pte;
 875                 }
 876                 kvm_flush_remote_tlbs(kvm);
 877                 return;
 878         }
 879
 880         for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 881                 ent = pt[i];
 882
 883                 if (is_shadow_present_pte(ent)) {
 884                         if (!is_large_pte(ent)) {
 885                                 ent &= PT64_BASE_ADDR_MASK;
 886                                 mmu_page_remove_parent_pte(page_header(ent),
 887                                                            &pt[i]);
 888                         } else {
 889                                 --kvm->stat.lpages;
 890                                 rmap_remove(kvm, &pt[i]);
 891                         }
 892                 }
 893                 pt[i] = shadow_trap_nonpresent_pte;
 894         }
 895         kvm_flush_remote_tlbs(kvm);
 896 }
 897
 898 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
 899 {
 900         mmu_page_remove_parent_pte(sp, parent_pte);
 901 }
 902
 903 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
 904 {
 905         int i;
 906
 907         for (i = 0; i < KVM_MAX_VCPUS; ++i)
 908                 if (kvm->vcpus[i])
 909                         kvm->vcpus[i]->arch.last_pte_updated = NULL;
 910 }
 911
 912 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
 913 {
 914         u64 *parent_pte;
 915
 916         ++kvm->stat.mmu_shadow_zapped;
 917         while (sp->multimapped || sp->parent_pte) {
 918                 if (!sp->multimapped)
 919                         parent_pte = sp->parent_pte;
 920                 else {
 921                         struct kvm_pte_chain *chain;
 922
 923                         chain = container_of(sp->parent_ptes.first,
 924                                              struct kvm_pte_chain, link);
 925                         parent_pte = chain->parent_ptes[0];
 926                 }
 927                 BUG_ON(!parent_pte);
 928                 kvm_mmu_put_page(sp, parent_pte);
 929                 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
 930         }
 931         kvm_mmu_page_unlink_children(kvm, sp);
 932         if (!sp->root_count) {
 933                 if (!sp->role.metaphysical)
 934                         unaccount_shadowed(kvm, sp->gfn);
 935                 hlist_del(&sp->hash_link);
 936                 kvm_mmu_free_page(kvm, sp);
 937         } else {
 938                 list_move(&sp->link, &kvm->arch.active_mmu_pages);
 939                 sp->role.invalid = 1;
 940                 kvm_reload_remote_mmus(kvm);
 941         }
 942         kvm_mmu_reset_last_pte_updated(kvm);
 943 }
 944
 945 /*
 946  * Changing the number of mmu pages allocated to the vm
 947  * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
 948  */
 949 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
 950 {
 951         /*
 952          * If we set the number of mmu pages to be smaller be than the
 953          * number of actived pages , we must to free some mmu pages before we
 954          * change the value
 955          */
 956
 957         if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
 958             kvm_nr_mmu_pages) {
 959                 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
 960                                        - kvm->arch.n_free_mmu_pages;
 961
 962                 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
 963                         struct kvm_mmu_page *page;
 964
 965                         page = container_of(kvm->arch.active_mmu_pages.prev,
 966                                             struct kvm_mmu_page, link);
 967                         kvm_mmu_zap_page(kvm, page);
 968                         n_used_mmu_pages--;
 969                 }
 970                 kvm->arch.n_free_mmu_pages = 0;
 971         }
 972         else
 973                 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
 974                                          - kvm->arch.n_alloc_mmu_pages;
 975
 976         kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
 977 }
 978
 979 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
 980 {
 981         unsigned index;
 982         struct hlist_head *bucket;
 983         struct kvm_mmu_page *sp;
 984         struct hlist_node *node, *n;
 985         int r;
 986
 987         pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
 988         r = 0;
 989         index = kvm_page_table_hashfn(gfn);
 990         bucket = &kvm->arch.mmu_page_hash[index];
 991         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
 992                 if (sp->gfn == gfn && !sp->role.metaphysical) {
 993                         pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
 994                                  sp->role.word);
 995                         kvm_mmu_zap_page(kvm, sp);
 996                         r = 1;
 997                 }
 998         return r;
 999 }
1000
1001 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1002 {
1003         struct kvm_mmu_page *sp;
1004
1005         while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1006                 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1007                 kvm_mmu_zap_page(kvm, sp);
1008         }
1009 }
1010
1011 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1012 {
1013         int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1014         struct kvm_mmu_page *sp = page_header(__pa(pte));
1015
1016         __set_bit(slot, &sp->slot_bitmap);
1017 }
1018
1019 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1020 {
1021         struct page *page;
1022
1023         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1024
1025         if (gpa == UNMAPPED_GVA)
1026                 return NULL;
1027
1028         down_read(&current->mm->mmap_sem);
1029         page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1030         up_read(&current->mm->mmap_sem);
1031
1032         return page;
1033 }
1034
1035 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1036                          unsigned pt_access, unsigned pte_access,
1037                          int user_fault, int write_fault, int dirty,
1038                          int *ptwrite, int largepage, gfn_t gfn,
1039                          pfn_t pfn, bool speculative)
1040 {
1041         u64 spte;
1042         int was_rmapped = 0;
1043         int was_writeble = is_writeble_pte(*shadow_pte);
1044
1045         pgprintk("%s: spte %llx access %x write_fault %d"
1046                  " user_fault %d gfn %lx\n",
1047                  __func__, *shadow_pte, pt_access,
1048                  write_fault, user_fault, gfn);
1049
1050         if (is_rmap_pte(*shadow_pte)) {
1051                 /*
1052                  * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1053                  * the parent of the now unreachable PTE.
1054                  */
1055                 if (largepage && !is_large_pte(*shadow_pte)) {
1056                         struct kvm_mmu_page *child;
1057                         u64 pte = *shadow_pte;
1058
1059                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1060                         mmu_page_remove_parent_pte(child, shadow_pte);
1061                 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1062                         pgprintk("hfn old %lx new %lx\n",
1063                                  spte_to_pfn(*shadow_pte), pfn);
1064                         rmap_remove(vcpu->kvm, shadow_pte);
1065                 } else {
1066                         if (largepage)
1067                                 was_rmapped = is_large_pte(*shadow_pte);
1068                         else
1069                                 was_rmapped = 1;
1070                 }
1071         }
1072
1073         /*
1074          * We don't set the accessed bit, since we sometimes want to see
1075          * whether the guest actually used the pte (in order to detect
1076          * demand paging).
1077          */
1078         spte = shadow_base_present_pte | shadow_dirty_mask;
1079         if (!speculative)
1080                 pte_access |= PT_ACCESSED_MASK;
1081         if (!dirty)
1082                 pte_access &= ~ACC_WRITE_MASK;
1083         if (pte_access & ACC_EXEC_MASK)
1084                 spte |= shadow_x_mask;
1085         else
1086                 spte |= shadow_nx_mask;
1087         if (pte_access & ACC_USER_MASK)
1088                 spte |= shadow_user_mask;
1089         if (largepage)
1090                 spte |= PT_PAGE_SIZE_MASK;
1091
1092         spte |= (u64)pfn << PAGE_SHIFT;
1093
1094         if ((pte_access & ACC_WRITE_MASK)
1095             || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1096                 struct kvm_mmu_page *shadow;
1097
1098                 spte |= PT_WRITABLE_MASK;
1099
1100                 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1101                 if (shadow ||
1102                    (largepage && has_wrprotected_page(vcpu->kvm, gfn))) {
1103                         pgprintk("%s: found shadow page for %lx, marking ro\n",
1104                                  __func__, gfn);
1105                         pte_access &= ~ACC_WRITE_MASK;
1106                         if (is_writeble_pte(spte)) {
1107                                 spte &= ~PT_WRITABLE_MASK;
1108                                 kvm_x86_ops->tlb_flush(vcpu);
1109                         }
1110                         if (write_fault)
1111                                 *ptwrite = 1;
1112                 }
1113         }
1114
1115         if (pte_access & ACC_WRITE_MASK)
1116                 mark_page_dirty(vcpu->kvm, gfn);
1117
1118         pgprintk("%s: setting spte %llx\n", __func__, spte);
1119         pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1120                  (spte&PT_PAGE_SIZE_MASK)? "2MB" : "4kB",
1121                  (spte&PT_WRITABLE_MASK)?"RW":"R", gfn, spte, shadow_pte);
1122         set_shadow_pte(shadow_pte, spte);
1123         if (!was_rmapped && (spte & PT_PAGE_SIZE_MASK)
1124             && (spte & PT_PRESENT_MASK))
1125                 ++vcpu->kvm->stat.lpages;
1126
1127         page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1128         if (!was_rmapped) {
1129                 rmap_add(vcpu, shadow_pte, gfn, largepage);
1130                 if (!is_rmap_pte(*shadow_pte))
1131                         kvm_release_pfn_clean(pfn);
1132         } else {
1133                 if (was_writeble)
1134                         kvm_release_pfn_dirty(pfn);
1135                 else
1136                         kvm_release_pfn_clean(pfn);
1137         }
1138         if (speculative) {
1139                 vcpu->arch.last_pte_updated = shadow_pte;
1140                 vcpu->arch.last_pte_gfn = gfn;
1141         }
1142 }
1143
1144 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1145 {
1146 }
1147
1148 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1149                            int largepage, gfn_t gfn, pfn_t pfn,
1150                            int level)
1151 {
1152         hpa_t table_addr = vcpu->arch.mmu.root_hpa;
1153         int pt_write = 0;
1154
1155         for (; ; level--) {
1156                 u32 index = PT64_INDEX(v, level);
1157                 u64 *table;
1158
1159                 ASSERT(VALID_PAGE(table_addr));
1160                 table = __va(table_addr);
1161
1162                 if (level == 1) {
1163                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1164                                      0, write, 1, &pt_write, 0, gfn, pfn, false);
1165                         return pt_write;
1166                 }
1167
1168                 if (largepage && level == 2) {
1169                         mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1170                                      0, write, 1, &pt_write, 1, gfn, pfn, false);
1171                         return pt_write;
1172                 }
1173
1174                 if (table[index] == shadow_trap_nonpresent_pte) {
1175                         struct kvm_mmu_page *new_table;
1176                         gfn_t pseudo_gfn;
1177
1178                         pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1179                                 >> PAGE_SHIFT;
1180                         new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1181                                                      v, level - 1,
1182                                                      1, ACC_ALL, &table[index]);
1183                         if (!new_table) {
1184                                 pgprintk("nonpaging_map: ENOMEM\n");
1185                                 kvm_release_pfn_clean(pfn);
1186                                 return -ENOMEM;
1187                         }
1188
1189                         table[index] = __pa(new_table->spt)
1190                                 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1191                                 | shadow_user_mask | shadow_x_mask;
1192                 }
1193                 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1194         }
1195 }
1196
1197 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1198 {
1199         int r;
1200         int largepage = 0;
1201         pfn_t pfn;
1202
1203         down_read(&current->mm->mmap_sem);
1204         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1205                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1206                 largepage = 1;
1207         }
1208
1209         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1210         up_read(&current->mm->mmap_sem);
1211
1212         /* mmio */
1213         if (is_error_pfn(pfn)) {
1214                 kvm_release_pfn_clean(pfn);
1215                 return 1;
1216         }
1217
1218         spin_lock(&vcpu->kvm->mmu_lock);
1219         kvm_mmu_free_some_pages(vcpu);
1220         r = __direct_map(vcpu, v, write, largepage, gfn, pfn,
1221                          PT32E_ROOT_LEVEL);
1222         spin_unlock(&vcpu->kvm->mmu_lock);
1223
1224
1225         return r;
1226 }
1227
1228
1229 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1230 {
1231         int i;
1232         struct kvm_mmu_page *sp;
1233
1234         if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1235                 return;
1236         spin_lock(&vcpu->kvm->mmu_lock);
1237         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1238                 hpa_t root = vcpu->arch.mmu.root_hpa;
1239
1240                 sp = page_header(root);
1241                 --sp->root_count;
1242                 if (!sp->root_count && sp->role.invalid)
1243                         kvm_mmu_zap_page(vcpu->kvm, sp);
1244                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1245                 spin_unlock(&vcpu->kvm->mmu_lock);
1246                 return;
1247         }
1248         for (i = 0; i < 4; ++i) {
1249                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1250
1251                 if (root) {
1252                         root &= PT64_BASE_ADDR_MASK;
1253                         sp = page_header(root);
1254                         --sp->root_count;
1255                         if (!sp->root_count && sp->role.invalid)
1256                                 kvm_mmu_zap_page(vcpu->kvm, sp);
1257                 }
1258                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1259         }
1260         spin_unlock(&vcpu->kvm->mmu_lock);
1261         vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1262 }
1263
1264 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1265 {
1266         int i;
1267         gfn_t root_gfn;
1268         struct kvm_mmu_page *sp;
1269         int metaphysical = 0;
1270
1271         root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1272
1273         if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1274                 hpa_t root = vcpu->arch.mmu.root_hpa;
1275
1276                 ASSERT(!VALID_PAGE(root));
1277                 if (tdp_enabled)
1278                         metaphysical = 1;
1279                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1280                                       PT64_ROOT_LEVEL, metaphysical,
1281                                       ACC_ALL, NULL);
1282                 root = __pa(sp->spt);
1283                 ++sp->root_count;
1284                 vcpu->arch.mmu.root_hpa = root;
1285                 return;
1286         }
1287         metaphysical = !is_paging(vcpu);
1288         if (tdp_enabled)
1289                 metaphysical = 1;
1290         for (i = 0; i < 4; ++i) {
1291                 hpa_t root = vcpu->arch.mmu.pae_root[i];
1292
1293                 ASSERT(!VALID_PAGE(root));
1294                 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1295                         if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1296                                 vcpu->arch.mmu.pae_root[i] = 0;
1297                                 continue;
1298                         }
1299                         root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1300                 } else if (vcpu->arch.mmu.root_level == 0)
1301                         root_gfn = 0;
1302                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1303                                       PT32_ROOT_LEVEL, metaphysical,
1304                                       ACC_ALL, NULL);
1305                 root = __pa(sp->spt);
1306                 ++sp->root_count;
1307                 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1308         }
1309         vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1310 }
1311
1312 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1313 {
1314         return vaddr;
1315 }
1316
1317 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1318                                 u32 error_code)
1319 {
1320         gfn_t gfn;
1321         int r;
1322
1323         pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1324         r = mmu_topup_memory_caches(vcpu);
1325         if (r)
1326                 return r;
1327
1328         ASSERT(vcpu);
1329         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1330
1331         gfn = gva >> PAGE_SHIFT;
1332
1333         return nonpaging_map(vcpu, gva & PAGE_MASK,
1334                              error_code & PFERR_WRITE_MASK, gfn);
1335 }
1336
1337 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1338                                 u32 error_code)
1339 {
1340         pfn_t pfn;
1341         int r;
1342         int largepage = 0;
1343         gfn_t gfn = gpa >> PAGE_SHIFT;
1344
1345         ASSERT(vcpu);
1346         ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1347
1348         r = mmu_topup_memory_caches(vcpu);
1349         if (r)
1350                 return r;
1351
1352         down_read(&current->mm->mmap_sem);
1353         if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1354                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1355                 largepage = 1;
1356         }
1357         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1358         up_read(&current->mm->mmap_sem);
1359         if (is_error_pfn(pfn)) {
1360                 kvm_release_pfn_clean(pfn);
1361                 return 1;
1362         }
1363         spin_lock(&vcpu->kvm->mmu_lock);
1364         kvm_mmu_free_some_pages(vcpu);
1365         r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1366                          largepage, gfn, pfn, kvm_x86_ops->get_tdp_level());
1367         spin_unlock(&vcpu->kvm->mmu_lock);
1368
1369         return r;
1370 }
1371
1372 static void nonpaging_free(struct kvm_vcpu *vcpu)
1373 {
1374         mmu_free_roots(vcpu);
1375 }
1376
1377 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1378 {
1379         struct kvm_mmu *context = &vcpu->arch.mmu;
1380
1381         context->new_cr3 = nonpaging_new_cr3;
1382         context->page_fault = nonpaging_page_fault;
1383         context->gva_to_gpa = nonpaging_gva_to_gpa;
1384         context->free = nonpaging_free;
1385         context->prefetch_page = nonpaging_prefetch_page;
1386         context->root_level = 0;
1387         context->shadow_root_level = PT32E_ROOT_LEVEL;
1388         context->root_hpa = INVALID_PAGE;
1389         return 0;
1390 }
1391
1392 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1393 {
1394         ++vcpu->stat.tlb_flush;
1395         kvm_x86_ops->tlb_flush(vcpu);
1396 }
1397
1398 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1399 {
1400         pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1401         mmu_free_roots(vcpu);
1402 }
1403
1404 static void inject_page_fault(struct kvm_vcpu *vcpu,
1405                               u64 addr,
1406                               u32 err_code)
1407 {
1408         kvm_inject_page_fault(vcpu, addr, err_code);
1409 }
1410
1411 static void paging_free(struct kvm_vcpu *vcpu)
1412 {
1413         nonpaging_free(vcpu);
1414 }
1415
1416 #define PTTYPE 64
1417 #include "paging_tmpl.h"
1418 #undef PTTYPE
1419
1420 #define PTTYPE 32
1421 #include "paging_tmpl.h"
1422 #undef PTTYPE
1423
1424 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1425 {
1426         struct kvm_mmu *context = &vcpu->arch.mmu;
1427
1428         ASSERT(is_pae(vcpu));
1429         context->new_cr3 = paging_new_cr3;
1430         context->page_fault = paging64_page_fault;
1431         context->gva_to_gpa = paging64_gva_to_gpa;
1432         context->prefetch_page = paging64_prefetch_page;
1433         context->free = paging_free;
1434         context->root_level = level;
1435         context->shadow_root_level = level;
1436         context->root_hpa = INVALID_PAGE;
1437         return 0;
1438 }
1439
1440 static int paging64_init_context(struct kvm_vcpu *vcpu)
1441 {
1442         return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1443 }
1444
1445 static int paging32_init_context(struct kvm_vcpu *vcpu)
1446 {
1447         struct kvm_mmu *context = &vcpu->arch.mmu;
1448
1449         context->new_cr3 = paging_new_cr3;
1450         context->page_fault = paging32_page_fault;
1451         context->gva_to_gpa = paging32_gva_to_gpa;
1452         context->free = paging_free;
1453         context->prefetch_page = paging32_prefetch_page;
1454         context->root_level = PT32_ROOT_LEVEL;
1455         context->shadow_root_level = PT32E_ROOT_LEVEL;
1456         context->root_hpa = INVALID_PAGE;
1457         return 0;
1458 }
1459
1460 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1461 {
1462         return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1463 }
1464
1465 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1466 {
1467         struct kvm_mmu *context = &vcpu->arch.mmu;
1468
1469         context->new_cr3 = nonpaging_new_cr3;
1470         context->page_fault = tdp_page_fault;
1471         context->free = nonpaging_free;
1472         context->prefetch_page = nonpaging_prefetch_page;
1473         context->shadow_root_level = kvm_x86_ops->get_tdp_level();
1474         context->root_hpa = INVALID_PAGE;
1475
1476         if (!is_paging(vcpu)) {
1477                 context->gva_to_gpa = nonpaging_gva_to_gpa;
1478                 context->root_level = 0;
1479         } else if (is_long_mode(vcpu)) {
1480                 context->gva_to_gpa = paging64_gva_to_gpa;
1481                 context->root_level = PT64_ROOT_LEVEL;
1482         } else if (is_pae(vcpu)) {
1483                 context->gva_to_gpa = paging64_gva_to_gpa;
1484                 context->root_level = PT32E_ROOT_LEVEL;
1485         } else {
1486                 context->gva_to_gpa = paging32_gva_to_gpa;
1487                 context->root_level = PT32_ROOT_LEVEL;
1488         }
1489
1490         return 0;
1491 }
1492
1493 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1494 {
1495         ASSERT(vcpu);
1496         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1497
1498         if (!is_paging(vcpu))
1499                 return nonpaging_init_context(vcpu);
1500         else if (is_long_mode(vcpu))
1501                 return paging64_init_context(vcpu);
1502         else if (is_pae(vcpu))
1503                 return paging32E_init_context(vcpu);
1504         else
1505                 return paging32_init_context(vcpu);
1506 }
1507
1508 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1509 {
1510         vcpu->arch.update_pte.pfn = bad_pfn;
1511
1512         if (tdp_enabled)
1513                 return init_kvm_tdp_mmu(vcpu);
1514         else
1515                 return init_kvm_softmmu(vcpu);
1516 }
1517
1518 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1519 {
1520         ASSERT(vcpu);
1521         if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1522                 vcpu->arch.mmu.free(vcpu);
1523                 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1524         }
1525 }
1526
1527 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1528 {
1529         destroy_kvm_mmu(vcpu);
1530         return init_kvm_mmu(vcpu);
1531 }
1532 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1533
1534 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1535 {
1536         int r;
1537
1538         r = mmu_topup_memory_caches(vcpu);
1539         if (r)
1540                 goto out;
1541         spin_lock(&vcpu->kvm->mmu_lock);
1542         kvm_mmu_free_some_pages(vcpu);
1543         mmu_alloc_roots(vcpu);
1544         spin_unlock(&vcpu->kvm->mmu_lock);
1545         kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1546         kvm_mmu_flush_tlb(vcpu);
1547 out:
1548         return r;
1549 }
1550 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1551
1552 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1553 {
1554         mmu_free_roots(vcpu);
1555 }
1556
1557 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1558                                   struct kvm_mmu_page *sp,
1559                                   u64 *spte)
1560 {
1561         u64 pte;
1562         struct kvm_mmu_page *child;
1563
1564         pte = *spte;
1565         if (is_shadow_present_pte(pte)) {
1566                 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
1567                     is_large_pte(pte))
1568                         rmap_remove(vcpu->kvm, spte);
1569                 else {
1570                         child = page_header(pte & PT64_BASE_ADDR_MASK);
1571                         mmu_page_remove_parent_pte(child, spte);
1572                 }
1573         }
1574         set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1575         if (is_large_pte(pte))
1576                 --vcpu->kvm->stat.lpages;
1577 }
1578
1579 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1580                                   struct kvm_mmu_page *sp,
1581                                   u64 *spte,
1582                                   const void *new)
1583 {
1584         if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1585                 if (!vcpu->arch.update_pte.largepage ||
1586                     sp->role.glevels == PT32_ROOT_LEVEL) {
1587                         ++vcpu->kvm->stat.mmu_pde_zapped;
1588                         return;
1589                 }
1590         }
1591
1592         ++vcpu->kvm->stat.mmu_pte_updated;
1593         if (sp->role.glevels == PT32_ROOT_LEVEL)
1594                 paging32_update_pte(vcpu, sp, spte, new);
1595         else
1596                 paging64_update_pte(vcpu, sp, spte, new);
1597 }
1598
1599 static bool need_remote_flush(u64 old, u64 new)
1600 {
1601         if (!is_shadow_present_pte(old))
1602                 return false;
1603         if (!is_shadow_present_pte(new))
1604                 return true;
1605         if ((old ^ new) & PT64_BASE_ADDR_MASK)
1606                 return true;
1607         old ^= PT64_NX_MASK;
1608         new ^= PT64_NX_MASK;
1609         return (old & ~new & PT64_PERM_MASK) != 0;
1610 }
1611
1612 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1613 {
1614         if (need_remote_flush(old, new))
1615                 kvm_flush_remote_tlbs(vcpu->kvm);
1616         else
1617                 kvm_mmu_flush_tlb(vcpu);
1618 }
1619
1620 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1621 {
1622         u64 *spte = vcpu->arch.last_pte_updated;
1623
1624         return !!(spte && (*spte & shadow_accessed_mask));
1625 }
1626
1627 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1628                                           const u8 *new, int bytes)
1629 {
1630         gfn_t gfn;
1631         int r;
1632         u64 gpte = 0;
1633         pfn_t pfn;
1634
1635         vcpu->arch.update_pte.largepage = 0;
1636
1637         if (bytes != 4 && bytes != 8)
1638                 return;
1639
1640         /*
1641          * Assume that the pte write on a page table of the same type
1642          * as the current vcpu paging mode.  This is nearly always true
1643          * (might be false while changing modes).  Note it is verified later
1644          * by update_pte().
1645          */
1646         if (is_pae(vcpu)) {
1647                 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1648                 if ((bytes == 4) && (gpa % 4 == 0)) {
1649                         r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1650                         if (r)
1651                                 return;
1652                         memcpy((void *)&gpte + (gpa % 8), new, 4);
1653                 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1654                         memcpy((void *)&gpte, new, 8);
1655                 }
1656         } else {
1657                 if ((bytes == 4) && (gpa % 4 == 0))
1658                         memcpy((void *)&gpte, new, 4);
1659         }
1660         if (!is_present_pte(gpte))
1661                 return;
1662         gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1663
1664         down_read(&current->mm->mmap_sem);
1665         if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
1666                 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1667                 vcpu->arch.update_pte.largepage = 1;
1668         }
1669         pfn = gfn_to_pfn(vcpu->kvm, gfn);
1670         up_read(&current->mm->mmap_sem);
1671
1672         if (is_error_pfn(pfn)) {
1673                 kvm_release_pfn_clean(pfn);
1674                 return;
1675         }
1676         vcpu->arch.update_pte.gfn = gfn;
1677         vcpu->arch.update_pte.pfn = pfn;
1678 }
1679
1680 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
1681 {
1682         u64 *spte = vcpu->arch.last_pte_updated;
1683
1684         if (spte
1685             && vcpu->arch.last_pte_gfn == gfn
1686             && shadow_accessed_mask
1687             && !(*spte & shadow_accessed_mask)
1688             && is_shadow_present_pte(*spte))
1689                 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
1690 }
1691
1692 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1693                        const u8 *new, int bytes)
1694 {
1695         gfn_t gfn = gpa >> PAGE_SHIFT;
1696         struct kvm_mmu_page *sp;
1697         struct hlist_node *node, *n;
1698         struct hlist_head *bucket;
1699         unsigned index;
1700         u64 entry, gentry;
1701         u64 *spte;
1702         unsigned offset = offset_in_page(gpa);
1703         unsigned pte_size;
1704         unsigned page_offset;
1705         unsigned misaligned;
1706         unsigned quadrant;
1707         int level;
1708         int flooded = 0;
1709         int npte;
1710         int r;
1711
1712         pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
1713         mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1714         spin_lock(&vcpu->kvm->mmu_lock);
1715         kvm_mmu_access_page(vcpu, gfn);
1716         kvm_mmu_free_some_pages(vcpu);
1717         ++vcpu->kvm->stat.mmu_pte_write;
1718         kvm_mmu_audit(vcpu, "pre pte write");
1719         if (gfn == vcpu->arch.last_pt_write_gfn
1720             && !last_updated_pte_accessed(vcpu)) {
1721                 ++vcpu->arch.last_pt_write_count;
1722                 if (vcpu->arch.last_pt_write_count >= 3)
1723                         flooded = 1;
1724         } else {
1725                 vcpu->arch.last_pt_write_gfn = gfn;
1726                 vcpu->arch.last_pt_write_count = 1;
1727                 vcpu->arch.last_pte_updated = NULL;
1728         }
1729         index = kvm_page_table_hashfn(gfn);
1730         bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1731         hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1732                 if (sp->gfn != gfn || sp->role.metaphysical)
1733                         continue;
1734                 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1735                 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1736                 misaligned |= bytes < 4;
1737                 if (misaligned || flooded) {
1738                         /*
1739                          * Misaligned accesses are too much trouble to fix
1740                          * up; also, they usually indicate a page is not used
1741                          * as a page table.
1742                          *
1743                          * If we're seeing too many writes to a page,
1744                          * it may no longer be a page table, or we may be
1745                          * forking, in which case it is better to unmap the
1746                          * page.
1747                          */
1748                         pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1749                                  gpa, bytes, sp->role.word);
1750                         kvm_mmu_zap_page(vcpu->kvm, sp);
1751                         ++vcpu->kvm->stat.mmu_flooded;
1752                         continue;
1753                 }
1754                 page_offset = offset;
1755                 level = sp->role.level;
1756                 npte = 1;
1757                 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1758                         page_offset <<= 1;      /* 32->64 */
1759                         /*
1760                          * A 32-bit pde maps 4MB while the shadow pdes map
1761                          * only 2MB.  So we need to double the offset again
1762                          * and zap two pdes instead of one.
1763                          */
1764                         if (level == PT32_ROOT_LEVEL) {
1765                                 page_offset &= ~7; /* kill rounding error */
1766                                 page_offset <<= 1;
1767                                 npte = 2;
1768                         }
1769                         quadrant = page_offset >> PAGE_SHIFT;
1770                         page_offset &= ~PAGE_MASK;
1771                         if (quadrant != sp->role.quadrant)
1772                                 continue;
1773                 }
1774                 spte = &sp->spt[page_offset / sizeof(*spte)];
1775                 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1776                         gentry = 0;
1777                         r = kvm_read_guest_atomic(vcpu->kvm,
1778                                                   gpa & ~(u64)(pte_size - 1),
1779                                                   &gentry, pte_size);
1780                         new = (const void *)&gentry;
1781                         if (r < 0)
1782                                 new = NULL;
1783                 }
1784                 while (npte--) {
1785                         entry = *spte;
1786                         mmu_pte_write_zap_pte(vcpu, sp, spte);
1787                         if (new)
1788                                 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1789                         mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1790                         ++spte;
1791                 }
1792         }
1793         kvm_mmu_audit(vcpu, "post pte write");
1794         spin_unlock(&vcpu->kvm->mmu_lock);
1795         if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
1796                 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
1797                 vcpu->arch.update_pte.pfn = bad_pfn;
1798         }
1799 }
1800
1801 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1802 {
1803         gpa_t gpa;
1804         int r;
1805
1806         gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1807
1808         spin_lock(&vcpu->kvm->mmu_lock);
1809         r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1810         spin_unlock(&vcpu->kvm->mmu_lock);
1811         return r;
1812 }
1813
1814 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1815 {
1816         while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1817                 struct kvm_mmu_page *sp;
1818
1819                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1820                                   struct kvm_mmu_page, link);
1821                 kvm_mmu_zap_page(vcpu->kvm, sp);
1822                 ++vcpu->kvm->stat.mmu_recycled;
1823         }
1824 }
1825
1826 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1827 {
1828         int r;
1829         enum emulation_result er;
1830
1831         r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1832         if (r < 0)
1833                 goto out;
1834
1835         if (!r) {
1836                 r = 1;
1837                 goto out;
1838         }
1839
1840         r = mmu_topup_memory_caches(vcpu);
1841         if (r)
1842                 goto out;
1843
1844         er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1845
1846         switch (er) {
1847         case EMULATE_DONE:
1848                 return 1;
1849         case EMULATE_DO_MMIO:
1850                 ++vcpu->stat.mmio_exits;
1851                 return 0;
1852         case EMULATE_FAIL:
1853                 kvm_report_emulation_failure(vcpu, "pagetable");
1854                 return 1;
1855         default:
1856                 BUG();
1857         }
1858 out:
1859         return r;
1860 }
1861 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1862
1863 void kvm_enable_tdp(void)
1864 {
1865         tdp_enabled = true;
1866 }
1867 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
1868
1869 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1870 {
1871         struct kvm_mmu_page *sp;
1872
1873         while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1874                 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1875                                   struct kvm_mmu_page, link);
1876                 kvm_mmu_zap_page(vcpu->kvm, sp);
1877                 cond_resched();
1878         }
1879         free_page((unsigned long)vcpu->arch.mmu.pae_root);
1880 }
1881
1882 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1883 {
1884         struct page *page;
1885         int i;
1886
1887         ASSERT(vcpu);
1888
1889         if (vcpu->kvm->arch.n_requested_mmu_pages)
1890                 vcpu->kvm->arch.n_free_mmu_pages =
1891                                         vcpu->kvm->arch.n_requested_mmu_pages;
1892         else
1893                 vcpu->kvm->arch.n_free_mmu_pages =
1894                                         vcpu->kvm->arch.n_alloc_mmu_pages;
1895         /*
1896          * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1897          * Therefore we need to allocate shadow page tables in the first
1898          * 4GB of memory, which happens to fit the DMA32 zone.
1899          */
1900         page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1901         if (!page)
1902                 goto error_1;
1903         vcpu->arch.mmu.pae_root = page_address(page);
1904         for (i = 0; i < 4; ++i)
1905                 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1906
1907         return 0;
1908
1909 error_1:
1910         free_mmu_pages(vcpu);
1911         return -ENOMEM;
1912 }
1913
1914 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1915 {
1916         ASSERT(vcpu);
1917         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1918
1919         return alloc_mmu_pages(vcpu);
1920 }
1921
1922 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1923 {
1924         ASSERT(vcpu);
1925         ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1926
1927         return init_kvm_mmu(vcpu);
1928 }
1929
1930 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1931 {
1932         ASSERT(vcpu);
1933
1934         destroy_kvm_mmu(vcpu);
1935         free_mmu_pages(vcpu);
1936         mmu_free_memory_caches(vcpu);
1937 }
1938
1939 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1940 {
1941         struct kvm_mmu_page *sp;
1942
1943         list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1944                 int i;
1945                 u64 *pt;
1946
1947                 if (!test_bit(slot, &sp->slot_bitmap))
1948                         continue;
1949
1950                 pt = sp->spt;
1951                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1952                         /* avoid RMW */
1953                         if (pt[i] & PT_WRITABLE_MASK)
1954                                 pt[i] &= ~PT_WRITABLE_MASK;
1955         }
1956 }
1957
1958 void kvm_mmu_zap_all(struct kvm *kvm)
1959 {
1960         struct kvm_mmu_page *sp, *node;
1961
1962         spin_lock(&kvm->mmu_lock);
1963         list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1964                 kvm_mmu_zap_page(kvm, sp);
1965         spin_unlock(&kvm->mmu_lock);
1966
1967         kvm_flush_remote_tlbs(kvm);
1968 }
1969
1970 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
1971 {
1972         struct kvm_mmu_page *page;
1973
1974         page = container_of(kvm->arch.active_mmu_pages.prev,
1975                             struct kvm_mmu_page, link);
1976         kvm_mmu_zap_page(kvm, page);
1977 }
1978
1979 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
1980 {
1981         struct kvm *kvm;
1982         struct kvm *kvm_freed = NULL;
1983         int cache_count = 0;
1984
1985         spin_lock(&kvm_lock);
1986
1987         list_for_each_entry(kvm, &vm_list, vm_list) {
1988                 int npages;
1989
1990                 spin_lock(&kvm->mmu_lock);
1991                 npages = kvm->arch.n_alloc_mmu_pages -
1992                          kvm->arch.n_free_mmu_pages;
1993                 cache_count += npages;
1994                 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
1995                         kvm_mmu_remove_one_alloc_mmu_page(kvm);
1996                         cache_count--;
1997                         kvm_freed = kvm;
1998                 }
1999                 nr_to_scan--;
2000
2001                 spin_unlock(&kvm->mmu_lock);
2002         }
2003         if (kvm_freed)
2004                 list_move_tail(&kvm_freed->vm_list, &vm_list);
2005
2006         spin_unlock(&kvm_lock);
2007
2008         return cache_count;
2009 }
2010
2011 static struct shrinker mmu_shrinker = {
2012         .shrink = mmu_shrink,
2013         .seeks = DEFAULT_SEEKS * 10,
2014 };
2015
2016 static void mmu_destroy_caches(void)
2017 {
2018         if (pte_chain_cache)
2019                 kmem_cache_destroy(pte_chain_cache);
2020         if (rmap_desc_cache)
2021                 kmem_cache_destroy(rmap_desc_cache);
2022         if (mmu_page_header_cache)
2023                 kmem_cache_destroy(mmu_page_header_cache);
2024 }
2025
2026 void kvm_mmu_module_exit(void)
2027 {
2028         mmu_destroy_caches();
2029         unregister_shrinker(&mmu_shrinker);
2030 }
2031
2032 int kvm_mmu_module_init(void)
2033 {
2034         pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2035                                             sizeof(struct kvm_pte_chain),
2036                                             0, 0, NULL);
2037         if (!pte_chain_cache)
2038                 goto nomem;
2039         rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2040                                             sizeof(struct kvm_rmap_desc),
2041                                             0, 0, NULL);
2042         if (!rmap_desc_cache)
2043                 goto nomem;
2044
2045         mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2046                                                   sizeof(struct kvm_mmu_page),
2047                                                   0, 0, NULL);
2048         if (!mmu_page_header_cache)
2049                 goto nomem;
2050
2051         register_shrinker(&mmu_shrinker);
2052
2053         return 0;
2054
2055 nomem:
2056         mmu_destroy_caches();
2057         return -ENOMEM;
2058 }
2059
2060 /*
2061  * Caculate mmu pages needed for kvm.
2062  */
2063 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2064 {
2065         int i;
2066         unsigned int nr_mmu_pages;
2067         unsigned int  nr_pages = 0;
2068
2069         for (i = 0; i < kvm->nmemslots; i++)
2070                 nr_pages += kvm->memslots[i].npages;
2071
2072         nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2073         nr_mmu_pages = max(nr_mmu_pages,
2074                         (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2075
2076         return nr_mmu_pages;
2077 }
2078
2079 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2080                                 unsigned len)
2081 {
2082         if (len > buffer->len)
2083                 return NULL;
2084         return buffer->ptr;
2085 }
2086
2087 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2088                                 unsigned len)
2089 {
2090         void *ret;
2091
2092         ret = pv_mmu_peek_buffer(buffer, len);
2093         if (!ret)
2094                 return ret;
2095         buffer->ptr += len;
2096         buffer->len -= len;
2097         buffer->processed += len;
2098         return ret;
2099 }
2100
2101 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2102                              gpa_t addr, gpa_t value)
2103 {
2104         int bytes = 8;
2105         int r;
2106
2107         if (!is_long_mode(vcpu) && !is_pae(vcpu))
2108                 bytes = 4;
2109
2110         r = mmu_topup_memory_caches(vcpu);
2111         if (r)
2112                 return r;
2113
2114         if (!emulator_write_phys(vcpu, addr, &value, bytes))
2115                 return -EFAULT;
2116
2117         return 1;
2118 }
2119
2120 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2121 {
2122         kvm_x86_ops->tlb_flush(vcpu);
2123         return 1;
2124 }
2125
2126 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2127 {
2128         spin_lock(&vcpu->kvm->mmu_lock);
2129         mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2130         spin_unlock(&vcpu->kvm->mmu_lock);
2131         return 1;
2132 }
2133
2134 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2135                              struct kvm_pv_mmu_op_buffer *buffer)
2136 {
2137         struct kvm_mmu_op_header *header;
2138
2139         header = pv_mmu_peek_buffer(buffer, sizeof *header);
2140         if (!header)
2141                 return 0;
2142         switch (header->op) {
2143         case KVM_MMU_OP_WRITE_PTE: {
2144                 struct kvm_mmu_op_write_pte *wpte;
2145
2146                 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2147                 if (!wpte)
2148                         return 0;
2149                 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2150                                         wpte->pte_val);
2151         }
2152         case KVM_MMU_OP_FLUSH_TLB: {
2153                 struct kvm_mmu_op_flush_tlb *ftlb;
2154
2155                 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2156                 if (!ftlb)
2157                         return 0;
2158                 return kvm_pv_mmu_flush_tlb(vcpu);
2159         }
2160         case KVM_MMU_OP_RELEASE_PT: {
2161                 struct kvm_mmu_op_release_pt *rpt;
2162
2163                 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2164                 if (!rpt)
2165                         return 0;
2166                 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2167         }
2168         default: return 0;
2169         }
2170 }
2171
2172 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2173                   gpa_t addr, unsigned long *ret)
2174 {
2175         int r;
2176         struct kvm_pv_mmu_op_buffer buffer;
2177
2178         buffer.ptr = buffer.buf;
2179         buffer.len = min_t(unsigned long, bytes, sizeof buffer.buf);
2180         buffer.processed = 0;
2181
2182         r = kvm_read_guest(vcpu->kvm, addr, buffer.buf, buffer.len);
2183         if (r)
2184                 goto out;
2185
2186         while (buffer.len) {
2187                 r = kvm_pv_mmu_op_one(vcpu, &buffer);
2188                 if (r < 0)
2189                         goto out;
2190                 if (r == 0)
2191                         break;
2192         }
2193
2194         r = 1;
2195 out:
2196         *ret = buffer.processed;
2197         return r;
2198 }
2199
2200 #ifdef AUDIT
2201
2202 static const char *audit_msg;
2203
2204 static gva_t canonicalize(gva_t gva)
2205 {
2206 #ifdef CONFIG_X86_64
2207         gva = (long long)(gva << 16) >> 16;
2208 #endif
2209         return gva;
2210 }
2211
2212 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2213                                 gva_t va, int level)
2214 {
2215         u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2216         int i;
2217         gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2218
2219         for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2220                 u64 ent = pt[i];
2221
2222                 if (ent == shadow_trap_nonpresent_pte)
2223                         continue;
2224
2225                 va = canonicalize(va);
2226                 if (level > 1) {
2227                         if (ent == shadow_notrap_nonpresent_pte)
2228                                 printk(KERN_ERR "audit: (%s) nontrapping pte"
2229                                        " in nonleaf level: levels %d gva %lx"
2230                                        " level %d pte %llx\n", audit_msg,
2231                                        vcpu->arch.mmu.root_level, va, level, ent);
2232
2233                         audit_mappings_page(vcpu, ent, va, level - 1);
2234                 } else {
2235                         gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2236                         hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2237
2238                         if (is_shadow_present_pte(ent)
2239                             && (ent & PT64_BASE_ADDR_MASK) != hpa)
2240                                 printk(KERN_ERR "xx audit error: (%s) levels %d"
2241                                        " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2242                                        audit_msg, vcpu->arch.mmu.root_level,
2243                                        va, gpa, hpa, ent,
2244                                        is_shadow_present_pte(ent));
2245                         else if (ent == shadow_notrap_nonpresent_pte
2246                                  && !is_error_hpa(hpa))
2247                                 printk(KERN_ERR "audit: (%s) notrap shadow,"
2248                                        " valid guest gva %lx\n", audit_msg, va);
2249                         kvm_release_pfn_clean(pfn);
2250
2251                 }
2252         }
2253 }
2254
2255 static void audit_mappings(struct kvm_vcpu *vcpu)
2256 {
2257         unsigned i;
2258
2259         if (vcpu->arch.mmu.root_level == 4)
2260                 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2261         else
2262                 for (i = 0; i < 4; ++i)
2263                         if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2264                                 audit_mappings_page(vcpu,
2265                                                     vcpu->arch.mmu.pae_root[i],
2266                                                     i << 30,
2267                                                     2);
2268 }
2269
2270 static int count_rmaps(struct kvm_vcpu *vcpu)
2271 {
2272         int nmaps = 0;
2273         int i, j, k;
2274
2275         for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2276                 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2277                 struct kvm_rmap_desc *d;
2278
2279                 for (j = 0; j < m->npages; ++j) {
2280                         unsigned long *rmapp = &m->rmap[j];
2281
2282                         if (!*rmapp)
2283                                 continue;
2284                         if (!(*rmapp & 1)) {
2285                                 ++nmaps;
2286                                 continue;
2287                         }
2288                         d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2289                         while (d) {
2290                                 for (k = 0; k < RMAP_EXT; ++k)
2291                                         if (d->shadow_ptes[k])
2292                                                 ++nmaps;
2293                                         else
2294                                                 break;
2295                                 d = d->more;
2296                         }
2297                 }
2298         }
2299         return nmaps;
2300 }
2301
2302 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2303 {
2304         int nmaps = 0;
2305         struct kvm_mmu_page *sp;
2306         int i;
2307
2308         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2309                 u64 *pt = sp->spt;
2310
2311                 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2312                         continue;
2313
2314                 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2315                         u64 ent = pt[i];
2316
2317                         if (!(ent & PT_PRESENT_MASK))
2318                                 continue;
2319                         if (!(ent & PT_WRITABLE_MASK))
2320                                 continue;
2321                         ++nmaps;
2322                 }
2323         }
2324         return nmaps;
2325 }
2326
2327 static void audit_rmap(struct kvm_vcpu *vcpu)
2328 {
2329         int n_rmap = count_rmaps(vcpu);
2330         int n_actual = count_writable_mappings(vcpu);
2331
2332         if (n_rmap != n_actual)
2333                 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2334                        __func__, audit_msg, n_rmap, n_actual);
2335 }
2336
2337 static void audit_write_protection(struct kvm_vcpu *vcpu)
2338 {
2339         struct kvm_mmu_page *sp;
2340         struct kvm_memory_slot *slot;
2341         unsigned long *rmapp;
2342         gfn_t gfn;
2343
2344         list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2345                 if (sp->role.metaphysical)
2346                         continue;
2347
2348                 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2349                 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2350                 rmapp = &slot->rmap[gfn - slot->base_gfn];
2351                 if (*rmapp)
2352                         printk(KERN_ERR "%s: (%s) shadow page has writable"
2353                                " mappings: gfn %lx role %x\n",
2354                                __func__, audit_msg, sp->gfn,
2355                                sp->role.word);
2356         }
2357 }
2358
2359 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2360 {
2361         int olddbg = dbg;
2362
2363         dbg = 0;
2364         audit_msg = msg;
2365         audit_rmap(vcpu);
2366         audit_write_protection(vcpu);
2367         audit_mappings(vcpu);
2368         dbg = olddbg;
2369 }
2370
2371 #endif